Diagnostic prediction model for bacteraemia: an updated systematic review and meta-analysis

Bacteraemia represents a life-threatening condition requiring timely recognition, but the clinical utility of existing prediction models remains uncertain. We evaluated the predictive performance of clinical prediction models for bacteraemia. We searched MEDLINE, Cochrane CENTRAL, EMBASE, WHO International Clinical Trials Registry Platform, and ClinicalTrials.gov until February 2025 for development and validation studies of diagnostic prediction models for bacteraemia in adults and children undergoing blood culture testing. Risk of bias was assessed using PROBAST + AI. Random-effects meta-analyses were conducted for models with three or more external validations. Among 155 studies reporting 403 models, 229 were internal validations (192 adults, 34 paediatrics, 2 both, 1 not reported) and 174 were external validations (132 adults, 39 paediatrics, 3 both). Most external validation models (85%) had high risk of bias, primarily due to insufficient calibration reporting. Of 22 models eligible for meta-analysis, only six had sufficient calibration data. ID-BactER (adult) and EsVan (paediatric) were the only models with both discrimination and calibration data including at least one low risk of bias study. ID-BactER showed C-statistic 0.74 (95% CI 0.73–0.75; prediction interval 0.70–0.78) and O: E ratio 0.99 (95% CI 0.81–1.23; prediction interval 0.53–1.86). EsVan demonstrated C-statistic 0.77 (95% CI 0.69–0.83; prediction interval 0.55–0.90) and O: E ratio 0.93 (95% CI 0.68–1.28; prediction interval 0.33–2.64). Most externally validated models were at high risk of bias. Wide prediction intervals suggest considerable between-setting variability. Current bacteraemia prediction models require further rigorous external validation before clinical implementation. Not applicable.