Does a purpose-built Convolutional Neural Network improve the automated detection of cardiac abnormalities from ECG images compared to established models?
A purpose-built deep convolutional neural network achieved 97.2% accuracy in classifying ECG images into four diagnostic categories, outperforming established models like VGG16 and ResNet50.
The global health burden of cardiovascular diseases, responsible for significant worldwide mortality, underscores the urgent requirement for advanced diagnostic solutions. The electrocardiogram (ECG) remains a cornerstone of cardiac assessment, but its reliance on manual analysis introduces delays and interpretive variability. To bridge this gap, our research introduces a novel deep-learning framework designed for the automated, multi-class identification of cardiac conditions directly from ECG image data. We engineered a dedicated Convolutional Neural Network (CNN) from the ground up and trained it on a curated public dataset encompassing four key diagnostic categories: Normal rhythm, Myocardial Infarction, Abnormal Heartbeat, and History of MI. Our comprehensive preprocessing pipeline standardized input images through grayscaling, resizing, and pixel normalization, while strategic data augmentation fortified the model's ability to generalize. This purpose-built CNN, which leverages batch normalization and dropout for stability, attained a benchmark test accuracy of 97.2% and a weighted F1-score of 97.3%, surpassing the performance of established models like VGG16 and ResNet50. Crucially, attention mapping validated that the network's decisions align with diagnostically significant waveform segments, building credibility for clinical use. This high level of performance confirms the system's viability as a dependable tool, poised to expedite diagnosis, support healthcare providers, and enhance cardiac care pathways in diverse medical environments.
Kumar et al. (Fri,) studied this question.