Combining stacked polygenic scores with clinical risk factors improves cardiovascular risk prediction in people with type 2 diabetes

BackgroundRecommended CVD prediction models do not perform well in people with diabetes. We aimed to determine whether models combining polygenic scores (PGS) with clinical risk factors could more accurately predict 10-year risk of six facets of CVD, including: coronary heart disease (CHD), heart failure (HF), and atrial fibrillation (AF). MethodsThree groups were selected from the UK Biobank: 143,459 control participants without diabetes or a history of CVD, 5,229 with diabetes but without CVD, and 1,621 with diabetes and a history of CVD. Data from 29 phenotype-specific polygenic scores (PGS) were stacked and combined with clinical risk-factors. Performance was evaluated using a 20% independent hold-out sample, with results stratified on duration of diabetes. ResultsIn people without diabetes combining the stacked PGS with clinical risk factor modestly outperformed models that exclusively used clinical risk factors, with the largest improvement observed for AF (c-statistic difference: 0.03). In people with diabetes, models that combined the stacked PGS with clinical risk factors showed marked improved performance compared to the risk factor only models. This difference was largest in people with newly diagnosed diabetes (without a history of CVD), with a PGS + clinical risk factor model c-statistic: 0.83 (95%CI 0.83; 0.84) for CHD and 0.84 (95%CI 0.82; 0.85) for HF, compared to a clinical risk factor model c-statistic: 0.68 (95%CI 0.68; 0.69) and 0.60 (95%CI 0.58; 0.62) for CHD and HF respectively. ConclusionsCombining PGS with clinical risk factors improves CVD risk prediction in people with diabetes. Research in contextO_ST_ABSWhat is already known about this subject?C_ST_ABSO_LICardiovascular disease (CVD) remains a prominent cause of morbidity and mortality for people with type 2 diabetes. The currently available CVD prediction models do not provide sufficiently accurate prediction in people with diabetes, prohibiting much-needed personalization of management strategies. C_LIO_LIIn the general population, phenotype-specific polygenic scores (PGS) have shown to modestly improve CVD risk prediction. However, models for CVD prediction in the general population are often already highly accurate, limiting the scope for PGS to further improve performance. C_LIO_LIGiven the multifactorial etiology of CVD, combining information (stacking) from multiple trait-specific PGS (e.g., on CHD, LDL-C and blood pressure) is expected to improve performance. C_LI What is the key question?O_LIWhat is the added benefit of incorporating PGS with conventional clinical risk factors in CVD prediction for people with type 2 diabetes? C_LI What are the new findings?O_LIIn people with diabetes, models that combined the stacked PGS with clinical risk factors showed marked improved performance compared to the risk factor-only models. C_LIO_LIWhile age was the predominant risk factor in people without diabetes, in people with diabetes the contribution of age was outranked by our stacked PGS. C_LIO_LIModel performance depended on the duration of diabetes, with models performing better in people with a recent diagnosis, for example in this group the c-statistic for CHD was 0.83 (95%CI 0.83; 0.84), and for HF 0.84 (95%CI 0.82; 0.85). C_LI How might this impact on clinical practice in the foreseeable future?O_LICombining PGS with clinical risk factors improves CVD risk prediction in people with diabetes. Incorporating PGS in risk prediction models may offer unique possibilities to reliably identify people with a meaningful risk of developing CVD. C_LI ACRONYMS O_TBL View this table: org.highwire.dtl.DTLVardef@17737e3org.highwire.dtl.DTLVardef@1f63dc3org.highwire.dtl.DTLVardef@150e6d5org.highwire.dtl.DTLVardef@62b8fcorg.highwire.dtl.DTLVardef@15ebe21_HPS_FORMAT_FIGEXP M_TBL C_TBL