ASCLEPIUS

Autonomous crew health monitoring for deep-space missions
–

Crew status

Cabin environmentECLSS

Space radiation · liveNOAA GOES

Loading live GOES proton flux…

ISS position · live–

Shaded ellipse: South Atlantic Anomaly, where trapped-proton dose rates spike.

Vitals vs. personalised expected band (±2σ)last 4 h

The shaded band is what this astronaut's body should be doing right now. It is built from their own baseline, the crew timeline (sleep, work, exercise, EVA) and their circadian phase. Alerts fire on deviation from the band, not on fixed limits, so a 150 bpm workout is normal but 95 bpm at rest can be an early warning.

Onboard differential diagnosissignature match

NASA evidence for this findingNASA NTRS

Long-duration risk outlookmodel estimates

Autonomy

–
–

Alerts & onboard actions

No alerts. Inject a medical event from the top bar to see the system respond.

Does it work? Monte Carlo validation

Each trial simulates one astronaut's 10-hour shift with a random physiology, a randomly timed exercise session and realistic sensor noise. In 60% of trials a medical event of random type is injected at a random time with random severity (50–110% of nominal). The rest are healthy controls. ASCLEPIUS sees the sensor stream exactly as it would in flight. Its calibration, time constants and circadian phase are deliberately imperfect, so it never knows the true physiology. We compare it with standard fixed alarm limits (HR >120 or <40, SpO₂ <92%, RR >24, T >38.0 °C, SBP <90 or >160), muted during exercise to be fair.

Why this matters

On the ISS, a flight surgeon in Houston is a voice call away. On the way to Mars, a one-way signal takes 3 to 22 minutes, so a question and answer can take 45 minutes or more, and communication blackouts can last for weeks. NASA's roadmap calls for Earth-independent medical operations. The crew and their software must detect, triage and act alone. Fixed alarm limits built for hospital beds fail in space for two reasons. Exercise, which astronauts do for about 2 hours a day, constantly triggers them. And subtle early warnings, like a fever starting or slow dehydration, stay "in range" until it is late. ASCLEPIUS learns each astronaut's own normal for every activity and time of day, detects deviations early, explains them, names the likely cause, and gives the onboard protocol immediately.

What is real and what is simulated

  • Live and real: GOES ≥10 MeV and ≥100 MeV proton flux (NOAA SWPC, the radiation environment right now), live ISS position and South Atlantic Anomaly crossings, and NASA GIBS Blue Marble imagery.
  • Official NASA research: every diagnosis and long-duration risk links to NASA's own reports from the NASA Technical Reports Server (NTRS), including Human Research Program evidence reports, fetched live through a serverless proxy, with a built-in snapshot for offline use.
  • Grounded in NASA research: mission dose rates (ISS, lunar surface, Mars cruise), the 600 mSv career limit, bone-loss rates, in-flight core-temperature rise, and communication delays.
  • Simulated: crew vital signs. Real astronaut medical telemetry is not public, so a physiological model generates it. The detection software would run unchanged on real wearable feeds such as Bio-Monitor, Astroskin or ECG patches.

Pipeline (runs in the browser, about 1 ms per crew-minute)

  1. Ingest 6 vitals per astronaut per minute: HR, HRV (RMSSD), SpO₂, respiration, core temperature, systolic BP.
  2. Predict the expected value from the personal baseline, the timeline-driven activity response (first-order lag), the circadian term and a slow adaptive bias.
  3. Compute uncertainty-aware z-scores. Uncertainty grows with exertion and during transitions.
  4. Detect with a smoothed multivariate score for acute events and a two-sided CUSUM for slow drifts, plus hard safety floors.
  5. Diagnose by matching the deviation pattern against known pathophysiology signatures, fused with cabin sensors.
  6. Act: show the onboard protocol now, and queue an Earth consult with an ETA that accounts for light-time delay.

Detection maths

expected_v(t) = base_v + lag( activity_offset_v(timeline), τ_v ) + circadian_v(t) + bias_v σ_v(t) = √( σ_sensor² + (0.17·|activity_offset|)² + (0.35·|transition gap|)² + σ_calib² ) z_v = EWMA( (x_v − expected_v) / σ_v ) S = max( max|z| / 3.3 , rms(z) / 1.7 ) → alert if S > 1 for 3 min CUSUM_v = max(0, C + z·√((1−φ)/(1+φ)) − 0.5) (two-sided, φ≈0.8 noise autocorrelation) → drift alert if C > 10 Hard floors : SpO₂ < 90%, T > 38.5 °C, HR > 170 or < 40, SBP < 85 → critical

Differential diagnosis

Each condition has a physiological signature, the direction and size of its effect on the six vitals, expressed in units of σ. The observed z-vector is compared with every signature by cosine similarity. For example, hypoxia is SpO₂↓↓ with HR↑ and RR↑; atrial fibrillation is HR↑ with HRV↑↑ (an irregular rhythm) and BP↓; heat stress is T↑↑ with HR↑↑. Cabin CO₂ sensor data raises or lowers the CO₂ hypothesis. The output is a ranked differential, not a black box.

Long-duration risks (NASA Human Research Program)

Roadmap

ASCLEPIUS

The doctor on board, when Earth is too far away. On the way to Mars a message to a doctor on Earth takes up to 22 minutes each way, so this software watches a crew of four and acts first.

  1. Left: crew status, cabin air and live space radiation.
  2. Centre: the selected astronaut's vitals. The shaded band is what their body should be doing right now. Leaving the band is the warning sign.
  3. Right: alerts with the likely cause and the action to take.

Try it: use "Inject a medical event" in the top bar. Proof that it works is in the Validation tab.

ASCLEPIUS

The doctor on board, when Earth is too far away.

asclepius-crew-health.vercel.app · NASA Space Apps Challenge 2026