ECG image analysis

Result: the app turns an ECG image into a digitized signal, rhythm, axis, PR, and QRS, then produces a clinical format report with a review flag. Validation metrics are add digitization error and flag rate.

Problem

ECG images are easy to share and hard to compute on. A report that looks confident but is wrong is worse than a report that asks for review. The product decision was to make uncertainty a first class output, not an error toast.

Data

The ground truth source is add source and license. The validation set size is add image count and patient count if available. Images must be synthetic or public and de identified. No patient data is used in the demo.

Approach and why

Clean the image and remove text and grid artifacts where possible.
Calibrate the grid so pixel distances can become time and voltage.
Extract the trace, then detect R peaks.
Estimate rhythm, axis, PR, and QRS from the trace and the lead assumptions.
Generate a report with a review flag when signal quality or confidence is low.

The calibration step is explicit because axis and intervals are meaningless without a stated scale. The review flag is the safety valve. It is not a decoration.

Results

Check	Result	Notes
Digitization error	add pixel or millimeter error	Against the reference trace
R peak sensitivity	add result	Detection on the validation set
PR interval error	add result in milliseconds	add tolerance
Review flag rate	add rate	On the validation set

Evaluation

Leakage: no patient appears in both the calibration and validation sets. The split is by patient where possible.
Baseline: a simple threshold based trace extractor, so the added steps have to earn their place.
Validation: held out images with a reference signal or reference measurements.
Metric: digitization error, R peak sensitivity, interval error, and flag rate are reported separately. A single accuracy number would hide the failure mode.
Reproducibility: fixed image preprocessing settings, a recorded calibration assumption, and a saved validation manifest.

A reviewer should challenge any clinical claim, any accuracy number without a validation set, and any axis value without a stated lead assumption.

Limitations

The app is not a medical device. It does not diagnose. Image quality, cropping, and grid spacing can break calibration. The review flag threshold is add threshold and rationale. The demo uses a synthetic sample and is labeled as such.

Next steps

Build a validation set with a reference signal, calibrate the flag threshold, add a cardiologist review pass, and publish a model card with intended use and out of scope use.