Generation and validation of a classification model to diagnose familial hypercholesterolaemia in adults

The early diagnosis of familial hypercholesterolaemia is associated with a significant reduction in cardiovascular disease (CVD) risk. While the recent use of statistical and machine learning algorithms has shown promising results in comparison with traditional clinical criteria, when applied to screening of potential FH cases in large cohorts, most studies in this field are developed using a single cohort of patients, which may hamper the application of such algorithms to other populations. In the current study, a logistic regression (LR) based algorithm was developed combining observations from three different national FH cohorts, from Portugal, Brazil and Sweden. Independent samples from these cohorts were then used to test the model, as well as an external dataset from Italy.

The area under the receiver operating characteristics (AUROC) and precision-recall (AUPRC) curves was used to assess the discriminatory ability among the different samples. Comparisons between the LR model and Dutch Lipid Clinic Network (DLCN) clinical criteria were performed by means of McNemar tests, and by the calculation of several operating characteristics.

AUROC and AUPRC values were generally higher for all testing sets when compared to the training set. Compared with DLCN criteria, a significantly higher number of correctly classified observations were identified for the Brazilian (p < 0.01), Swedish (p < 0.01), and Italian testing sets (p < 0.01). Higher accuracy (Acc), G mean and F₁ score values were also observed for all testing sets.

Compared to DLCN criteria, the LR model revealed improved ability to correctly classify observations, and was able to retain a similar number of FH cases, with less false positive retention. Generalization of the LR model was very good across all testing samples, suggesting it can be an effective screening tool if applied to different populations.