Boosting Cardiovascular Disease Prediction Accuracy: A Hybrid AI Strategy Integrating LLM-Generated Risk Scores

Abstract

Predictive modeling for cardiovascular disease faces challenges, as traditional machine learning (ML) models often miss semantic details in clinical data, while standalone Large Language Models (LLMs) lack clinical reliability. This study evaluates a hybrid approach to address this, integrating a risk score generated by an LLM as an augmented feature into an XGBoost model. This method combines the semantic processing of LLMs with the predictive robustness of established ML algorithms. The hybrid model demonstrated a significantly higher F1-score (0.898) compared to the standalone XGBoost model (0.853). Further analyses confirmed this enhancement was attributable to the incremental information from the LLM feature, which SHAP analysis identified as the single most impactful predictor. The method's general applicability was also shown, improving performance for 13 out of 14 tested machine learning algorithms. This work presents a practical strategy for leveraging LLMs as feature extractors to effectively and reliably enhance clinical prediction models.

Ähnliche Arbeiten