Performance of prediction models for incident heart failure and heart failure hospitalization in individuals with type 2 diabetes: a systematic review and meta-analysis.
Alexandra Mircescu, William Allan, Mohammad Haris et al.
✦ AI-curated · Sources linked
In 30 seconds
This systematic review and Bayesian meta-analysis assessed multivariable models predicting incident heart failure (HF) or HF hospitalization (HFH) in people with type 2 diabetes. It included 65 studies describing 56 HF and 44 HFH models. After excluding studies at high risk of bias, two HF models and two HFH models showed acceptable discrimination, while DM-CURE showed good discrimination for HFH (0.837, 95% CI 0.757-0.913). None of these models reported a prediction horizon under ten years. Across all models, none had prospective testing and only three had clinical utility analyses.
Key findings
- Among models from studies not at high risk of bias, DM-CURE had good HFH discrimination (0.837, 95% CI 0.757-0.913).
- RECODe (0.711, 95% CI 0.651-0.767) and DMRS (0.758, 95% CI 0.684-0.827) showed acceptable discrimination for incident HF.
- WATCH-DM 2022 (r) (0.705, 95% CI, 0.624-0.795) and WATCH-DM 2022(i) (0.718, 95% CI 0.587-0.850) showed acceptable HFH discrimination.
- None of the models was prospectively tested, and only three underwent clinical utility analysis.
Why it matters
Prediction models may help identify people with diabetes at risk of HF or HFH, but discrimination alone does not establish that using a model improves clinical decisions or outcomes. The limited prospective evaluation and utility analyses mean that reported performance should be interpreted as evidence of predictive ability, not proof of clinical benefit.
What to keep in mind
The review found no prospective testing of the models and clinical utility analyses for only three; the evaluated models also lacked prediction horizons under ten years.