Predicting Ventricular Arrhythmia in Myocardial Ischemia Using Deep Learning

Figures

• 

  Figure 1. Schematic representation of the temporal relationship between PVCs and intervention periods with induced myocardial ischemia. The figure illustrates 2 intervention episodes (Intervention 1 and Intervention 2), each consisting of sequences of beats leading up to PVC events. PVCs are marked by vertical red lines (PVC 1, PVC 2, PVC 3 in Intervention 1; PVC 1, PVC 2 in Intervention 2). Colored segments represent beats preceding PVCs within each intervention period, with arrows indicating progression toward the next PVC. A recovery period separates the 2 interventions. Crossed-out segments denote periods excluded from analysis. PVC = premature ventricular contractions.

• 

  Figure 2. LSTM-based framework for time-to-PVC prediction from epicardial electrograms. A: Model architecture: sequences of 3, 5, or 7 consecutive non-PVC epicardial beats are processed through a 6-layer stacked LSTM network with an attention mechanism for temporal weighting. The final representation passed through dropout regularization and a fully connected layer to pro-duce a continuous time-to-PVC prediction in seconds. B: Paradigm A (Intra-Experiment Cross-Validation): All samples from 21 experiments (n = 2252 PVCs; 85,000–88,000 total samples) were pooled and stratified randomly into training (80%), validation (10%), and test (10%) sets, ensuring all experiments contribute to each partition. C: Paradigm B (Subject-Specific Fine-Tuning): Leave-1-experiment-out validation protocol, where the model was pre-trained on 20 experiments for 75 epochs, then fine-tuned on 10%–15% of the held-out experiment for 20 additional epochs, with the remaining 85%–90% reserved for testing. This approach simulates clinical adaptation to individual patients with limited subject-specific data. LSTM = long short-term memory; PVC = premature ventricular contractions.

• 

  Figure 3. Pooled 80–10–10 predicted vs true time-to-PVC for 3, 5, and 7 beats. The dashed line is the identity (yˆ = y); proximity indicates calibration. PVC = premature ventricular contractions.

• 

  Figure 4. Prediction error as a function of horizon. Curves show mean absolute error (MAE, s) computed in decile bins of the true time-to-PVC; the x-axis is logarithmic and points are plotted at bin midpoints. Top-left: Paradigm A for input contexts of 3, 5, and 7 beats. Top-right/bottom panels: Paradigm B for 3- (top-right), 5- (bottom-left), and 7-beat (bottom-right) models, comparing fine-tuning with 10% vs 15% of subject data. Shaded ribbons denote ±1 standard error within each bin. Fewer samples at the longest horizons yield wider bands. PVC = premature ventricular contractions.

• 

  Figure 5. Paradigm B, fine-tuned with 15% of the held-out experiment. Predicted vs true time-to-PVC for 3-, 5-, and 7-beat inputs shown for a single illustrative fold (fold 3 of 21). The dashed line marks the identity (yˆ = y); the solid line is an ordinary least squares fit with the equation inset. These panels are examples; variability across folds is observed.

• 

  Figure 6. Distribution of test MAE across the 21 leave-1-experiment-out folds in Paradigm B, stratified by input context (3/5/7 beats) and fine-tuning percentage (10% or 15%). For each condition, a single box summarizes the 21-fold-level MAEs, where each fold’s MAE was computed on that fold’s test set after fine-tuning on the specified fraction of the held-out experiment. Notched boxes indicate an approximate 95% confidence interval for the median (black line); the box spans the interquartile range (IQR); whiskers follow the Tukey rule (1.5 × IQR); orange diamonds denote the mean; gray points are outliers. Fine-tuning percentages refer to the proportion of data from the held-out experiment used for adaptation; the remainder was used exclusively for testing.

• 

  Figure 7. Qualitative examples from the test set (Paradigm B, 15% fine-tuning, fold 4). Each panel shows a short epicardial electrogram segment (Potential, mV, vs Time, ms). Top: large-error case with true time-to-PVC = 57.541 s, predicted = 151.262 s (absolute error = 93.721 s; relative error = 162.9%). Bottom: small-error case with true = 56.291 s, predicted = 73.163 s (absolute error = 16.872 s; relative error = 30.0%). Despite nearly identical true horizons (≈ 56–58 s), the traces do not exhibit obvious visual differences that would explain the disparity in error, underscoring the subtlety of the predictive cues.