Global performance of predictive models for dengue severity, hospitalization and mortality: a systematic review and meta-analysis of 146 studies

OBJECTIVES: Predictive models are increasingly used to support the clinical management of dengue, but their performance varies widely across settings. We aimed to evaluate multivariable prediction models for dengue severity, mortality, and hospitalization, and to summarize the predictors most consistently associated with these outcomes. METHODS: We searched five databases for studies that developed or validated multivariable models predicting severity, hospitalization, or mortality in dengue populations. Two reviewers independently selected studies and extracted data, and risk of bias was assessed with PROBAST. We pooled diagnostic accuracy estimates using bivariate models and synthesized predictor effects with random-effects meta-analyses. RESULTS: A total of 146 studies were included: 109 addressed severity, 42 mortality, and 12 hospitalizations. For severity, pooled sensitivity was 0·85 (95% CI 0·82-0·87) and the overall AUC was 0·93 (0·91-0·94), with machine learning models slightly outperforming traditional regression (AUC 0·93 vs 0·90). PROBAST classified 112 of 146 studies as high risk of bias. Bleeding, shock, and hypoalbuminemia were the predictors most consistently associated with adverse outcomes. CONCLUSIONS: Dengue prediction models, especially those based on machine learning, show good discrimination for severity, but the evidence is limited by high risk of bias and scarce external validation.