Explainable and Adaptive Deep Learning for Cardiac Ischemia Detection Using Surface ECGs: A Transfer Learning Approach

Background & Aims: The timely identification of cardiac ischemia within an acute cardiovascular care setting is crucial for initiating prompt therapeutic interventions. Traditional electrocardiogram (ECG) interpretation predominantly relies on manual analysis, which is logistically challenging and may fail to detect subtle ischemic indicators potentially resulting in delayed treatment initiation. This study aims to develop an advanced deep learning model for ischemia detection using ECG data by taking advantage of existing models via transfer learning to improve diagnostic precision and operational efficiency.

Methods & Results: We developed an advanced deep learning model by optimizing and adapting published, well-established architectures validated for predicting myocardial infarction. Our model was then trained and validated using the MIMIC-IV-ED and MIMIC-IV-ECG databases, which include emergency department (ED) admissions at the Beth Israel Deaconess Medical Center. We utilized the dataset, focusing on all patients who presented to the ED with suspected acute coronary syndrome (ACS), totaling 27,302 ECGs, including 1,074 cases of ischemia. The median age of the patients was 56 years, and 51.9% were female. The dataset was split into training, validation, and testing subsets in a 70:15:15 ratio. Results are reported for the testing subset. Leveraging transfer learning paradigms, we optimized the model's generalizability and robustness. The following techniques were applied: stratified 10-fold cross-validation ensured consistent performance across diverse data subsets; class weighting in the loss function addressed class imbalance; early stopping mechanisms prevented over-fitting; and a learning rate scheduler (ReduceLROnPlateau) optimized the training trajectory. The model architecture was augmented with residual blocks, batch normalization, and dropout layers, thereby improving feature extraction capabilities and mitigating over-fitting. We further employed mixed precision training to elevate computational efficiency and utilized an optimizer with L2 regularization (weight decay) to further enhance generalization. Our model achieved a weighted average accuracy of 0.85 in predicting cardiac ischemia, with a specificity of 0.89. To ensure model transparency and explainability, we implemented Gradient-weighted Class Activation Mapping (Grad-CAM), facilitating the visualization of ECG features that influenced the model's predictions (see Figure).

Conclusions: Our deep learning model, integrated with Grad-CAM explainability, offers a transparent and interpretable tool for ischemia detection. Adapting existing models to our clinical setting while optimizing for generalizability and robustness allowed us to quickly achieve acceptable model performance. Providing physicians with visualized insights into the model's decision-making mechanisms fosters trust and enables informed clinical decisions, moving beyond traditional black-box solutions. Further studies will be performed to understand and enhance decision-making and validate the overall reliability of the model.