Fig. 1: The workflow of AI-enabled CHD detection with ECG data.

Hand-crafted human-concept features (a) were computed with some rules (corresponding formulas were in the Supplementary Materials) on pediatric ECG-waveform data (b), while the wavelet coefficient energy characteristics (as wavelet features (c)) obtained by performing wavelet transformation on the pediatric ECG-waveform data. Features of these three types were fed into the proposed AI model (d) for automatic fusion and CHD detection. e and f Illustrated the receiver operator characteristic curves and precision-recall curves of the AI model’s CHD detection performances on a test set and two external test sets. g Illustrated the CHD detection effects (by the net reclassification index (NRI)) of the AI model and cardiologists assisted by the AI model, compared with cardiologists without any assistance as a baseline, across 10 randomly sampled test data groups (from the Center-A test set). The analysis of NRI(+) and NRI(−) was included in the Supplementary Materials. Source data are provided as a Source Data file.