Advancing Explainability and Predictive Accuracy in AI-Driven Cardiovascular Risk Assessment: A Hybrid Statistical and Neural Network Approach

Abstract

Cardiovascular disease (CVD) remains a major global health challenge, making accurate and timely risk prediction essential for improving patient outcomes. Traditional statistical approaches, such as the Cox Proportional Hazards Model, provide interpretable results but are limited in their ability to capture complex and nonlinear relationships in clinical data. In contrast, machine learning models, particularly Feedforward Neural Networks (FFNNs), offer strong predictive performance but lack interpretability, which restricts their clinical applicability.

This thesis investigates two complementary approaches for cardiovascular risk prediction. First, FFNN models are developed and evaluated to assess their ability to model nonlinear relationships in structured clinical data. Second, a novel hybrid computational framework is proposed, integrating Winner-Takes-All (WTA) competitive learning for patient clustering with cluster-specific Cox proportional hazards models. This approach enables subgroup-level risk estimation and captures heterogeneity within the patient population.

To further enhance interpretability, a new cumulative risk measure, termed the Cumulative Prevalence Ratio (CPR), is introduced. CPR summarises time-dependent hazard information into a single interpretable value, facilitating direct comparison of long-term cardiovascular risk across patient clusters.

The proposed methods are evaluated using a large-scale cardiovascular dataset, with performance assessed using the Concordance Index and ROC-based measures. Comparative analysis highlights the trade-off between the predictive strength of neural networks and the interpretability of the hybrid framework. The results indicate that the proposed hybrid approach achieves a balance between predictive performance and clinical interpretability.

Overall, this work contributes to the development of transparent, subgroup-aware, and clinically meaningful models for cardiovascular risk prediction.