Introduction: Prediction of arrhythmic events during myocardial ischemia could improve outcomes, yet represents a complex challenge, particularly for traditional models. Deep learning, specifically Long Short-Term Memory (LSTM) networks, offers a solution by capturing sequential dependencies. This study introduces an LSTM-based approach to predict the time to the next premature ventricular contraction (PVC), leveraging high-resolution in situ data from a myocardial ischemia model. Our goal is to develop a robust pipeline for real-time arrhythmia prediction, with the end goal of advancing clinical risk stratification.
Methods: We analyzed electrograms from 11 large-animal experiments and identified 1832 PVC events. We then computed the time-to-PVC across 15-second runs, sampled at 1 kHz, of the intervention protocol within each experiment. An LSTM model (247 inputs, 1024 hidden units, softplus activation) was trained on 10 experiments via PyTorch, using a hybrid loss (mean squared log error (MSLE) + mean absolute error (MAE)), with one experiment held out for testing. Hyperparameter tuning via sweeping was used to optimize the performance of the network.
Results & Conclusion: Training achieved a validation MAE of 8.6 seconds, while testing on the held-out experiment yielded an average MAE of 135 seconds and an average loss of 68.5. Scatter plots of predicted versus true time-to-PVC showed a strong, positive correlation for validation during training and a similar overall trend (albeit weaker correlation) for testing. Future enhancements will focus on reducing the MAE and average loss for the held-out data and improving overall model generalizability. This study demonstrates the baseline feasibility of such an ML approach to arrhythmia event prediction in the context of myocardial ischemia.