Development and external validation of a novel clinical tool to predict low-density lipoprotein cholesterol (LDL-C) response for Statin OPtimisation (SOP Tool) in parallel population-cohorts in the United Kingdom and Hong Kong
One in two individuals prescribed statins for primary prevention of cardiovascular disease do not derive the intended therapeutic benefit of lowered cholesterol levels. Despite the vast number of people on statins and their high variability in cholesterol-lowering response, approaches to identify these individuals and predict their therapeutic response are lacking. The aim of this large, international study was to determine factors associated with sub-optimal LDL-C response to statins; and to derive and externally validate a clinical algorithm to predict sub-optimal LDL-C response to statins.
354,187 patients treated with statins (9/3/1990 – 6/7/2016) were identified from electronic health record (EHR) databases, from the United Kingdom (UK) Clinical Practice Research Datalink (CPRD) and the Hong Kong (HK) Clinical Data Analysis & Reporting System (CDARS). Patients were free from cardiovascular disease at baseline, with baseline LDL-C measured within 12 months prior to the initiation of statins, and at least one follow-up LDL-C measured within 24 months after initiation of statins. Two methods were used to develop multivariable models: standard logistic regression, and a deep-learning machine-learning algorithm, using data from the UK CPRD. The models were validated using two separate cohorts of patients in UK CPRD and HK CDARS. Sub-optimal statin response at 24 months, defined as less than a 40% reduction in baseline LDL-C level (based on national guidelines). Model performance was assessed by discrimination, measured by area under the receiver operating characteristic curve (AUC).
Study cohorts comprised 183,283 patients from UK and 170,904 from HK for external validation. Incidence rates for cardiovascular disease (CVD) where higher in sub-optimal statin responders (UK: 18.97 per 1000 person-years, 95% CI 18.33 – 19.63; HK: 17.36 per 1000 person-years, 95% CI 17.06 – 17.67) compared to optimal statin responders (UK: 16.18 per 1000 person-years, 95% CI 15.59 – 16.79; HK: 16.68 per 1000 person-years, 95% CI 16.28 – 17.09). Age, atrial fibrillation, diabetes, dyslipidemias, statin potency, treated hypertension, number of prescribed medications, baseline LDL-C level, corticosteroids and other lipid lowering medications were associated with LDL-C response. Discrimination for the logistic model were 0.703 (95% CI, 0.699 to 0.707) for the UK validation cohort and 0.677 (95% CI 0.675 to 0.680) for the HK validation cohort. The deep-learning model using the same risk factors showed similar discrimination to the logistic model in UK (0.703, 95% CI 0.699 to 0.707) and HK (0.680, 95% CI 0.678 to 0.683) validation cohorts.
Patients prescribed statins who do not achieve optimal LDL-C response have an increased incidence of CVD. A clinical algorithm developed and validated internationally across populations, has shown consistency and accuracy in predicting sub-optimal LDL-C response to statins. The algorithm uses predictors routinely available in EHRs and could enhance appropriate cholesterol management by identifying those less likely to benefit from statins.