NASA OSDR OSD-575PMData wearables

What we tested and what is still unproven.

Real NASA blood data and a simulation you can rerun.

–of healthy simulated days flagged by fixed Earth-style rules (five-scenario average)
–of days flagged by AEROVYNT action alerts.
–of SIMULATED events caught at Watch level
–real NASA post-flight lab flags vs the number expected by chance

Simulation shows trade-offs, not clinical accuracy. NASA labs: 4 crew, 3 days; 168 of 168 values copied correctly from NASA’s file (a data-integrity check, not a validation of the engine). Wearables: 13 people (PMData).

NASA Inspiration4 blood panels

Personal baseline vs lab range

Do personal baselines see changes that population ranges miss?

A post-flight draw counts as a personal change when it moves more than the 95% reference change value from that person’s pre-flight mean. Any such rule flags some draws by chance, so it was also run on 20,000 simulated studies with no real change.

Real wearables (PMData)

Simulation benchmark SIMULATED

120 crews per scenario, 140 days each, half with an injected event. Fixed Earth-style rules vs AEROVYNT; brackets are 95% bootstrap intervals. Reproduce: python aerovynt_engine.py in research/.

Evidence ladder
  1. 1 · Simulation benchmarksReached. 120 crews per scenario, five scenarios, stress tests on fresh seeds. Shows trade-offs under assumptions we defined.
  2. 2 · Real wearable noise checkReached. 13 people from PMData (no illness labels) used to calibrate noise and bound false alarms.
  3. 3 · Real NASA lab dataReached, with a caveat. Inspiration4 OSD-575: 4 crew, 3-day flight. Shows the mechanism; cannot settle detection.
  4. 4 · Clinician reviewNot reached. The reviewer kit is ready; reviews appear below with role and date.
  5. 5 · Replay on illness-labelled dataNot reached. research/aerovynt_replay.py exists and has not been run on real labelled data.
  6. 6 · Analog mission or bed-rest trialNot reached. This would show the Co-Pilot works with real crews.
Clinician and expert feedback
Limitations
1
No validation on real illness labels. Real data is unlabelled (PMData) or 4 crew over 3 days (OSD-575). Detection is shown only on simulated events.
2
Simulated benchmarks. Event shapes are ours: rates show trade-offs, not real-world accuracy. Slow 25–45 day cycles remain a weakness. The stress indicator was added after the benchmark and is not part of it.
3
Thresholds were set on the simulator. The engine’s limits were frozen using the same simulator that produces the benchmark, so benchmark detection rates are optimistic. The simulator’s sleep noise (0.35 h) is lower than real wearable users (median 0.93 h); a stress test that resamples real PMData sleep noise is included, and detection there is 30–99% by scenario (43% for subtle, 30% for slow drift). With slow 25–45 day cycles, or all stressors at once, AEROVYNT action false alarms rise to about 35–40% and the plain personal z-score (about 11%) is better.
3b
Illustrative thresholds. Limits follow common clinical conventions; the fast/slow clock rule is a design choice, not clinically tested.
4
Not a medical device. It never replaces the Flight Surgeon. Demo days 1–28 are simulated; entries are processed locally.
➜
Next: replay on infection-labelled wearable data, flight-surgeon review.
Clinical review · none yet

No flight surgeon has reviewed this prototype. The reviewer sheet lists every limit and crew-facing sentence with blank verdict columns.

Sources: NASA OSDR study OSD-575 (SpaceX Inspiration4), file LSDS-8 Comprehensive Metabolic Panel; Sanders et al. (2024) npj Microgravity 10:56, doi:10.1038/s41526-024-00393-5. OSDR Inspiration4 records are U.S. Government Works. PMData: Thambawita et al. (2020), datasets.simula.no/pmdata, CC BY 4.0. Barger et al. (2014) Lancet Neurology 13:904–912. Van Dongen et al. (2003) Sleep 26:117–126. Watson et al. (2015) Sleep 38:843–844.

Build 15 Nov 2026 · AEROVYNT · NASA Space Apps Challenge · Research prototype, not medical advice · Team: Muhammad Salman and Sayed Hamza · Built with SI assistance (Claude, Anthropic), see SI_DISCLOSURE.md