Fatigue Modeling As A Tool For Managing Fatigue
Fatigue Modeling As A Tool For Managing Fatigue
Fatigue Modeling As A Tool For Managing Fatigue
LIST OF FIGURES
Figure 1 .........................................................................................................................................4
Figure 2 .........................................................................................................................................5
Figure 3 .........................................................................................................................................6
Figure 4 .........................................................................................................................................8
Figure 5 .......................................................................................................................................11
Figure 6 .......................................................................................................................................14
-3-
PURPOSE
The purpose of the FAST development effort has been to develop a user-friendly,
computerized tool for operational planners and schedulers based on a highly researched and
recognized model of human sleep and cognitive performance. The Fatigue Avoidance
Scheduling Tool (FAST) allows a user to predict cognitive performance efficiency based on the
timing and amount of sleep an individual receives prior to and during the period. For
transportation applications, FAST provides the scheduler the ability to assess fatigue
consequences of alternative schedules.
The effort has built on a newly developed model of sleep and performance invented by
the first author called the Sleep, Activity, Fatigue, and Task Effectiveness (SAFTE) Model [1].
This model predicts human cognitive performance based on 20 years of sleep and circadian
rhythm research. Hursh invented the first sleep and performance model for the Walter Reed
Army Institute of Research and the current model is an advanced modification of that Army
model. The current version of the model makes valid predictions of performance under a broad
range of schedule conditions, from minimal to complete sleep deprivation, at any time of day and
for normal adult subjects ranging in age from the early twenties to mid-fifties. The model is
homeostatic and adjusts its predictions of future performance based on the recent sleep history of
the projected population or specific individuals. In the model, a circadian process influences
both performance and sleep regulation. Sleep regulation is dependent on hours of sleep, hours of
wakefulness, current sleep debt, the circadian process and sleep fragmentation (awakenings
during a period of sleep) that reduce sleep quality. Performance is dependent on the current
balance of the sleep regulation process, the circadian process, and sleep inertia. An additional
benefit of SAFTE is that it can be easily enhanced by future studies to refine fatigue effects on
specific subject populations, specific aspects of operator performance, and the effects of
interventions.
The initial phase of the FAST development effort incorporated the SAFTE Model into a
software tool for scheduling operators to evaluate alternative schedules for their effects on
performance capacity, as degraded by fatigue and circadian variation. The tool incorporates
-4-
interpretive tools for visualizing performance changes over time and the capability to
simultaneously compare multiple schedules on the basis of predicted changes in cognitive
capacity. FAST allows the user to view the effects of pre-programmed and user-defined
sleep/wake schedules on predicted performance effectiveness. The tool provides a simple, user
interface enabling rapid visual and quantitative estimates of the effects of a variety of factors on
the cognitive performance of operators. Figure 1 shows an actual screen from the current FAST
program comparing two schedules simultaneously. Schedules may be viewed in a window, and
two or more windows may be overlaid or tiled for comparison. They may be copied to another
program or directly printed. The tool allows the user to load pre-programmed sleep schedules,
edit them using keyboard and mouse commands, and save edited schedules.
Figure 1: This is a screen image of the FAST main window and shows performance by a railroad
engineer based on a log of sleep and on-duty time. The top window shows predicted
performance based on actual sleep. The bottom window shows potential improved performance
based on the addition of increased sleep during off-duty periods. The dashboard window shows a
ten percent improvement in predicted performance effectiveness at the day and time indicated by
the vertical cursor.
has been implemented, along with export to other programs, such as a spreadsheet or
presentation.
Several options have been developed to aid the interpretation of performance changes.
One valuable option to aid comparison of several schedules is the overlay of a table of interval
statistics. This table shows the average “Performance Effectiveness” for successive hours while
awake and while working. In addition to average effectiveness, the program computes the
percent of time below a selectable criterion, such as 70% effectiveness. The percent of time
below criterion (% BCL) gives an estimate of the time spent at higher risk of error. These tables
can be printed or copied to the clipboard for inclusion in other documents.
Figure 2: This is a screen shot of the dashboard display with FAST showing the levels of five
fatigue factors and five performance metrics. At this time in the schedule (0503 hrs), two fatigue
factors are at dangerous levels (red flags).
A second valuable tool is the fatigue indicators dashboard shown in figure 2. This
window provides a summary of critical fatigue factors operating at any time in a schedule. The
user can place the cursor at any time in the schedule using the graphical screen and the
dashboard will summarize five fatigue factors and indicate if any factors are at a potentially
dangerous level. The fatigue factors are: amount of sleep in the last 24hrs, chronic sleep debt,
number of hours awake, time of day, and circadian phase desynchrony. The dashboard also
displays alternative performance metrics such as lapse likelihood and reaction time changes. The
dashboard feature was specifically added to the FAST software to aid with the analysis of factors
that might contribute to an operator error or accident. Whenever effectiveness declines below
approximately 90%, the dashboard can be consulted for factors that are responsible for that
-6-
decline in performance. This helps to explain the deficit and points to potential strategies to
correct the problem.
SAIC has created an algorithm for shift-work phase adjustment and transmeridian
relocation within the SAFTE model. The model contains logic to detect the change in
work/sleep patterns and to readjust the phase of the circadian rhythm depending on whether the
new pattern is indicative of a change in time zone or shift in work schedule (shift rotations). The
phase adjustment feature permits the circadian process to predict “Jet Lag” based on travel from
east to west and west to east, illustrated in Figure 3. This feature also permits the software to
properly adjust the circadian rhythm for shift-work schedules typical of many transportation and
industrial operations.
Figure 3: Disruptions of performance following eastward (upper pane)l and westward (lower
panel) travel across 6 time zones. Note that the model predicts greater on the job disruptions
(red portion of line) following eastward travel and a longer period of adjustment.
For aviation applications, the software can display waypoints along the travel route. The
program computes lighting conditions and approximate interpolated geographic positions. The
user can print a Mission Timeline to guide the crew during the performance of the flight.
-7-
The general architecture of the current SAFTE model is shown in Figure 4. A circadian
process influences both performance and sleep regulation. Sleep regulation is dependent on
hours of sleep, hours of wakefulness, current sleep debt, the circadian process, and fragmentation
(awakenings during a period of sleep). Performance is dependent on the current balance of the
-8-
sleep regulation process, the circadian process, and sleep inertia. Although developed
independently, the resulting model has structural similarity to the scheme suggested by
Acherman and Borbely [3] and when the simulation is integrated over time approximates
(ignoring circadian influences in the model) the mathematics of the homeostatic model of
Folkard and Akerstedt [4]. However, the new model has been optimized to predict changes in
cognitive performance and incorporates features not included in any prior comprehensive model.
These features are: a multi-oscillator circadian process, a circadian sleep propensity process, a
sleep fragmentation process, and a circadian phase adjusting feature for time zone changes. Each
component will be discussed in detail.
ACTIVITY
CIRCADIAN OSCILLATORS ADJUSTED
SLEEP
PHASE
REGULATION 12
PERFORMANCE
MODULATION
SLEEP DEBT
FEEDBACK EFFECTIVENESS
LOOP
SLEEP INTENSITY
SLEEP ACCUMULATION
(Reservoir Fill) SLEEP RESERVOIR INERTIA
SLEEP “QUALITY”
FRAGMENTATION
PERFORMANCE USE
(Reservoir Depletion)
Biomedical Modeling and Analysis Program
Circadian Oscillators
Performance while awake and the drive to sleep are both controlled, in part, by a
circadian process [4], [5]. Performance and alertness reach a major peak in the early evening,
about 2000 hours, and fall to a minimum at about 0400 hours. There is a secondary minimum in
the early afternoon, about 1400 hours, and a secondary morning peak at about 1000 hours.
Correlated with this pattern is a rising tendency to fall asleep that reaches a peak at about the
same time performance and alertness reach their minima. The existence of both a major and a
minor peak in performance and two corresponding minima at other times suggests that at least
two oscillators are involved in the circadian process.
-9-
The sleep and performance model incorporates a circadian process that is composed of
the sum of two cosine waves, one with a period of 24 hours and one with a period of 12 hours.
This arousal oscillator drives both variations in predicted cognitive effectiveness and sleep
propensity. These two translations of the oscillator have identical frequency and phase
components and differ only in amplitude and sign; a rise in arousal produces an increase in
performance and a decrease in propensity to sleep. The circadian process is depicted in the large
rectangle shown in the diagram of the SAFTE model, Figure 4. In addition, based on
observations that the amplitude of circadian variation increased with hours of sleep deprivation,
the amplitude of the performance rhythm is a linear function that increases from a minimum to a
maximum depending on the level of sleep debt (reservoir capacity minus current reservoir level).
Cognitive Effectiveness
Consistent with the approach proposed by [5] and [3], the SAFTE model stipulates that
cognitive effectiveness and alertness are primarily dependent on variations in the two processes
just described: the endogenous circadian rhythm (reflected in body temperature) and current
sleep reservoir balance resulting from the sleep-wake cycle. A third factor, called sleep inertia,
is the temporary disturbance in performance that often occurs immediately following awakening,
see [10]. The predictions of the model are normally in terms of changes from cognitive
effectiveness, expressed as percent of baseline performance when well rested. This measure
corresponds to performance on a psychomotor vigilance task. In addition, the parameters of the
- 11 -
Equilibrium States
A homeostatic representation of sleep regulation leads to an important implication: if a
subject is scheduled to take less than an optimal amount of sleep each night, for example, four
hours per day, the reservoir initially loses more units during the awake period than are made up
during the sleep period. This results in a sleep debt at the end of the sleep period that
accumulates over days. However, since the rate of sleep accumulation increases with sleep debt,
eventually, the rate of sleep accumulation increases such that four hours of sleep makes up for
twenty hours awake. At this point, the reservoir reaches an equilibrium state and no further debt
is accumulated, although the initial deficit remains as long as the person remains on this
schedule. The sleep homeostat is not infinitely elastic; any schedule that provides less than 4
hours of sleep per day (for the average person) will not reach an equilibrium state and
performance capacity will gradually deplete to zero.
- 13 -
Sleep Timing
The model is sensitive to the time of day of the sleep period. For an individual given
eight hours of sleep per day, starting at 1200 hours (noon) each day, performance reaches a peak
of 100 % at the start of each awake period (2000 hours); performance then rapidly declines
during the late night and early morning hours to a strong dip at about 0400 hours. Minimum
predicted performance could be as low as 66% compared to minimum performance under a
normal sleep schedule of 90%. This alteration in pattern results from two factors. First, sleep
intensity is initially less for sleep periods starting at noon. This results in a small accumulated
debt that is quickly offset by the homeostatic sleep mechanism. The second, more persistent
effect is the circadian oscillation of performance that reaches its minimum in the early morning
hours. This pattern has strong implications for performance under shift schedules that require
daytime sleep. It is well documented that most mistakes on the night shift occur during the early
morning hours ([16], [17], and [18]).
PVT Speed
Actual Sleep Times - Revised SAFTE Model
110
95
(as a % of Baseline)
Mean Speed 9 Hr
80 7 Hr
5 Hr
3 Hr
65
SAFTE/FAST
R2 = 0.94
50
0 T1 T2 B E1 E2 E3 E4 E5 E6 E7 R1 R2 R3
Day
Figure 6: Fit of the current SAFTE model to the PVT results of the sleep dose response study
based on actual sleep durations (Balkin, et al., 2000).
Often demanding civilian schedules provide less than the optimal eight hours of sleep a
day for extended periods of time. These schedules provided chronic restricted amounts of sleep.
A recent study of chronic sleep restriction conducted at the Walter Reed Army Institute of
Research in cooperation with the Department of Transportation provided data on schedules of
seven, five, and three hours of time in bed over seven days [20]. The latest version of the
SAFTE Model predicts both the performance degradation effects and rate of recovery from
those schedules with an R2 of 0.94. The data are shown as symbols in Figure 6 and the
predictions of the model are shown as the heavy lines. The first three points were from baseline
days with eight hours time in bed; the next seven points were from the experimental days with
time in bed set to the values shown in the legend; the last three days were with recovery sleep of
eight hours time in bed.
One important outcome of this study was a quantification of individual variability in
sensitivity to sleep restriction. Based on that finding, FAST can plot a line that represents some
lower bound of the population variance, such as the lowest 20% of subjects. FAST may be the
only fatigue model that estimates prediction error based on population variance.
- 15 -
of 700 human factors accidents and a similar size sample of nonhuman factors accidents will be
analyzed, as well as a similar sample of non-alcohol or drug related engineer decertification
events. This large sample of about 2100 accidents and events should provide sufficient statistical
power to 1) verify if a fatigue model based solely on work schedule data can detect a tendency
for human factors accidents to occur with a higher likelihood when performance is predicted to
be low (predictive validation), and 2) determine at what level of performance this tendency
occurs (calibration).
The final initiative underway is to provide a standard human factors fatigue questionnaire
to aid accident investigators in gathering the multiple threads of information necessary to
evaluate fatigue as a contributing factor. The system will permit the investigator to create a data
base of information for each accident. As the investigator gathers data and interviews workers
and witnesses, the program prompts the investigator with questions and permits electronic
recording of answers. The program guides the investigator to ask about the person’s typical
sleep and rest patterns, about medical conditions and medications being taken, and about the
particular events leading up to the accident under investigation. The focus is on garnering data
necessary to do a fatigue analysis and should be administered to subjects close to the time of the
accident when memory for events prior to the accident is fresh. The program organizes all the
responses into a searchable data base and ports the schedule information into FAST for fatigue
analysis.
CONCLUSIONS
Fatigue management in transportation should be an iterative process involving all the
stakeholders in the problem: management, labor, and government regulators. At each stage in
the process – analysis, understanding, commitment, change, and evaluation – tools are needed to
objectively assess potential fatigue and design alternatives that work to reduce that fatigue. A
major new initiative involves using mathematical models of fatigue to serve as that objective
metric. The models take information about work schedules and sleep schedules, if available, and
project the impact of sleep duration and timing on cognitive capacity at different times of the
day. When only work schedule data are available, another algorithm is used to estimate likely
sleep patterns under the work schedule. The US Department of Defense has sponsored the
development of such a model – the SAFTE model – and a software tool based on the model –
the Fatigue Avoidance Scheduling Tool, or FAST – that has been validated against laboratory
measures of cognitive performance. The Federal Railroad Administration has sponsored work to
enhance the FAST software with several features to adapt it to the transportation environment
and to collect data to test the ability of a fatigue model such as SAFTE/FAST to predict fatigue
related accidents. Studies are underway to test the predictive validity of the tool and to calibrate
it for fatigue related railway accidents and incidents. Other projects will develop associated tools
to improve usability and interoperability with other scheduling software.
- 18 -
REFERENCES
1. Hursh, S.R. (1998) Modeling Sleep and Performance within the Integrated Unit
Simulation System (IUSS). Final Report for the United States Army Soldier Systems
Command; Natick Research, Development and Engineering Center, Natick, Massachusetts
01760-5020; Science and Technology Directorate; Technical Report: Natick/TR-98/026L.
2. Pollard, J. K. (1996) “Locomotive Engineer’s Activity Diary.” U.S. Department of
Transportation, Federal Railroad Administration, Office of Policy and Program
Development and Office of Research and Development, Washington, D.C. 20590.
3. Achermann, P. and A. Borbely, (1991). Combining different models of sleep
regulation. Journal of Sleep Research, 1, 144-147. Borbely, A. and P. Achermann, (1992).
Concepts and models of sleep regulation: an overview, Journal of Sleep Research, 1, 63-79.
4. Folkard, S. and T. Akerstedt (1991). A three process model of the regulation of
alertness and sleepiness. In Ogilvie R., Broughton R., eds. Sleep, arousal and performance
problems and promises. Boston: Birkhauser, 11, 26.
5. Monk, T. (1982). The arousal model of time of day effects in human performance
efficiency. Chronobiologia, 49-54.
6. Klein, K. Wegmann, H., Athanassenas G. et al. (1976). Air operations and circadian
performance rhythms. Aviation Space and Environmental Medicine. 47(3), 221-229.
7. Graeber, R.C. (1980). Recent studies relative to the airlifting of military units across
time zones. In L.E. Scheving and F. Halberg (Ed.), Chronobiology: Principles and
Applications to Shifts in Schedules (pp. 353-370). Rockville, MD: Sijthoff & Noordhoff.
8. Haus & Halberg (1980). The circadian time structure. In L.E. Scheving & F. Halberg,
eds., Chronobiology: Principals & Applications to Shifts in Schedules. (pp. 47-94).
Rockfield:MD Sijthoff & Noordhoff.
9. Akerstedt, T. and S. Folkard, (1991). Validation of the S and C Components of the
Three-Process Model of Alertness Regulation. Sleep, 19(18), 1-6.
10. Folkard, S. and T. Akerstedt (1987). Towards a Model for the prediction of alertness
and/or fatigue on different sleep/wake schedules in Contemporary Advances in Shiftwork
Research (Eds. A. Oginski, J. Pokorski & J. Rutenfranz), 231-240, Medical Academy,
Krakow.
11. Monk, T. (1991). Circadian aspects of subjective sleepiness: a behavioral messenger?
In T. Monk (Ed) Sleep, Sleepiness and Performance. Wiley, Chinchester, 1991, 39-63.
12. Monk, T. and D. Embry (1981). A field study of circadian rhythms in actual and
interpolated task performance. In Night and Shift Work: Biological and Social Aspects (Eds.
A. Reinberg, N. Vieux & P. Andlauer), 473-480, Pergamon Press, Oxford.
13. Lavie, P. (1991). The 24-hour sleep propensity function (SPF): practical and
theoretical implications. In: T. Monk (Ed) Sleep, Sleepiness and Performance. Wiley,
Chinchester, 1991: 65-93.
14. Voigt, E., P. Engel, & H. Klein (1968). Int. Z. angew. Physiol. einschl. Arbetisphysiol.,
25, 1-12.
15. Folkard, S. and T. Monk (1979). Shiftwork and performance. Human Factors, 21,
483-492.
16. Harris, W. (1977). Fatigue, circadian rhythm and truck accidents in Vigilance (Ed. R.
Mackie), 133-147, Plenum Press, New York.
- 19 -
17. Hamelin, P. (1987). Lorry drivers times habits in work and their involvement in traffic
accidents. Ergonomics, 30, 1323-1333. [19] Angus, R. & R. Heslegrave (1985). Effects of
sleep loss on sustained cognitive performance during a command and control simulation.
Behavior Research Methods, Instruments, & Computers, 17(1), 55-67.
18. Ribak, J., I. Ashkenzia, A. Klepfish, et. al (1983). Diurnal rhythmicity and Airforce
flight accidents due to pilot error, Aviation, Space, and Environmental Medicine, 54, 1096-
1099.
19. Angus, R. and R. Heslegrave (1985). Effects of sleep loss on sustained cognitive
performance during a command and control simulation. Behavior Research Methods,
Instruments, and Computers, 17(1), 55-67.
20. Balkin, T., Thorne, D., Sing, H., Thomas, M., Redmond, D., Wesensten, N., Williams,
J., Hall, S., and Belenky, G. (2000). Effects of sleep schedules on commercial driver
performance. (Report No. DOT-MC-00-133). Washington, DC: U.S. Department of
Transportation, Federal Motor Carrier Safety Administration.