Incorporating multimorbidity into cardiovascular competing risk models: performance in a national cohort

Abstract Background Multimorbidity, defined as the presence of two or more long-term conditions, increases cardiovascular disease (CVD) risk and affects life expectancy and clinical management, yet most population-level CVD risk equations do not incorporate multimorbidity or the increased competing risk of non-CVD mortality. New Zealand’s (NZ) primary prevention policy-level equations are used for health planning and identifying high-risk subpopulations, but the impact of incorporating multimorbidity has not been previously evaluated. Aim To assess the incremental value of adding the diagnosis-based M3 multimorbidity index to national five-year CVD risk equations and to compare performance with competing risk-adjusted models. Methods An NZ national health contact cohort of 2.15 million adults aged 30–84 years without prior CVD (1.16 million women; 991,000 men) was analysed. Cox proportional hazards models based on the NZ Health Contact Cohort equations were compared with Fine–Gray competing risk models, each developed with and without the categorical M3 index (containing 55 long-term condition categories). Model fit, discrimination, calibration, and reclassification at clinically relevant thresholds were evaluated. Sensitivity analyses assessed alternative multimorbidity measures: simple disease count and the Charlson Comorbidity Index (CCI). Results Adding the M3 index improved model fit and led to non-significant gains in discrimination, reflected by C-statistics increasing from 0.827 to 0.830 in women and 0.777 to 0.779 in men in both Cox and Fine–Gray models. Calibration improved within multimorbidity subgroups, and models without M3 consistently underestimated CVD risk for individuals with any long-term conditions (M30) across all risk thresholds. Competing risk adjustment reduced overprediction overall and yielded modest additional improvements for people living in more socioeconomically deprived areas, Ma¯ori and Pacific populations, and those with multimorbidity (M32). Reclassification gains were slight but generally aligned predicted and observed risk more closely at key treatment thresholds. In sensitivity analyses, the M3 index outperformed the CCI and was comparable or slightly superior to a disease count, although all measures showed limited incremental value because most of the population had no recorded long-term conditions using hospital discharge data. Conclusions Excluding multimorbidity from policy-level CVD equations leads to underestimation of risk for people with multimorbidity. Adding the M3 index improved prediction within these subgroups, and competing risk adjustment enhanced calibration in several high-risk groups. However, because the M3 index relies on hospitalisation data, its distribution is highly skewed, and overall improvements in model performance were modest. There is a need to explore additional data sources to improve long-term condition capture and refine New Zealand’s policy-level CVD risk models.