Developing a new LDLc regressor machine learning model to predict LDLc levels and comparing with conventional LDLc formula: our experience from clinical laboratory dataset.

BACKGROUND: The treatment goals of cardiovascular diseases are focused primarily on LDLc. ML-based LDLc estimation methods offer a superior, reliable alternative to traditional formulas. This study aims to evaluate the performance of ML-based LDLc estimation models and compare their accuracy to direct LDLc and conventionally used formulas. METHODS: ML models based on Elastic Net based Formula, Elastic Net based Formula, Random Forest(RDF), eXtreme Gradient Boost, Multiple Layer Perceptron Regressor (MLP) were developed based on 61,444 patient datapoints i.e. 36,866-internal dataset and 24,578-external dataset. The performance of model evaluated in data subsets based on Total Cholesterol and TG levels. These subsets were evaluated i.e. adjusted R2, MSE, RMSE, Mean Absolute Error, % Bias along with evaluation matrix score (EMS)-based ranking and Subset-Integrated EMS (SI-EMS). RESULTS: Based on the EMS and SI-EMS, MLP model was depicted to be the best model for overall dataset as well as all subsets. Bland Altmann plot showed all ML models Highly Reliable based on Bias and Trend (r). Among the formulas, Martin's is only Highly Reliable. Classification Validation Score (CVS) depicted better performance of ML models than their counterparts except when classifying Very High LDL. MLP was in top five methods with equivalent CVS to top models. CONCLUSION: Based on the SI-EMS, ML models and formula were evaluated, MLP neural network performed the best to estimate LDLc levels and the web application constructed by utilizing best model could be a potential aid to clinicians and laboratories in estimating LDL values closer to the actual measurement.