Aeronautical medical fitness assessment

01 Problem and context

Determining a pilot's medical fitness requires cross-referencing clinical and psychological evaluations against extensive national and international aeronautical and medical regulations. It is slow, fragmented, and high-stakes: an error carries clinical and legal consequences. Furthermore, sending sensitive medical data to an external AI service is unacceptable.

Aeromedical certification decides who is cleared to fly and under what conditions. The judgment is not purely clinical: it cross-references lab results and psychological evaluations against DGAC Chile and ICAO rules, and a mistake carries clinical and legal weight at the same time. It is also a domain where the data is sensitive by definition and where the safety culture holds that automation never gets the final word. The problem was identified from inside it: years operating professional flight simulators with real procedures surfaced a friction an outside engineer would not have seen.

02 Solution

A system that ingests medical tests and real-time sensor signals, reasoning via a model grounded in a dataset built from the regulations. Every verdict includes its source and justification, not a model opinion, along with associated risks and specialist referrals.

01 Research · 02 Dataset · 03 Training · 04 Testing · 05 Web application · 06 Validation

03 Results

Verifiable figures in the repository and its documentation:

04 How it's built

A decision-support system that evaluates pilot medical fitness. It ingests lab test results, psychological evaluations, and real-time sensor signals (blood pressure, heart rate, oxygenation) and reasons using a model anchored to a dataset built from national and international aviation and medical regulations. It delivers a verdict (fit / fit with restrictions / unfit) complete with sources, rationale, associated risks, and specialist referrals. It has commercial potential and is currently under discussion with the DGAC.

05 Ethical decisions

• ✦The pilot decides, with one toggle, whether their medical data is stored or automatically wiped at logout. Opt out and the system analyzes the data, returns the result, and deletes everything.
• ✦Even when the verdict is «fit», the system still refers the pilot to a named specialist: an algorithm should not be the last word on someone's body.
• ✦History exists only because the pilot asked for it. Whoever opts in signs back in and compares progress against earlier results; nothing is retained by default.
• ✦No verdict travels alone: it carries its regulatory source, its rationale, and the associated risks.

06 Highlights

• ✦Multimodal ingestion: lab tests and psychological evaluations + real-time sensor signals (blood pressure, heart rate, oxygenation).
• ✦Model grounded in a dataset built from national and international aviation and medical regulations; every response includes its source and rationale.
• ✦Comprehensive output: verdict, associated risks, and specialist referrals, not just an isolated number.
• ✦Six-phase pipeline: research, dataset, training, testing, web application, and validation.
• ✦The project evaluator established contact with DGAC, and discussions with the agency are currently underway.

07 Limitations

The dataset was built from aviation and medical regulation. It is not a corpus of real pilot clinical records, since open aeromedical data barely exists. That sets the ceiling: the system reasons over regulatory criteria and does not replace the judgment of an aviation medical examiner. Sensor signals come from consumer devices, whose accuracy is not clinical grade; they work as trend and context, not diagnosis. Regulation changes, so the dataset needs versioning and periodic review. And validation with DGAC is still in progress: none of this amounts to approval by the agency yet.