Population-level cardiovascular risk prediction models including biochemical predictors in 800 000 individuals

Most countries lack the clinical cohort data required to create cardiovascular disease (CVD) risk prediction models and rely on models developed elsewhere. However, individual-person linkage of health administrative databases enables the creation of local models.

This study aimed to explore whether the addition of routine biochemical tests could improve the performance of 'administrative data-based' models.

Estimated glomerular filtration rate, haemoglobin A1c, total cholesterol/high-density lipoprotein cholesterol ratio, and triglyceride tests performed on one-third of New Zealand adults were identified and linked to national administrative health datasets (11 564 665 tests). Sex-specific Cox models estimating 5-year CVD in people aged 30-74 without prior CVD were developed, internally and temporally validated, and compared, including and excluding the four biochemical predictors. 805 817 individuals were included, of whom 64% had all four examined test results available. 1.8% of women and 3.4% of men experienced a CVD event during 4.7 years mean follow-up. All biochemical predictors except triglyceride level in men were statistically significant independent predictors of CVD risk. The addition of these predictors improved global model performance metrics over models without biochemical predictors and identified 34% of men and 28% of women in the cohort who accounted for 75% of all CVD events, respectively.

Biochemical predictors improved the performance of administrative data-based CVD risk prediction models, illustrating the untapped potential of widely available databases. Similar models could be developed in many countries and healthcare organizations if health administrative databases were linked. These models could be remotely applied across populations to inform CVD prevention policy and practices.

Novel CVD risk prediction models have recently been developed in New Zealand and Denmark from predictors available in health administrative data, but their performance is limited by their parsimony. Biochemical predictors are now linkable to New Zealand administrative datasets and we show that they improve the predictive ability of previous 'administrative data-based' models. These new models were remotely deployed across a large New Zealand region, identifying 27% of individuals who accounted for 74% of subsequent CVD events. Similar models could be created in many countries and healthcare organizations to aid population health planning efforts, investigate equity gaps, and pre-screen populations who experience difficulties accessing traditional health services.