MACHINE LEARNING-BASED PERSONALIZED STATIN THERAPY FOR DYSLIPIDEMIA MANAGEMENT

초록

Background: Dyslipidemia increases cardiovascular disease (CVD) risk. Although statins treat dyslipidemia, side effects like myalgia and diabetes risk can be deterrents. While guidelines suggest LDL-C level targets for statin use based on patient profiles, they may overlook individual statin response. This underscores the importance of personalized statin therapy. We aimed to design a machine-learning algorithm using electronic health records to suggest optimal statin treatment based on individual CVD risk. Methods: Data was sourced from patients starting statins between 2003-2022 across three tertiary university hospitals. Excluded were patients with evident atherosclerotic CVD, those without direct LDL-C measurements, and those potentially non-compliant with statin therapy. We utilized data from the largest hospital (9,300 cases) for model creation and internal validation. The remaining hospitals provided 7,855 and 4,345 cases for external validation. Predictive models employed multiple linear regression, KNN, SVM, Random Forest, XGB, and MLP with 5-fold cross-validation. Results: Many initial statin prescriptions missed the target LDL-C level, indicating empirical prescription limitations. Feature importance showed diabetes, SCORE2/SCORE2-OP, baseline LDL-C, Hba1c, and age as vital for target LDL-C attainment. XGBoost had the best predictive power, with an AUROC of 0.84, 0.86 accuracy, 0.80 precision, 0.52 recall, and 0.63 F1-score during 5-fold cross-validation. For external validation, the XGB model showed 0.81 AUROC, 0.82 accuracy, 0.71 precision, 0.48 recall, and 0.57 F1-score. Using XGB in clinics can boost the likelihood of achieving LDL-C targets to 71%-80% with initial prescriptions. Conclusion: Machine learning offers promise for tailored statin therapy in dyslipidemia. Future research will validate the algorithm’s efficacy and improve its scalability.