This investigates the application of deep learning for abnormality classification in 12-lead electrocardiograms (ECGs). We present the approach and findings of Team IRS in the PhysioNet/Computing in Cardiology (CinC) challenge 2024.
Our study leveraged the PTB-XL ECG dataset to train deep learning models based on the VGG architecture. We employed a holdout test set within PTB-XL for initial assessment and further validated performance on entirely new datasets. The model was built on VGG16 architecture with ImageNet pre-trained weights. The original classification layer was removed and replaced with a classification head consisting of a global average pooling 2D layer, and a dropout layer for training, followed by a fully connected layer with one output and sigmoid activation. The classification head was initially trained for up to 10 epochs with early stopping, while all other layers were frozen. The entire model was then unfrozen, and trained until no further drop in validation loss was seen (early stopping with patience of 6). For most training instances, binary cross-entropy loss was used.
The model was evaluated using an F-score. While our participation (Team IRS) achieved a rank of 40th among officially ranked teams in the unofficial submission phase, the F-score obtained was 0.46. This indicates the potential of the VGG-based deep learning approach for ECG classification, but also highlights the need for further refinement.
The results suggest that deep learning models hold promise for ECG abnormality classification. However, compared to top-ranking teams, our approach requires further optimization to achieve competitive performance. Future work may explore techniques to address data imbalance and enhance the model interpretability.