Creation of a Prognostic System for Predicting Adverse Events in Patients with Myocardial Infarction based on Large Language Models

Developing prognostic models to identify patients at high risk of major adverse cardiac events (MACEs) after myocardial infarction (MI) is a crucial task for personalized medicine. A study conducted on data from 218 MI patients demonstrated that using machine learning algorithms, including Catboost and LightGBM, effectively predicts MACEs. Key risk predictors included statin dosage and VEGFR-2 genotype, where low statin doses and the presence of the risk T-allele significantly increased the likelihood of adverse outcomes. The language model ChatGPT-4-Turbo was utilized to generate text data, which was then used to test and fine-tune machine learning models, including RuBioBert. The Catboost model, trained on a reduced set of features, achieved high predictive accuracy (ROC AUC = 0.813), underscoring the importance of integrating genetic information into clinical practice to improve MACEs risk prediction. The RuBioBert model, fine-tuned on text data derived from the reduced feature set, showed similar performance (ROC AUC = 0.811), demonstrating the applicability of language models to patient medical records for this task.