BACKGROUND
Health forecasting predicts trends in future health events at a population level. This is achieved through risk prediction algorithms which can estimate the risk of future disease developing. Screening algorithms can systematically identify patients with a high probability of undiagnosed diseases for diagnostic testing. Predisease is of particular interest as a precursor of chronic morbidity.
OBJECTIVE
Prediabetes and prehypertension are precursors to chronic morbidity, namely hypertension, and diabetes. In such patients, algorithms that predict risk of progression to cardiovascular morbidity can support targeted preventative care. This potential is being increasingly realised by availability of electronic health record (EHR) data. To understand current capabilities, we systematically review the predictive performance of existing risk algorithms for prehypertension and prediabetes.
METHODS
We describe a dual domain systematic review and meta-analysis of the accuracy of available risk tools to (1) predict prehypertensive deterioration to cardiovascular morbidity, & (2) predict prediabetes deterioration to diabetic morbidity. The primary outcome was the accuracy of the risk scores, and the secondary outcomes were the reporting quality and risk of bias.
RESULTS
Accuracy of risk prediction in prehypertension and prediabetes was high: the pooled C statistic for All Cause Cardiovascular Disease was 0.77 (CI 0.71, 0.84) and the pooled Sensitivity for All Cause Diabetic Disease Spectrum risk was 0.68 (CI 0.65, 0.7). However, we found high risk of bias, with inconsistent reporting in both prehypertension and prediabetes papers. We found that predictive performance was generally accurate. However, there remain limitations due to confounders and methodological inconsistency, such as timeframe, which undermines comparison. The systematic and safe deployment of risk algorithms into clinical use requires attention paid to policy and governance, as well as technical aspects of data and deployment infrastructure. We propose nine recommendations for policymakers and commissioners, organised under an "A to I" framework.
CONCLUSIONS
Risk algorithms show great promise, but further work is required to enable fair and responsible deployment.
CLINICALTRIAL
PROSPERO registration (IDs 425686 & 425683)