Evaluating the AI-predictive potential of the PIC database for pediatric ICU mortality

Background Reliable Artificial Intelligence (AI) tools in healthcare depend on both data quality and model design. However, the predictive integrity of datasets is rarely assessed. Publicly available resources for critically ill children remain scarce. The Pediatric Intensive Care (PIC) database remains, to the best of current knowledge, the only open-access resource focused on pediatric critical care research. Methods This study evaluates the predictive potential of the PIC database. After applying strict cohort selection and data preprocessing, mortality risk was analyzed using statistical analysis through univariate Odds Ratio (OR) to measure associations, and machine learning (ML) models interpreted through SHAP values to assess feature importance and model behavior. Results ML experiments showed moderate but limited predictive performance for mortality. CatBoost performed best, with an AUC of 0.8521 (0.8414–0.8627) on the hold-out test set, but low sensitivity (42%) limited its ability to identify non-survivors. These findings suggest dataset-related constraints that reduce clinical reliability. Most relevant variables were minimum lactate, maximum lactate, mean respiratory rate, maximum hematocrit, and minimum urea. Conclusion Analysis showed that PIC lacks essential physiological data, highlighting the limits of current pediatric databases and the need for richer multicenter datasets to enable more robust, clinically applicable AI models. Impact Provides the first comprehensive evaluation of the PIC database, assessing its suitability for machine learning-based predictive modeling. Identifies key clinical variables and data gaps that influence the reliability of AI–based mortality prediction and model performance. Underscores the need for richer, multicenter pediatric ICU databases for research that incorporate essential physiological and neurological indicators. Demonstrates the value of combining statistical association and AI interpretability methods to assess the predictive integrity of clinical datasets.