A visually interpretable detection method combines 3-D ECG with a multi-VGG neural network for myocardial infarction identification
Overview
Identifying myocardial infarction from a 12-lead electrocardiogram requires recognizing changes in waveform shape and their relationships across leads. Neural classifiers can assist this process, but their predictions are difficult to assess when the supporting signal regions are hidden. This work develops an image-based ECG representation that supports both automated classification and visual inspection of the evidence behind a prediction.
The method converts minimally denoised ECG recordings into top-view images of a three-dimensional ECG representation. A multi-VGG architecture analyzes complementary views of the heartbeat: the QRS complex, the ST region, and the complete beat. Combining these branches allows the classifier to use localized abnormalities alongside broader waveform information. Grad-CAM++ then maps the classification evidence back to image regions and ECG leads, making it possible to examine whether the network attends to medically relevant features.
Experiments report inter-patient classification accuracies of 95.65% on PTB and 97.23% on PTB-XL. The accompanying visual analysis compares highlighted regions with established ECG characteristics of myocardial infarction. Together, these evaluations examine predictive performance and the interpretability of the learned features on the studied datasets.
Citation
Rui Fang, Chih-Cheng Lu, Cheng-Ta Chuang, and Wen-Han Chang. (2022). "A visually interpretable detection method combines 3-D ECG with a multi-VGG neural network for myocardial infarction identification." Computer Methods and Programs in Biomedicine, 219, 106762.