Multi-scale heart simulation augments the explainability of artificial intelligence-enabled electrocardiogram through provision of an electrocardiogram database labelled with cellular pathologies.

BACKGROUND AND OBJECTIVES: Although artificial-intelligence-enhanced electrocardiograms (AI-ECGs) offer prediction and diagnosis capabilities superior to those of humans, they exhibit poor explainability and interpretability because of their complex-neural-network-derived black-box characteristics. To augment the explainability of artificial intelligence-enhanced electrocardiogram (AI-ECG) analysis, we proposed a method combining AI-ECG and synthetic ECG database created by a multi-scale heart simulator. METHODS: Using the "UT-Heart" multi-scale heart simulator running on the supercomputer Fugaku, we simulated 30,720 12-lead ECG recordings. This dataset comprises an exhaustive combination of 12 cellular and subcellular pathologies reportedly associated with heart failure and was analysed using a previously developed AI-ECG system that accurately classifies ECGs into New York Heart Association (NYHA) functional classes. By analysing the relationship between HF severity and labelled pathology in each simulated ECG recording, we elucidated the origin of abnormalities detected using AI-ECG. RESULTS: AI-ECG classified 30,618 ECGs (excluding 102 arrhythmia cases) into 2234 control and 28,384 HF cases. A separate three-group classification identified 2234 control, 18,444 NYHA I/II, and 9940 NYHA III/IV cases. In the two-group classification, significant differences (p < 0.01) were observed in sodium (Na) and Na-calcium exchanger currents and the transmural distribution of distinct cell types. Although the three-group classification revealed a severity-dependent progression of the Na current abnormality, the cell distribution in NYHA III/IV was closer to that of normal cases than to that of NYHA I/II. These findings did not explain the changes in the ECG waveform that the AI-ECG identified as notable features of heart failure in the heatmap analysis. CONCLUSIONS: The ECG dataset generated using the multi-scale heart simulator can enhance the explainability of AI-ECGs by elucidating the mechanisms underlying HF-severity-specific changes in ECGs of heart failure.