Dataset dependency of low-density lipoprotein-cholesterol estimation by machine learning.

Area of Science:

• Biochemistry
• Medical Informatics
• Machine Learning

Background:

• Accurate low-density lipoprotein-cholesterol (LDL-C) estimation is crucial for cardiovascular disease risk assessment.
• Traditional methods like the Friedewald formula have limitations, prompting exploration of advanced techniques.

Purpose of the Study:

• To evaluate the applicability of machine learning (ML) for LDL-C estimation.
• To assess the impact of training dataset characteristics on ML model performance.

Main Methods:

• Nine ML models were developed using hyperparameter tuning and cross-validation.
• Models were trained on three distinct datasets: health check-up participants and two clinical patient cohorts.
• Performance was validated against the Friedewald and Martin methods using a separate clinical patient test set.

Main Results:

• ML models trained on clinical patient data demonstrated superior performance compared to those trained on health check-up data, exceeding the Martin method's accuracy.
• Models trained on health check-up data were comparable or inferior to the Martin method and tended to overestimate LDL-C levels.
• ML models trained on clinical data showed better convergence and smaller differences compared to direct LDL-C measurement.

Conclusions:

• Machine learning offers a valuable approach for LDL-C estimation.
• Training datasets must possess characteristics aligned with the target population for optimal ML model performance.
• The versatility of ML methods necessitates careful consideration of dataset selection.