Matrix Gating-based Long Short-Term Memory Deep Learning for Cardiovascular Disease Risk Stratification with Intraplaque Neovascularization Biomarker

Cardiovascular disease (CVD) is a leading cause of mortality and disability worldwide, highlighting the need for accurate and scalable risk stratification systems. Conventional machine learning (ML) models often struggle to capture complex nonlinear interactions among heterogeneous biomarkers. On the contrary, the standard Long Short-Term Memory (LSTM) networks may be limited by conventional gating when modeling multi-dimensional dependencies for multi-class risk prediction. In this study, we propose an “extended LSTM with matrix memory gating (xLSTMmg)” to enhance dependency learning for biomarker-driven CVD risk stratification. The dataset consisted of 500 patients with 39 biomarkers (features), including demographic, clinical, laboratory, medication-use variables, and radiomics-based plaque markers, with AngioScore (0–3) as ground truth. Random Forest Regression (RFR) is used for feature importance analysis, followed by benchmarking against Logistic Regression (LR), Gaussian Naïve Bayes (GNB), Linear Discriminant Analysis (LDA), and AdaBoost, as well as a conventional LSTM baseline. Model performance is evaluated using K5 cross-validation, external validation, statistical modelling and Receiver Operating Characteristic (ROC) analysis with Area Under the Curve (AUC) as the primary discrimination metric. The proposed xLSTMmg achieves an AUC of 0.9372, outperforming GNB (0.7486), AdaBoost (0.7653), LDA (0.8059), LR (0.8124), and conventional LSTM (0.8758). These results demonstrate that matrix memory gating significantly improves multi-class CVD risk discrimination and provides a robust framework for biomarker-based clinical decision support.