DConv-LSTM-Net: A Novel Architecture for Single- and 12-Lead ECG Anomaly Detection

Electrocardiograms (ECGs) can be considered a viable method for cardiovascular disease (CVD) diagnosis. Recently, machine learning algorithms such as deep neural networks trained on ECG signals have demonstrated the capability to identify CVDs. However, existing models for ECG anomaly detection learn from relatively long (60 s) ECG signals and tend to be heavily parameterized. Thus, they require large time and computational resources during training. To address this, we propose a novel deep learning architecture that exploits dilated convolution layers. Our architecture benefits from a classical ResNet-like formulation, and we introduce a recurrent component to better leverage temporal information in the data, while also benefiting from the dilated convolution operation. Our proposed architecture is capable of learning from single- and 12-lead ECG signals and thus offers a flexible solution for CVD diagnosis. In our experiments, we perform subject-independent ten-fold cross-validations (CVs) and compare our results with two existing benchmark models using the PhysioNet atrial fibrillation (AF) challenge dataset, the China Physiological challenge, the PTB-XL repository from PhysioNet, and the Georgia dataset. For all the four datasets, our model archives state-of-the-art performance, with an upto 8% F1 score gain achieved. Our neural conduction plots demonstrate the effectiveness of having convolution layers with varying dilation factors and the use of recurrent networks to capture rhythmic patterns. Our architecture is explainable and has the ability to learn from short ECG segments. Using neural conductance, we reveal interesting hidden patterns learned by our model, which reflect the medical phenomena/characteristics associated with CVD. Code is publically available here.