Validation and Improvement of a Machine-learning-based LDL Prediction Model Using Retrospective Lipid Profile Data Retrospektif Lipid Profili Verileri Kullanılarak Makine Öğrenmesi Tabanlı LDL Tahmin Modelinin Doğrulanması ve İyileştirilmesi

Creative Commons License

DEMİRCİ F., Emeç M., ÖZCANHAN M. H., GÜRSOY DORUK Ö., AKAN P.

Anatolian Journal of General Medical Research, cilt.35, sa.3, ss.307-320, 2025 (Scopus, TRDizin)

Özet

Objective: Direct measurement of low-density lipoprotein cholesterol (LDL-C) is time-consuming and expensive when triglycerides (TG) exceed 400 mg/dL. We sought to validate and refine a machine-learning (ML) model for rapid estimation of LDL-C in hypertriglyceridemic sera. Methods: We extracted 25.991 lipid profiles (TG: 400-800 mg/dL) collected between 2010 and 2022 from two Turkish university hospitals. After an 80/20 split, seven ML algorithms were trained; the top two (random forest and XGBoost) were stacked with a decision tree meta-learner (model-3). Performance on the external test set (n=1.279) was compared with that of direct homogeneous LDL-C assays and the Sampson's formula (NIH-Equ-2) using balanced accuracy, precision, recall, F1 score, specificity, Pearson correlation coefficient, and Bland-Altman analysis, following International Federation of Clinical Chemistry and Laboratory Medicine analytical performance specifications. Results: Model-3 yielded balanced accuracy =99.3%, precision =98.9%, recall =98.9%, and specificity =99.8%. Predicted LDL-C correlated strongly with direct measurement (r=0.996, p<0.001) and reduced the mean absolute error by 54% compared with NIH-Equ-2. Only 0.39% of cases were underclassified relative to the European Society of Cardiology/European Atherosclerosis Society LDL-C risk categories. Bland-Altman plots demonstrated no significant proportional bias across the LDL-C range (mean bias =-0.2 mg/dL; 95% limits of agreement:-7.8 to+7.4 mg/dL). Conclusion: A stacked ensemble ML model delivers near-assay accuracy for LDL-C prediction in high-TG samples and markedly outperforms current formula. Implementation could enable same day, low-cost LDL-C reporting without extra laboratory procedures, supporting faster dyslipidaemia management.