Explainable ECG Analysis by Explicit Information Disentanglement With VAEs

OBJECTIVE: The interpretation of electrocardiogram (ECG) signals is vital for diagnosis of cardiac conditions. Traditional methods rely on expert knowledge, which is time consuming, costly and potentially misses subtle features. AI has shown promise in ECG interpretation, but clinically desired model explainability is often lacking in literature. METHODS: We introduce an explainable AI method for ECG classification by partitioning the variational autoencoder (VAE) latent space into a label-specific and a non-label-specific subset. By optimizing both subsets for signal reconstruction and one subset also for prediction while constraining the other from learning label-specific information with an adversarial network, the latent space is disentangled in a supervised manner. This latent space is leveraged to create enhanced visualizations for ECG feature interpretation by means of attribute manipulation. As a proof of concept, we predict the left ventricular function (LVF), a critical prognostic determinant in cardiac disease, from the ECG. RESULTS: Our study demonstrates the effective segregation of LVF-specific information within a single dimension of the VAE latent space, without compromising classification performance. We show that the proposed model improves state-of-the-art VAE methods (AUC 0.832 vs. 0.790, F1 0.688 vs. 0.640) in prediction and performs comparable to ground truth LVF (concordance 0.72 vs.0.72) in predicting survival. CONCLUSION: The model facilitates the interpretation of LVF predictions by providing visual context to ECG signals, offering a general explainable and predictive AI method. SIGNIFICANCE: Our explainable AI model can potentially reduce time and expertise required for ECG analysis.