Virtual genetic diagnosis for familial hypercholesterolemia powered by machine learning

Area of Science:

• Cardiovascular Genetics
• Medical Informatics
• Lipid Metabolism

Background:

• Familial hypercholesterolemia (FH) is a prevalent genetic disorder affecting lipid metabolism.
• Current gold standard diagnosis relies on genetic testing, often limited to specialized centers.
• Clinical scores like the Dutch Lipid Score serve as accessible but less precise alternatives.

Purpose of the Study:

• To develop a more reliable diagnostic approach for FH using machine learning.
• To create a
• virtual
• genetic test for FH.
• To enhance diagnostic accuracy beyond existing clinical scores.

Main Methods:

• Employed three machine learning algorithms: classification tree (CT), gradient boosting machine (GBM), and neural network (NN).
• Trained and tested algorithms on two independent FH cohorts (Gothenburg and Milan).
• Compared algorithm performance against the Dutch Lipid Score using AUROC curves.

Main Results:

• Machine learning algorithms outperformed the Dutch Lipid Score in predicting FH-causative mutations.
• AUROC values for ML algorithms ranged from 0.70 to 0.83, exceeding the Dutch Lipid Score's 0.64-0.68.
• Gradient boosting and neural networks showed particularly strong predictive performance.

Conclusions:

• Machine learning approaches provide a superior alternative to the Dutch Lipid Score for FH diagnosis.
• These algorithms can facilitate a virtual genetic test for FH.
• This technology is crucial for improving FH diagnosis in settings without direct genetic testing access.