Risk prediction models for new heart failure diagnosis demonstrated acceptable discrimination with pooled C-statistics of 0.82 for men and 0.85 for women, but showed high bias risk.
Do multivariable risk prediction models accurately predict a new diagnosis of heart failure within 5 years in community-based populations?
Risk prediction models for incident heart failure in the community show acceptable to excellent discrimination but are currently limited by a high risk of bias and a lack of external validation.
Absolute Event Rate: 0% vs 0%
Objectives Earlier heart failure (HF) diagnosis in the community could allow timely treatment initiation and prevent unnecessary hospitalisation, but identifying those at risk remains challenging. We aimed to summarise the performance of risk prediction models for a new diagnosis of HF. Design Systematic review of multivariable incident HF risk prediction models in the community setting. Data sources MEDLINE and Embase were searched from inception to 9 November 2023. Eligibility criteria Observational, community-based studies reporting prediction model performance for incident HF within a 5-year time horizon. Data extraction and synthesis Two reviewers independently screened and extracted data. Where possible, C-statistics (or area under the receiver operating characteristic curve) with 95% CIs were extracted. Risk of bias was assessed using the Prediction model Risk Of Bias ASsessment Tool and certainty of evidence by the Grading of Recommendations, Assessment, Development and Evaluation. Results Eighteen studies described 45 prediction models, 27 used traditional statistical methods and 18 applied machine learning. Most (39/45) demonstrated acceptable discrimination (C-statistic >0.70). Overall, C-statistics ranged from 0.675 to 0.954, typically with narrow 95% CIs. External validation was performed for 31 models, but only two—the modified PCP-HF models for white men and women—were validated in three cohorts, the highest among all the models. Exploratory random-effects meta-analysis of these models showed pooled C-statistics of 0.82 (95% CI 0.82 to 0.82) for men and 0.85 (95% CI 0.82 to 0.88) for women, indicating excellent discrimination but more heterogenous performance among women. Model performance was at high risk of bias due to unreported or inappropriate handling of missing data, and the certainty of evidence was very low. Conclusion Risk prediction models for a new diagnosis of HF in the community performed well, but were at high risk of bias and lacked external validation. Future model development requires appropriate data sources, robust handling of missing data, external validation and clinical testing to assess their impact on earlier HF diagnosis and outcomes. PROSPERO registration number CRD42022347120.
Thaitirarot et al. (Thu,) reported a other. Risk prediction models for new heart failure diagnosis demonstrated acceptable discrimination with pooled C-statistics of 0.82 for men and 0.85 for women, but showed high bias risk.