Sudden cardiac death (SCD) from arrhythmia is a leading cause of mortality globally. While implantable cardioverter-defibrillators (ICDs) prevent SCD from ventricular arrhythmias, current clinical criteria (LVEF <30-35%) identify only 20% of all SCD cases, highlighting the need for improved risk assessment tools. Existing risk stratification approaches often broadly stratify populations, failing to personalize predictions based on individual patient characteristics and the time evolution of the disease. Previous methods also lack reliable uncertainty estimates for their predictions. This research aims to address these limitations by developing a deep learning model that leverages raw contrast-enhanced cardiac magnetic resonance (CMR) images, which visualize scar distribution, alongside clinical covariates to predict patient-specific survival curves and associated uncertainty, providing a more accurate and personalized risk assessment for arrhythmic sudden cardiac death (SCDA). The use of raw imaging data avoids the need for time-consuming manual processing and arbitrary thresholds, allowing the model to learn complex features directly from the images. This approach offers the potential for rapid and accurate risk stratification in a large population.

Several studies have identified risk factors for SCDA, and various risk stratification approaches have been developed attempting to improve upon the limitations of solely using left ventricular ejection fraction (LVEF). However, these approaches have faced challenges in clinical implementation due to limitations like broad population stratification, ignoring time-dependent risk evolution, and providing inadequate confidence estimates for predictions. Previous work using image-derived computational models of cardiac electrical function has shown promise but is computationally intensive, making it impractical for large-scale screening. While some efforts have incorporated imaging-derived features into SCDA risk stratification, these often rely on manual processing or coarse features, limiting their effectiveness. Deep learning has shown potential in other areas of healthcare, but its application to contrast-enhanced cardiac images for arrhythmia risk assessment remains limited. The current study builds upon this prior work by creating a more advanced deep learning-based approach that addresses the aforementioned limitations.

The researchers developed a deep learning framework called Survival Study of Cardiac Arrhythmia Risk (SSCAR). SSCAR uses a combined approach involving two neural networks: a 3D convolutional neural network that processes raw, unsegmented late gadolinium enhancement (LGE)-CMR images to learn features related to scar distribution, and a dense neural network that analyzes 22 clinical covariates. Both networks independently predict patient-specific survival curves. The outputs of these sub-networks are then combined using a learned linear combination (ensembling) to provide a final, more comprehensive prediction. The model uses a log-logistic distribution to model the time to SCDA, with parameters learned directly from the data. Two parameters characterize each patient's time-to-SCDA probability distribution: the location (µ) representing the most probable time, and the scale (σ) representing the uncertainty. The model's performance was evaluated using Harrell's concordance index (c-index) and the integrated Brier score (IBS), both adjusted for censoring. The study included an internal validation set (156 patients from the LVSPSCD study) and an independent external test set (113 patients from the PRE-DETERMINE and DETERMINE studies). Gradient-based sensitivity analysis was used to interpret the learned features from both the CMR and covariate networks. Data preprocessing involved segmentation of the left ventricle myocardium from the LGE-CMR images using a previously developed convolutional neural network, followed by normalization and augmentation of image data and standardization of clinical covariates.

SSCAR demonstrated excellent performance in both internal validation and external testing. In the internal validation set, SSCAR achieved c-indexes ranging from 0.82 to 0.89 and IBS values from 0.04 to 0.12 for time points up to 10 years, indicating strong risk discrimination and calibration. The area under the ROC curve was 0.87, and the area under the PR curve was 0.93. The model's performance transferred well to the external test set, with c-indexes ranging from 0.71 to 0.77 and IBS values from 0.03 to 0.14. The area under the ROC curve was 0.72, and the area under the PR curve was 0.73, showing good generalizability. The CMR-only sub-network in SSCAR outperformed a standard Cox proportional hazards model built using clinical covariates, which included manually engineered CMR features. When compared with a Cox proportional hazards model using only clinical covariates, SSCAR showed significant improvement across several metrics (c-index, balanced accuracy, F-score, and IBS). Gradient-based sensitivity analysis revealed that the CMR network learned nuanced features beyond simple identification of enhanced tissue, and the covariate network identified expected relationships between clinical factors and risk of SCDA. A significant positive correlation was observed between prediction error and the scale parameter (Pearson's r=0.42, P<0.001), indicating that SSCAR appropriately quantifies uncertainty in its predictions.

The SSCAR framework provides a significant advancement in SCDA risk assessment by leveraging deep learning to analyze raw CMR images and clinical covariates simultaneously, resulting in improved accuracy and personalized risk prediction. The model's ability to predict patient-specific survival curves with uncertainty estimates addresses crucial limitations of previous approaches. The strong performance on an independent external dataset demonstrates the model's generalizability and potential for clinical translation. The ability to quantify uncertainty is particularly important, offering a mechanism for self-correction and reducing reliance on potentially erroneous high-confidence predictions, a known limitation of some deep learning models. The findings suggest that integrating advanced image analysis techniques within a survival model can enhance the accuracy and personalization of risk prediction, moving beyond simple at-risk/not-at-risk classifications to provide a more nuanced understanding of patient-specific risk trajectories.

SSCAR represents a significant step towards personalized SCDA risk prediction. Its superior performance, generalizability, and ability to quantify uncertainty make it a promising tool for clinical decision-making. Future work could focus on incorporating additional data types, such as right ventricle CMR information, expanding the range of cardiomyopathies studied, and further refining interpretability through more advanced explainable AI techniques. Validation in larger, more diverse patient populations is also crucial for broad clinical implementation.

The study's relatively small dataset, while mitigated by rigorous cross-validation and external testing, could limit the model's generalizability to even larger and more diverse populations. The list of clinical covariates used was not entirely comprehensive due to data harmonization between the internal and external cohorts. The model does not explicitly account for competing risks of death, although future extensions could address this. Differences in CMR acquisition protocols between the internal and external cohorts may have introduced variability.