Ana Pina from the Centro de Estudos de Doencas Cronicas at NOVA Medical School in Portugal, and her colleagues recently published an innovative study regarding the diagnosis of familial hypercholesterolemia (FH).

This study was published in the European Journal of Preventive Cardiology. 

The Study

These researchers investigated three different types of machine learning algorithms to see if they were a better or worse diagnostic tool for FH.

To do so, they collected information from 248 FH patients in Gothenburg and 364 FH patients in Milan. The researchers split the Gothenburg group so that 174 were used for training data and 74 were used for internal testing. All patient data in the Milan cohort were utilized for the team’s external testing.  

Machine Learning

All of the machine learning algorithms were trained to identify the gene mutations which cause FH; PCSK9, apolipoprotein B, and LDL receptor. The three different algorithms are listed below.

1. Classification Tree
2. Gradient Boosting Machine
3. Neural Network

Algorithms 2 and 3 are considered two of the leading algorithms. The classification tree interprets results based on thresholds established from the inputted variables.

All three of these algorithms were better able to predict the mutations than the standard Dutch Lipid Score. Similar results were found for both cohorts.

The researchers measured the area under the receiver operating curve or AUROC. Below are the results for each cohort.

Gothenburg

1. .79
2. .83
3. .83

Milan

1. .7
2. .78
3. .76

Dutch Lipid Score

The Dutch Lipid Score considers a patients family history, physical exam, clinical history, and untreated levels of LDL.

The researchers found that the AUROC for this score was .68 and .64 for the Gothenburg and Milan cohort, respectively.

The Comparison 

It is certainly noteworthy that all of the machine learning mechanisms had a better performance than Dutch Lipid Score. Family history and physical exams are not a part of the machine learning process and yet the tool was still found to be accurate. With just age and lipoprotein profile, these algorithms were able to properly diagnose FH patients.

Ultimately, the researchers contend that machine learning should not necessarily replace the standard Dutch Lipid Score, at least for analyzing FH at the population level. However, in specialized clinics, this tool may be beneficial for the diagnosis of individual patients.

You can read more about this study here.