Understanding the effect of bias on machine learning models: the case of low-density lipoprotein cholesterol (LDL-C) estimation

Laboratory data variability is a significant challenge to the generalizability of machine learning (ML) models in clinical settings. This study aimed to evaluate how acceptable levels of analytical bias affect an ML-based LDL-C estimation model and to assess its feasibility for routine clinical use.

A lipid profile dataset from 28,480 patients with directly measured LDL-C as the reference standard was used to develop an eXtreme Gradient Boosting (XGBoost) model. Data were split 80/20 for training and testing. Eight bias simulations based on biological variation-derived desirable analytical performance specifications were applied to test data in positive and negative directions for total cholesterol (TC: ±4.54%), HDL-cholesterol (HDL-C: ±6.43%), and triglycerides (TG: ±11.23%).

The unbiased XGBoost model achieved R² = 0.965 and RMSE = 0.189 mmol/L. Under the worst-case bias scenario (TC: −4.54%, HDL-C: +6.43%, TG: +11.23%), R² declined to 0.859 and RMSE increased to 0.382 mmol/L. Critically, 13.2% of patients truly exceeding the 2.6 mmol/L clinical decision threshold were misclassified below it.

Even analytically acceptable bias levels can introduce clinically meaningful variability in ML model predictions. Laboratory professionals must play an active role in developing, validating, and establishing regulatory frameworks for clinical ML models.