SkinDeepRESEARCHSteve Seguin

Use case · Idea to test

ECG labels without a written report

Return structured predictions from a heart trace.

  1. Recorded ECG
  2. Signal model + label head
  3. Predictions for review
Illustration of the proposed workflow.

A research system could read a recorded ECG and return defined rhythm or diagnostic labels for review, without generating a narrative report. More than one label may apply to a recording.

Where the small model fits

An ECG encoder supplies signal features to a small, potentially multi-label classifier. An experimental early-exit version could check intermediate layers and continue when uncertain. This would reduce model computation, not shorten the recording needed to observe an event.

What would need to work

Compare with a conventional ECG classifier at matched quality; structured outputs alone are not a new capability. Use patient-separated PTB-XL cases, then other devices and populations. Report sensitivity by finding, false alarms, uncertain cases and latency. No clinical accuracy or safe stopping rule is established here.

Research proposal only. SkinDeep has no clinical model, patient-use demo or validated medical performance for this task.

A dataset to start with

PTB-XL contains annotated 12-lead, 10-second ECG recordings and recommended evaluation splits. PTB-XL dataset and evaluation splits.

Related experiments and evidence