Three machine learning approaches outperformed the Dutch Lipid Score to diagnose patients with genetic mutations that cause familial hypercholesterolemia, according to a study published in the European Journal of Preventive Cardiology.
“Our results do not imply that machine learning can replace the Dutch Lipid Score at a population level; rather, that machine learning may help the genetic diagnosis of familial hypercholesterolemia, especially in the context of specialized lipid clinics,” Ana Pina, MD, PhD student at Centro de Estudos de Doencas Cronicas at NOVA Medical School/Faculdade de Ciencias Medicas at Universidade Nova de Lisboa in Lisbon, Portugal, and colleagues wrote.
Researchers analyzed data from 248 patients from the FH Gothenburg cohort and 364 patients from the FH-CEGP Milan cohort with complete target traits profiles. Patients from the FH Gothenburg cohort were split into training data (n = 174) and internal test groups (n = 74), whereas all patients from the FH-CEGP cohort were used for external testing.
Three machine learning algorithms were used to identify familial hypercholesterolemia mutations, specifically apolipoprotein B, PCSK9 or LDL receptor. All of the algorithms were trained using triglyceride/LDL, HDL and LDL/age.
“First, we used a classification tree, whose results could be interpreted based on thresholds of the input variables,” Pina and colleagues wrote. “Additionally, we used a gradient boosting machine and a neural network, which are some of the best performing machine learning algorithms, even though they do not provide explicit information on how they performed the classification.”
The three algorithms were then compared with the Dutch Lipid Score, which takes into consideration the patient’s clinical history, family history of premature CVD in first-degree relatives, untreated LDL levels and physical examination.
Compared with the Dutch Lipid Score, a gradient boosting machine (area under the receiver operating curve [AUROC] = 0.83), a classification tree (AUROC = 0.79) and a neural network (AUROC = 0.83) performed better in predicting carriers of mutations associated with familial hypercholesterolemia in the Gothenburg cohort. This improved performance was also observed in the Milan cohort for a classification tree (AUROC = 0.7), a neural network (AUROC = 0.76) and a gradient boosting machine (AUROC = 0.78).
The AUROC for the Dutch Lipid Score was 0.68 for the Gothenburg cohort and 0.64 for the Milan cohort.
“All machine learning algorithms performed better than the Dutch Lipid Score,” Pina and colleagues wrote. “This is especially interesting considering that these algorithms only make use of personal information (age and the widely assessed lipoprotein profile), thus not depending, for example, on family history awareness and physical examination.” – by Darlene Dobkowski
Disclosures: The study was supported in part by a project grant from Amgen and Sanofi Aventis. Pina reports no relevant financial disclosures. Please see the study for all other authors’ relevant financial disclosures.