MIDNet: Myocardial Infarction Detection Network using Attention-BiLSTM from ECG

The 12-lead electrocardiogram (ECG) is usually the first test a clinician uses to check for myocardial infarction, but reading every trace by hand is slow and screening queues build up. Automated tools that read the ECG accurately could take some of that load off clinicians and flag infarction sooner. We compare two deep learning models on this problem using the public PTB-XL 12-lead ECG dataset, both trained to tell confirmed myocardial infarction apart from confirmed normal recordings. The first is a 1D residual network (ResNet) with eight residual blocks that works directly on the raw signal. The second, which we call MIDNet, is a convolutional front end followed by a bidirectional LSTM (BiLSTM) with an attention layer that weights the most informative parts of the signal before pooling. Both models use the same PTB-XL split, the same light augmentation, and a class-weighted cross-entropy loss. On the held-out test fold the ResNet reaches 0.964 accuracy and 0.981 AUROC, while MIDNet reaches 0.953 accuracy and 0.981 AUROC with about a third of the parameters; both beat a strong handcrafted-feature baseline. Each model trains in under two minutes on a single GPU. These numbers suggest that small residual and attention-BiLSTM models are already accurate enough to be useful for ECG triage.