Machine learning prediction of mortality in pediatric fungemia using the Candida score

Pediatric fungemia in pediatric intensive care units (PICUs) carries high mortality. We evaluated whether the Candida Score, combined with clinical variables, predicts mortality after diagnosis using a prespecified multivariable logistic regression (primary model) and benchmarked discrimination against Random Forest and Gradient Boosting Machine. We analyzed 85 pediatric fungemia cases from a PICU (2016-2020). The prespecified primary model was multivariable logistic regression with predefined covariates; Random Forest and Gradient Boosting Machine were exploratory comparators. Discrimination was evaluated on a held-out test set and by 10-fold cross-validation and bootstrapping. In 85 cases, the median age was 6 months and median weight 4.8 kg; 62.4% were male. Candida albicans was the most prevalent species (37.6%). Of the subjects, 39 (45.9%) died and 46 (54.1%) survived. On the held-out test set (n = 17), logistic regression achieved accuracy 0.735 and AUC 0.800. Random Forest achieved AUC 0.861 (precision 1.000; recall 0.778), and Gradient Boosting achieved AUC 0.847 (precision 0.875; recall 0.778). Internal validation (10-fold cross-validation and bootstrap resampling) supported model stability.Conclusion Integrating the Candida Score with clinical predictors shows potential for mortality risk stratification after fungemia diagnosis. In this single-center cohort, Random Forest yielded the highest discrimination on the test set. Findings are exploratory and require external validation in larger, multicenter studies before clinical use.