Performance comparison of different classification algorithms applied to the diagnosis of familial hypercholesterolemia in paediatric subjects

Abstract

Abstract Familial Hypercholesterolemia (FH) is an inherited disorder of lipid metabolism, characterized by increased low density lipoprotein cholesterol (LDLc) levels. The main purpose of the current work was to explore alternative classification methods to traditional clinical criteria for FH diagnosis, based on several biochemical and biological indicators. Logistic regression (LR), decision tree (DT), random forest (RF) and naive Bayes (NB) algorithms were developed for this purpose, and thresholds were optimized by maximization of Youden index (YI). All models presented similar accuracy (Acc), specificity (Spec) and positive predictive values (PPV). Sensitivity (Sens) and G-mean values were significantly higher in LR and RF models, compared to the DT. When compared to Simon Broome (SB) biochemical criteria for FH diagnosis, all models presented significantly higher Acc, Spec and G-mean values (p < 0.01), and lower negative predictive value (NPV, p < 0.05). Moreover, LR and RF models presented comparable Sens values. Adjustment of the cut-off point by maximizing YI significantly increased Sens values, with no significant loss in Acc. The obtained results suggest such classification algorithms can be a viable alternative to be used as a widespread screening method. An online application has been developed to assess the performance of the LR model in a wider population.