OxyDrop
Does your blood oxygen look lower than your own baseline? A phone-camera check, no extra hardware.
- Runs on your phone
- Video never uploaded
- Works offline
It is a research demo validated on only 6 people. It does not measure SpO₂, it is not a medical device, and it must never replace a pulse oximeter or a clinician. If you feel short of breath, confused, or unwell, get medical help now.
What it does
Rest for 60 seconds so it learns your normal. The first answer arrives at 70 seconds (the calibration plus the first 10-second window); after that, every 2 seconds, it shows one of three results: Normal vs your baseline, Possible low oxygen: re-check with a pulse oximeter (a drop toward SpO₂ below 90%), or Unsure, re-measure. It shows “unsure” and no heart-rate number whenever the signal is poor, you move, the camera stutters, or the pulse cannot be locked.
- Works only seated at rest. Exercise raises heart rate and causes false alarms.
- Detects a drop from your own first minute. It cannot tell if you are already low at the start.
- Validated on only 6 people, and it ties a much simpler logistic chain. It cannot follow a resting heart rate below about 51 bpm (it says “unsure”).
- Your camera video never leaves your phone. Everything runs in this page.
- Normalvs your baseline
- Possibly lowre-check with an oximeter
- Can't tellit says so instead of guessing
How it works
Default model (since build v3.6): PhotonDistill, a small trained neural network (39 inputs, two hidden layers of 32 or 64 units, three training seeds) that runs in this page. It reads 39 numbers per 10 s window computed from the average colour per frame, relative to your first minute (heart-rate and colour-ratio changes, signal quality, breathing), and gives three outputs: the probability of SpO₂ below 90%, an SpO₂ estimate (debug view only, not validated) and its own confidence. When its confidence is too low it says Unsure; otherwise a probability above 0.5 means Possible low oxygen. The page shows that output as it is. It was trained from two simple teacher models fitted on the same people. Held out on the 6 Hoffman people it scores a time-matched balanced accuracy of 0.784 (95% interval 0.596 to 0.834), a tie with the logistic chain (paired difference -0.012, interval -0.091 to 0.012), answering 69.0% of windows with 93.3% of those answers right. In a simulation of a phone whose flash sits farther from the lens (the hardest level, X1) it scores 0.682 against 0.508 for the chain, which was its purpose. Add ?engine=hybrid to the address for the previous model described below.
The previous model (?engine=hybrid) has three parts that all run in this page, with no server. (1) A physics chain: average colour per frame, a band-pass filter, beat detection (heart rate, variability, breathing, pulse strength), red/green/blue light-absorption ratios and the subtraction of your first-minute median, scored by two logistic regressions. (2) A small neural network (OxyNet, 167,588 parameters, three training seeds) that reads the same filtered signal together with the physics channels relative to your own first minute and learns a correction to the chain’s score. (3) A meta-learner adds the two into one score and conformal thresholds turn it into Normal, Possible low oxygen or Unsure. The heart rate shown is the average of the network and a spectral estimate when they agree within 5 bpm, otherwise the network’s. The quality checks (no pulse lock, movement, dropped frames, saturation) come first and are unchanged. The only data set with real SpO₂ below 90% and phone video that we found is UW Hoffman 2022 (6 people), so the low-oxygen part is trained on very little. Phone-camera low-oxygen detection is not new (Hoffman et al. 2022 and others); this is an on-device, personalised re-implementation, not a new sensing principle. Add ?engine=logistic to the address to run the logistic chain alone (the baseline).
Held-out results on those 6 people (each tested with models fitted on the other 5), for the exact previous model (the fixed residual hybrid): time-matched balanced accuracy is 0.809 (95% interval 0.644 to 0.863), against 0.796 (0.635 to 0.856) for the logistic chain alone. The 0.809 is optimistic, because this hybrid variant was picked after its scores on those same 6 people had been seen; the nested estimate for the whole selection procedure is 0.779 (0.638 to 0.831). A chain refit with the same tuning and no network scores 0.816, and at the default confidence level the chain alone answers 92.5% of windows at 90.3% accuracy, against 74.5% of windows at 89.5% accuracy for the hybrid. So the simpler chain has the higher point estimates and answers more windows. All the intervals overlap, so this is a tie and not a loss, and the network does not beat the chain at detecting low oxygen. At the default confidence level the hybrid answers 74.5% of windows, 89.5% of those answers are right, and 70.0% of the windows with low oxygen raise an alert; these confidence-level figures were also read off the hybrid variant chosen with those 6 people in view, so they are optimistic too.
Heart rate is only an exploratory result. The hybrid heart-rate rule is within 5 bpm of the reference in 48.0% of windows on 100 people (BUT PPG and MTHS), against 41.9% for the spectral estimator. That rule was chosen after the pooled table had been seen, and the gain comes from BUT PPG (41.9% against 30.4%). On the Hoffman inner folds (pooled out-of-fold windows of the 6 Hoffman people that selected the rule, not an outer test result) the spectral estimator alone does better (88.3% against 82.6% for the rule), and on MTHS it is nominally better (63.3% against 59.4%). One person (resting heart rate 43 bpm) never locks, so every window for them is “unsure”. The app does not show an SpO₂ number: the network’s SpO₂ output is only in the debug view, labelled experimental, because on a clinical contact sensor it ties the textbook ratio-of-ratios (mean error 2.94 against 3.07 points) and on phone data it shows no skill. Exercise falsely flags many windows.
The app is built to work on any phone (it tests each lens, adapts exposure and checks signal quality), but it was trained on six people filmed with a few phones: untested Androids and other phones are a limitation. A generic weak-pulse guard answers “unsure” when the pulse is weaker than the weakest 0.1% of the training windows. Measured on two other studies’ phone videos it would refuse 6.7% (MTHS) and 8.8% (BUT PPG) of windows, against 0.1% of the training windows, so a phone with a weak flash will see more “unsure” answers; a refused window still shows a spectral heart rate. See the project page for the full audit.
Prototype. Not medical advice. Not a medical device.
Place your fingertip
One finger, flat across the camera bump, covering the lens and the flash. Light pressure, like resting it on a table.
- Lens the app picks the best one
- Flash cover it with the same finger
- Light pressure flat pad, do not press hard
- Sit down and rest for 2 minutes. Rest the phone and your hand on a table. The first minute becomes your baseline, so only start if you feel well: the app cannot know whether your oxygen was already low at the start.
- Lay your fingertip flat across the whole camera plateau, covering the lens the app picks and, if you can, the flash under the same finger. Light pressure, do not press hard. Phones differ: on newer iPhone Pro models the flash is 2-3 cm from the main lens, so a flat finger over both gives a weaker but usable glow.
- Keep still. Don’t talk. Keep the screen on. After you tap the button the app tests every camera for about 3 seconds each and picks the one with the strongest pulse. If no back lens works it switches to the front camera with a white screen.
Per-lens scores
Light source
The app picks the lens with the strongest pulse. You can override it here.
Signal quality
Device checks
Debug view (for troubleshooting)
Measurement
Learning your normal
Rest for about 60 seconds. Keep the finger flat and still, don't talk, keep the screen on.
How confident is this?
Answers are held to a fixed confidence level.
The model scores every 10-second window and compares that score with thresholds fixed in advance on held-out data. When one window is not clearly on either side of them, the app says Unsure, re-measure instead of picking the more likely answer. Signal quality, movement and frame drops are checked before the model is consulted at all.
Oxygen status at start unknown: every answer compares you with your own first minute, so it cannot tell whether you were already low when you began.
Signal and recent readings
- normal
- possible low oxygen
- unsure
- unsure (moving)
Debug view (for troubleshooting)
Prototype. Not medical advice. Not a medical device.