Informatics in Medicine Unlocked · Published 2026-04-17 · DOI 10.1016/j.imu.2026.101760
This research presents a Neuro-Symbolic Clinical Agent for cardiovascular risk assessment that integrates ensemble machine learning with tightly coupled counterfactual simulation and targeted guideline retrieval. The framework is trained and validated on the CAIR-CVD-2025 dataset using a hybrid stacking architecture, where CatBoost, Random Forest, and a specialized neural network implement the neural risk prediction module of the proposed system. The ensemble achieved the highest discriminative performance with an ROC-AUC of 0.872 (95% CI: 0.836–0.910) and demonstrated excellent calibration with a Brier Score of 0.148, consistent with established benefits of ensemble-based risk prediction. Beyond prediction, the system functions as an active reasoning agent that performs simulation-guided inference: a local SHAP layer identifies patient-specific risk drivers, while a rule-based simulation engine executes “what-if” experiments to quantify the risk reduction achievable through specific interventions (e.g., lipid management). To address the “reasoning gap”, the reporting module utilizes a “Safe RAG” protocol, guiding guideline retrieval through SHAP-identified risk drivers to prevent hallucinations. Automated assessment confirmed that the LLaMA-3.3-70B generation layer achieved high faithfulness (0.89) to ESC/AHA guidelines. These findings indicate that predictive accuracy, probabilistic reliability, and evidence-based reasoning can coexist within a transparent, agentic decision-support system.
Abstract from DOAJ. Public domain (CC0 1.0).
Read the article at the publisher →