What these devices detect
Fatigue detection devices look for signals consistent with reduced alertness during an activity. NIOSH distinguishes tools that predict future fatigue from information such as prior rest from those detecting possible current manifestations through biological or performance measures. An isolated reading is not a medical fitness assessment.
This entry covers eye-tracking cameras, wearable devices and systems combining information from a vehicle or workstation. Their preventive value depends on a specific question: what situation will they help identify, and what safe decision can follow? Buying a sensor without arranging an organisational response leaves the intervention incomplete. The useful outcome is better control of exposure, rather than the accumulation of individual scores. Selection should therefore start with the work problem and the people able to change it.
Sources: [1]
Fatigue, sleepiness and distraction are different labels
Before comparing products, ask for an operational definition of each warning. One system may classify eye closure, another gaze deviation and another a driving pattern. Grouping them under a single fatigue label makes the information harder to interpret. Contracts and training should preserve the name of the observed variable and distinguish it from the proposed explanation.
As an editorial assessment criterion, a display should distinguish a valid signal, an insufficient signal and an interpreted alert. Losing sight of a face means information is missing; it does not establish that the person is rested. Nor should it automatically become a violation attributed to the driver. Reviewing the episode requires context, the software version and an effective opportunity to challenge an interpretation.
Evidence and validation before procurement
NIOSH recommends examining validity, reliability, generalisability and the ability to distinguish true from false readings. It also warns against making technology the primary measure for reducing fatigue. Assessment needs to match the intended operating conditions, rather than relying solely on a sales demonstration.
A proposed procurement dossier can request the study population, task, lighting conditions, reference criterion and results separated by error type. Ask whether evaluation data concern people different from those used to develop the algorithm. An overall accuracy figure is insufficient for deciding what happens when a system misses a dangerous episode or repeatedly interrupts safe work. Uncertainties should remain visible in the technical acceptance record, together with responsibilities for investigating unexpected results after installation.
Sources: [1]
Eye tracking and wearable devices
For eye tracking, document precisely what is recorded: images, gaze positions, closure durations or a calculated index. For a wearable device, establish which sensor feeds the result, how incorrect placement is identified and what happens when the battery or connection fails. Capabilities should not be attributed to every device simply because they share a commercial category.
For a pilot, we propose a matrix of workplace situations: different permitted glasses, normal lighting changes, movements needed for the job and compatible protective equipment. The purpose is to test the system, not to induce sleep deprivation or expose fatigued people to dangerous tasks. Trials that could affect safety need a controlled setting, competent responsibility and predefined stopping criteria. Missing or unusable measurements should remain identifiable in the resulting dataset.
Sources: [1]
Organising a response within a fleet
Australia’s NHVR highlights the relationship of trust between company and driver and the practical ability to stop driving in its 2024 guidance announcement. This reference concerns Australian heavy transport; it does not replace other countries’ obligations or certify a particular product.
As a preventive application, a fleet procedure should identify who receives a warning, how a safe stop is supported in the circumstances and how help is obtained without pressure to continue. It should also address a driver’s own report of tiredness when the device gives no warning. Delivery planning, relief arrangements and available rest opportunities belong to the solution. A message requiring the driver to make up the stopping time afterwards undermines the protective purpose of the response.
Sources: [3]
Applications for operators and air traffic controllers
The joint ICAO, CANSO and IFATCA guide for air traffic service providers integrates measurement, work organisation and fatigue risk management. Its measurement appendix considers different information sources; it does not provide blanket approval to replace professional judgement with a camera.
Transferring a road detector to a control room requires a fresh assessment. Looking successively at several displays may be part of correct task performance. A project needs to examine concurrent alarms, team coordination and handovers before assigning meaning to gaze. The same applies to necessary inspection movements in an industrial workstation. This entry proposes treating such applications as specific uses requiring their own evidence, rather than automatic extensions of vehicle validation. Neither an industrial nor an aviation application should inherit an unsupported assurance of effectiveness.
Sources: [4]
Privacy and decisions affecting a person
The Spanish data protection authority AEPD explains safeguards against solely automated decisions with legal or similarly significant effects, including decisions involving performance and health. Legal exceptions do not make every workplace processing activity permissible. The analysis must concern the actual purpose and data involved.
In the proposed design, prevention staff, data protection specialists and worker representatives agree what information is necessary, who can access it and for how long. Operational warnings should be distinguished from clinical information and disciplinary assessment. A device score should not function as a medical certificate or an automatic verdict on an individual. Disagreements require meaningful human review, with enough information to correct errors and explain the decision. Access should reflect responsibilities rather than simple technical availability.
Sources: [5]
A pilot with useful acceptance criteria
NIOSH recommends defining objectives before choosing technology and using available information to guide its deployment. An objective may be to assess a preventive intervention rather than simply increase the number of sensors.
We propose starting with a written decision about the problem, scope and expected response. Acceptance criteria can then cover signal quality, availability, reviewable warnings and the ability to activate protective measures. Record a comparable baseline and invite reports of usability problems. The review team should be able to suspend the pilot if it creates distraction, disproportionate surveillance or pressure to work while tired. A significant algorithm update requires renewed checks of aspects that may change; purchasing the system does not close the assessment. Record the reasons for both continuation and suspension decisions.
Sources: [2]
Hypothetical example of organisational improvement
A logistics company observes repeated warnings during the final part of several overnight routes. In this hypothetical example, it initially does not know whether they reflect tiredness, camera adjustment problems or both. It reviews a sample with drivers, identifies signal losses associated with one installation and keeps these episodes separate from confirmed alerts.
The company also changes a schedule that accumulated waiting time before deliveries and organises support to stop or hand over the service. It then compares similar routes and periods, including working hours and system availability. It does not claim to have cured fatigue or that every decline in warnings proves lower risk. The learning lies in connecting technical information, workers’ experience and verifiable changes to work organisation, while retaining uncertainty about results that cannot yet be explained.
Indicators for learning without punishment
A monitoring dashboard can show the proportion of time with a usable signal, warnings reviewed, time taken to receive support and completed organisational measures. Totals should be accompanied by the relevant exposure: driving hours, shifts or operating time. Ten warnings in a hundred hours and ten in ten thousand hours lead to different interpretations.
We also propose recording voluntary reports, challenges to erroneous results and real opportunities for rest. A fall in warnings accompanied by disconnected cameras is not an improvement. An initial rise in reports may reflect greater trust. Aggregate analysis should seek situations that can be corrected, avoiding public rankings of the most fatigued drivers. Incidents need independent investigation: an alarm record alone does not establish the cause of a crash.
Relationship to OSH and management records
Technological detection can inform fatigue management, occupational road safety and cognitive ergonomics when responsibilities are clear. This entry connects with wearables, computer vision, privacy and professional driver risks. It does not replace workstation assessment, working-time planning or healthcare when needed.
A management system can document the pilot purpose, equipment version, accepted limitations, training, reviews and corrective actions. Before integrating data, confirm which functions the chosen platform actually provides. This proposal does not attribute a particular biometric detector to Sabentis. Completion should be judged by whether people receive support and working conditions are corrected; obtaining a new chart is only an intermediate result. Keeping the original question visible prevents a technical deployment from becoming an end in itself.
Frequently asked questions
Does a wearable determine medical fitness?
No. Its output is not a clinical assessment and should not become an automatic certificate of fitness.
Does having no warnings make tired driving safe?
No. Absence of an alert does not demonstrate absence of risk; people need a route to report tiredness and receive support.
Can the same algorithm cover driving and air traffic control?
That cannot be assumed. Tasks, movements and consequences differ; each use requires its own evidence and assessment.
What is a useful pilot outcome?
A verifiable protective response, usable data, reviewable errors and improvements to work, rather than simply more warnings.
