Document resource
Objectives To review the application of prediction models and risk factors identified by prediction models for invasive fungal infection (IFI) in children, and assess model performance, methodological rigour and applicability.Design This is a systematic review of diagnostic prediction models and a meta-analysis of the risk factors. This study was registered on PROSPERO and performed according to the Preferred Reporting Items for Systematic Reviews and Meta-analysis and Prediction model risk of bias assessment tool.Data sources PubMed, Embase (Ovid), Medline, Cochrane Library and four Chinese Databases were searched on 10 Mar 2025.Eligibility criteria We included original studies that developed diagnostic prediction models for IFI in children and excluded the informal records.Data extraction and synthesis Odds ratio (OR) with 95% confidence interval (CI) was calculated for risk factors, and a random-effects meta-analysis was applied to factors reported in at least two studies. For prediction models, a descriptive analysis was conducted to summarise model characteristics, model performance and the risk of bias.Results Nine studies were included from 4069 articles. Nine studies developed ten diagnostic prediction models, and logistic regression was the most commonly used method. The predictive performance showed an area under receiver operating curves (AUROC) ranging from 0.76 to 0.95, but meta-analysis of AUROC was not conducted due to heterogeneity. All studies were identified as having a high risk of bias in critical appraisal, particularly in the analysis, mainly due to the lack of validation, as well as the failure to appropriately evaluate model performance and overfitting. Only two of nine studies that developed prediction models used internal or external validation.Conclusions Logistic regression is a common method for predicting IFI in children, although machine learning methods have been popular in prediction models. Our study identified all studies as high risk of bias. To reduce bias, studies should use calibration measures, internal and external validation more frequently, and consider shrinkage methods when developing models.