Enhancing ICD Patient Selection: A Convolutional Neural Network Approach to ECG-Based Risk Stratification for Ventricular Arrhythmias

Background: Implantable Cardioverter Defibrillators (ICDs) are life-saving devices for patients at risk of sudden death due to ventricular arrhythmias (VAs). However, the identification of individuals who benefit most from ICD therapy remains challenging. This study seeks to leverage the interpretative capabilities of convolutional neural networks (CNNs) applied to 12-lead ECG data, aiming to improve risk stratification by predicting VAs and segregating primary from secondary prevention cohorts. By doing so, it endeavors to uncover nuanced markers of arrhythmia susceptibility, thus providing a more robust predictive framework for arrhythmic risk.

Methods: We analyzed 2079 patients who received ICDs between 2010 and 2023, with adequate follow-up data being available for 1039 patients. In the followed-up population, raw 12-lead ECG data (represented as 12x5000 matrices for 12 leads at 500Hz over 10 seconds) was available for 1223 recordings from 619 patients, collected within 90 days prior to ICD implantation. The primary outcome for the model was the prediction of future VAs necessitating ICD intervention, either by shock or antitachycardic pacing. We employed a secondary model to distinguish patients with primary from those with secondary prevention indications, bypassing the need for follow-up data, with 2563 available ECG recordings from 1257 patients. For model training, a 5-fold cross-validation was performed, in which the data were split patient-wise to ensure that individual patient data were not mixed between folds. Patient groups underwent a 70/10/20 split into training, validation, and test sets for each repetition. ECG data was processed through the CNN architecture InceptionTime using PyTorch with cyclical learning rate adaptation and an early stopping criterion to prevent overfitting. Model performance was evaluated using ROC curve analysis and area under the curve (AUC).

Results: The first model predicted VAs during follow-up with an AUC of 0.65, based on n=257 ECGs from N=123 patients who experienced VAs and n=966 ECGs from N=496 patients who did not. Although this performance surpasses the threshold of chance (AUC of 0.5), indicating some level of validity in identifying at-risk patients, the results also suggest that the model's predictive capacity may be limited by the available sample size. In comparison, the secondary CNN model, which categorizes patients into primary prevention (n/N = 1784/897) and secondary prevention indications (n/N = 779/360), demonstrated a substantially higher AUC of 0.77, indicating a more reliable distinction between patient cohorts.

Discussion: The investigation into the potential of ECG raw data for predicting VAs through CNN models has revealed several key insights. Although the model for VA prediction demonstrated only moderate performance, it notably surpassed random chance, indicating that ECG data does indeed contain predictive elements. The substantially higher performance of the secondary model accentuates the potential of machine learning to decipher complex ECG patterns indicative of a susceptibility to VAs. The evidence points to the presence of discernible ECG features predictive of arrhythmic events, offering a promising addition to current stratification methods. However, further research with expanded datasets is imperative to improve the clinical applicability of these predictive models and ultimately aid clinicians in the complex decision-making process of ICD implantation.