Machine Learning With Novel Lipid Parameters Improved FH Classification

An expert with >15 years' experience, 2 trained junior clinicians, and a fully automated convolutional neural network trained on 599 multicenter disease cases measured LV chamber volumes, mass, and LVEF.
Machine learning was used to develop a model incorporating multiple biomarkers that improved the classification of familial hypercholesterolemia.

Incorporation of apoB/apoA‐I, triglyceride/apoB, and low-density lipoprotein (LDL) 1 into machine-learning models could better classify familial hypercholesterolemia (FH), according to a study published inScientific Reports.

The study included a dataset of 211 children aged 2 to 17 years for whom a basic set of lipid parameters was available for analysis. A molecular study classified 88 patients as FH-positive and 123 as FH-negative. FH status was defined by the presence or absence of known FH casual variants in the LDLR, APOB, and/or PCSK9 genes. Various blood lipid biomarkers were collected and used to train classification models based on logistic regression. A training set was used to generate a model while posterior validation was performed in the testing set.

The investigators found that LDL1, apoC-III, total cholesterol/high-density lipoprotein cholesterol and LDL consistently demonstrated high relevancy for distinguishing between patients who were FH-positive and FH-negative. In addition, these same parameters improved the identification of patients who were monogenic. Developed models that only used total cholesterol and LDL cholesterol levels increased the specificity of classification compared with simple cut‐off values.

The investigators noted that a set of models, including a best-ranked model, will require biochemical parameters that are not frequently available in clinical practice but can be provided by clinical laboratories. Current models that are typically only available in specialized laboratories include ApoA-II, ApoC-II, ApoC-III, and small dense LDL values. Models that use Lipoprint parameters, including LDL1, MIDC, or MIDB are currently only used for research purposes.

Identified models could have applications in clinical practice to prioritize patients for genetic testing and may be more reliable in predicting FH positive-status than Simon Broome clinical criteria, the study researchers concluded.


Correia M, Kagenaar E, van Schalkwijk DB, Bourbon M, Gama-Carvalho M. Machine learning modelling of blood lipid biomarkers in familial hypercholesterolaemia versus polygenic/environmental dyslipidaemia. Sci Rep. 2021;11(1):3801. doi:10.1038/s41598-021-83392-w