On-board health monitoring for astronauts on long-duration missions
It learns each crew member's own normal, spots problems early, explains them and says what to do, even when Earth is 14 minutes away.
Quick start · Features · Results · How it works · Limitations
Important
HELIOS is a research prototype built for a NASA Space Apps hackathon challenge. It is not a medical device. The demo crew's daily data is simulated. No clinician has reviewed the rules or protocols. Don't use it to make medical decisions.
On a Mars mission, Earth can be up to 20 minutes away each way, and solar conjunction can cut contact for about two weeks. Radiation, isolation, altered gravity and a closed cabin slowly change the body and mind, and the ground can't watch over the crew in real time. Astronauts have to evaluate and act on their own health.
Fixed medical limits don't solve this. A resting heart rate of 64 can be normal for one person and a warning sign for another.
- Gathers wearable data, daily check-ins, cognition tests, lab panels, dosimetry and cabin readings across 24 indicators in 7 health domains.
- Learns each person's baseline, from preflight data first and then slowly in flight, so a developing problem doesn't become the new "normal".
- Detects sudden jumps, sustained shifts and slow multi-week drifts.
- Explains each finding with its evidence and who should be involved.
- Guides action with on-board protocols, offline triage, an EVA go/no-go check and a ready-to-send message for the flight surgeon.
There's nothing to install and no build step. You need Python 3 to serve the files; Node.js is only needed for the tests and tools.
git clone https://github.com/WillBettleheim/Helios.gitcd Heliosnpm startThen open http://localhost:8173. If you don't have npm, run python3 -m http.server 8173 instead.
For a quick demo, open http://localhost:8173/?judge. It skips the intro and plays the mission replay.
To use your own data, open My profile or select Make it mine on the dashboard. Your data stays in your browser.
| Feature | What it does |
|---|---|
| My health | Readiness score, health-domain cards, ranked findings with "do now" steps, a body map, today's tasks and cabin environment |
| Where your points went | Explains the readiness score area by area |
| What changed | Compares findings and score with 1, 7 or 30 days ago |
| Data coverage | Shows which readings are current or overdue, and gives no score when too little is current |
| Trends | Each indicator charted against your normal range, limits and mission events, with a table view and CSV export |
| Weekly trends | Separate multi-week analysis that finds slow drifts and estimates when they began |
| Lab panels | Blood count and metabolic panels scored against your own preflight draws |
| Live sensor | Heart rate and HRV from a standard Bluetooth heart-rate strap, with a simulator for browsers without Web Bluetooth |
| Mission replay | Replays days 90–147 and compares when HELIOS flagged each problem with when a fixed limit was crossed |
| Feature | What it does |
|---|---|
| Daily check-in | Two-minute report on sleep, mood, energy, stress, symptoms and vitals, with a privacy choice and a private journal |
| Alertness test | PVT-B reaction test, scored for reaction time, lapses and false starts |
| Cognition tests | Digit span and math processing, compared with your own history |
| Day planner | 24-hour plan for light, caffeine, peak work, naps, exercise and sleep, adjusted to your sleep debt and alertness |
| Feature | What it does |
|---|---|
| Protocols | 11 step-by-step procedures, such as storm shelter, sleep recovery, CO₂ headache and bone countermeasures, each with escalation criteria |
| What works for you | Checks 2–7 days after a protocol whether its target indicators improved |
| Symptom triage | Offline yes/no guides for chest pain, vision change, head injury, burns and allergic reactions |
| Medical kit | Inventory with expiry dates, stock forecasts and a per-person medication log |
| EVA readiness | GO / CAUTION / NO-GO before a spacewalk, with reasons and what would change the verdict |
| Space weather | Simulated solar-storm warning with a countdown, shelter timer and dose avoided |
| Blackout mode | During a simulated loss of contact, reroutes ground escalations to the Crew Medical Officer and queues reports |
| Feature | What it does |
|---|---|
| Ask the flight surgeon | Drafts one complete SBAR message, since every question costs a long round trip. You edit it, and nothing leaves until you queue it |
| Crew (CMO) view | Crew status and findings for the medical officer, with private entries hidden |
| Crew health map | The whole crew over 12 weeks, from measurements only |
| Downlink report | Summary for the flight surgeon: queue it, export it as JSON or print it |
| FHIR export | HL7 FHIR R4 bundle for ground medical systems |
| Encrypted sync | Optional relay that only ever stores ciphertext and can simulate light-time delay |
| Security status | Shows in Settings what's protected on this device, and offers encrypted backups |
| Model validation | Every result on real Inspiration4 data, including the weak ones |
- Works offline: installable as a PWA. Voice control is the one feature that needs a connection.
- Your own data: imports Apple Health exports and CSV files from Fitbit, Garmin, Oura, Whoop or a spreadsheet.
- Private: no accounts or analytics, and optional AES-256-GCM encryption with a PIN.
- Accessible: screen-reader support, a glove mode for larger controls and reduced-motion support.
- Readable anywhere: dark, day and red low-blue-light sleep themes.
- Multilingual: navigation, headings and common controls in English, French, German, Japanese, Russian and Hindi. Other text appears in English.
![]() Mission replay |
![]() Ask the flight surgeon |
![]() Crew health map |
![]() Weekly trends |
Each result below says what kind of data it comes from, because that limits what it can prove.
| Test | Data | Result |
|---|---|---|
| Detection benchmark | Simulated: 200 held-out people | 95.2% of problems caught, with 2.7 false alarms per person-month (was 9.96). Median delay rose from 8 to 12 days |
| Alert load | Real: PMData, 16 healthy people, about 4 months each | Alerts down 72% (7.2 → 2.1 per person-month). Time in alert down from 35.6% to 6.3% |
| Lab models | Real: NASA OSDR Inspiration4, 4 astronauts | Ensemble AUROC 0.85 (p = 0.012). Nested estimate 0.68 (p = 0.15) |
| HRV from sensor data | Real: PhysioNet MMASH, 19 nights | r = 0.98 against a reference computation, mean difference 1.6 ms |
| Mission replay | Simulated: demo mission | 19 problems flagged, 13 of which never crossed a fixed limit. Of the 6 that did, 2 were flagged earlier (by up to 25 days) and 4 on the same day |
Note
Correction. An earlier version of this README said detector v2 caught 98.8% of problems. A scoring flaw counted alarms that were already running when a problem began. Scored correctly, v2 caught 72% and spent 29% of the time in false alarm. Detector v3 replaced it.
Detection benchmark details
tools/benchmark.js generates randomized simulated astronauts. Between-person spread comes from real MMASH data, daily noise is autocorrelated, and each person has 10 preflight readings. 40% stay healthy, and the rest develop problems with a random indicator, onset, speed and size. Detector variants were compared on seeds 1–200, and the chosen one was confirmed on unseen seeds 1001–1200.
Only a new alarm that starts on or after a problem begins counts as a detection. False alarms are measured both as separate alarms and as time spent in alarm, so a stuck alarm can't hide.
| Detector (held-out people) | Caught | Median delay | False alarms per person-month | Time in false alarm |
|---|---|---|---|---|
| v1: 3-day mean + 14-day projection + all limits | 94.8% | 8 d | 9.96 | 10.3% |
| v3 (current) | 95.2% | 12 d | 2.71 | 3.1% |
Two further tests:
- Weekly labs and scans: no rule won outright. The default catches 94% with 2.0 false alarms a month. A confirm-twice rule catches 90% with 0.97, and it's what the Conservative setting uses.
- Problems spanning several measurements: a combined pattern score added 1.2 points of detection for about 11% more false alarms. It's used only when no single measurement has alarmed.
node tools/benchmark.js 200 "v1: original daily engine|HELIOS engine (as shipped, v3)"node tools/benchmark.js --sparsenode tools/benchmark.js --syndromesReal multi-week data (PMData)
Each person's first 30 days served as their baseline, and the remaining months were judged day by day. The raw data isn't stored in this repository; the script downloads the small daily files (about 9 MB) and checks their published checksums.
node tools/fetch-pmdata.js && node tools/pmdata-check.jsThis test exposed a design flaw: the astronaut exercise prescription was being enforced like a safety limit, which kept ordinary people in alert 74–90% of the time. It's now a target that applies only to someone whose normal meets it.
These are healthy people living normal lives, with no labelled health events, so this test measures alert load. It doesn't show that real problems get caught, and not every alert is false.
Real astronaut blood data (NASA OSDR)
tools/evaluate.js uses open clinical data from the four SpaceX Inspiration4 crew members: OSD-569 (complete blood count) and OSD-575 (metabolic, cardiovascular and immune panels). The CSV files are in data/osdr/.
Task: using each person's preflight draws as a baseline, single out the draw from the day after landing. There are 4 positives and 24 negatives, tested leave-one-crew-out.
| Model | AUROC | Post-flight only | Day-after-landing draw ranked first |
|---|---|---|---|
| Clinical reference ranges | 0.44 | 0.45 | 1 of 4 |
| HELIOS rules (personal baseline, ≥ 3 SD) | 0.34 | 0.38 | 0 of 4 |
| Spaceflight-signature model (hand-picked run) | 0.79 | 0.92 | 3 of 4 |
| Ensemble of 24 configurations | 0.85 | 0.96 | 3 of 4 |
| Nested search (best of 24, chosen on training crew only) | 0.68 | 0.65 | 1 of 4 |
The nested search is the fairest estimate of picking one best model. The ensemble was designed after seeing earlier runs, so treat it as promising, not proven. With four people, no result stays significant after correcting for multiple comparisons. At the threshold learned from the other crew, sensitivity is only 0–25%, so the models rank draws well but aren't calibrated to raise alarms.
node tools/evaluate.jsReal wearable data (PhysioNet MMASH)
This tests HELIOS's sensor and import code on real 24-hour recordings. Each recording covers about one day, so it doesn't test multi-week detection. The data isn't stored in this repository.
node tools/fetch-mmash.js && node tools/mmash-check.js| Check | Result |
|---|---|
| Bluetooth heart-rate packet decoding | Round-trip error ≤ 0.48 ms, within the protocol's resolution |
| HRV (RMSSD) vs a reference computation | r = 0.98, mean difference 1.6 ms, 19 nights |
| Noisy recording (31% artefacts) | Flagged as too noisy instead of reporting a wrong value |
| Real sleep files through the CSV importer | 21 of 21 imported correctly; one empty file rejected cleanly |
| Real people vs default ranges | 6 of 19 outside the default resting heart-rate band, and 8 of 19 outside the HRV band, which is why baselines are personal |
flowchart LR
I[Wearables, check-ins,<br/>tests, labs, habitat] --> S[Sanitizer and<br/>plausibility checks]
S --> B[Personal baseline]
B --> E[Detector v3]
E --> R[Rules: explained findings]
R --> A[Protocols, triage,<br/>EVA check]
R --> G[Consult, report,<br/>FHIR]
Detecting a change (detector v3):
| Signal | Method | Result |
|---|---|---|
| Sustained shift | Robust EWMA of deviation from your baseline (λ = 0.2, adjusted for autocorrelation, each day capped at ±4 SD) | Monitor at 3.25, Action at 4.75 |
| Sudden jump | A single reading ≥ 4 SD from your baseline | Action, the same day |
| Slow drift | Mann–Kendall test on the last 8 weekly averages (p < 0.01) with Sen's slope | Alert, with rate and estimated start week |
| Safety limit | Fixed medical limits | Action and Urgent limits always alarm |
| Weekly labs and scans | 2 and 3 SD rules with a 14-day projection | Monitor or Action |
| Several small changes | Combined score across related indicators | Pattern finding, only when nothing has alarmed alone |
Personal baselines start from preflight readings, or from your first 7 readings if you have none. After that they adapt with a 30-day half-life, using only readings more than 14 days old. Your spread is blended with a population reference, so a few readings can't create an unrealistically tight band.
Readiness is 100 minus a capped penalty for each health domain. It's a summary for the crew, not a clinically validated score.
Alert sensitivity can be set to Sensitive, Standard or Conservative on a personal profile.
| Domain | Indicators include |
|---|---|
| Cardiovascular | Resting heart rate, HRV, systolic blood pressure |
| Bone & Muscle | Estimated hip bone density change, urinary calcium, exercise, grip strength |
| Immune | Lymphocytes, CRP, latent-virus shedding, core temperature |
| Radiation | Daily dose rate, career dose projected to mission end |
| Sleep & Behavioral | Sleep duration and efficiency, mood, stress, crew connection, reaction time |
| Neuro-ocular | Retinal nerve-fiber change, headache days |
| Nutrition & Hydration | Body mass, fluid and energy intake |
Cabin CO₂, temperature, humidity and noise are tracked separately as shared exposures that can explain symptoms.
The demo is set on mission day 147 of a simulated 400-day Mars mission.
| Crew member | Readiness | Story |
|---|---|---|
| Ren Nakamura | 49 | Weeks of repair work: less sleep, slower reactions, missed exercise, faster bone loss and headaches after a CO₂ scrubber fault |
| Alex Reyes | 72 | Gradual rise in blood pressure and resting heart rate, with career radiation dose projected near the limit |
| Dr. Kai Lindqvist | 100 | Healthy, as a check against false alarms |
| Sami Haddad | 68 | Latent-virus reactivation with falling lymphocytes and rising CRP |
- On the device: there are no accounts or analytics, and nothing leaves unless you export, sync or send it. Voice control uses the browser's online speech recognition.
- Injection and data theft: a strict Content Security Policy blocks injected scripts and sending data to other sites. Script files carry integrity hashes, and stored text is escaped.
- Encryption at rest: optional AES-256-GCM with a key derived from your PIN (PBKDF2, 600,000 iterations). Changing or removing the lock needs the current PIN, wrong attempts are rate-limited, and the app auto-locks. A forgotten PIN can't be recovered.
- Encrypted backups: a passphrase-protected backup file, available in Settings.
- Encrypted sync: the relay only sees ciphertext and rejects malformed or excessive requests. The passphrase is held in memory only, and a crewmate's record is never merged into yours.
- Private by choice: private behavioural answers and the journal stay out of the CMO view, reports and FHIR exports. Your own backup file includes the journal.
- Validated input: saved data, backups and synced bundles all pass through a sanitizer, and physiologically impossible values are rejected.
Settings → Security status shows what's protected on your device. See SECURITY.md for the full list, the known limits and how to report a problem. The security design hasn't been independently audited, and no software is impossible to hack.
index.html, sw.js, manifest.webmanifest App shell and offline support
css/styles.css Themes and layout
js/
data.js Mission, crew, indicators, simulator
engine.js Detector, baselines, rules, readiness
weekly.js Multi-week trend detection
guard.js, sanitize.js, vault.js Page guards, input validation, encryption
app.js Core screens and feature API
features/ 23 self-contained feature modules
data/osdr/ NASA OSDR Inspiration4 lab data
tools/ Evaluation, benchmark, data checks, sync relay
test/ Node tests
docs/ Pitch kit, one-pager, architecture
Features plug in through a small API (H.app) and event bus (H.bus). See docs/ARCHITECTURE.md for diagrams.
Run the asset-version check and the tests with Node.js 20 or later:
npm testThe MMASH and PMData tests run only after those datasets have been downloaded.
After changing any JavaScript or CSS, update the asset hashes. Otherwise the tests fail, and the browser refuses to load the changed file:
npm run versionAfter editing js/features/terms.js, regenerate TERMS.md:
npm run termsTo run the encrypted sync relay:
npm run sync-serverTwo checks run from the browser console while the app is being served:
// Visits every screen, presses safe buttons and reports errors or broken text
const r = await (await import('/tools/smoke.js')).sweep(); console.table(r.issues);// Checks labels, contrast, heading order and ARIA references on the current screen
await import('/tools/a11y-audit.js').then((m) => m.audit());- Detection isn't clinically validated. The catch rate comes from simulation, and the real multi-week data has no labelled health events.
- The people tested aren't representative. PMData and MMASH cover healthy people on Earth, and the astronaut data covers four people on a three-day mission.
- The demo mission is simulated, and the replay figures show the method, not clinical performance.
- Fewer false alarms costs some speed: the median detection delay is about 4 days longer than v1's.
- Not yet tested: a real screen reader, real Apple or Fitbit exports, a physical Bluetooth strap, and use by real crew.
- Not yet reviewed: the medical content by a clinician, the terms by a lawyer, and the security design by an auditor. See SECURITY.md for what isn't protected.
- Clinical review of the rules and protocols by flight-medicine specialists
- Validation on analog-mission data with labelled health events
- Adapters for flight hardware such as biomonitors, actigraphy and dosimeters
- DTN transport for downlink over delay-tolerant networks
- User testing with people in isolation or analog settings
- Pitch script and video storyboard
- One-page summary
- Architecture
- Security policy and known limits
- Terms & Conditions, which every user accepts on first run. They are a starting point, not legal advice.
- NASA Open Science Data Repository: Inspiration4 crew clinical data, OSD-569 and OSD-575.
- PMData: Thambawita V. et al., "PMData: a sports logging dataset", ACM MMSys 2020, doi:10.1145/3339825.3394926, CC BY 4.0. Not redistributed here.
- MMASH: Rossi A. et al., "Multilevel Monitoring of Activity and Sleep in Healthy People" (v1.0.0), PhysioNet, 2020, doi:10.13026/cerq-fc86, ODbL v1.0. Results are a Produced Work from this database, which is not redistributed here.
HELIOS is an independent project. It isn't affiliated with or endorsed by NASA, SpaceX or any space agency.
No open-source license has been chosen yet, so all rights are reserved by default. Open an issue if you'd like to reuse the code. The datasets above keep their own licenses.




