24Hr emergency line +44 (0)20 8476 4099
Join Balpa

Can software predict fatigue?

by Dr Andy Whale BALPA Biomathematical Modeller

Our members tell us consistently that fatigue is one of the top concerns that they have, both in terms of the potential consequences for flight safety, and for the general health and wellbeing of pilots. Aviation probably represents one of the most challenging environments for the management of fatigue. While pilots’ working hours are regulated by prescriptive limits on the number of permissible working and flying hours in a given time scale, these limits do not prevent the creation of rosters that are likely to be grossly fatiguing. 

In addition to the prescriptive rules, the regulations also contain an overarching requirement that airlines should not assign pilots rosters that are likely to be dangerously fatiguing, and that pilots should not fly when they are experiencing a potentially unsafe level of fatigue. In order to address this second point of regulation, more and more airlines are turning to software packages that claim to be able to determine, with scientific accuracy, how fatiguing a roster will be. But how do these systems work, and can they really determine how fatigued a pilot will be during a given roster? 

There are around half a dozen commercial systems currently on the market, which can integrate with airlines’ existing rostering systems. The Australian regulator (CASA) conducted a comprehensive review of these systems in 2007– with one exception (which no airline in the UK currently uses), all of these systems are based on a scientific model of sleep regulation, known as the ‘two-process model’ (occasionally the three-process model).

This model was first developed in the 1980s to predict when people were most likely to sleep and awaken, however it was soon apparent that the model could also be used to estimate fatigue and its consequences. In brief, a pilot’s roster is entered into the model as a series of work duties, specifying the timing and location of duties, including any time zone changes. The model will then make a series of predictions about when a person flying those duties is most likely to sleep, and for how long. Likely fatigue levels are then estimated throughout the roster, based on the pilot’s estimated sleep-wake history and their circadian rhythm. These are the two processes that are most reliably correlated with fatigue and is where the two-process model gets its name (a third process, sleep inertia, is sometimes included). 

While the commercial systems that airlines use are based on these scientific principles, they each conduct their own ongoing testing and validation of their own system, and make some adjustments to the model, such as an adjustment for workload, based on the number of sectors a pilot has flown that day. This presents a difficulty, since the commercial providers are generally reluctant to share the nature of these adjustments, or even the methods and data they have used to calculate them. 

While it is easy to understand the commercial justification for this stance, it makes it impossible for these systems to be independently checked for accuracy. Given that these systems are increasingly being used to inform decisions that could have safety implications, this lack of independent verification represents a major concern. 

BALPA also has reservations about the level of understanding of how these systems work within airlines. We have spoken to a number of pilots, who have reported a fatigue concern to their airline, only to be met by the response ‘the fatigue model says the roster is not fatiguing, so you should be fine’, occasionally accompanied by the more sinister assertion that, since the model says the roster is not fatiguing, this indicates that the individual in question must be mis-managing their rest. 

Such thinking represents a gross misunderstanding of how fatigue predictions models work, and the scientific principles upon which they are based. By their very nature, models make predictions based on what the ‘average’ person is most likely to experience under a specific set of circumstances. Such predictions should never be assumed to be valid for a specific individual, since everyone will experience fatigue differently. It is also important to bear in mind where the limits of fatigue predictions are; while the two processes of sleep-wake history and circadian rhythm are the most consistent, reliable, and more importantly, predictable contributors to fatigue, there a myriad of other factors that could influence how fatigued a given individual is at any given moment, such as genetics, medical issues, and even whether an individual has a young child at home. Predictive fatigue models are unlikely to ever have the fidelity to account for such factors. 

In answer to the question, software can give an indication of how much fatigue is likely for a given roster and could be used to provide an overall estimate of the likely levels of fatigue present in an airline’s schedule (e.g. over the course of a season). They should never be used however, to make a statement regarding the level of fatigue a specific individual is, or should, experience for a given roster, and they should certainly never be treated as equal to a pilots’ own judgement about their own level of fatigue, and whether it presents a flight safety risk. 

Both in scientific terms, and in the legislation, the only person in a position to make that determination is the pilot themselves.