Familial Hypercholesterolemia Identification by Machine Learning Using Lipid Profile Data Performs as Well as...

Area of Science:

• Cardiology
• Genetics
• Artificial Intelligence

Background:

• Familial hypercholesterolemia (FH) is a prevalent genetic disorder often undiagnosed, leading to premature cardiovascular disease.
• Machine learning (ML) presents a novel approach to enhance the identification of individuals with FH.

Purpose of the Study:

• To develop an ML model using basic lipid profile data for improved FH screening.
• To achieve diagnostic performance comparable to existing clinical criteria.

Main Methods:

• A hybrid ML model combining logistic regression, deep learning, and random forest was developed.
• Model performance was evaluated using Area Under the Receiver Operator Characteristic (AUROC) curves against LDL-C cutoffs and the Dutch Lipid Clinic Network criteria.
• Validation was performed on internal, external, and lower-prevalence datasets, with genetic mutation identification as the gold standard.

Main Results:

• The ML model achieved a superior AUROC of 0.711 on an external dataset compared to LDL-C cutoffs (0.642).
• Its performance was comparable to the Dutch Lipid Clinic Network criteria (0.705).
• The model demonstrated enhanced accuracy on lower prevalence datasets (AUROCs of 0.801 and 0.856).

Conclusions:

• The ML model effectively identifies genetically confirmed FH, outperforming LDL-C cutoffs and matching established criteria, even without clinical data.
• Its accuracy increases in lower prevalence populations, highlighting its potential for broad application.
• ML offers a promising tool for both screening and diagnosing FH.