Does this heart sound need an echo?
This tool does not diagnose. It listens to a heart-sound recording and says whether a murmur is audible enough to warrant an echocardiogram. A murmur is not rheumatic heart disease, and the decision to refer belongs to a clinician.
Pick a recording
Five recordings from the CirCor DigiScope dataset. The reference annotation for each is hidden until after the model has run.
Or record from a microphone
Hold the microphone firmly against the chest, over the upper left side of the sternum, in as quiet a room as you can find. Stay still.
The held-out numbers on this model come from clinical recordings made with a digital stethoscope. Performance on a phone microphone is unmeasured, so treat a microphone result as a demonstration, not a screen.
What the numbers are
Measured once, on patients the model never saw during training or threshold selection. Nothing here was tuned after it was read.
Held-out performance
How to read that
The split is by patient, never by recording. Each patient contributes several recordings from different auscultation points, so splitting by recording would put the same heart on both sides of the wall and the held-out numbers would be flattering nonsense.
The threshold was chosen on a separate validation split and then applied unchanged to the test set. It sits where it does because a missed murmur sends a child home and an over-referral spends an echo slot. Those costs are not equal.
Data and credit
Trained on the CirCor DigiScope Phonocardiogram Dataset, distributed by PhysioNet under the Open Data Commons Attribution Licence (ODC-By 1.0), which requires this attribution.
Oliveira J, Renna F, Costa PD, Nogueira M, Oliveira C, Ferreira C, Jorge A, Mattos S, Hatem T, Tavares T, Elola A, Rad AB, Sameni R, Clifford GD, Coimbra MT. The CirCor DigiScope Dataset: From Murmur Detection to Murmur Classification. IEEE Journal of Biomedical and Health Informatics, 2022.
Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PCh, Mark RG, Mietus JE, Moody GB, Peng C-K, Stanley HE. PhysioBank, PhysioToolkit, and PhysioNet. Circulation 101(23):e215-e220, 2000.
How it works
Four stages, no hidden ones. A recording goes in, a single number comes out, and that number is compared against one threshold that was chosen deliberately.
The pipeline
What this cannot tell you
The honest boundary of the tool, written down rather than left for somebody to discover.
Limits
- A murmur is not rheumatic heart disease. It is one audible sign that something may be worth imaging. Many children have innocent flow murmurs and perfectly normal valves.
- Silence is not a clean bill of health. Valve damage can be present with no murmur an ear or a microphone can find. That is part of why echocardiography exists, and why listening alone cannot settle the question either way.
- The score is a model output, not a risk. A score of 0.90 does not mean a 90 per cent chance of disease. It means the model placed this recording at 0.90 on the scale it learned, and the only decision attached to that is where it falls relative to the threshold.
- The setting is unmeasured beyond the source data. The model was trained on recordings from a screening campaign in Brazil, captured with a digital stethoscope by trained operators. Performance on phone microphones, in noisier rooms, with other populations, has not been measured. Nobody should assume it carries over.
- Recordings the annotators could not call were excluded. Human experts labelled some recordings Unknown because they could not hear the answer. Those were kept out of training, and the model has no way to say "I cannot tell" in their place.
Finished and not finished
- Finished. The model, the held-out evaluation, the browser feature pipeline (checked cell by cell against the Python that produced the measured numbers), and this interface. Inference runs entirely on the device.
- Not finished. Any validation outside the source dataset. No prospective use, no phone-microphone evaluation, no clinical sign-off. This is a demonstration of a method, not an instrument anyone should screen a child with.