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

We evaluated the applicability of a machine learning-based low-density lipoprotein-cholesterol (LDL-C) estimation method and the influence of the characteristics of the training datasets.

Three training datasets were chosen from training datasets: health check-up participants at the Resource Center for Health Science (N = 2664), clinical patients at Gifu University Hospital (N = 7409), and clinical patients at Fujita Health University Hospital (N = 14,842). Nine different machine learning models were constructed through hyperparameter tuning and 10-fold cross-validation. Another test dataset of another 3711 clinical patients at Fujita Health University Hospital was selected as the test set used for comparing and validating the model against the Friedewald formula and the Martin method.

The coefficients of determination of the models trained on the health check-up dataset produced coefficients of determination that were equal to or inferior to those of the Martin method. In contrast, the coefficients of determination of several models trained on clinical patients exceeded those of the Martin method. The means of the differences and the convergences to the direct method were higher for the models trained on the clinical patients' dataset than for those trained on the health check-up participants' dataset. The models trained on the latter dataset tended to overestimate the 2019 ESC/EAS Guideline for LDL-cholesterol classification.

Although machine learning models provide valuable method for LDL-C estimates, they should be trained on datasets with matched characteristics. The versatility of machine learning methods is another important consideration.