Objective Early identification of mortality risk in critically ill children with suspected infection remains challenging. Whether severity scores (SIRS, pSOFA, Phoenix) add incremental value beyond data-derived physiologic features in machine-learning models is uncertain. We compared the predictive value of physiologic dynamics, treatment features, and continuous laboratory trends expert-derived severity representations for 24-hour PICU mortality prediction, and derived a parsimonious, interpretable model. Study Design We analyzed 3,310 PICU encounters (107 deaths, 3.2% mortality) using first-24-hour data. A staged, domain-structured feature selection strategy was used to build sequential XGBoost models evaluated via 5-fold stratified cross-validation. Performance was assessed by AUROC, AUPRC, Brier score, and specificity at 80% sensitivity, with SHAP-guided pruning at each stage. Results Vital sign extrema provided baseline discrimination. Incorporation of variability and slope features improved calibration and specificity. Vasopressor and supplemental oxygen timing and intensity features produced substantial gains in AUROC and AUPRC. Continuous laboratory extrema and temporal trends reflecting evolving organ dysfunction yielded further improvement. Expert-threshold-derived metrics, including SIRS, pSOFA, and Phoenix composite scores, did not improve performance beyond data-derived features. Conclusions Routinely collected hourly PICU data from the first-24-hour are sufficient for XGBoost to extract deterioration patterns strongly predictive of early mortality, without requiring severity thresholds.
Velez et al. (Mon,) studied this question.