category: literaturenote citekey: munozdeescalonaairtrafficcontrollerscommunication2025 title: "Air traffic controllers communication analysis as a proxy of task demand and mental workload: Using voice recording markers for safety critical task" authors: "Muñoz-de-Escalona, Enrique; Leva, Maria Chiara; Gianini, Gabriele; de Frutos, Patricia Lopez; Jadronova, Martina" year: 2025 date: 2025-00-00 2025 doi: 10.1016/j.ssci.2025.106928 publication: Safety Science url: "https://doi.org/10.1016/j.ssci.2025.106928" zotero_key: UE3BR9CC zotero_storage: 9BGASUZ7 collections: transpordi_planeerimine folder: 001_artiklid firstAuthor: "Muñoz-de-Escalona, Enrique"

status: converted

Contents lists available at ScienceDirect

Safety Science

journal homepage: www.elsevier.com/locate/safety

Air traffic controllers communication analysis as a proxy of task demand and mental workload: Using voice recording markers for safety critical task

Enrique Munoz-de-Escalona ˜ a,* , Maria Chiara Leva ^a , Gabriele Gianini ^b , Patricia Lopez de Frutos ^c , Martina Jadronova ^d

^a Environmental Sustainability and Health Institute, Technological University Dublin, Dublin, Ireland
^b Department of Informatics, Systems and Communication, Universita ` degli Studi di Milano-Bicocca, Italy
^c CRIDA A.I.E. ATM R&D + Innovation Reference Centre, Madrid, Spain
^d Mind, Brain and Behaviour Research Centre, University of Granada, Granada, Spain

ARTICLE INFO

Keywords: Air Traffic Management Mental Workload Human Performance & Reliability Behavioural Markers Risk Management Speech-Based Workload Estimation High Mental workload Detection Prediction models Classification models

ABSTRACT

The anticipated increase in air traffic over the coming decades will lead to a more congested airspace, thereby heightening the mental workload of air traffic controllers, who are responsible for maintaining the safety, reliability, and efficiency of air travel. Sustained overload is recognized in the literature as a significant factor influencing performance. Developing computational models for real-time monitoring of air traffic controllers' mental workload is critical for devising effective task support strategies to prevent the detrimental effects of declines in human performance. This research stems from consideration related to simple mental workload model for air traffic controllers based on behavioural markers that can be unobtrusively collected in real time from their voice recordings. These markers may reflect the task load and complexity experienced by a human operator, which in turn can be used in this study as a proxy indicator of mental workload. In this study over 80 h of voice recordings from real air traffic controllers were analysed to examine changes in behaviours and strategies that they might employ in response to variations in task demands while performing their tasks. Results indicate that traffic density levels can be predicted through behavioural measures derived from their voice communications. This suggests that variations in task demands lead to changes in their behaviours and strategies to effectively manage the situation and, thereby, reflect their mental workload. This early-stage model will be cross-validated and refined in a subsequent research stage using well-established psychophysiological and subjective measurements.

1. Introduction

The global air traffic industry is projected to undergo significant growth over the following decades. Airports Council International (ACI) World forecasts that in 2024 global passenger traffic will reach 9.4 billion passengers, surpassing the pre-pandemic levels recorded in 2019, when passenger volume was 9.2 billion. Additionally, further analysis suggests that, despite macroeconomic challenges such as inflation, GDP fluctuations, and geopolitical tensions, the projected growth will reach double the levels by 2042, with a compound annual growth rate (CAGR) of 4.3 %. Furthermore, the widespread use of Unmanned Aerial Vehicles (UAVs) or drones, which will share the skies with manned aircraft, is expected to further contribute to airspace congestion (Morales-Ferre

et al., 2019). A higher number of flights and potential conflicts will increase the demand for operations, leading to greater the mental workload (MW) and stress levels of air traffic controllers (ATCos), who are responsible for maintaining safe, reliable, and efficient air travel. ATCos play a vital role in making decisions related to air traffic safety, and their role is considered the most critical for flight safety (Hedayati et al., 2021). They monitor air traffic, assess potential hazards, develop strategies to prevent future conflicts, and implement these strategies (Pawlak et al., 1996). MW can generally be understood as a relationship between the mental resources required by a task (which depend on the complexity of the task) and the resources available to the operator (such as activation, motivation, etc.) (Young et al., 2015). MW and human performance are closely connected. Research across various fields has

E-mail addresses: Enrique.MunozDeEscalonaFernandez@tudublin.ie (E. Munoz-de-Escalona), ˜ mariachiara.leva@tudublin.ie (M.C. Leva), gabriele.gianini@ unimib.it (G. Gianini), pmldefrutos@e-crida.enaire.es (P.L. de Frutos), jadronova@ugr.es (M. Jadronova).

^* Corresponding author at: C/ Vicente Aleixandre, 29, Peligros, Granada, Spain.

consistently demonstrated that both mental overload and underload can lead to a decline in performance (e.g. Omolayo and Omole 2013; da Silva 2014; Suarez ´ et al., 2024). Hence, it is crucial to monitor and assess to what extent an ATCos may experience sustained high task demand and MW considering the detrimental effects it can have on human reliability (Canas ˜ et al., 2019). Traditionally, researchers have focused on identifying the factors that determine task complexity (e.g. Grossberg 1989; Histon et al., 2002; Radiˇsi´c et al., 2020). The fundamental idea is straightforward: higher task complexity results in higher MW levels for the ATCo, leading to efforts to predict MW based on a linear combination of task demand factors. Therefore, task complexity seems to be very useful as a proxy indicator of MW. However, as mentioned earlier, MW is influenced not only by task complexity but also by its interplay with individual factors (e.g. Xie and Salvendy, 2000; Triyanti et al., 2021). Therefore, assessing MW extends beyond evaluating task complexity and necessitates a shift toward human-centred modelling of MW. ATCos' MW is likely to remain the single greatest functional limitation on the capacity of the ATM system (Majumdar and Polak 2001). Efforts to estimate and monitor MW in the ATM domain have utilized subjective, psychophysiological, and behavioural measures (Yazgan et al., 2021). Nonetheless, no single universally accepted method currently exists in the industry to assess and monitor MW and its impact on performance. Subjective and psychophysiological measures are often criticized for being intrusive and impractical in real work scenarios (e.g. Longo and Leva 2021). These methods interfere with ATCos' performance: online MW measurement demands attentional resources for introspection each time subjective MW reports are requested, and most psychophysiological measures require intrusive equipment, which can disrupt task performance and even affect experienced MW (Moray, 2013). Additionally, subjective measures can be biased (Hancock, 2017).

In facing these challenges, the present work was a feasibility study prompted by CRIDA A.I.E., the Spanish Reference Center for Research, Development and Innovation in ATM. There was an interest in exploring possible behavioural measures that can be unobtrusively deployed in a real-world environment for predicting and classifying high MW conditions.

Suitable behavioural measures include (1) voice recordings from ATCos (communication with pilots and other ATCos) and (2) their interactions with ATM technical systems. The primary advantage of this type of model is its ability to overcome the limitations of psychophysiological and subjective methodologies. Communication patterns and interactions with ATM technical systems can be indirectly analysed in real-time using logs from the existing ATM technical system automation, which utilizes equipment already integrated into ATCos' tasks. Furthermore, these behavioural measures are not susceptible to distortion. In this study, the objective is to develop a simple model able to use behavioural markers that can be unobtrusively collected in real time from their voice recordings to detect high task demand conditions, which in turn can be used as a proxy indicator of MW. Currently this is not done, and we explored how it could be achieved and what markers and label are worth deploying to do so. The resulting early-stage model will then be cross-validated and refined in a subsequent stage using wellestablished psychophysiological measurements and subjective MW reports.

2. Related work

2.1. Measuring MW in the ATM domain

Mental workload can be understood as a relationship between the mental resources required by a task and the resources available to the operator (Young et al., 2015). MW assessment has been a longstanding focus for ATM researchers, as ATCos' MW is considered the primary functional limitation of the ATM system's capacity (Majumdar and Polak, 2001). Improving safety standards necessitates a precise and thorough evaluation of ATCos' MW (Socha et al., 2020). To ensure

ATCos can perform their duties effectively, it is crucial to manage their MW, whether excessive or insufficient, as both can lead to performance impairment. Various methodologies have been developed to quantify MW, including three main methodologies: subjective, psychophysiological, and behavioural measures (see Hilburn and Jorna 2001).

Subjective measurements are the most predominant for several reasons: they are cost-effective, easy to use, and better accepted by ATCos, compared to psychophysiological measures. However, subjective measurements have drawbacks such as memory bias, social desirability, and unwillingness to report. The NASA-TLX is one of the most widely used subjective measurements in ATM (e.g. Brookings and Wilson 1994; Willems and Stein 2001; Arico ` et al., 2015; Triyanti et al., 2021), along with the Air Traffic Workload Input Technique (e.g. Leighbody et al., 1992; Willems and Stein 2001; Fincannon and Ahlstrom 2016). To measure subjective MW in real-time, the Instantaneous Self-Assessment (ISA) scale is the most widely used method, originally developed for the ATM field (Jordan 1992) and validated after (Leggatt 2005). Additionally, Studies such as the one by Fürstenau and Radüntz (2021) were developed that provides evidence of a power law relationship between the ISA workload measure and ATCOs' radio communication, used as an objective indicator of task demand. The findings show that increased traffic flow positively correlates with both the frequency and duration of ATCOs' radio calls and their self-reported ISA workload levels.

Regarding psychophysiological measures, significant effort has also been made to use them in the ATM field due to their considerable advantages, mainly their objectivity and sensitivity. Psychophysiological measures refer to techniques that assess the relationship between psychological processes and physiological responses, which can be influenced, for example, by stress, mental workload, or fatigue. However, these methods are often rejected by ATCos because they are highly intrusive. Table 1 provides a summary of key psychophysiological measures commonly used to assess MW.

ATM researchers have also attempted to identify behavioural measures of ATCOs' MW, as these measures are much less intrusive than psychophysiological methods and more objective than subjective assessments. Additionally, some behavioural measures could potentially be automated and integrated into ATM technical systems. For instance, Schmidt (1976) identified the most demanding ATCo tasks by developing a difficulty index based on execution time and decision frequency. Morrow et al. (1993) studied the interactions between total radio communication time and ATCos' MW. Communication demands were also found to correlate with ATCos' performance (Hilburn 2004). Other studies linked communication performance with shift work (Corradini and Cacciari, 2002) and subjective MW with the number and duration of communication events (Manning et al., 2002). Lastly, some researchers studied the relationship between ATCos' strategies and decision-making activities and their experienced MW (Pawlak et al., 1996; Loft et al., 2007). Although behavioural measures offer practical advantages, such as being unobtrusive and applicable in real-world environments, they also present several challenges. Specifically, they are not always effective in capturing rapid fluctuations in mental workload, require indirect interpretation to infer cognitive states, and do not provide a complete view of how the controller truly feels or thinks. Their interpretability is also highly dependent on task and contextual factors.

2.2. Cognitive strategies that adapt workload to task demands

Research in ATM has traditionally posited that task complexity triggers MW, thereby limiting the capacity of en-route sectors (e.g. Christien et al., 2002). However, many authors agree that the functional relationship between task-load complexity and MW is largely unknown (e.g. Ath`enes et al., 2002). MW is not merely a function of task complexity, as it is influenced by numerous human factors. Prevalent models of ATCo's MW suggest that various factors within the ATM system (including traffic, airspace, and operational constraints) impose certain task demands on ATCos, but the level of MW experienced by

Table 1 Summary of key psychophysiological metrics used to assess MW.

Category	Subcategory	Effects	References
Brain electrical activity	The P300 amplitude in Event-Related Potentials (ERPs). (ERPs refer to brain responses that are triggered by specific stimuli and can be measured using	Decreased P300 amplitude with heightened MW	Brookings et al., 1996; Cabon et al., 1997; Wilson and Russell, 2003
	EEG). Composite indexes derived from combinations of EEG measures	Task Load Index (TLI) and EEG based Workload index as reliable measures of MW: absolute power of theta waves proves to be a good parameter for evaluating MW at different levels of air traffic control task. Significant correlations between MW and EEG signals from the frontal, parietal, and	Hamann, 2023 Arico ` et al., 2019; Laybidi et al., 2023 John et al., 2022
Brain metabolic measures	Functional Near Infrared Spectroscopy (fNIRs) – (non invasive technique that measures brain activity by detecting changes in blood oxygenation).	occipital regions. Oxygenation levels increase with higher traffic density fNIRs shows faster detection of MW changes compared to subjective reports	Ayaz et al., 2011; Ayaz et al., 2012; Harrison et al., 2013 Harrison et al., 2014
	Haemoglobin concentration	Need further exploration	Ayaz et al., 2012; Bernhardt et al., 2019
Heart and breathing measures	Heart rate (HR)	Increased heart rate (HR) with greater task demands	Lille and Burnod, 1983; Rose and Fogg, 1993; Backs, Navidzadeh and Xu, 2000; Radüntz et al., 2019, Radüntz et al., 2021
		HR more sensitive than HRV Others report	Kaber et al., 2007; Radüntz et al., 2019 Brookings et al.,
		show mixed results	1996; Metzger and Parasuraman, 2001; Kaber et al., 2007, Xiong et al., 2023
	Heart rate variability (HRV) based on spectral analysis Respiratory rate	Significant for reflecting MW HRV more sensitive than HR Consistently increases under higher task	Socha et al., 2020 Nikolova and Danev 2011 Brookings et al., 1996; Backs et al., 2000
Psychogalvanic response		demands Sensitivity to variations in task demands, correlating well with MW indexes	Averty et al., 2002

Table 1 (continued )

Category	Subcategory	Effects	References
Ocular measures	Pupil size	like the Traffic Load Index and NASA-TLX Increases with greater MW during ATM tasks	Ahlstrom and Friedman-Berg, 2006; Martin et al., 2011; Truschzinski and Wirzberger, 2017; Bernhardt et al., 2019; Munoz-de Escalona and Canas, 2017, Munoz-de- ˜ Escalona and Canas, ˜ 2018; Munoz-de- ˜ Escalona et al., 2019; Munoz-de- ˜
	Blink parameters	Blink duration	Escalona et al., 2020; John et al., 2022 Brookings et al.,
		and blink rate tend to decrease as task demands increase	1996; John et al., 2022
		The decrease in blink rate is not observed with higher task demands	Ahlstrom and Friedman-Berg, 2006
	Fixation parameters	No significant results have been found	Hyun et al., 2006; Marchitto et al., 2016
	Saccadic movement parameters	The correlation between MW and saccadic parameters was not found	Di Stasi et al.,2010
		In contrast, identified an inverse relationship between saccadic peak velocity and task demands, though they found no significant differences in saccade velocity and duration	Brookings et al., 1996
	Gaze duration	It was found that gaze duration, the number of fixation points, and saccade duration increased almost linearly with the rise in air traffic (number of aircraft in the airspace)	Wang et al.,2021

them depends on their experience, skills, and the strategies they employ to manage these demands. In this context, according to the model proposed by Loft et al. (2007) the MW experienced by ATCos is influenced not only by ATC task-load factors but also by the strategies selected by the ATCo to manage the situation and this selection is based on the prioritization of objectives. For example, a strategy refers to a specific type of behaviours that involves a set of control activities to prioritize safety over orderliness or expeditiousness. This could involve adjusting

aircraft trajectories in a less efficient manner, but one that reduces the need for future monitoring, control, and coordination, thereby reducing the workload. The study by Kontogiannis and Malakis (2013) provides valuable insights into complexity mitigation strategies in air traffic control. The authors find that, at low complexity levels, controllers simplify problems by perceiving aircraft as groups, reducing clutter on the radar screen, and anticipating potential threats such as weather or airspace congestion. Experienced controllers adopt a critical stance and continuously assess their understanding to adjust their plans according to changes in the situation. Additionally, they regulate their performance and work pace, for example, by choosing vertical separation over horizontal, even when both options are valid. Finally, the authors highlight that controllers streamline communications, such as by contacting other ATCos to share essential information or reaching out to nearby airports. The metacognitive processes of the ATCo play a critical role in guiding the prioritization of strategies during task performance. Metacognition is closely related to situational awareness (SA), involving an awareness of potential future task demands, available time to address these demands (time pressure), and knowledge of one's own skills, experience, and capacity. These factors help the ATCo prioritize strategies based on the evolving needs of the scenario. In other words, strategy selection is driven by how the ATCo's prioritizes objectives as task demands change over time (Kirwan and Flynn, 2002; Histon and Hansman, 2008). In addition, perceived MW impacts on perceived available time. For example, Balta et al. (2024) discovered that operators tend to underestimate time when performing high-workload tasks. This suggests that operators perceive less time available to respond, increasing subjective time pressure and prompting them to adjust their strategies to cope with task demands. Perceived MW is also shaped by factors such as the "peak effect," where the intensity of workload at its highest point influences ATCos' predictions about future MW (Qiao et al., 2021). In summary, metacognition can trigger a change in strategy to address any imbalance between the work to be done and the actual work being accomplished. Considering the above, we could say that it could be possible to infer ATCos' MW by observing shifts in strategies in response to variations in demand.

2.3. Inferring ATCos' MW based on how humans' behaviour shifts in response to task-demands

To understand how human behaviour changes when performing tasks and facing problems under varying time pressures and task demands, we can also examine two well-known models: Hollnagel's Contextual Control Model (COCOM) (1993) and Rasmussen's Skills, Rules, and Knowledge (SRK) model (1983). The COCOM model suggests that humans adopt different problem-solving strategies based on the ratio between the available time to take an action and the time required to complete that action, which involves assessing the situation, selecting a response, and executing that response. The model identifies four behavioural patterns: strategic, tactical, opportunistic, and scrambled. When a person has sufficient time to act, they follow a strategic behavioural pattern. As time pressure increases, their behaviour may shift through these stages, ultimately reaching a scrambled pattern under extreme time pressure. Consequently, MW increases as the individual progresses through these stages. Strategic and tactical behaviours are proactive, while opportunistic and scrambled behaviours are reactive. ATCos aim to maintain strategic or tactical patterns and avoid opportunistic and scrambled ones, although this is not always feasible.

Interviews with ATCos conducted by various research groups have shown that the MW associated with conflict resolution is heavily influenced by the specific knowledge ATCos accumulate over time and with experience (e.g. Kallus et al., 1999; Neal et al., 1998). This knowledge is often referred to as a "conflict-solving library,". When a conflict arises, the ATCo compares the situation with previous experiences and adapts existing solutions to the new context. If the conflict cannot be resolved with previous solutions in this way, ATCos must develop a new solution, which increases cognitive effort and, therefore, the MW. In this manner, MW is influenced not only by the time required to perform the action but also by the cognitive cost imposed by task demands.

To better understand this, we can refer to Rasmussen's classic Skill, Rule, Knowledge (SRK) model (1983). The SRK model describes human behaviour and associated errors in human–machine environments by categorizing behaviour into three levels based on the cognitive effort required for the task. Depending on the task demands and the person's available mental resources (such as activation level, experience, and skills), one or more levels may be engaged during task performance and decision-making. These three behavioural categories are known as skillbased, rule-based, and knowledge-based levels. According to the SRK model, the higher the level of control when performing a task, the greater the consciousness and control over the actions taken, and the higher the cognitive cost. Skill-level behaviour is efficient and economical in terms of cognitive cost but is difficult to control because it is automatic and unconscious. In contrast, knowledge-level behaviour demands significant cognitive processing but offers high controllability. The choice of control level depends largely on task demands and the difficulty these demands present to an individual. As task difficulty increases for an individual, a higher control level is required, leading to greater cognitive complexity and associated MW. The shift between control levels occurs due to an imbalance between task demands and the mental abstraction needed to understand and resolve the situation. In the context of ATM, nominal situations are typically managed using automatic, skill-based behaviour. However, if the situation becomes more challenging, the ATCo must find the optimal solution to prevent negative outcomes.

2.4. A simple model to explain the relationship between ATCos' MW and their behavioural patterns

In view of the above, time pressure and task demands (among other metacognitive factors) would determine the different strategies and behaviours followed by ATCos when performing their control activity. In addition, the perceived difficulty, not the complexity, would be what would determine MW level and would trigger a change in strategy and behaviours to meet the goals and to maintain MW under acceptable levels. Therefore, we can assume that a change in MW level can be reflected as a change in strategies and behaviours of the ATCo when performing their tasks.

We will now outline a very simple model explaining the relationship between MW and ATCo's behaviour. ATCos would normally operate under low to moderate MW conditions. In these scenarios, time pressure and task demands are both low to moderate and SA is high. The ATCo would normally follow a strategic control mode but shifting, eventually, to the tactical control mode. Furthermore, most of their behaviours would be skill-based (low cognitive cost) but, again, reaching sometimes the rule-based control level for solving more complex problems. Under these nominal situations ATCos would have enough time to plan and, therefore, they would be acting proactively. Moreover, most of the situations would be straightforward and solved with few cognitive resources, but some others more difficult would involve greater cognitive processing. ATCos would follow strategies to keep orderliness and expeditiousness, as safety is not compromised. For example, they could issue more shortcuts to traffics, since this strategy enhances flight efficiency.

During overload situations, time pressure and task demands will increase significatively, leading to a tactical control mode, and even opportunistic one if the difficulty increases further. SA is moderate. Their behaviours would be rule and knowledge-based depending on the problems they need to solve. Under these non-nominal situations ATCos have very limited time, so they cannot plan and, therefore, they could start working reactively. This would require a higher level of abstraction that would consume many cognitive resources, especially when developing new solutions. This increase in cognitive cost will be reflected in

the behaviours and strategies of the ATCo to cope with task demands. For example, with higher cognitive demands ATCos would start talking more frequently and, even, faster to communicate more pieces of information in a shorter time. In addition, they could be more confused and start hesitating more frequently and making more speech and content mistakes. Regarding cognitive strategies, they could try increasing their reduced SA, for example, posing more information requests to pilots. Safety would be prioritized over orderliness and expeditiousness; therefore, they could become more conservative and respect flights' air flight plan and not give shortcuts to flights.

Finally, during degraded situations, ATCos would be facing extraordinary time pressure and task demands, working reactively and with low SA. behaviours. The control mode becomes opportunistic or even disordered in the worst cases, and ATCos may resort to irrational trial-and-error behaviours if the situation exceeds their cognitive limits. Basically, the ATCo cannot manage the situation anymore and the risk is very high. For example, hesitations, speech and content mistakes would increase even more. Inexpert ATCos could even get so confused that they could get blocked during a while, requiring immediate assistance and help to recover the situation. Regarding cognitive strategies, safety would be the only priority. For instance, ATCos would change their strategy to focus all their cognitive resources on solving single problems sequentially, in order of priority, to try to recover the situation. However, this sort of degradation is normally never reached in the ATM sphere, since there are control systems that would impede the possibility of reaching such critical scenarios.

In summary, comprehending ATCos' MW entails understanding the diverse strategies and behaviours they adopt to address external task demands. As discussed earlier, shifts in strategy and behaviour are guided by the ATCo's metacognition: their ability to simultaneously process various factors such as time pressure, current and future task demands, and their own skills and experience. Although we cannot directly observe their metacognition, we can observe their behavioural patterns to infer their internal states and, consequently, their MW.

3. Methods

This study aims to develop a MW simple model for ATCos based on behavioural measures that can be unobtrusively and in real-time collected from ATM automation systems. Fig. 1 outlines the methodological steps taken to achieve our objectives.

3.1. Case study analysed, materials and instruments

3.1.1. Area control centre ATCos

ATM tasks involve decision-making by ATCos to ensure smooth traffic flow while maintaining aircraft safety by avoiding potential conflicts (Abdesslem et al., 1999). In essence, ATM tasks consist of sequencing and spacing aircraft to maintain a safe distance between them at all times (McGee et al., 1997). ATCos provide three main types of services: Area Control Centre (ACC) service, Approach (APP) service and Tower (TWR) service. For this study, we focus on modelling the workload of ACC-ATCos. ACC-ATCos spend most of their working time (1) interacting with two main tools: the radar (a technical system that monitors and manages aircraft, displaying their spatial position with a label containing key information for quick visualization) and the "Strip Board" (it complements the radar by displaying "Flight Progress Strips," summarizing key flight details for a clear overview of air traffic) and (2) managing communications with both pilots and other ATCos. In ACC-ATCo operations, the strips are used primarily to detect and resolve potential aircraft conflicts, while in APP and TWR services, they assist in managing the sequencing of arrivals, departures, and the overall visual traffic at and around airports. Effective communication with pilots and other controllers is critical for ensuring the smooth flow of air traffic and safety. ACC-ATCos work in pairs within the same airspace: one serves as the Executive-ATCo (EXE) managing traffic within the sector, handling tactical tasks and issuing instructions to pilots, while the other, the Planner-ATCo (PLN), coordinates with neighbouring sectors to ensure smooth traffic entry and exit (see Fig. 2). Consequently, the PLN-ATCo's duties primarily involve communication with other ATCos. A more detailed breakdown of the tasks performed by EXE and PLN-ATCos is provided in Table 2.

Finally, it is important to emphasize that, despite their distinct roles, EXE and PLN ATCos monitor each other's actions to reduce errors and ensure safety, order, and efficiency. For example, the PLN-ATCo may notice an error in a clearance from the EXE-ATCo and alert them for correction. Likewise, the EXE-ATCo might suggest a more effective traffic plan and discuss it with the PLN-ATCo. Both ATCos can also share tasks to manage their workload.

3.1.2. Voice communication switching system and digital recording system The Voice Communication Switching System (VCSS) is a computerized system enabling ATCos to communicate with pilots or other ATCos

Fig. 1. Overview of the methodological approach followed in this work.

Fig. 2. Executive and Planner ATCos interacting with the radar system and each other.

Table 2 General task distribution of executive and planner ATCos.

Executive ATCo	Planner ATCo
Ensure safe separation between aircrafts	Identify potential safety risks in advance (e.g., risks associated with incoming air traffic)
Communicate directly with pilots	Develop plans to resolve risky situations, with the option to discuss them with the EXE-ATCo
Issue instructions and clearances to pilots (e.g., flight level changes)	Manage entry and exit conditions to ensure a proper organization of air traffic
Provide relevant information to pilots (e. g., weather forecast)	Coordinate deviations from standard procedures
Implement the plan developed by the PLN-ATCo, with the flexibility to make changes if necessary	Alert the supervisor to non-standard or risky situations
Monitor the air traffic within the sector	Notify nearby sectors of non-standard or risky situations
Ensure that aircraft follow issued instructions and clearances	Coordinate with other authorities regarding operational air traffic (e.g., handling non-standard situations involving military aircraft)
Inform the PLN-ATCo of any situations they may not have been aware of	Inform the EXE-ATCo of any situations they may not have been aware of

telephone lines. VCSS systems are situated in Air Traffic Control centres and Air Traffic Control towers. Consequently, most of the communications made by ATCos are transmitted via the VCSS (coordination with other ATCos can also occur through direct communication with those located in the same control position). All communications are automatically recorded and stored in all control centres worldwide (using the Digital Recording System (DRS)), primarily for security reasons, allowing retrieval and analysis in case of incident or accident. In this research, voice communications from various ATCos operators spanning different time periods were extracted from official voice storage servers to compile the voice communications database analysed.

3.1.3. Voice communications database

Voice recordings of real ACC-ATCos were extracted from the DRS servers at intervals ranging from 20 to 90 min, resulting in a database of over 80 h of voice communications from 77 randomly selected participants. All recordings correspond to the same operational sector, "DGU," to ensure consistent comparison among ATCOs, as tasks can vary significantly across different sectors. The DGU sector is classified as medium-to-high complexity at the European level and is considered representative of Spanish en-route sectors due to its operational characteristics. It exhibits significant variability in traffic density and complexity throughout the day, reflecting a wide range of task-demand and communication scenarios. Future studies may extend the analysis to include additional sectors.

The recordings were completely anonymized for the analysis, to comply with the data and research ethics protocols imposed by the participating organization. This was to ensure participant confidentiality, particularly given the risk of re-identification in the specific operational setting. The absence of demographic variables limits the capacity to directly assess inter-individual variability in responses to mental workload. Nevertheless, the principal objective of this study is to advance the development of a generalizable framework for detecting behavioral indicators of MW in ATCos, irrespective of individual characteristics. This approach reflects operational realities, where demographic information is often unavailable, and systems must perform consistently across diverse user populations. Although individual differences could not be explicitly analyzed, the dataset comprises a heterogeneous sample of ATCos, with data collected across varied temporal and contextual conditions. This variability provides partial mitigation for the absence of demographic detail and strengthens the generalizability of the identified workload indicators. Each participant was assigned an anonymous ID, therefore, demographic information could not be obtained. Subsequently, a specialized tool developed by the ATM research centre CRIDA A.I.E was employed to process the raw data, dividing the voice recording intervals into fragments corresponding to each individual communication. Each communication represents the verbal expressions made by ATCos, separated by periods of silence. While each audio fragment typically consisted of a single communication, there were instances where a fragment contained more than one communication. Therefore, each audio fragment was reviewed to verify and manually split the communications if necessary, ensuring the database's accuracy. This study was developed according to the recommendations of the local ethical guidelines of the Research Ethics and Integrity Committee of Technological University Dublin (REIC-21-223) while the data collection was carried out with the internal approval of

the research ethics of the Spanish ATM research centre CRIDA A.I.E.

3.1.4. Raw data from the voice communications datasets

The automatic tool utilized for fragmenting the voice recordings automatically generates a.xml file, which provides specific data about each communication (see Fig. 3). The fundamental unit of analysis is the communication itself, with each row representing a single communication. The following elements of information are automatically computed and displayed for each communication:

File name: Name assigned to the voice recording fragment file.
Channel: VCSS channel used for the communication.
ATCo_ID: Anonymous ID assigned to the Target-ATCo.
Date: Date of the recording.
Time: Time at which the ATCo verbalizes the communication.
Duration: Duration of the communication in milliseconds.
Flight ID: Call sign of the aircraft, if mentioned in the communication.
Control event: Type of clearance, instruction, or information provided to the pilot.
Message: Explicit message conveyed to the pilot.

Some of this information was not relevant for the development of the MW model and was only useful for organizing the data (such as ATCos' ID, flight call sign, date, and communication channel). However, the time of communication, duration, type of control event, and explicit message issued were measures that could reflect behavioural changes of the ATCo in response to certain task demands and levels of MW. Control events encompass various types of clearances, instructions, and useful information provided by ATCos to pilots which as they result from are crucial as the cognitive processes involved in managing air traffic and problem-solving. While we cannot directly observe or measure the cognitive processes, we can observe, measure, and compute their outcomes, such as voice commands. The Table A in Appendix A summarizes the various control events that an ATCo can issue while performing their ATM tasks. Finally, it is important to highlight that the databases and datasets that were used during this research constitute sensitive private data property of CRIDA A.I.E. This data contains confidential communications and behavioural information from real ATCos and therefore will not be accessible for public verification or reuse.

3.2. From raw data to feature variables

All data, comprising 80.7 h of voice recordings were processed organizing them in 5-minute intervals with manual relabelling and verification, resulting in a total of 968 intervals. Each communication was listened to twice, to ensure accurate transcription of the audio and proper identification by the automated tool of the control events issued by the ATCo and manually corrected if necessary. Simultaneously, manual labelling was performed by quantifying the number of occurrences for each of the following variables: repetition of content within the same communication, hesitations, speech errors, content errors, confusions and conflict-resolution related phraseology. Subsequently, another set of variables was manually tallied including frequency of communications with pilots, frequency of coordination with other ATCos, frequency of single control events and frequency of grouped control events were recorded. Additionally, further calculations were necessary to determine the following variables: total control events, control events range, control events combination, duration of communications with pilots, duration of coordination with other ATCos, speech rate and adherence to flight plan. To ensure the accuracy and reliability of the manual labelling process, we sought the expertise of a field specialist, who reviewed the labelling process and provided confirmation on key segments. A summary of all the variables considered in this study can be found in Table B in Appendix B.

3.2.1. Target variable

In this study, we utilized "Airspace Sector Flight Occupancy" as an indicator of task demand and, consequently, as a proxy for the mental MW levels experienced by ATCos. This measure is defined as the maximum number of aircraft within an airspace sector during a given period. While traffic density alone does not fully determine MW (e.g., Lee and Prevot, 2012), there is substantial evidence in the scientific literature indicating a strong correlation between traffic density and the complexity of ATM tasks performed by ATCos (e.g., Benoit, 1997). The ATM field has shown significant interest in identifying various factors contributing to task complexity. Key variables include the ratio of the standard deviation of speed to average speed, variability in headings and speeds, aircraft trajectories, weather conditions, incidents or operational errors, horizontal and vertical proximity, and the cognitive nature of the task (e.g., Prandini et al., 2011; Hilburn, 2004; Djokic et al., 2010). For this study, the only real, empirical, and valid observable variable available from the ATM log recordings was "Airspace Sector Flight Occupancy." Despite its limitations, it is a valid indicator of task demand. It is reasonable to infer that as the number of aircraft under an ATCo's control increases, so does the level of task demand, thereby elevating their experienced MW. Moreover, many ATM centres in Europe currently use occupancy as the primary metric for Air Traffic Flow and Capacity Management (ATFCM). Consequently, when occupancy exceeds predetermined thresholds for each airspace sector, the ATFCM function takes action to safeguard them either by modifying airspace configurations or by implementing actions that affect flight plans. The overarching goal of this process is to prevent ATCOs from experiencing

Fig. 3. A snapshot of the raw Target-ATCo communication dataset.

excessively high MW. Currently, this metric acts as an indicator of ATCo MW. However, it is important to emphasize that using occupancy as an indicator of task demand is an assumption and should only be considered a proxy, not a direct measurement of task demand. Traffic density levels of the different sectors are automatically recorded and stored in control centres worldwide. As a result, alongside voice recordings, occupancy data can be extracted for specific time intervals. By aligning the voice communication dataset with the occupancy data, we were able to establish an objective proxy of task demand for model development.

3.2.2. Predictor variables

A change in MW would manifest as a shift in a set of variables representing the behaviour and strategies of ATCos during the execution of their ATM tasks. We began with a set of variables previously reported in a study by Munoz-de-Escalona ˜ et al. (2023). In that study, extensive literature research was conducted to identify suitable behavioural variables capable of reflecting ATCos' MW. They were then analysed for their feasibility and suitability for extraction from ATM system logs. The reader can find more information on how the variables were quantified, the expected trends, and other characteristics in Table B in Appendix B.

Form of communications

The set of variables reported in Table 3 below encompasses various features of communication that reflect specific verbalization styles but are unrelated to the content of the message.

Performance of ATCos' communications

These variables, reported in Table 4, reflect certain aspects of ATCos' performance during the execution of their ATM tasks. The overarching

Table 3

Variables	Expected changes with increase in MW
Frequency of communications with pilots (COM_Freq_comm_pilots)	More frequent communications with pilots, leading to an overall increase in communication volume to cope with higher task-demands
Duration of communications with pilots (COM_Duration_comm_pilots)	Longer durations of overall communication per interval to cope with higher task-demands
Frequency of coordinations with other ATCos (COM_Freq_coord_ATCos)	As MW transitions from low to moderate levels, the frequency of EXE-ATCo's coordinations is expected to rise to cope with higher task demands. However, a higher MW would result in a reduction, since this task will be assumed by the PLN-ATCo
Duration of coordinations with other ATCos (COM_Duration_comm_pilots)	As MW transitions from low to moderate levels, the duration of EXE ATCo's coordinations is expected to rise to cope with higher task demands. However, a higher MW would result in a reduction, since this task will be assumed by the PLN-ATCo
Speech rate (COM_Speech_rate)	The speech rate rises because the ATCo must convey more information in less time. However, in overload situations, the speech rate would decrease to ensure message accuracy and avoid potential repetitions.
Control events grouping (COM_Ctrl_events_grouping)	Since every communication is conducted to issue at least one control event, any further control events issued within the same communication are defined as "extra control events". At moderate levels of MW, ATCos tend to group several control events. However, a sharp increase in MW would lead ATCos to keep the communication more essential and avoid amalgamate many control events within a single communication.

Table 4 Variables of Performance of ATCos' communications along with the expected changes.

Variables			Expected changes with increase in MW
Hesitations (PER_Hesitations) Silences Easy to identify, typically appearing as sounds like "uh," "mmh," or "ehhm.	Fillers Pauses in speech perceived as hesitation, identified by disruptions in the natural flow of speech.	Lengthening Prolonged syllables signaling hesitation, identified by comparing them to surrounding segments. These are manually annotated to avoid confusion with other lengthenings. Trained annotators can distinguish hesitation-related lengthenings with high accuracy due to pitch and syllable position (Betz et al., 2023).	An increase in MW would result in an increase in the frequency of hesitations produced in the speech of the ATCo, similar to what found in other spheres (Betz et al., 2023; Montaci´e and Caraty 2014; Yin and Chen 2007; Arnold et al., 2007).
	Speech errors (PER_Speech_errors)		An increase in

A speech error is a mismatch between what is intended and what is said, affecting sounds, words, or grammar. Examples include: (1) "Ryanair 3278 you are on contact radar" (instead of "radar contact") and (2) "Transavia 2876 climb and maintain flight level two fix zero" (instead of "two six zero").

Content errors (PER_Content_errors) An increase in

A content error occurs when the information is inaccurate, incomplete, or misleading, leading to misunderstandings or mistakes in task execution. For example, in air traffic control, a content error could occur if an ATCo gives incorrect altitude instructions, like "Descend to flight level 130″ instead of 150.

speech errors is expected since limited attention and working memory make it harder for ATCos to monitor and self-correct their speech. content errors is expected, because high MW impairs memory retrieval and decisionmaking processes. When cognitive resources are overloaded, ATCos may struggle to recall and structure messages correctly, leading to content mistakes. Some of the trigger phraseology used for tallying this variable were: disregard, correction and apologies Confusions (PER_Confusions) An increase in the frequency of confusions is expected, as high MW weakens attention, working memory, and information processing, reducing ATCo's situational awareness. Some of the trigger phraseology used for tallying this variable were say again and report position

assumption is that increased MW would result in performance deterioration for ATCos' communications, which would be evident across various variables.

Content of the communications

This set of variables, reported in Table 5, pertains to the specific content conveyed through communications, irrespective of the way the message is delivered.

Adherence to flight plan

Adherence to flight plan reflects the degree of adjustment between the anticipated trajectory outlined in the flight plan and the actual trajectory followed by the aircraft. According to Histon and Hansman (2008) during periods of low MW, ATCos would adopt an efficiencybased strategy, resulting in an increased issuance of shortcuts to aircrafts to enhance flight efficiency, potentially leading to decreased adherence to flight plan. Conversely, in high MW circumstances, ATCos would likely shift to a security-oriented strategy, reducing the number of shortcuts issued to ensure aircraft adhere closely to their expected trajectories. This variable was calculated by measuring the difference in Nautical Miles between radar traces and flight plan expected trajectories of the aircrafts passing through the sector. We anticipated that adherence to flight plan would follow a downward trend.

4. Exploratory data analytics and prediction model learning

In this section we describe the dataset, the exploratory analytics and finally the regression analysis.

4.1. The dataset

Overall, the 968 record dataset consists of 22 columns: the integer target variable Occupancy and of 21 features (each a potential predictor variable). The latter are all numerical: most of them take a small number of integer values and have been encoded as integers for the analysis, the reminder have been encoded as floating-point numbers (ADH_to_flight_plan, COM_Duration_comm_pilots, COM_Speech_rate, and COM_Duration_coord_ATCos).

4.2. Variable association analysis

We recall that Occupancy was used as a dependent variable since it is a proxy of the level of task demand.

In the exploratory phase of the analysis, we have estimated the association of the individual predictor features with the target variable, and the association of features to one another.

Already the visual inspection reveals that all the integer valued

Table 5 Variables of content of the communications along with the expected changes.

Repetition of content within the same communication (CNT_Repet_content) Conflict-resolution related phraseology

(CNT_Conflict_res_phrslgy) Total control events (CNT_Total_control_events) Control events range (CNT_Ctrl_events_range)

Variable Expected changes with increase in MW

It is expected that ATCos tend to reiterate crucial instructions to pilots, aiming to prevent pilot confusion and streamline future operations Increased MW would prompt the ATCo to employ phraseology associated with conflict resolution. Some examples of specific phraseology

include expedite, immediately, for spacing, due to traffic, to avoid traffic, for traffic separation, and non-standard

An increase in task-demands would lead the ATCo to issue a higher amount of control events to effectively manage the air traffic

As task-demand increases ATCos are likely to encounter a greater variability in traffic scenarios that they must manage. This will manifest in an expansion of the range of control events employed. A higher variability of control events also implies a greater variability of cognitive operations being performed

Control events combination An increase in task demands could lead the ATCo to adjust the types of control events used in their communications. In this research, we did not consider isolated control events but rather combinations of control events that could reflect changes in the MW levels perceived by the ATCo

CE combi.

Assume flight þ transfer flight (CNT_COMB_flight_transfer) An increase in the frequency of this variable is expected, as it Comprises the aggregate of two control events: 1) assuming a flight (entering their sector) and 2) transferring a flight (leaving their sector). We posit that this combination of control events reflects the number of flights monitored and managed by the ATCo during a specific interval

Control events issued to delay a response from the ATCo to pilots.

Control events issued to resolve potential conflicts detected between flights

Control events issued by the ATCo to request information from pilots

Control events issued by the ATCo intended to alter the flight's configuration (e.g. heading change, altitude change, vectors, etc.). Repetitions (CNT_COMB_repetitions) A higher number of these control events indicates an increase

Control events that had to be repeated by the ATCo, primarily because they were not acknowledged by the pilot and required repetition

represents the overall volume of traffic handled by the ATCo during a specific interval

Standby (CNT_COMB_standby) An increase in the frequency of this variable is expected, as it groups control events intended to delay the ATCo's response to pilots, indicating the need for more time to address the situation due to urgent tasks or the resolution of complex situations that require cognitive resources and processing time

Conflict resolution (CNT_COMB_confl_res) We anticipate a higher frequency of this category, as it includes control events to resolve conflicts between flights, which increases cognitive load and stress, thereby raising the level of MW

Information requests (CNT_COMB_info_req) An increase in levels of this CE combination is expected. In situations of higher task demands, the ATCo's working memory approaches its limit, potentially compromising situational awareness. Consequencly, the ATCo may need to request more information to enhance their situational awareness.

Commands An increase in the frequency of this variable is expected, as it (CNT_COMB_commands) helps manage a large volume of traffic efficiently

in MW, as the ATCo must ensure that their clearances are acknowledged and executed to release that task, thereby freeing cognitive resources to handle other concurrent tasks. Inability to release tasks will create higher time pressure, which would increase the MW perceived by the ATCo.

predictors as well as COM_Duration_comm_pilots display a loose comonotonous relationship with the target, characterized by high variability around an ideal positive slope line: we will indicate this subset of variables as group-A variables. Examples are shown in Fig. 4. The reminder variables (ADH_to_flight_plan, COM_Speech_rate, and COM_-Duration_coord_ATCos), which we assigned to group-B, are not compatible with a linear or even a co-monotonic dependence. To quantify the association between the predictors and the target we computed Pearson's ρ (looking for linear or polynomial dependence, we also ran various types of regressions of target vs single variable) and Mutual Information coefficients (to detect associations not amenable to the previous cases).

As to the Pearson Correlation Coefficient ρ, all group-A variables display some linear correlation with the target (at a very good significance level, p < 10⁻ ⁵ , except of COM_Freq_coord_ATCos, which, still has p < 0.011). Seven of them have corr > 0.40 (see the first column of the heatmap in Fig. 5): they are COM_Freq_comm_pilots (ρ = 0.629), CNT_COMB_flight_transfer (ρ = 0.618), COM_Duration_comm_pilots (ρ = 0.612), CNT_Total_control_events (ρ = 0.598), PER_Hesitations (ρ = 0.524), CNT_Ctrl_events_range (ρ = 0.491), CNT_COMB_commands (ρ = 0.408).

No group-B variable shows a significant linear correlation with the target. To better characterize the relationship of the individual features with the target we tried for every predictor a target-vs-predictor linear regression, quadratic polynomial regression and Poisson regression (this, relied on the hypothesis that the integer valued target data were generated by a Poisson random variable): linear regression turned out to neatly outperform Poisson regression in terms of Coefficient of Determination R2 ; polynomial regression did not fare significantly better than linear regression: the improvements on the third digit of R2 where undermined by a loss if significance in the coefficients. All these elements prompted for seeking for a simple linear model regressing the target vs. the set of features in group-A, which we will describe in the next section.

We also computed the Mutual Information (MI) to investigate nonlinear forms of associations between the features and the target: since MI is the difference between the entropies of joint and marginal distributions, its definition does not contain explicitly the variables values. Consequently, MI represents an agnostic characterization agnostic with respect to any functional dependence between variables. To have an index taking values in the interval [0,1] we computed the Mutual Information Coefficient (MIC) defined in equation1.

$$MIC(X,Y) = \sqrt{1 - \exp(-2MI(X,Y))}, \text{ with } MI(X,Y),$$ (1)

The MIC definition represent the MI between the variables X and Y (which we determined using also the Miller-Madow correction for small samples), i.e. MI(X, Y) = H(X) + H(Y)− H(X, Y), where H() is the entropy functional. Since the values Pearson's coefficient (ranging from [− 1,1]) and the values of MIC are not directly comparable, we analysed their ranking of predictor features. We found that the two rankings substantially agreed, except for two features of group-B (thus not linearly correlated with the target) which had an irrelevant ρ but an

Fig. 4. Examples of dependencies between predictor variables and the target variable Occupancy. Top: two variables from group-A, where one can observe an approximate linear relationship; the red line corresponds to the result of the Ordinary Least Square linear fit. Bottom: two variables of group-B, not linearly correlated with the target, but still carrying information about it: e.g. it is apparent that for COM_Speech_rate > 4, the target is almost always less or equal to 7, whereas for COM_Speech_rate < 1, it is always less or equal to 10. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 5. Heatmap of the Pearson Correlation coefficients between variables pairs. The first column corresponds to the correlation with the target variable. Notice that some variables such as ADH_to_flight_plan, COM_freq_coord_ATCos, and COM_Speech_rate are not significantly correlated with any other variables, whereas the variables of the group most correlated with the target are also strongly correlated to one another.

appreciable MIC: ADH_to_flight_plan and COM_Speech_rate (both at 0.05 significance level).

Looking for associations between predictors we noticed a high coefficient ρ for some pairs (see Fig. 5), especially for those variables which also display a stronger correlation with the target itself. Notable cases are given by the set COM_Freq_comm_pilots, CNT_Total_control_events, COM_Duration_comm_pilots and CNT_Ctrl_events_range. Thus, the best individual predictors carry considerable redundancy, which must be considered when building the global prediction model. The association between predictors quantified using the MIC display a similar landscape (not reported here for space reasons).

In summary, the preliminary analysis showed that the variables in group-A lend themselves to linear modelling in relation to the target, provided that one suitably considers feature correlations. In contrast, two features of group-B, despite not exhibiting even an approximate linear relationship with the target, contain some information that could help improve the prediction to some extent.

4.3. Regression models for occupancy prediction

Our main purpose consisted in building a model that maximized prediction accuracy but also granted interpretability. For this reason, and in the light of the exploratory analysis we chose to use a linear model. To address the considerable correlation, present between some pairs of predictor variables, we experimented different alternative solutions, involving feature selection and combination, and corresponding to a family of related methods. Specifically, we assessed linear regression with several regularization schemes (LASSO and Ridge Regression), Bayesian Ridge Regression, Partial Least Square Regression, Stepwise Linear Regression and Poisson Regression. We quantified the crossvalidate performance of each approach in terms of the coefficient of determination R2 . The results, reported in some detail below, show the different methods lead to comparable performances, but exploit different although largely overlapping sets of features, by assigning them different importance.

Finally, experimented with a number of advanced Machine Learning models. They did not outperform the linear regression approaches.

4.3.1. Linear and Poisson regression

We experimented several variants of linear and Poisson regression. The 5-fold cross validated performance outcomes are summarized in Table 6 in terms of the coefficient of determination R2 , measuring the portion of the original variance of the target variable explained by the prediction model. One can observe that the different performances are compatible, except for Poisson Regression, which falls significantly behind. Hereafter we discuss the advantages and drawbacks of the different models and provide a more detailed account only for those which we found to be more valuable in terms of performance, significance and interpretability.

Ordinary Least Square regression tries to use all the available predictors and find the coefficients that minimize the square of the sum of the residuals. It does not apply any regularization to prevent large

Table 6 Coefficient of determination for the different regression models.

5-fold Cross Validated regression	R2 and error
Ordinary Least Square (OLS) Regression	0.467 ±
	0.013
Linear Stepwise Regression	0.468 ±
	0.011
Partial Least Square (PLS) Regression (3 components)	0.472 ±
	0.010
Least Absolute Shrinkage and Selection Operator (LASSO)	0.470 ±
Regression	0.011
Ridge Regression	0.469 ±
	0.013
Bayesian Ridge Regression	0.471 ±
	0.012
Poisson Regression	0.422 ±
	0.018

coefficients and is not robust to collinear predictors. With our data we found that the resulting coefficients were characterized by large uncertainties: approximately half of them were not significant at the p < 0.05 level.

An approach to address the predictor collinearity problem consists in trying to build orthogonal predictors by linearly combining the original predictors, in analogy to what is done in Principal Component Analysis. Partial Least Square (PLS) Regression extracts orthogonal components, but unlike PCA, it does so while maximizing the covariance with the target variable. Running PLS we have found that the first 3 component explain about 99 % of the variance. PLS appears to have a slightly better performance than other methods, however the minor improvement is counterbalanced by the loss of interpretability of the overall model, because each component is a linear combination of all the original variables (the 21 coefficients of the combination are reported in Ap pendix C).

Another effective approach to address collinearity consists in performing feature selection during linear regression. This can be achieved in several ways. A traditional way consists in using the Stepwise Linear regression, e.g. Forward Stepwise Linear Regression. The latter is a greedy process by which the predictors are selected one by one from the set of features based on a predefined criterion, such as minimizing the residual sum of squares or maximizing the model's adjusted R2, until no further significant improvement is achieved. We ran this process requiring a significance p < 0.05. Each fold in the cross-validated runs yields a slightly different set (due to feature redundancy, some feature can be replaced by few others in the role of selected predictor). The noncross-validated run selects the features shown in Table 7, along with their coefficients. Notice that when running Backward Stepwise Linear Regression (which starts from the whole set of features and removes the least performing or non-significant from the predictor set) we obtained the same model.

$$Occupancy = 5.73 + 0.2b + 0.55c + 0.29d - 0.30e - 0.18f + 0.49g - 0.55h$$ (2)

where b = Frequency of communication with pilot, c = CE combi (assume flight + transfer flight), d = hesitations, e = control events

grouping, f = Speech rate, g = conflict resolution related phraseology, h = CE Combi. (information requests).

The magnitude of the coefficients in the above model reflects not only their importance but also their unit of measure. To assess the relative importance of the variables in a unit measure independent way we fitted the prediction model with standardized variables. The resulting ranking of the absolute values of the coefficients in decreasing order of importance is the following: CNT_COMB_flight_transfer, COM_-Freq_com_pilots, PER_Hesitations, COM_Ctrl_events_grouping, CNT_Conflict_res_phrslgy, CNT_COMB_info_req, COM_Speech_rate. The absolute value of the coefficients thus obtained are reported in the rightmost column of Table 7.

Another, more recently developed, method for performing feature selection at the same time as linear regression to address collinearities is represented by the LASSO (Least Absolute Shrinkage and Selection Operator regression) regularization technique. This performs the regression minimizing the sum of the residual squares but also adds an L1 regularization term: the sum of the absolute values of the coefficients. This forces the least important coefficients to zero and keeps only the most relevant ones. We have chosen the regularization strength weight λ by cross-validation. In our case the LASSO regression selects 12 features out of 21 as predictors (not reported here). However, all except for 3 coefficients are not significant to the p < 0.05 level. Similar to LASSO regression is Ridge Regression, which however uses an L2 regularization term (the sum of the square of the coefficients, again the weight λ was chosen by cross-validation): this, by construction does not force coefficients to zero. We ran Ridge Regression, obtaining results comparable to LASSO, also in terms of issues about the coefficients: only 5 out of 21 coefficients have a p < 0.05. When running Bayesian Ridge Regression (assuming a Gaussian prior on the coefficients and a Gamma prior on noise variance), which chooses the regularization strength automatically from the data, which normally works well with low statistics, we were confronted with the same coefficient significance issue.

Poisson Regression assumes a linear relationship between the logarithm of the target variable mean and the predictors and adopts a loss function that assumes a Poisson distribution of the target data around that mean. This type of regression model was the least performing among the considered models, when assessed by cross-validation.

In synthesis, the Poisson regression uses a functional dependence which is not supported by the data at the same level as the linear dependence; the linear regression methods, which allow all the features, or a good portion of them, to be used as predictors, run the risk of overfitting the data. Out of the models that drastically reduce the number of features, Partial Least Square regression performs reasonably well, but by construction does not provide interpretability, whereas Stepwise Linear Regression does not overfit the data and provides a reasonably performing model, still providing interpretability. Under all the criteria we stated earlier, Stepwise Linear Regression is the best model.

From the signs of the coefficients in this model we see that: the larger COM_Freq_com_pilots, CNT_COMB_flight_transfer, PER_Hesitations, and CNT_Conflict_res_phrslgy, the larger Occupancy; whereas the larger COM_Ctrl_events_grouping, COM_Speech_rate, and CNT_COMB_info_req, the smaller Occupancy.

Table 7 The model obtained by Forward Stepwise Linear Regression that selected 7 important features.

	coeff	stderr	t	p> t	[0.025	0.975]	Importance
(a) const	5.731	0.251	22.829	0.000	5.238	6.223	9.90
(b) COM_Freq_com_pilots	0.283	0.044	6.379	0.000	0.196	0.370	1.21
(c) CNT_COMB_flight_transfer	0.556	0.067	8.324	0.000	0.425	0.687	1.31
(d) PER_Hesitations	0.291	0.042	6.944	0.000	0.209	0.373	0.88
(e) COM_Ctrl_events_grouping	− 0.298	0.072	− 4.139	0.000	− 0.439	− 0.157	0.49
(f) COM_Speech_rate	− 0.178	0.050	− 3.539	0.000	− 0.276	− 0.079	0.33
(g) CNT_Conflict_res_phrslgy	0.489	0.156	3.121	0.002	0.181	0.795	0.34
(h) CNT_COMB_info_req	− 0.550	0.178	− 3.091	0.002	− 0.899	− 0.201	0.33

4.3.2. Other machine learning models

We used an AutoML library (the python library pycaret-3.3.2) with 25 regression models and automatic search in hyperparameter space to compare different not necessarily interpretable predictors. This confirmed that the prevalence of the linear models in terms of R2 . Even the best of those regression models, i.e. Gradient Boosting (R2 = 0.431), Random Forest (R2 = 0.423), AdaBoost (R2 = 0.393), Lightgbm (0.380), Extra Tree (R2 = 0.395)) did not outperform the linear regression approaches discussed in the previous section.

4.4. Classification models for threshold exceedance

From a practical point of view often it is not strictly necessary to predict the exact value of a target variable but rather is important to predict when that variable passes a predefined safety threshold. Based on this consideration we considered the task of learning a classifier, able to tell when the target variable passed a given threshold. Mindful of the fact that the target variable we are adopting is just a proxy of the task demand, we investigated the problem of creating a classification model, based on the available features, able to predict threshold exceedance and endowed with interpretability.

We experimented with several thresholds, but discuss here only a representative one, defined by the 3rd quartile of the target variable distribution (this corresponds to the value of Occupancy equal to 13): data relative to other thresholds are reported in the Appendix D. The target values were assigned to the class Critical if Occupancy ≥ 13, (value Critical = True) and to the complementary class otherwise (value Critical = False). We experimented with 15 classification models (using the AutoML python library pycaret-3.3.2), assessing them primarily in terms of Accuracy, then in terms Precision, Recall, F1 Score and Mattew Correlation Coefficient. The best performing model in terms of Accuracy was Logistic Regression (LR), which is also a straightforwardly interpretable model. It corresponds to the fit of the functional dependence

$$\log\left[\frac{\Pr(Y = \mathit{True})}{\Pr(Y = \mathit{False})}\right] = \beta0 + \sum\nolimits{j=1}^n \beta_j X_j$$

where Pr(Y = True) and Pr(Y = False) are the probabilities of the target variable taking the two possible values (the argument of log are the odds), β0 is the intercept, βj with j > 0 are the other coefficients and Xj are the predictor variables. Aware of the effectiveness of the stepwise approach in terms of feature selection and performance, we refined the model by running a stepwise Logistic Regression procedure which issued a minimalistic model, involving only 6 predictor variables but exceeding 80 % accuracy and 70 % precision.

In detail, the 5-fold cross-validated model achieved Accuracy = 0.809 ± 0.004, Precision = 0.703 ± 0.024, Recall = 0.495 ± 0.023, F1 Score = 0.578 ± 0.013, MCC = 0.473 ± 0.012. The non-cross validated model involves only 6 predictors, of which 4 were also present in the linear regression model, two were not used by it, they are: CNT_COMB_standby and CNT_COMB_confl_res. The details are shown in Table 8. In synthesis the model is

$$\log[Odds] = -1.36 + 1.28b + 1.03c + 0.28d - 0.07e - 0.14i + 0.06j$$

with Odds ≡ Pr(Critical=True) Pr(Critical=False) .

The mapping between letters and variables is given in Table 8.

To support the interpretation of the results, the Table 8 reports also for each variable, the Odd Ratio, which is defined as the exponential of the corresponding coefficient, and represents the odds change in the probability of Critical = True for a 1-unit change in the predictor variable, holding other variables constant. For example, a 1-unit increase in COM_Freq_com_pilots increases the odds of Critical = True by a factor of 3.59.

5. Discussion

This study aimed to develop a model of the mental MW of ATCos based on behavioural measures that could be gathered unobtrusively and in real time from the ATM system automation, primarily focusing on their voice communications. To achieve this, we started with a set of variables previously reported in the initial study of this postdoctoral research project (Munoz-de-Escalona ˜ et al., 2023). This initial pool of potential variables was then narrowed down to those that were both feasible and practical to extract from the actual logs of the ATM systems. This resulting set of variables was categorized into four groups: "form of communications," "performance of ATCo's communications," "content of communications," and "adherence to flight plan.". The latter is the only variable not directly linked to ATCos' communications that was included for analysis.

For the model development, we conducted various forms of linear and a Poisson regression analysis utilizing the target occupancy as our best available proxy for ATCos' MW. This variable has demonstrated a robust correlation with the overall complexity of ATM tasks undertaken by ATCos. This choice was made as there were no alternative MW indicators available for extraction from the logs of the ATM systems.

5.1. Consideration regarding the linear regression model

The best linear regression analysis was a stepwise linear regression, and it yielded a Coefficient of Determination of R2 = 0.468.

It is worth noting that in psychological research, it's typical to see low R2 values due to the complexity of human behaviour. Factors like individual differences, environmental influences, and measurement error can lead to models that explain only a small proportion of variance in outcomes. R2 values as low as 10–30 % are often considered meaningful in psychology (Cohen, 1992). An R2 between 30 % and 50 % suggests that a significant portion of the variance in the dependent variable is explained by the independent variables. This range is typically seen as respectable in psychological research, where the complexity of human behaviour means, we often cannot expect to explain all variance with a limited set of predictors (Kline, 2015). R2 values above 50–60 % are rare but can be seen in psychological research, however, very high R2 values (above 70 %) might indicate overfitting, where the model explains not just true variance but also noise in the data (Frost, 2017). Based on this literature, an R2 value of 0.468 can be considered highly promising for research in this field.

The results from the stepwise linear regression analysis (see Table 7)

Table 8 Coefficients, standard errors and Odd Ration for the Logistic Regression model produced by stepwise LR Regression. The rightmost column reports the unit of measure independent variable importance indicator (obtained by standardizing the predictor variables and computing the absolute value of the corresponding coefficient).

Feature	Coefficient	StdError	OddsRatio	Variable Importance
(a) Intercept	− 1.364	0.098	0.26	1.364
(b) COM_Freq_com_pilots	1.277	0.303	3.59	0.689
(c) CNT_COMB_flight_transfer	1.034	0.230	2.81	0.641
(d) PER_Hesitations	0.284	0.053	1.33	0.667
(e) COM_Ctrl_events_grouping	− 0.071	0.036	0.93	0.214
(i) CNT_COMB_standby	− 0.137	0.061	0.87	0.226
(j) CNT_COMB_confl_res	0.059	0.021	1.06	0.253

revealed Frequency of communication with pilots as one of the most significant predictive variables for Occupancy. This result is not surprising in our view, since a higher traffic density would inevitably lead the ATCo to communicate more with pilots to effectively manage the air traffic. Furthermore, literature research has shown a strong correlation between frequency and duration of communications of the ATCos and their level of perceived MW (e.g. Cardosi 1993; Porterfield 1997; Manning et al., 2002; Manning et al., 2003; Wang et al., 2011; Ucl´es and García, 2014). ATCo communications can be seen as the result of the whole cognitive process that they must follow for managing the air traffic. This process involves 4 primary tasks: monitoring, evaluating, formulating a plan and implementing such plan. According to Pawlak's model (Pawlak et al., 1996), it is the combination of the first three primary tasks (monitoring, evaluating and formulating a plan) what drives ATCo's mental effort, but the only primary task that could be observed and, therefore, quantified is the implementation process, and this would be reflected mainly as the communications performed by ATCos.

The presence of CE Combi. (assume flight + transfer flight), also aligns logically with the expectations. This variable reflects the total number of flights that the ATCo had to monitor and manage at any given moment during the considered interval. It is reasonable to infer that this variable is a reflection of the occupancy variable itself, as it represents the overall traffic volume handled by the ATCo in that specific sector and at that interval. A higher value in this variable implies increased perceived complexity due to the rise in the total number of aircrafts managed by the ATCo, consequently leading to a higher MW.

The variable Hesitations is one of the most intriguing findings of this research. As previously mentioned, literature has shown a link between hesitations and MW (Betz et al., 2023; Montaci´e and Caraty 2014; Yin and Chen 2007; Arnold et al., 2007), however, to our knowledge, this variable has never been analysed in the ATM domain as a possible predictor for MW. The assumption underpinning this finding is that higher MW levels would require increased cognitive processing to manage complex problems, which would manifest as longer response times. In other words, the ATCo would be "buying" time to process the information and formulate a solution, leading to a higher frequency of hesitations in their speech.

Also, the presence of the variable Conflict resolution related phraseology aligns well with our expectations. It is reasonable to assume that conflict resolution increases cognitive load and stress, consequently increasing MW levels. A high task-demand scenario, with a high level of potential conflicts to resolve, would prompt the ATCo to employ phraseology associated with conflict resolution, thereby indicating an increase in MW levels.

The predictor variable CE grouping, exhibited a negative value. This indicated that higher occupancy values would lead the ATCo to avoid grouping control events when communicating with pilots. Instead, the ATCo would prefer to issue a single control event per communication. These results aligned with our expectations, as we hypothesized that the ATCo's information processing strategy would change based on the perceived complexity. In lower task-demand circumstances, the ATCo could process simultaneous tasks and resolve situations in parallel. However, in higher task-demand circumstances, the ATCo would focus on one task at a time due to cognitive overload, which leaves no remaining capacity to solve multiple problems simultaneously. In other words, the ATCo needs to automize, that is, simplify and break down information into single communications. This involves solving one problem (issuing the corresponding control event) and ensuring that the control event is acknowledge (receiving the corresponding readback) and effectively handled to release cognitive load before moving on to the next task.

Also, the variable CE Combi. (information requests), exhibits a negative value. This indicates that higher occupancy levels tend to decrease the frequency of control events issued by the ATCo to request information from pilots. This outcome contrasts with our initial hypothesis, as we expected that heightened task demands would prompt the ATCo to request more information from pilots, potentially due to a diminished situational awareness. However, our results showed the opposite. One possible explanation for this result is that as task demands increase, so does time pressure. Consequently, the ATCo may prioritize addressing necessary task and resolving any air traffic potential conflicts by issuing clearances to manage the flow. This leaves insufficient time for requesting additional information to pilots, thereby resulting in a decrease in the frequency of control events issued for information requests.

The variable Speech rate, although not linearly correlated with the Occupancy (ρ = − 0.064, p < 0.05), carries, as we observed previously, an appreciable amount of information about the target in a different form (MIC = 0.277, p < 0.05), thus it is selected by the stepwise regression algorithm as able to significantly predict the residuals, after the target has been regressed on the previous variables.

It is worth noticing that not all the features most strongly correlated with the target appear in the final linear model: indeed the variables Duration of communications with pilots, Total control events, and Control events range, are not included in the final model, despite being highly correlated with the target (their correlation with Occupancy being respectively ρ = 0.612, ρ = 0.598, and ρ = 0.491, all with p < 10⁻ ⁵ ). This exclusion was determined by the stepwise linear regression approach chosen and it is potentially to be justified considering that these variables significantly overlap with Frequency of communication with pilots, showing correlation coefficients of ρ = 0.933, ρ = 0.966, and ρ = 0.850, respectively (all with p < 10⁻ 10). Including multiple highly collinear predictors would not enhance predictive performance and could introduce multicollinearity issues. Therefore, only the most informative and non-redundant features were retained in the final model. This approach ensures that the model remains both parsimonious and robust (Table 9).

5.2. Consideration regarding the classification model

For long-term application, a classification model would likely be more useful than a predictive model, as it can help determine when and for how long an ATCo's task load or MW exceeds a critical safety threshold. Defining this threshold would require a dedicated experimental design, incorporating self-assessment metrics such as ISA and triangulation with psychophysiological measures. The results of the classification model based on the available features in this study are highly promising. During model development, various thresholds were tested, and the most representative was determined to be the 3rd quartile of the target variable distribution, corresponding to an Occupancy value of 13. Additionally, multiple classification models were evaluated in terms of Accuracy, Precision, and Recall. The Linear Regression (LR) model demonstrated the best performance in terms of Accuracy, with four out of the six variables selected through the stepwise approach aligning with those used in the linear regression models discussed in the previous section. These variables include Frequency of communication with pilots, CE Combi. (assume flight + transfer flight), Hesitations and CE grouping.

The two additional variables incorporated in the classification model were: CE Combi. Standy (a variable identifying control events issued to delay an ATCo's response to pilots) and CE Combi. Conflict Resolution (a variable capturing control events aimed at resolving potential conflicts between flights). Both variables play a crucial role in identifying the resolution of complex situations that demand significant cognitive resources and processing time. Therefore the classification model is also providing a promising modelling direction.

5.3. Conclusions, limitations and future directions

This study aimed to develop a simple MW model for ATCos based on behavioural measures that can be collected in real time from ATM system logs, primarily through communications conducted by ATCos. The ultimate goal is to create an automated tool capable of predicting MW

Table 9 Summary of the variables included in the final stepwise regression model.

Most relevant Variables	Hypothesis validation	Explanation
Frequency of communications with pilots	✓	Higher air traffic density leads to an increase in communication with pilots, as well as in the frequency and duration of such communications. Our results aligns with literature (Cardosi, 1993; Porterfield, 1997; Manning et al., 2002; Manning et al., 2003; Wang et al., 2011; Ucl´es and García, 2014)
CE Combi. Assume flight + transfer flight	✓	This variable is a reflection of the occupancy variable itself, as it represents the overall traffic volume handled by the ATCo in that specific sector and at that interval. A higher value in this variable implies increased perceived complexity due to the rise in the total number of aircrafts managed by the ATCo, consequently leading to a higher MW
Hesitations	✓	High levels of mental MW require greater cognitive processing, which is reflected in longer response times. Literature has linked hesitations with MW in other domains (Betz et al., 2023; Montaci´e and Caraty 2014; Yin and Chen 2007; Arnold et al., 2007)
Control events grouping	✓	The observed negative value indicated that higher occupancy values would lead the ATCo to avoid grouping control events when communicating with pilots, preferring to issue a single event per communication. This confirms our hypothesis that, in low demand conditions, the ATCo handles tasks simultaneously, while in high demand conditions, they focus on one task at a time due to cognitive overload.
Speech rate	×	Although not linearly correlated with Occupancy, this variable was selected by the stepwise regression algorithm as a significant predictor of the residuals after accounting for the other variables. Previous studies (e.g., Luig & Sontacchi, 2010) have associated speech rate with stress and time pressure. However, in this study, we could not identify a clear explanation for the observed speech rate trend.
Conflict-resolution related phraseology	✓	Conflict resolution increases cognitive load and stress, consequently increasing MW levels. A high task demand scenario, with a high level of potential conflicts to resolve, would prompt the ATCo to employ phraseology associated with conflict resolution, thereby indicating an increase in MW levels.
CE Combi. Information requests	×	Contrary to our initial hypothesis, increased task demands did not lead to more information requests from pilots. Instead, the results suggest that higher demands and time pressure push ATCos to prioritize essential tasks, such as issuing clearances to manage traffic flow. This urgency likely reduces the time available for requesting additional information, leading to a decrease in such control events.

levels in real time by analysing their voice communications. This tool would offer the advantage of making unobtrusive and immediate predictions, which would help improve ATCos' health and efficiency at work, as well as aid in developing and testing mechanisms to support MW variations and mitigate the negative effects of performance drops. Our results indicate that occupancy levels can be predicted through behavioural measures derived from ATCOs' voice communications. This suggests that variations in task demands lead to changes in ATCOs' behaviours and strategies to effectively manage the situation. Therefore, we can infer ATCOs' MW by observing shifts in their behaviours and strategies in response to task demands.

In recent years, automation tools have been developed that have successfully reduced the MW of ATCos, leading to significant improvements in ATM. One notable example is the decision-support tool for conflict detection and resolution, which aids ATCos in rapidly identifying and resolving high-density conflicts in advance (Sui and Zhang, 2022). This tool employs the Monte Carlo tree search (MCTS) algorithm to tackle the challenge of large search spaces and quickly generate effective solutions for resolving pairwise conflicts. By autonomously detecting conflicts and offering feasible resolution strategies, the tool aims to reduce ATCos' workload and enhance their ability to manage conflicts during en-route operations in a timely manner. Research has demonstrated that higher levels of automation result in a greater reduction in workload (e.g. Edwards et al., 2016; Wang et al., 2021). Based on this, we believe our tool could assist in optimizing task allocation between automation systems and ATCos. When high MW is detected, automation tools could take on a greater share of tasks, thereby reducing cognitive load on ATCos and preventing potential declines in human performance due to overload.

The main limitation of this study was using occupancy as the only indicator of air ATCos' MW. While traffic density is strongly correlated with task complexity, many other factors affect task complexity, meaning high traffic density can sometimes be associated with low complexity, and vice versa. Additionally, MW depends not only on task complexity but also on individual factors. Ideally, we would have included a subjective measure of MW, such as the Instantaneous Self-Assessment scale, or an objective measure, such as ocular parameters or EEG. However, the scope of the study was to test the feasibility of using only real observable variable currently collectable from existing logs. Subjective and psychophysiological measurements from ATCos are not available in actual operational environments. In any case, we are pleased to find that our results showed several behavioural MW indices significantly correlate with occupancy. This means that, assuming the positive correlation between occupancy and MW that has been found in the literature, we can say that according to our results, MW levels are indeed reflected in the behaviour of the ATCos. In any case, these results must be taken cautiously and must be cross validated using wellstablished MW subjective and psychophysiological measures. Another key limitation is the need for manual labelling of several variables that require expert judgment, such as hesitations and speech errors. This hinders the possibility of applying the model in real-time. A related limitation is the relatively small sample size analysed. Specifically, we believe that 80 h of voice recordings are insufficient to fully capture the significant interpersonal variability in ATCos' behaviour during task performance. However, both limitations could be overcome by developing software capable of automatically labelling the variables that currently require manual input. Such a tool would provide access to a virtually unlimited amount of data, significantly enhancing the validity of the findings. Furthermore, developing this tool would be essential for measuring ATCos' MW in real-time, as their voice communications would need to be processed instantly for this purpose. However, it would probably be not error free and it will be questions regarding the tool reliability and the resulting labelling accuracy. We believe that natural language processing would be the key to solving this problem in a notso-distant future. Indeed, recent advancements in technology have led to the development of prototypes for speech recognition and

understanding in the ATM sector. For example, Ohneiser et al. (2023) recently validated the Assistant-Based Speech Recognition (ABSR) system for radiotelephony communication in remote tower controller working positions. This system integrates a conventional speech recognizer with an assistant system and has demonstrated adequate recognition quality for use in ATC applications (Kleinert et al., 2019; Ohneiser et al., 2023). While the current results are not yet fully satisfactory, this system shows great potential for reducing workload and improving ATCos' situational awareness. We believe that our labelling work presented in this study could contribute to the further development of these tools by incorporating new variables that have proven relevant in predicting MW.

The next phase of this research project seeks to validate the model developed in this study using data obtained from simulations involving real ATCos. Specifically, ATCos will perform ATM tasks under varying task-demand conditions while both objective and subjective MW measures are recorded. More precisely, we will process voice recordings to extract the set of variables considered in this study. Additionally, we will incorporate other well-established MW measures to serve as indicators of ATCos' MW, including subjective assessments through the "Instantaneous Self-Assessment Scale" and objective psychophysiological measures such as ocular parameters and EEG, comparable with studies carried out in simulators by other researchers such as Li et al. (2023) that have used EEG and eye-tracking features in air traffic control tasks to explore MW and its link to situational awareness. It is important to highlight that such studies are typically only feasible in simulation environments, where ATCos activities can be influenced by the use of physiological recording methods (e.g., EEG, eye trackers) or the presence of external observers. These methods often require interruptions to administer questions or questionnaires for assessing situational awareness. In contrast, the novelty of our study lies in the fact that our approach—analysing voice recordings—offers an unobtrusive method that uses data already available to ATM operators. This allows us to gain valuable insights into the MW experienced by ATCos without relying on external observers or additional physiological or subjective measurements.

Ethics considerations

This study was developed according to the recommendations of the local ethical guidelines of the Research Ethics and Integrity Committee of Technological University Dublin (REIC-21-223) as well as with the internal approval of the ATM research center CRIDA A.I.E.

CRediT authorship contribution statement

Enrique Munoz-de-Escalona: ˜ Writing – review & editing, Writing – original draft, Visualization, Validation, Project administration, Methodology, Investigation, Formal analysis, Data curation. Maria Chiara Leva: Writing – review & editing, Validation, Supervision, Methodology, Investigation, Funding acquisition, Formal analysis, Conceptualization. Gabriele Gianini: Writing – review & editing, Validation, Resources, Methodology, Formal analysis. Patricia Lopez de Frutos: Validation, Supervision, Resources, Methodology, Investigation, Funding acquisition, Formal analysis. Martina Jadronova: Writing – review & editing, Writing – original draft, Visualization, Methodology, Investigation, Formal analysis.

Funding

The research conducted in this publication was funded by the Irish Research Council under grant number [EPSPD/2022/151] in collaboration with CRIDA A.I.E. The work was partially supported by the NextGenerationEU program through the project MUSA (National Recovery and Resilience Plan Mission 4 Component 2 Investment Line 1.5 (CUPG43C22001370007, CodeECS00000037), and by the Italian University and Research Ministry, within the Excellence Departments program 2023–2027 (law 232/2016) through the ReGAInS project 2023- NAZ-0207.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

The authors wish to thank CRIDA A.I.E for providing all the data used in this study. The authors wish to thank ENAIRE, for providing picture found in Fig. 2

Appendix A

Table A List of most common control events.

Control event code	Control event description
CTE	Sector entry (assume flight)
CS	Sector change (transfer flight)
CO	Inter-sector controller-controller coordination
A1	Change/level restriction by separation
A2	Speed change
A3	Direct for separation
A6	Authorization denial/delay
S2	Heading change
S3	Diversion approval
AC5	Flight level change
AC6	Constraint removal
AC7	Authorized to approach
AC8	Direct
AC9	Essential information
AC10	Important information
AC12	Squawk assignment
AC13	STAR assignment

(continued on next page)

Table A (continued )

Control event code	Control event description
H1	Holding stack
RE_AC5	Flight level request
RE_AC8	Direct request
RE_S2	Heading request
RE_A2	Velocity request

Appendix B

Table B Summary of variables considered for model development.

Variable	Type of variable	Hypothesis	Calculation method	Trend prediction	Feasibility of use in real context
Airspace Sector Flight Occupancy	Independent variable for model development	Higher Occupancy a ` Increase task-demands and MW consequently	(Maximum number of aircrafts within an airspace sector) per interval	−	High
Frequency of communications with pilots	Dependent Variable (Form of communication)	Higher MW ` a Increase in frequency of communications	(No. of communications) per interval	Upward	High
Duration of communications with pilots	Dependent Variable (Form of communication)	Higher MW ` a Increase in duration of communications	(Duration of communications) per interval	Upward	High
Frequency of coordinations with other ATCos	Dependent Variable (Form of communication)	• Low to moderate MW `<br>a Increase in<br>frequency of<br>coordinations<br>• Moderate to high<br>MW` a Decrease in frequency of coordinations	(No. of coordinations) per interval	Quadratic	High
Duration of coordinations with other ATCos	Dependent Variable (Form of communication)	• Low to moderate MW `<br>a Increase in<br>duration of<br>coordinations<br>• Moderate to high<br>MW` a Decrease in duration of coordinations	(Duration of coordinations) per interval	Quadratic	High
Speech rate	Dependent Variable (Form of communication)	• Low to moderate MW `<br>a Increase<br>speech rate<br>• Moderate to high<br>MW` a Decrease in speech rate	( ) No. of words per interval Duration of communications	Quadratic	High
Control events grouping	Dependent Variable (Form of communication)	• Low to moderate MW `<br>a Increase in CE<br>grouping<br>• Moderate to high<br>MW` a Decrease in CE grouping	(No. of extra control events) per interval	Quadratic	High
Hesitations	Dependent Variable (Performance of ATCos' communications)	Higher MW ` a Increase in hesitations	(No. of hesitations) per interval	Upward	Low
Speech errors	Dependent Variable (Performance of ATCos'	Higher MW ` a Increase in speech errors	(No. of speech errors) per interval	Upward	Low
Content errors	communications) Dependent Variable (Performance of ATCos' communications)	Higher MW ` a Increase in content errors	(No. of content errors) per interval	Upward	Moderate
Confusions	Dependent Variable (Performance of ATCos' communications)	Higher MW ` a Increase in confusions	(No. of confusions) per interval	Upward	Moderate
Repetition of content within the same communication	Dependent Variable (Content of the communications)	Higher MW ` a Increase in reiteration of contents	(No. of content reiterations) per interval	Upward	High

(continued on next page)

Table B (continued )

Variable	Type of variable	Hypothesis	Calculation method	Trend prediction	Feasibility of use in real
					context
Conflict-resolution	Dependent Variable	Higher MW ` a Increase	(No. of conflict resolution related phraseology) per interval	Upward	High
related phraseology	(Content of the	in conflict-resolution
	communications)	related phraseology
Total control events	Dependent Variable	Higher MW ` aIncrease	(No. of total control events) per interval	Upward	High
	(Content of the communications)	in total control events
Control events range	Dependent Variable	Higher MW ` a Increase	(Range of control events) per interval	Upward	High
	(Content of the	in control events range
	communications)
Control events	Dependent Variable	Higher MW ` a Increase	(No. of CE combi. (assume flight + transfer flight) per interval	Upward	High
combination	(Content of the	in CE combi. (assume
(assume flight + transfer flight)	communications)	flight + transfer flight)
Control events	Dependent Variable	Higher MW ` a Increase	(No. of CE combi. (standby) per interval	Upward	High
combination	(Content of the	in CE combi. (standby)
(standby)	communications)
Control events	Dependent Variable	Higher MW ` a Increase	(No. of CE combi. (conflict resolution) per interval	Upward	High
combination	(Content of the	in CE combi. (conflict
(conflict resolution)	communications)	resolution)
Control events	Dependent Variable	Higher MW ` a Increase	(No. of CE combi. (information requests) per interval	Upward	High
combination	(Content of the	in CE combi.
(information	communications)	(information requests)
requests)
Control events	Dependent Variable	Higher MW ` a Increase	(No. of CE combi. (commands) per interval	Upward	High
combination (commands)	(Content of the communications)	in CE combi. (commands)
Control events	Dependent Variable	Higher MW ` a Increase	(No. of CE combi. (repetitions) per interval	Upward	High
combination	(Content of the	in CE combi.
(repetitions)	communications)	(repetitions)

Appendix C

Table C The coefficients of the first three components from Partial Least Square regression, explaining 99 % of the target variable variance.

	Component 1	Component 2	Component 3
ADH_to_flight_plan	− 0.022853	− 0.086209	0.443138
CNT_COMB_commands	0.247403	− 0.211918	− 0.077404
CNT_COMB_confl_res	0.163263	− 0.118154	0.330587
CNT_COMB_flight_transfer	0.374274	0.511303	− 0.034523
CNT_COMB_info_req	0.105341	− 0.361095	0.006514
CNT_COMB_repetitions	0.111298	− 0.160867	− 0.073243
CNT_COMB_standby	0.122419	− 0.303003	0.165505
CNT_Conflict_res_phrslgy	0.192960	− 0.144922	0.385264
CNT_Ctrl_events_range	0.297311	− 0.187019	0.083921
CNT_Repet_content	0.154971	− 0.026670	− 0.087393
CNT_Total_control_events	0.362257	− 0.016843	− 0.087998
COM_Ctrl_events_grouping	0.185245	− 0.450183	− 0.375955
COM_Dur_com_pilots	0.370818	0.073192	− 0.213813
COM_Dur_coord_ATCos	0.027521	− 0.024107	0.019683
COM_Freq_com_pilots	0.381184	0.151084	0.033924
COM_Freq_coord_ATCos	0.049758	− 0.080464	− 0.273117
COM_Speech_rate	− 0.039024	− 0.260367	0.207064
PER_Confusions	0.091655	− 0.064945	0.155075
PER_Content_errors	0.115881	− 0.082320	− 0.141835
PER_Hesitations	0.317431	0.220057	0.308790
PER_Speech_errors	0.087857	0.045832	− 0.194744

Appendix D

Table D

Classification metrics from stepwise Logistic Regression for different values of threshold. They are obtained from 5-fold cross-validation. For each value of the threshold the positive class is assigned to values of Occupancy higher or equal to the threshold. The lowest threshold (Occupancy = 8) corresponds to the median of the distribution, the highest corresponds to the 90 % quantile.

Threshold	Accuracy	Precision	Recall	F1 Score	MCC
8	0.750 ± 0.012	0.794 ± 0.017	0.849 ± 0.015	0.820 ± 0.010	0.416 ± 0.024
9	0.743 ± 0.002	0.784 ± 0.012	0.787 ± 0.017	0.784 ± 0.003	0.468 ± 0.006
10	0.723 ± 0.005	0.741 ± 0.014	0.710 ± 0.024	0.723 ± 0.008	0.450 ± 0.008
11	0.764 ± 0.019	0.756 ± 0.017	0.647 ± 0.038	0.696 ± 0.025	0.512 ± 0.039
12	0.784 ± 0.011	0.736 ± 0.022	0.550 ± 0.030	0.628 ± 0.022	0.492 ± 0.027
13	0.809 ± 0.004	0.703 ± 0.024	0.495 ± 0.023	0.578 ± 0.013	0.473 ± 0.012
14	0.863 ± 0.010	0.742 ± 0.057	0.494 ± 0.034	0.586 ± 0.028	0.528 ± 0.034
15	0.886 ± 0.009	0.753 ± 0.051	0.472 ± 0.037	0.571 ± 0.030	0.534 ± 0.028
16	0.898 ± 0.009	0.624 ± 0.051	0.324 ± 0.051	0.408 ± 0.045	0.392 ± 0.032
17	0.927 ± 0.011	0.511 ± 0.133	0.203 ± 0.029	0.275 ± 0.040	0.281 ± 0.057

Data availability

The data that has been used is confidential.

References

Abdesslem, S., Boudes, N., Bressolle, M.C., Capsie, C., Corredor, A.F., Leroux, M., 1999. Le cadre conceptuel pour le recueil et l'analyse des donn´ees [The conceptual frame for data collection and analysis] (No. R98 842/ICC). Centre d'Etudes de la Navigation Aerienne, Toulouse, France.
Ahlstrom, U., Friedman-Berg, F.J., 2006. Using eye movement activity as a correlate of cognitive workload. Int. J. Ind. Ergon. 36 (7), 623–636.
Arico, ` P., Borghini, G., Graziani, I., Imbert, J.P., Granger, G., Benhacene, R., Babiloni, F., 2015. Air-traffic-controllers (ATCO): Neurophysiological analysis of training and workload. Italian J. Aerospace Med. 12, 35.
Arico, ` P., Reynal, M., Di Flumeri, G., Borghini, G., Sciaraffa, N., Imbert, J.P., Babiloni, F., 2019. How neurophysiological measures can be used to enhance the evaluation of remote tower solutions. Front. Hum. Neurosci. 13, 303.
Arnold, J.E., Hudson Kam, C.L., Tanenhaus, M.K., 2007. If you say thee uh you are describing something hard: The on-line attribution of disfluency during reference comprehension. J. Exp. Psychol. Learn. Mem. Cogn. 33, 914.
Ath`enes, S., Averty, P., Puechmorel, S., Delahaye, D., Collet, C., 2002. ATC complexity and controller workload: Trying to bridge the gap. In: Proceedings of the International Conference on HCI in Aeronautics. AAAI Press Cambridge, MA, pp. 56–60.
Averty, P., Athenes, S., Collet, C., Dittmar, A., 2002. Evaluating a new index of mental workload in real ATC situation using psychophysiological measures. In: Proceedings. The 21st Digital Avionics Systems Conference, vol. 2. IEEE, p. 7A4.
Ayaz, H., Willems, B., Bunce, S., Shewokis, P.A., Izzetoglu, K., Hah, S., Onaral, B., 2011. Estimation of cognitive workload during simulated air traffic control using optical brain imaging sensors. In: Foundations of Augmented Cognition. Directing the Future of Adaptive Systems: 6th International Conference, FAC 2011, Held as Part of HCI International 2011, Orlando, FL, USA, July 9–14 Proceedings 6. Springer Berlin Heidelberg, pp. 549–558.
Ayaz, H., Shewokis, P.A., Bunce, S., Izzetoglu, K., Willems, B., Onaral, B., 2012. Optical brain monitoring for operator training and mental workload assessment. Neuroimage 59 (1), 36–47.
Backs, R.W., Navidzadeh, H.T., Xu, X., 2000. Cardiorespiratory Indices of Mental Workload During Simulated Air Traffic Control Vol. 44(13), 89–92.
Balta, E., Psarrakis, A., Vatakis, A., 2024. The effects of increased mental workload of air traffic controllers on time perception: Behavioral and physiological evidence. Appl. Ergon. 115, 104162.https://doi.org/10.1016/j.apergo.2023.104162.
Benoit, A., 1997. Dynamic Control of Air Traffic: Criteria- Control Variables-Constraints. NASA (19980201658).
Bernhardt, K.A., Poltavski, D., Petros, T., Ferraro, F.R., Jorgenson, T., Carlson, C., Iseminger, C., 2019. The effects of dynamic workload and experience on commercially available EEG cognitive state metrics in a high-fidelity air traffic control environment. Appl. Ergon. 77, 83–91.
Betz, S., Bryhadyr, N., Türk, O., Wagner, P., 2023. Cognitive load increases spoken and gestural hesitation frequency. Languages 8 (1), 71.
Brookings, J.B., Wilson, J.F., 1994. Personality and family-environment predictors of self-reported eating attitudes and behaviors. J. Pers. Assess. 63 (2), 313–326.
Brookings, J.B., Wilson, G.F., Swain, C.R., 1996. Psychophysiological responses to changes in workload during simulated air traffic control. Biol. Psychol. 42 (3), 361–377.
Cabon, P., Mollard, R., Cointot, B., Martel, A., Besiot, P., 1997. Elaboration of a method for the assessment of psychophysiological states of ATCs in simulation. EEC Rep.
Canas, ˜ J. J., Ferreira, P., de Frutos, P. L., Puntero, E., Lopez, ´ E., Gomez-Comendador, ´ F., et al., 2019. Mental workload in the explanation of automation effects on ATC performance. In: Human Mental Workload: Models and Applications: Second International Symposium, H-WORKLOAD 2018, Amsterdam, The Netherlands, September 20-21, 2018, Revised Selected Papers 2 (pp. 202–221). Springer International Publishing.
Cardosi, K.M., 1993. Time required for transmission of time-critical air traffic control messages in an en route environment. Int. J. Aviat. Psychol. 3 (4), 303–313.
Christien, R., Benkouar, A., Chaboud, T., & Loubieres, P., 2002. Air traffic complexity indicators & ATC sectors classification. In: Proceedings. The 21st Digital Avionics Systems Conference (vol. 1, pp. 2D3–2D3). IEEE.
Cohen, J., 1992. A power primer. Psychol. Bull. 112 (1), 155–159.
Corradini, P., Cacciari, C., 2002. The effect of workload and workshift on air traffic control: A taxonomy of communicative problems. Cogn. Tech. Work 4, 229–239.
da Silva, F.P., 2014. Mental workload, task demand and driving performance: what relation? Procedia Soc. Behav. Sci. 162, 310–319.
Di Stasi, L.L., Renner, R., Staehr, P., Helmert, J.R., Velichkovsky, B.M., Canas, ˜ J.J., et al., 2010. Saccadic peak velocity sensitivity to variations in mental workload. Aviat.
Space Environ. Med. 81 (4), 413–417. https://doi.org/10.3357/ASEM.2579.2010. Djokic, J., Lorenz, B., Fricke, H., 2010. Air traffic control complexity as workload driver. Transportation research part C: emerging technologies 18 (6), 930–936.
Edwards, T., Homola, J., Mercer, J., Claudatos, L., 2016. Multifactor interactions and the air traffic controller: The interaction of situation awareness and workload in association with automation. IFAC-PapersOnLine 49 (19), 597–602.
Fincannon, T., Ahlstrom, V., 2016. Scale size of the air traffic workload input technique (ATWIT) a review of research. In: Proceedings of the Human Factors and Ergonomics Society Annual Meeting (vol. 60, no. 1, pp. 2078–2082). Sage CA: Los Angeles, CA: SAGE Publications.
Frost, J., 2017. How to interpret R-squared in regression analysis. Statistics by Jim 16. Fürstenau, N., Radüntz, T., 2021. Power law model for subjective mental workload and validation through air traffic control humanintheloop simulation. Cogn. Tech. Work 24, 291–315.https://doi.org/10.1007/s10111-021-00681-0.
Grossberg, M., 1989. Relation of sector complexity to operational errors. Quarterly report of the Federal Aviation Administration's Office of Air Traffic Evaluations and Analysis.
Hamann, A., 2023. A systematic investigation of EEG and fNIRs measures for the assessment of mental workload in the cockpit.
Hancock, P.A., 2017. Whither workload? Mapping a path for its future development. In: Human Mental Workload: Models and Applications: First International Symposium, H-WORKLOAD 2017, Dublin, Ireland, June 28–30, 2017, Revised Selected Papers 1 (pp. 3–17). Springer International Publishing.
Harrison, J., Izzetoglu, K., Ayaz, H., Willems, B., Hah, S., Woo, H., et al., 2013. Human performance assessment study in aviation using functional near infrared spectroscopy. In: Foundations of Augmented Cognition: 7th International Conference, AC 2013, Held as Part of HCI International 2013, Las Vegas, NV, USA, July 21–26, 2013. Proceedings 7 (pp. 433–442). Springer Berlin Heidelberg.
Harrison, J., ˙ Izzetoglu, ˘ K., Ayaz, H., Willems, B., Hah, S., Ahlstrom, U., Onaral, B., 2014. Cognitive workload and learning assessment during the implementation of a next generation air traffic control technology using functional near-infrared spectroscopy. IEEE Trans. Hum.-Mach. Syst. 44 (4), 429–440.
Hedayati, S., Sadeghi-Firoozabadi, V., Bagheri, M., Heidari, M., Sze, N.N., 2021. Evaluating differences in cognitive functions and personality traits among air traffic controllers with and without error history. Saf. Sci. 139, 105208.
Hilburn, B., 2004. Cognitive complexity in air traffic control: A literature review. EEC Note 4 (04), 1–80. Hilburn, B., Jorna, P.G.A.M., 2001. Workload and air traffic control. Stress, workload
and fatigue: theory, research and practice. PA Hancock and PA Desmond (Eds.).
Histon, J.M., Hansman, R.J., 2008. Mitigating complexity in air traffic control: the role of structure-based abstractions (Doctoral dissertation, Massachusetts Institute of Technology, Department of Aeronautics and Astronautics).

Histon, J.M., Hansman, R.J., Gottlieb, B., Kleinwaks, H., Yenson, S., Delahaye, D., Puechmorel, S., 2002. Structural considerations and cognitive complexity in air traffic control. In: Proceedings. The 21st Digital Avionics Systems Conference (vol. 1, pp. 1C2–1C2). IEEE.

Hollnagel, E., 1993. Human reliability analysis. Context and control.
Hyun, S.H., Kim, K.T., Song, S.M., Yoon, J.S., Lee, J.I., Cho, S.M., Sohn, Y.W., 2006. Effects of expertise and situation complexity on visual attention and action planning for air traffic control. In: Proceedings of the Human Factors and Ergonomics Society Annual Meeting (vol. 50, no. 3, pp. 420–423). Sage CA: Los Angeles, CA: SAGE Publications.
John, A.R., Singh, A.K., Do, T.N., Eidels, A., Nalivaiko, E., Gavgani, A.M., et al., 2022. Unraveling the physiological correlates of mental workload variations in tracking and collision prediction tasks. IEEE Trans. Neural Syst. Rehabil. Eng. 30, 770–781. https://doi.org/10.1109/TNSRE.2022.3157446. Epub 2022 Mar 29 PMID: 35259108.
Jordan, C.S., 1992. Experimental study of the effects of an instantaneous self assessment workload recorder on task performance, Defense Evaluation & 484 Research Agency, Report No. DRA/TM(CAD5)/92011.
Kaber, D.B., Perry, C.M., Segall, N., Sheik-Nainar, M.A., 2007. Workload state classification with automation during simulated air traffic control. Int. J. Aviat. Psychol. 17 (4), 371–390.
Kallus, K.W., Van Damme, D., Dittman, A., 1999. Integrated job and task analysis of air traffic controllers: Phase 2. Task analysis of en-route controllers (European Air Traffic Management Programme Rep. No. HUM. ET1. ST01. 1000-REP-04). EUROCONTROL, Brussels, Belgium.
Kirwan, B., Flynn, M., 2002. Towards a controller-based conflict resolution tool: a literature review (European Air Traffic Management Programme Rep. No. ASA. 01. CORA. 2. DEL04-A.LIT). EUROCONTROL, Brussels, Belgium.
Kleinert, M., Helmke, H., Moos, S., Hlousek, P., Windisch, C., Ohneiser, O., ... & Labreuil, A. (2019). Reducing controller workload by automatic speech recognition assisted radar label maintenance. Proceedings of the 9th SESAR Innovation Days, Athens, Greece, 2-5.
Kline, R.B., 2015. Principles and Practice of Structural Equation Modeling, 4th ed. Guilford Press.
Kontogiannis, T., Malakis, S., 2013. Strategies in coping with complexity: Development
of a behavioural marker system for air traffic controllers. Saf. Sci. 57, 27–34. Laybidi, M., Mazloumi, A., Nasl Saraji, J., Gharagozlou, F., Jafari, A.H., Shirzhiyan, Z., et al., 2023. Investigation of relationship between EEG theta power and mental
workload in air traffic control simulation. J. Health Saf. Work 13 (3), 459–473. Lee, P.U., Prevot, T., 2012. Prediction of traffic complexity and controller workload in mixed equipage NextGen environments. In Proceedings of the Human Factors and Ergonomics Society Annual Meeting (vol. 56, no. 1, pp. 100–104). Sage CA: Los Angeles, CA: SAGE Publications.
Leggatt, A., 2005. Validation of the ISA (Instantaneous self Assessment) subjective workload tool. In: Proceedings of the International Conference on Contemporary Ergonomics, pp. 74–78.
Leighbody, G., Beck, J., Amato, T., 1992. An operational evaluation of air traffic controller workload in a simulated en route environment. In: 37th Annual Air Traffic Control Association Conference Proceedings, pp. 122–130.
Li, Q., Ng, K.K., Simon, C.M., Yiu, C.Y., Lyu, M., 2023. Recognising situation awareness associated with different workloads using EEG and eye-tracking features in air traffic control tasks. Knowl.-Based Syst. 260, 110179.
Lille, F., Burnod, Y., 1983. Professional activity and physiological rhythms1. Adv. Biol. Psychiatry 11, 64–71.
Loft, S., Sanderson, P., Neal, A., Mooij, M., 2007. Modeling and predicting mental workload in en route air traffic control: Critical review and broader implications. Hum. Fact. 49 (3), 376–399.
Longo, L., Leva, M.C. (Eds.), 2021. Human Mental Workload: Models and Applications: 5th International Symposium, H-WORKLOAD 2021, Virtual Event, November 24–26,
2021, Proceedings. Springer Nature. Luig, J., Sontacchi, A., 2010. Workload monitoring through speech analysis: Towards a system for air traffic control. Conference: 27th Congress of the International Council
of Aeronautical Sciences. Majumdar, A., Polak, J., 2001. Estimating capacity of Europe's airspace using a simulation model of air traffic controller workload. Transp. Res. Rec. 1744 (1),
30–43. Manning, C., Fox, C., Pfleiderer, E., Mills, S., Mogilka, H., 2002. The relationship between air traffic control communication events and measures of controller taskload and workload. Air Traffic Control Quart. 10 (2), 69–83.
Manning, C., Fox, C., Pfleiderer, E., 2003. Relationships between measures of air traffic controller voice communications, taskload, and traffic complexity. In 5th USA/ Europe Air Traffic Management R&D Seminar.
Marchitto, M., Benedetto, S., Baccino, T., Canas, ˜ J.J., 2016. Air traffic control: Ocular metrics reflect cognitive complexity. Int. J. Ind. Ergon. 54, 120–130.
Martin, C., Cegarra, J., Averty, P., 2011. Analysis of mental workload during en-route air traffic control task execution based on eye-tracking technique. In: Engineering Psychology and Cognitive Ergonomics: 9th International Conference, EPCE 2011, Held as Part of HCI International 2011, Orlando, FL, USA, July 9–14, 2011. Proceedings 9 (pp. 592–597). Springer Berlin Heidelberg.
McGee, J.P., Mavor, A.S., Wickens, C.D. (Eds.), 1997. Flight to the Future: Human Factors in Air Traffic Control. National Academies Press.
Metzger, U., Parasuraman, R., 2001. The role of the air traffic controller in future air traffic management: An empirical study of active control versus passive monitoring. Hum. Fact. 43 (4), 519–528.
Montaci´e, C., Caraty, M.-J., 2014. High-level speech event analysis for cognitive load classification. Interspeech 2014, 731–735.

Morales-Ferre, R., Richter, P., Falletti, E., De La Fuente, A., Lohan, E.S., 2019. A survey on coping with intentional interference in satellite navigation for manned and unmanned aircraft. IEEE Commun. Surv. Tutor. 22 (1), 249–291.

Moray, N. (Ed.), 2013. Mental Workload: Its Theory and Measurement, Vol. 8. Springer Science & Business Media.
Morrow, D., Lee, A., Rodvold, M., 1993. Analysis of problems in routine controller-pilot communication. Int. J. Aviat. Psychol. 3 (4), 285–302.
Munoz-de-Escalona, E., Canas, J.J., 2017. Online measuring of available resources. H-Workload 2017: The First International Symposium on Human Mental Work.
Munoz-de-Escalona, ˜ E., Canas, ˜ J.J., 2018. Latency differences between mental workload measures in detecting workload changes. In: International Symposium on Human Mental Workload: Models and Applications. Springer, Cham, pp. 131–146. https:// doi.org/10.1007/978-3-030-14273-5\_8.
Munoz-de-Escalona, ˜ E., Canas, ˜ J.J., van Nes, J., 2019. Task demand transition rates of change effects on mental workload measures divergence. In: International Symposium on Human Mental Workload: Models and Applications. Springer, Cham, pp. 48–65. https://doi.org/10.1007/978-3-030-32423-0\_4.
Munoz-de-Escalona, ˜ E., Canas, ˜ J.J., Leva, C., Longo, L., 2020. Task demand transition peak point effects on mental workload measures divergence. In: International Symposium on Human Mental Workload: Models and Applications. Springer, Cham, pp. 207–226. https://doi.org/10.1007/978-3-030-62302-9\_13.
Munoz-de-Escalona, ˜ E., Leva, C., de Frutos, P.L., 2023. Developing a first-approach model of Air Traffic Controllers' Mental Workload based on behavioural measures: A Theory For Modelling Air Traffic Controllers' Mental Workload.
Neal, A., Griffin, M., Neale, M., Bamford, E., Boag, C., 1998. Human factors issues in the transition to a CNS/ATM environment: final report (report prepared for AirServices Australia). University of Queensland, Brisbane, Australia.
Nikolova, R., Danev, S., 2011. Effect of mental work load on autonomic, cardiovascular control. Amb 38 (2), 15.
Ohneiser, O., Helmke, H., Shetty, S., Kleinert, M., Ehr, H., Schier-Morgenthal, S., Cern ˇ a, ´ A., 2023. Assistant based speech recognition support for air traffic controllers in a multiple remote tower environment. Aerospace 10 (6), 560.
Omolayo, B.O., Omole, O.C., 2013. Influence of mental workload on job performance. Int. J. Humanit. Soc. Sci. 3 (15), 238–246.
Pawlak, W., Brinton, C., Crouch, K., Lancaster, K., 1996. A framework for the evaluation of air traffic control complexity. In: Guidance, Navigation, and Control Conference, p. 3856.
Porterfield, D.H., 1997. Evaluating controller communication time as a measure of workload. Int. J. Aviat. Psychol. 7 (2), 171–182.
Prandini, M., Piroddi, L., Puechmorel, S., Brazdilova, S.L., 2011. Toward air traffic complexity assessment in new generation air traffic management systems. IEEE transactions on intelligent transportation systems 12 (3), 809–818.
Qiao, H., Zhang, J., Zhang, L., Li, Y., Loft, S., 2021. Exploring the peak- end effects in air traffic controllers' mental workload ratings. Hum. Fact. Ergon. Soc. 64 (8), 1292–1305. https://doi.org/10.1177/0018720821994355.
Radiˇsi´c, T., Andraˇsi, P., Novak, D., Juriˇci´c, B., Antulov-Fantulin, B., 2020. Air traffic complexity as a source of risk in ATM. Risk Assessment in Air Traffic Management.
Radüntz, T., Mühlhausen, T., Fürstenau, N., Cheladze, E., Meffert, B., 2019. Cardiovascular parameters for mental workload detection of air traffic controllers. In: Intelligent Human Systems Integration 2019: Proceedings of the 2nd International Conference on Intelligent Human Systems Integration (IHSI 2019): Integrating People and Intelligent Systems, February 7–10, 2019, San Diego, California, USA. Springer International Publishing, San Diego, California, USA, pp. 789–794.
Radüntz, T., Mühlhausen, T., Freyer, M., Fürstenau, N., Meffert, B., 2021. Cardiovascular biomarkers' inherent timescales in mental workload assessment during simulated air traffic control tasks. Appl. Psychophysiol. Biofeedback 46 (1), 43–59. https://doi. org/10.1007/s10484-020-09490-z.
Rasmussen, J., 1983. Skills, rules, and knowledge; signals, signs, and symbols, and other distinctions in human performance models. IEEE Trans. Syst. Man Cybern. 3, 257–266.
Rose, R.M., Fogg, L.F., 1993. Definition of a responder: analysis of behavioral, cardiovascular, and endocrine responses to varied workload in air traffic controllers. Psychosom. Med. 55 (4), 325–338.
Schmidt, D.K., 1976. On modeling ATC work load and sector capacity. J. Aircr. 13 (7),
Socha, V., Han´ akov´ a, L., Valenta, V., Socha, L., Abela, ´ R., Kuˇsmírek, S., Tecl, J., 2020. Workload assessment of air traffic controllers. Transp. Res. Procedia 51, 243–251.
Su´ arez, M.Z., Vald´es, R.M.A., Moreno, F.P., Jurado, R.D.A., de Frutos, P.M.L., Comendador, V.F.G., 2024. Understanding the research on air traffic controller workload and its implications for safety: A science mapping-based analysis. Saf. Sci. 176, 106545.
Sui, D., Zhang, K., 2022. A tactical conflict detection and resolution method for en route conflicts in trajectory-based operations. J. Adv. Transp. 2022 (1), 9283143.
Triyanti, V., Azis, H.A., Prasetyawan, Y., Iridiastadi, H., Yassierli, 2021. Individual factors related to mental workload in air traffic controller. In: Convergence of Ergonomics and Design: Proceedings of ACED SEANES 2020. Springer International Publishing, pp. 271–278.
Truschzinski, M., Wirzberger, M., 2017. A dynamic process model for predicting workload in an air traffic controller task. CogSci.
Ucl´es, N.R., García, J.M.C., 2014. Relationship between workload and duration of ATC voice communications. 6th International Conference on Research in Air Transportation, Istanbul, Turkey.

Wang, Y., Hu, M., Bellot, P., Vormer, F., Duong, V., 2011. Spatial, temporal, and grouping behaviors in controller communication activities. 9th USA/Europe Air Traffic Management R&D Seminar, Berlin.

Wang, Y., Hu, R., Lin, S., Schultz, M., Delahaye, D., 2021. The impact of automation on air traffic controller's behaviors. Aerospace 8 (9), 260. https://doi.org/10.3390/ aerospace8090260.
Willems, B., Stein, E., 2001. Controller workload and visual scanning during simulations containing scripted midair conflicts and varied task load. Proceedings of the 11th Annual International Aviation Psychology Symposium.
Wilson, G.F., Russell, C.A., 2003. Real-time assessment of mental workload using psychophysiological measures and artificial neural networks. Hum. Fact. 45 (4),
Xie, B., Salvendy, G., 2000. Prediction of mental workload in single and multiple tasks environments. Int. J. Cogn. Ergon. 4 (3), 213–242.
Xiong, R., Wang, Y., Tang, P., Cooke, N.J., Ligda, S.V., Lieber, C.S., Liu, Y., 2023. Predicting separation errors of air traffic controllers through integrated sequence analysis of multimodal behaviour indicators. Adv. Eng. Inf. 55, 101894. https://doi. org/10.1016/j.aei.2023. 101894.
Yazgan, E., Sert, E., S¸ims¸ek, D., 2021. Overview of studies on the cognitive workload of the air traffic controller. Int. J. Aviat. Sci. Technol. 2 (01), 28–36.
Yin, B., Chen, F., 2007. Towards automatic cognitive load measurement from speech analysis. In: Jacko, J.A. (Ed.), Human-Computer Interaction. Interaction Design and Usability. Springer, Berlin and Heidelberg, pp. 1011–1020.
Young, M.S., Brookhuis, K.A., Wickens, C.D., Hancock, P.A., 2015. State of science: mental workload in ergonomics. Ergonomics 58 (1), 1–17.

munozdeescalonaairtrafficcontrollerscommunication2025.md 170 KB Riwayat Mentahan

status: converted

Contents lists available at ScienceDirect

Safety Science

Air traffic controllers communication analysis as a proxy of task demand and mental workload: Using voice recording markers for safety critical task

ARTICLE INFO

Keywords: Air Traffic Management Mental Workload Human Performance & Reliability Behavioural Markers Risk Management Speech-Based Workload Estimation High Mental workload Detection Prediction models Classification models

ABSTRACT

1. Introduction

2. Related work

2.1. Measuring MW in the ATM domain

2.2. Cognitive strategies that adapt workload to task demands

2.3. Inferring ATCos' MW based on how humans' behaviour shifts in response to task-demands

2.4. A simple model to explain the relationship between ATCos' MW and their behavioural patterns

3. Methods

3.1. Case study analysed, materials and instruments

3.1.1. Area control centre ATCos

3.1.3. Voice communications database

3.1.4. Raw data from the voice communications datasets

3.2. From raw data to feature variables

3.2.1. Target variable

3.2.2. Predictor variables

Form of communications

Performance of ATCos' communications

Content errors (PER_Content_errors) An increase in

Content of the communications

Adherence to flight plan

4. Exploratory data analytics and prediction model learning

4.1. The dataset

4.2. Variable association analysis

CE combi.

4.3. Regression models for occupancy prediction

4.3.1. Linear and Poisson regression

4.3.2. Other machine learning models

4.4. Classification models for threshold exceedance

5. Discussion

5.1. Consideration regarding the linear regression model

5.2. Consideration regarding the classification model

5.3. Conclusions, limitations and future directions

Ethics considerations

CRediT authorship contribution statement

Funding

Declaration of competing interest

Acknowledgements

Appendix A

Appendix B

Appendix C

Appendix D

Data availability

References

munozdeescalonaairtrafficcontrollerscommunication2025.md 170 KB

Riwayat Mentahan