Optimized Cardiovascular Disease Prediction Using Stochastic L1 Regularization and SHAP-Based Interpretability

Cardiovascular disease remains a leading cause of death worldwide. Thus, there is a need for predictive models with adequate accuracy for early diagnosis and intervention of the disease. Artificial neural networks have shown potential in this domain because they can handle complicated and nonlinear relationships in medical data. However, ANNs tend to overfit on high-dimensional datasets with small sample sizes, which can impact their ability to generalize. In this study, we present a novel regularization technique named Stochastic L1 regularization with layer-wise probabilities. This approach applies the L1 penalty randomly to weights, with each layer having its own probability, promoting adaptive sparsity. It enhances generalization and supports more effective feature selection. Furthermore, SHAP (SHapley Additive exPlanations) utilizes the contribution of each feature in the prediction, thereby promoting transparency and clinical use of any given model. On a common CVD dataset, the stochastic L1-regularized ANN attained an accuracy of 97%, a precision of 98%, a recall of 96%, and an F1 score of 97%.