A machine learning model to aid detection of familial hypercholesterolaemia

2. TEXT ABSTRACT AND KEYWORDS Background and Aims People with monogenic familial hypercholesterolaemia (FH) are at an increased risk of premature coronary heart disease and death. Currently there is no population screening strategy for FH, and most carriers are identified late in life, delaying timely and cost-effective interventions. The aim was to derive an algorithm to improve detection of people with monogenic FH. Methods A penalised (LASSO) logistic regression model was used to identify predictors that most accurately identified people with a higher probability of FH in 139,779 unrelated participants of the UK Biobank, including 488 FH carriers. Candidate predictors included information on medical and family history, anthropometric measures, blood biomarkers, and an LDL-C polygenic score (PGS). Model derivation and evaluation was performed using a random split of 80% training and 20% testing data. Results A 14-variable algorithm for FH was derived, where the top five variables included triglyceride, LDL-C, and apolipoprotein A1 concentrations, self-reported statin use, and an LDL-C PGS. Model evaluation in the test data resulted in an area under the curve (AUC) of 0.77 (95% CI: 0.71; 0.83), and appropriate calibration (calibration-in-the-large: -0.07 (95% CI: -0.28; 0.13); calibration slope: 1.02 (95% CI: 0.85; 1.19)). Employing this model to prioritise people with suspected monogenic FH is anticipated to reduce the number of people requiring sequencing by 88% compared to a population-wide sequencing screen, and by 18% compared to prioritisation based on LDL-C and statin use. Conclusions The detection of individuals with monogenic FH can be improved with the inclusion of additional non-genetic variables and a PGS for LDL-C.