Scientific ML: ECG image digitization

The agent gets images of ECG printouts, photographed, scanned, or rendered, and reconstructs all twelve time-aligned millivolt traces from each. It builds and trains offline and submits through a sealed leaderboard gateway.

Decades of cardiology exist only on paper: ECG printouts filed, scanned, or photographed, with the underlying voltage traces lost. Modern analysis runs on signals, not pictures. Recovering the original twelve millivolt traces from an image of the printout would unlock those archives for contemporary tooling.

Reconstructions are scored by aligned signal-to-noise ratio in decibels against the true traces; the measured ceiling of the metric is 60 dB rather than a normalized 1, and the reward normalizes the private score of the best-public submission against a measured floor.

Every run ended up in the same place: a U-Net mapping printout pixels straight to waveform. No staged pipeline of grid detection, trace separation, and calibration recovery appears anywhere in the experiment logs.

The strongest run got there by elimination. It submitted nothing for more than half its budget while it tested alternatives locally. It measured a classical dynamic-programming approach at negative signal-to-noise. It found a network that classified the trace over bins too diffuse to commit to a waveform. The recovered hours went to longer training, test-time augmentation, and an ensemble. The weakest run settled on its U-Net in the first half and spent the rest post-processing its output. Lead fusion, band tweaks, and sharpening bought gains measured in fractions of a decibel.

Even the best reconstruction stayed under ten decibels of aligned signal-to-noise against a sixty-decibel ceiling. The model family is right enough to start, and nearly all of the fidelity is still missing.