A systematic review and meta-analysis of prediction models for incident heart failure and chronic kidney disease in community-based electronic health records

Abstract

Abstract Background Heart failure (HF) and chronic kidney disease (CKD) share common pathophysiological pathways,(1) and disease-modifying therapies.(2-4) A model that identifies a high-risk population for HF and CKD in primary care electronic health records (EHRs) could facilitate population-based screening initiatives. Purpose To provide an overview of prediction models for incident HF and/or CKD that have been validated in community-based EHR cohorts, to determine common predictive variables, and synthesise discriminatory abilities. Methods We performed a systematic review of prediction models derived, validated and/or augmented for either HF and/or CKD prediction in community-based electronic health records cohorts using Medline and Embase through 25 September 2022. Measures of discrimination were extracted and pooled by Bayesian meta-analysis, with heterogeneity assessed through a 95% prediction interval (PI). Risk of bias was assessed using Prediction model Risk Of Bias Assessment Tool (PROBAST). Results From 15 106 unique records we included 29 studies. We identified 36 models that had been developed and/or validated for incident HF, and 16 for incident CKD, though not one had been simultaneously validated for both outcomes in the same study. Machine learning methodology had been used to develop 25 models for incident HF prediction, and 2 models for CKD. Overall, 71% of model results were at high risk of bias predominantly driven by high risk of bias in the analysis domain. Calibration was infrequently reported for prediction models for incident HF, especially those developed through machine learning, but well-reported for prediction models of incident CKD. Age was included in almost every model (Figure 1), hypertension and diabetes were included in more than 70% of models, and markers of vascular risk or manifest disease were also frequently used. For prediction models of incident HF the summary c-statistic was 0.814 (95% CI 0.707-0.883, n = 54 reports) with a 95% PI 0.612-0.949 demonstrating a large amount of heterogeneity between studies. For prediction models for incident CKD the summary c-statistic was 0.833 (95% CI 0.804-0.859, n = 27 reports) with a 95% PI 0.672-0.945 demonstrating a large amount of heterogeneity between studies. In sensitivity analysis where model results at high risk of bias were excluded, the summary c-statistic for prediction models for incident HF was 0.878 (95% CI 0.841-0.914, 95% PI 0.780-0.957, n = 5 reports), and for CKD was 0.856 (95% CI 0.814-0.892, 95% PI 0.685-0.968, n = 13 reports). Conclusions Models for prediction of incident HF and CKD in community-based electronic health records show good discrimination performance and high risk of bias. Common predictors for incident CKD and HF suggest there are opportunities to unify prediction for these two common and treatable diseases.Figure 1:Shared predictors