A lightweight graph neural network to predict long-term mortality in coronary artery disease patients: an interpretable causality-aware approach.

BACKGROUND

Coronary artery disease (CAD) causes substantial death toll in the United States and worldwide. While traditional methods for CAD mortality prediction are based on established risk factors, they have significant limitations in accuracy, adaptability to diverse populations, performance for individual risk prediction compared to group data, and incorporation of socioeconomic and lifestyle variations. Machine learning (ML) models have demonstrated superior performance in CAD prediction; however, they often struggle with capturing complex data interactions that can impact mortality.

METHODS

We proposed lightweight, interpretable graph neural network (GNN) models, utilizing data from a large trial of hypertensive patients with CAD to predict mortality using a concise set of critical features. While this smaller set of features can improve efficiency and implementation in clinical settings, the model's "lightweight" nature facilitates fast real-time applications. We utilized a hybrid approach, which first uses logistic regression (LR) to identify statistically significant features, followed by propensity score matching (PSM) to identify potentially causal features. These causal features, alongside demographic variables, were employed to create a graph of patients, drawing edges between patients with similar causal features. Accordingly, lightweight 5-layer graph convolutional network (GCN) and graph attention network (GAT) were designed for mortality prediction, followed by an interpretable method (i.e., GNNExplainer) to report the feature importance.

RESULTS

The proposed GCN achieved a recall of 93.02% and a negative predictive value (NPV) of 89.42%, higher than all other classifiers. Accordingly, a web-based decision support system (DSS), called CAD-SS, was developed, capable of predicting mortality and identifying risk factors and similar patients, guiding clinicians in reliable and informed decision-making.

CONCLUSIONS

Our proposed CAD-SS, which utilizes an interpretable and causality-aware lightweight GCN model, demonstrated reasonably high performance in predicting mortality due to CAD. This unique system can help identify the most vulnerable patients.