PICTOR versus PRISM-3: a comparative evaluation of mortality prediction models in a Pakistani paediatric PICU: a retrospective cohort study

Background : PRISM-3 and PIM-3 are the paediatric severity scores widely used to estimate mortality risk in the paediatric intensive care unit (PICU), yet both were derived from high-income country populations and have never been formally validated in Pakistani hospitals. Whether their predicted prob- abilities remain clinically trustworthy in lower-middle-income country (LMIC) settings is unknown. This study formally assessed PRISM-3 calibration in a Pak- istani PICU, identified clinical predictors of in-hospital mortality, and developed PICTOR (Pediatric Intensive Care Outcome Risk Predictor), a locally trained, age-stratified ensemble framework providing continuously updated mortality risk predictions at per-minute, 30-minute, and hourly horizons from routinely collected monitoring data. Methods : A retrospective cohort study of 1,586 patients (1,851 visits; 191 deaths) admitted to the PICU at Aga Khan University Hospital, Pakistan (September 2018–September 2021). PRISM-3 calibration was assessed using the Hosmer-Lemeshow goodness-of-fit test. Predictors of mortality were identified by Mann-Whitney U test and univariate logistic regression. PICTOR was trained within five paediatric age groups using clinician-guided variable selection, cluster- based oversampling within the training partition only, and ensemble stacking of machine learning and deep learning models. All hyperparameter optimisation was performed within the training set only. Results : PRISM-3 was miscalibrated (Hosmer-Lemeshow HL = 272.57, p < 0.001; confirmed in all five age groups individually), systematically underesti- mating observed mortality. The strongest predictors of death were hypercapnia (OR 1.87, 95% CI: 1.64–2.14, p < 0.001) and overall physiological severity (partial PRISM-3 score OR 1.97, 95% CI: 1.69–2.29, p < 0.001). PICTOR out- performed partial PRISM-3 on both primary metrics in every age group (F1: 0.367–0.580 vs. 0.238–0.360; AUPRC: 0.302–0.496 vs. 0.169–0.327), achieving an AUCROC of 0.871 (95% CI: 0.870–0.872) and F1-score of 0.441 (95% CI: 0.439–0.442) for minute-by-minute prediction, improving to 0.649 at one hour. Conclusions : PRISM-3 is miscalibrated in this Pakistani PICU cohort and should not be used as a standalone mortality estimator without local recalibra- tion. PICTOR, trained on local data and delivering multi-horizon continuous risk prediction from routine monitoring, provides a clinically meaningful alternative with direct implications for other LMIC settings where internationally derived severity tools may systematically underestimate mortality risk.