Atheera reads breathing and movement from the Wi-Fi already on board.
No wearables, no cameras. Here is how it works, how well, and what it does not claim.
- Step 1
Listen to the Wi-Fi
The station's access point measures how its signal bends around each person, on many frequencies at once.
- Step 2
Clean the signal
Spikes are removed and the breathing band, 6 to 36 breaths a minute, is kept from the frequencies that carry it best.
- Step 3
Read the body
A breathing rate with a clarity score, and a published, pretrained model that recognises activities and falls.
- Step 4
Decide
Rules compare each reading with that person's schedule and usual level. Unclear readings never raise an alarm.
- Step 5
Tell the flight surgeon
Alerts that show the evidence behind them, and a plain-language handover brief.
How well it works
of activity windows labelled correctly
Run scripts/eval_activity.py to fill this in.
falls caught
Run scripts/eval_activity.py to fill this in.
breaths a minute, average error against a chest belt
The WiRe-Breath recordings are still being processed.
What is inside
| What it does | Method | Learned from data? |
|---|---|---|
| Breathing rate | Signal processing, no learned model: a band-pass filter, the 30 clearest channels fused into one spectrum, and Viterbi tracking over its peaks | No |
| Breathing clarity | Isotonic regression from the strength of the spectral peak to the chance that the reading is within 2 breaths a minute | Yes, one number in and one out |
| Activity and falls | ResNet18 neural network, an ensemble of 4 (11.2 million parameters each) with calibrated probabilities | Yes, on whole UT-HAR recordings |
| Unusual movement | Robust z-score of a motion index against that person's own last 2 minutes | No |
| Alerts | Rules: context, normal ranges, a clarity gate, and a wait before alarming | No |
| Handover brief | Claude language model. It only summarises and never decides an alert | Pre-trained, with a template when offline |
Breathing has no learned rate on purpose: a model trained on typical breathing can quietly pull readings toward typical rates and hide the abnormal breathing this monitor exists to catch.
- No heart rate, sleep apnoea or seizure detection. Those are future research.
- It is not a medical device. It supports a flight surgeon; it does not diagnose.
- No data from orbit exists yet. Every recording was made on Earth.
- Activity results come from recordings held out of training, but every UT-HAR recording shares the same people and room, so accuracy in a new room will be lower.
- The breathing method was designed by comparing variants on these same recordings, so its error is probably a little optimistic. Only the clarity score is cross-checked by person.
- Breathing is scored on short recordings of people lying or sitting still. Whole nights and moving crew will be harder.
- The demo's slow-breathing event replays a real recording slowed down. No public dataset has genuinely slow breathing, and real slow breathing may look different.
- Breathing and activity come from different datasets and hardware, so each crew member in the demo uses one source.
- Each headline number was recomputed from the raw results (the confusion matrix and the group counts) and matches this page.
- 38 automated tests pass: known-answer breathing rates from 7 to 34 breaths a minute, pure noise gets no confidence, the alert rules, and the rest tracker.
- The published SenseFi score reproduces here, and this page shows why it is too high: its test windows overlap its training recordings.
- Everything can be rebuilt from the public datasets with the commands in the repository. The full report is docs/VALIDATION.md.
- WiRe-Breath: 5 people, 9 sleeping and sitting postures, quiet and noisy rooms, with a chest-belt reference. Ge et al., IEEE JBHI 2025, CC BY 4.0.
- UT-HAR via SenseFi: 7 activities including falls, Intel 5300 Wi-Fi cards. Yousefi et al. 2017; benchmark and weights by Yang et al., MIT licence.