OA37 Validating QRISK3 for cardiovascular disease risk prediction in rheumatoid arthritis in the UK Biobank

Abstract

Background/Aims Established cardiovascular disease (CVD) risk prediction tools, such as QRISK3, have been shown to be less accurate in patients with rheumatoid arthritis (RA). QRISK3 was developed for use in the general population, but study cohorts are not always representative of the wider public, which may explain some of these discrepancies. Our aim here, therefore, was to validate QRISK3 for RA CVD risk prediction in the UK Biobank (UKB), a large-scale health study of the general population. Methods Data from UKB participants was used to derive the necessary fields to calculate and validate QRISK3 separately per sex. Prevalent RA and incidence cases of CVD were defined relative to the initial UKB assessment date, using self-report questionnaires and hospital inpatient data. Prevalent CVD cases, as well as participants with a prescription for statins or who were missing data for the Townsend deprivation index, were excluded from the analysis. Missing data for other covariates were imputed via MICE, and the predicted QRISK3 probabilities were pooled using Rubin’s rules. QRISK3 has previously been shown to overestimate CVD risk in the UKB, and predictions were therefore re-calibrated by estimating new 10-year baseline hazards for the UKB cohort. Survival analysis was performed to validate QRISK3 in assessing 10-year CVD risk using Harrell’s C-statistic (discrimination) and calibration plots (calibration). Results Data for 412,899 participants was eligible for inclusion (∼57% women; N = 237,515), with 4570 prevalent cases of RA (0.01%) and 19,552 incidence cases of CVD (0.05%), over a median follow-up time of 10.8 years. With 358 incidence cases of CVD (∼8%), men and women with RA had significantly worse survival outcomes than the rest of the cohort (log-rank test p = 6.8x10-20 and p = 5.2x10-19, respectively). The RA subgroup additionally differed significantly across a variety of clinical CVD risk factors, including BMI (p = 5.7x10−22), systolic blood pressure (p = 8.2x10−17), and hypertension (p = 3.5x10−67). QRISK3 discrimination was modest and comparable across the two groups, with C-statistics of 0.68 for men, and 0.69 for women with RA, compared to 0.7 and 0.72 for the rest of the cohort. After re-calibration, QRISK3’s predictions are shown to be largely accurate across the cohort. However, issues persist in the RA subgroup, where predictions did not align with observed CVD risk, particularly at the tail end of predictions and in young women. Conclusion UKB participants are on average healthier than the general population, and re-calibration is therefore essential for accurately estimating a model’s predictive accuracy. Here we show that QRISK3 has poor calibration for RA patients; together with previous results this highlights the need for an updated QRISK version for use in RA. The re-calibrated QRISK3 model provides a baseline to further explore prediction modelling of CVD risk across different rheumatic and musculoskeletal diseases in the UKB. Disclosure M. Stadler: None. S. Zhao: None. A. Barton: Grants/research support; Pfizer, Galapagos, Scipher Medicine and Bristol Myers Squibb. J. Bowes: None.