LLMs for Cardiovascular Risk Prediction from Structured Clinical Data

Researchers developed a hybrid framework combining structured clinical data with large language models to predict coronary artery disease, achieving 94.61% fidelity in converting patient records to natural language narratives. While traditional machine learning outperformed LLMs in accuracy, the study demonstrates that LLM-based classification offers significant privacy advantages by eliminating exposure of sensitive numerical patient data in clinical prediction systems.

This research addresses a critical intersection between artificial intelligence advancement and healthcare privacy—two domains increasingly vital to institutional and individual stakeholders. The study demonstrates that LLMs can effectively transform structured clinical data into interpretable narratives while maintaining data fidelity, opening new pathways for machine learning applications in sensitive medical contexts.

The broader healthcare AI landscape has been dominated by traditional machine learning approaches that require direct access to raw patient data, including lab values, vital signs, and diagnostic codes. Privacy regulations like HIPAA and GDPR create compliance friction for these systems. This work builds on growing recognition that foundation models can serve as privacy-preserving intermediaries, processing sensitive information conceptually rather than storing it numerically.

For healthcare institutions and AI developers, the implications are substantial. Organizations can leverage LLM-based classification systems without maintaining extensive databases of raw patient metrics, reducing liability and regulatory burden. The finding that Random Forest outperformed zero-shot LLM classification is less significant than the privacy trade-off it reveals—practitioners now have a choice between marginal accuracy gains and substantial privacy improvements.

The validation pipeline achieving 94.61% consistency demonstrates that information loss during natural language conversion remains minimal, suggesting clinical adoption feasibility. Healthcare systems and medical AI vendors should monitor similar hybrid architectures, as they potentially enable deployment in privacy-sensitive jurisdictions where data localization requirements previously restricted model training. Future research should explore whether few-shot LLM performance improves with domain-specific tuning or larger clinical datasets.

• →LLM-based clinical prediction systems preserve patient privacy by eliminating numerical data exposure while maintaining 94.61% information fidelity.
• →Random Forest achieved superior accuracy compared to LLM classification, but privacy advantages may outweigh marginal accuracy trade-offs in real-world deployment.
• →Hybrid frameworks converting structured data to natural language narratives enable compliance with privacy regulations without sacrificing clinical prediction capability.
• →The study validates synthetic clinical narrative generation as a viable intermediate representation for sensitive healthcare applications.
• →Healthcare institutions can adopt LLM-based systems to reduce regulatory burden and data security liability in clinical decision support.