A multimodal risk stratification model using deep learning and tabular data

Abstract

Cardiovascular diseases (CVDs) and strokes are leading global causes of mortality, with existing diagnostic practices largely reactive—identifying conditions only after clinical manifestation. This underscores a critical need for predictive models that enable early, symptom-based risk assessment across multiple related diseases. To address this, we propose a deep learning–based framework for the simultaneous prediction of CVD and stroke risks using tabular health data. A merged dataset of 71,000 samples and 16 harmonized features was constructed by integrating heart disease and stroke datasets, expanded through CTGAN-based synthesis. Two architectures were evaluated: a Multi-Layer Perceptron (MLP) as a baseline and an adapted multi-task TabNet model with a shared attention encoder and dual output heads. Class imbalance was mitigated using SMOTE-NC and Focal Loss, and interpretability was enhanced through TabNet’s attention mechanism combined with SHAP explanations. Experimental results demonstrate that TabNet achieved an accuracy of 88.83% and an AUROC of 0.9124, outperforming the MLP baseline. External validation on a subset of the UK Biobank yielded an AUROC of 0.85, confirming the model’s generalizability. These findings highlight the potential of interpretable deep learning models to support early, data-driven risk stratification of cardiovascular and stroke risks directly from cross-sectional tabular data.