Leveraging transfer learning with convolutional neural networks for cardiovascular disease detection

Cardiovascular diseases (CVDs) are the top global cause of death. Early diagnosis of CVDs could mitigate their severity. A common method for diagnosis is electrocardiogram (ECG) signal analysis; however, it is time-consuming and inaccurate, even for expert physicians. To address these limitations, deep learning models offer a promising alternative for automated, high-accuracy analysis. This study aimed to determine whether parameter efficiency or a residual network architecture is more impactful in creating an accurate deep learning model. We hypothesized that residual networks would achieve superior accuracy in ECG-based CVD detection compared to other non-residual network architectures when leveraging transfer learning on small medical datasets. This advantage was attributed to the ability of residual networks to mitigate the vanishing gradient problem and preserve low-level signal details, whereas other non-residual network architectures instead prioritize parameter efficiency. To test this hypothesis, we trained EfficientNetV2-S and ResNet-50 on the same open-source dataset of 928 ECG images obtained from Ch. Pervaiz Elahi Institute of Cardiology in Multan, Pakistan to detect CVDs and potential signs of CVDs. The two models were evaluated on a test set. The findings supported our hypothesis, as the ResNet-50 model outperformed the EfficientNetV2-S model on all metrics. We concluded that architectural factors such as residual connections outweigh parameter efficiency in ECG analysis. Deep learning residual network architectures that leverage transfer learning could offer an efficient alternative to manual analysis.