Grow-and-Compress: Panel-Aware Discovery of Compact, Interpretable ECG Marker Panels

Interpretable clinical marker panels underperform black-box neural networks at electrocardiogram (ECG) risk prediction. On the mortality prediction task, we find that a substantial part of this gap reflects how markers are discovered: optimizing them one at a time collects correlated variants of a single physiological axis. We introduce grow-and-compress, which discovers markers as a panel. Grow proposes typed, executable measurement programs against the current panel's residual error and steers them toward physiological axes the panel has not yet captured, keeping only candidates that survive false-discovery-rate control and held-out replication; compress uses each marker's physiological axis label to remove redundancy and select a compact, diverse panel. On a $\sim 45$K-patient subset of MIMIC-IV-ECG (one-year mortality), a three-marker panel (lead-I QT, QRS voltage, RR variability) reaches $0.750$ test AUROC from the waveform alone, above the size-matched NHANES panel ($0.692$) and the best three markers from single-marker search ($0.717$), which stay confined to one axis. Across matched sizes the discovered panels add $0.03$--$0.04$ AUROC over single-marker aggregation and span three to four physiological axes. On this task the improvement comes from the panel-aware search rather than the black box: removing the CNN and the feature bank reproduces the same axes and accuracy, and the result holds across seeds, proposer models, and gating choices. The markers transfer to time-to-death, and the panel is a compact, auditable complement to the still-leading black box.