Accurate mortality risk prediction is crucial in pediatric intensive care to guide clinical decision-making, allocate resources, benchmark quality, and counsel families [1]. Traditional adult ICU scores (e.g. APACHE, SAPS) were developed in the 1980s but generally perform poorly in children due to differing physiology and disease patterns [2]. Consequently, pediatric-specific models have been developed. The Pediatric Risk of Mortality (PRISM) score is a physiology-based risk model that uses 14–17 commonly measured variables (e.g. vital signs, blood gases, laboratory values) in the first hours of PICU admission to estimate mortality risk [3]. The PRISM framework has undergone several updates (PRISM III, PRISM IV) to maintain accuracy as care improves. In practice, PRISM is widely used to calculate standardized mortality ratios and to adjust for illness severity in pediatric cohorts [4]. By contrast, the Acute Physiology and Chronic Health Evaluation (APACHE) II score was developed for adult ICU populations (1985) and combines 12 physiologic parameters plus age and chronic health points into a logistic regression model of mortality [5]. APACHE II has been extensively validated in adult cohorts, but it has not been widely applied or recalibrated for children [6,7]. A recent commentary noted that although APACHE II could be calculated in pediatric patients, only a few small studies have examined its performance, and adult-derived thresholds may not generalize to children. For example, one single-center PICU report (n=100) found APACHE II distinguished survivors from nonsurvivors (mean scores 16.6 vs 26.1) with an area under the ROC curve (AUC) of 0.889 [8]. However, APACHE II’s calibration and discrimination in broader pediatric cohorts remain uncertain. Thus, there is a clear need to directly compare PRISM and APACHE II in predicting PICU mortality. Existing pediatric studies usually focus on PRISM or the Pediatric Index of Mortality (PIM), with little direct head-to-head data on APACHE II. In PICUs, PRISM (often PRISM III or IV), PIM and the Pediatric Logistic Organ Dysfunction (PELOD) score are the standard models [4]. APACHE II’s adult origins and lack of pediatric recalibration leave an evidence gap. Our retrospective cohort study aims to evaluate and compare the discriminative ability and calibration of PRISM (III or IV) versus APACHE II in a modern PICU population. By quantifying each score’s predictive accuracy, we seek to inform the choice of mortality risk model for pediatric critical care.

In Pakistan’s resource-constrained healthcare landscape, pediatric intensive care units (PICUs) grapple with immense clinical and logistical challenges, yet limited data exist evaluating prognostic models in this context. This retrospective cohort study represents the first institutional head-to-head comparison of PRISM and APACHE II scores at Lahore General Hospital, thereby filling a critical knowledge gap in regional pediatric critical care. While prior national studies have independently validated severity indices, none have directly contrasted them within a single cohort, underscoring the novelty and relevance of our analysis. By identifying which scoring system more accurately stratifies mortality risk, our findings have tangible implications for optimizing triage, resource allocation, and outcomes in high-acuity pediatric care. The study was conducted in accordance with STROBE guidelines and received institutional ethics committee approval, with informed consent waived due to its retrospective nature—ensuring both methodological rigor and ethical integrity.

Study design and setting

This retrospective cohort study was conducted at the Pediatric Intensive Care Unit (PICU) of Lahore General Hospital, Lahore. The study spanned a period of 6 months.

Study population

Pediatric patients aged 0 to 18 years admitted to the PICU during the study period were eligible for inclusion. Inclusion criteria required that patients had complete data necessary for calculating both the Pediatric Risk of Mortality (PRISM) score and the Acute Physiology and Chronic Health Evaluation II (APACHE II) score within the first 24 hours of PICU admission. Patients were excluded if they were transferred from another intensive care unit or if essential data required to compute either scoring system were missing or incomplete.

Data collection

Data were extracted from patient medical records using a structured data collection form. Variables collected included:

• Demographic information: age, sex
• Clinical characteristics: comorbidities, reason for PICU admission
• Physiological parameters: heart rate, respiratory rate, mean arterial pressure, temperature, Glasgow Coma Scale, and other relevant clinical signs
• Laboratory findings: arterial pH, serum bicarbonate, lactate levels, and other variables as required by the PRISM and APACHE II scoring systems
• Outcomes and interventions: duration of PICU stay, requirement for mechanical ventilation, and in-hospital mortality

Scoring was performed retrospectively using standard criteria available at:

• PRISM: https://sfar.org/scores2/prism2.php
• APACHE II: https://www.mdcalc.com/calc/1868/apache-ii-score

Statistical analysis

All statistical analyses were performed using SPSS and R software. Descriptive statistics were used to summarize baseline demographic and clinical characteristics; continuous variables were presented as means ± standard deviation or medians with interquartile ranges, while categorical variables were expressed as frequencies and percentages.

The predictive performance of the PRISM and APACHE II scores was evaluated using Receiver Operating Characteristic (ROC) curve analysis. The area under the ROC curve (AUC) was calculated for each score to assess discriminatory ability. The DeLong test was used to compare the AUCs of the two models. Calibration was assessed using the Hosmer–Lemeshow goodness-of-fit test and Brier scores to evaluate the accuracy of predicted mortality against observed outcomes. Mean absolute error (MAE) and mean squared error (MSE) were also calculated. Statistical significance was defined as a p-value <0.05.

A total of 214 patients were enrolled in the study, with a nearly equal gender distribution. There were 111 females (50.7%) and 103 males (49.3%) (Table 1). Pearson correlation analysis revealed a very weak positive correlation between PRISM scores and mortality (r = 0.054), and a very weak negative correlation between APACHE II scores and mortality (r = –0.055). These results suggest that neither score demonstrated a strong linear association with patient outcome in this dataset. Multivariate logistic regression showed that neither the PRISM score (β = 0.0665, p = 0.521) nor the APACHE II score (β = –0.0549, p = 0.522) was a statistically significant predictor of mortality. The model intercept was statistically significant (β = –1.8839, p = 0.048). However, the overall model fit was poor (pseudo R² = 0.0102), and the log-likelihood ratio test was not statistically significant (p = 0.7286). ROC curve analysis revealed that the PRISM score had an area under the curve (AUC) of 0.656, while the APACHE II score had an AUC of 0.587. Although PRISM performed better numerically, DeLong’s test indicated no statistically significant difference between the two AUCs (Z = 0.864, p = 0.3877; 95% CI for AUC difference: –0.087 to 0.225). The Hosmer–Lemeshow goodness-of-fit test showed acceptable calibration for both models. For PRISM, χ² = 6.25 (df = 7, p = 0.5107), and for APACHE II, χ² = 9.47 (df = 8, p = 0.3046), indicating no significant difference between observed and predicted outcomes across risk deciles. The Brier score was marginally lower for PRISM (0.123) compared to APACHE II (0.127), suggesting slightly better predictive accuracy. However, mean absolute error (MAE) and mean squared error (MSE) favored APACHE II (MAE: 0.023; MSE: 0.00068) over PRISM (MAE: 0.040; MSE: 0.00193). The 90th percentile of absolute error was also lower for APACHE II (0.044) compared to PRISM (0.058), indicating more consistent calibration. Neither PRISM nor APACHE-II scores showed a statistically significant difference between the two groups (likely deceased vs survived), based on the p-values (p >0.05 for both). Although PRISM has a larger absolute mean difference (-2.339 vs -1.507), it is still not statistically significant, indicating insufficient evidence to conclude a real difference in severity scores between the groups using this test.

Figure 1. Age distribution histogram of PICU patients.

Figure 2. Receiver operating characteristic (ROC) curves comparing the predictive performance of PRISM and APACHE II scores for mortality in PICU patients.

Figure 3. Calibration plot for PRISM and APACHE II scores in predicting mortality among PICU patients.

Our study adds to the growing body of evidence that adult-derived severity indices, such as APACHE II, have limited applicability in pediatric populations. While PRISM demonstrated comparatively better calibration and discrimination, neither score achieved the level of predictive accuracy reported in larger multicenter cohorts, underscoring the importance of contextual validation. This finding is particularly relevant for low- and middle-income countries (LMICs), where resource allocation is constrained and clinicians often rely on pragmatic adaptations of available tools. By directly contrasting PRISM and APACHE II in a large South Asian cohort, our work highlights that adopting adult models without recalibration risks misclassification of mortality risk and potentially misinforms clinical decision-making in pediatric intensive care units.

Recent literature underscores that PRISM-based models generally offer good discrimination for PICU mortality, while calibration may vary across populations [9,10]. In the Brazilian multicenter cohort of Rodrigues-Santos et al., PRISM exhibited an AUROC of 0.86 (95% CI 0.83–0.89), indicating strong discrimination. However, it underestimated mortality across a broad range of risk having standardized mortality ratio of 1.3:1 and demonstrated poor calibration [9]. Similarly, Alkhalifah et al. in a Saudi tertiary PICU reported PRISM discrimination with AUC ~0.81 [11]. A systematic review by Shen et al. confirmed that PRISM has good discriminatory power with pooled AUC ranging between 0.8 to 0.9 in PICU patients [10], though quality of evidence was rated low to moderate. In these studies, PRISM consistently achieved AUCs generally in the 0.80–0.90 range for mortality, comparable to or slightly higher than other pediatric scores. For example, PRISM-III showed AUC of 0.81 versus PIM-3 AUC of 0.80 in one Saudi study [11]. By contrast, literature on APACHE II in pediatric settings is sparse. As an adult metric, APACHE II is not standard in PICUs [7]. The few available pediatric data suggest APACHE II can discriminate survivors from non-survivors, but usually in small or specialized cohorts. Sankar’s review notes one study reporting an AUC of 0.889 for APACHE II in 100 children [8]. However, without multicenter validation, generalization is unknown. No large PICU study has been published validating APACHE II. Adult ICU analyses have long shown APACHE II’s AUC around 0.85 for general populations, but its performance on unmodified pediatric physiology is likely inferior. For example, the Pediatric Index of Mortality and PELOD-2 achieve similarly high discrimination with simpler inputs [12]. In any case, APACHE II’s calibration to pediatric risk remains untested.

Clinically, PRISM scores are widely used in PICUs for research and benchmarking [4], whereas APACHE II would require adaptation. PRISM’s advantage is pediatric specificity; APACHE II’s potential advantage would be familiarity since adult ICUs use it but in children, chronic conditions are captured differently. Some authors argue that even newer machine-learning models or combined scores might eventually outperform static scores [13].

Our results suggest trends with modest effect sizes, highlighting the need for larger multicenter studies, they also fill an important evidence gap by providing one of the largest head-to-head comparisons of PRISM and APACHE II in pediatric critical care from an LMIC setting. The consistency of our findings with prior smaller reports strengthens the conclusion that PRISM retains greater clinical utility, while APACHE II should not be extrapolated uncritically to children. In line with previous studies, several investigations have demonstrated strong predictive performance of PRISM-based models in pediatric intensive care populations. In a multicenter prospective study conducted across four PICUs, Ekinci et al. evaluated multiple pediatric mortality prediction scores, including PRISM and PRISM-IV, and reported strong discriminatory performance for these models in critically ill children, supporting their utility for risk stratification in PICU settings [16]. Similarly, Kaur et al. assessed the performance of the PRISM III score in a tertiary PICU and found excellent discrimination for mortality prediction. In their analysis, each one-point increase in the PRISM III score was associated with a 1.25-fold increase in the odds of mortality, highlighting the strong prognostic relationship between physiologic derangement and clinical outcomes in critically ill pediatric patients [17]. Taken together, these findings from multicenter and single-center PICU cohorts reinforce the established prognostic utility of PRISM-based models, although variability in discrimination across different healthcare settings suggests that contextual validation remains essential, particularly in resource-limited environments. These findings highlight the need for pediatric-specific, locally validated risk prediction models and underscore the importance of future multicenter studies aimed at recalibrating existing scores or developing novel indices tailored to diverse healthcare settings. In doing so, our study not only cautions against the uncritical application of adult-derived scoring systems in pediatric populations but also contributes to the broader effort to improve mortality risk prediction in global pediatric critical care.

In this retrospective cohort of 214 critically ill children, PRISM demonstrated modestly better performance than APACHE II in predicting PICU mortality, although neither score showed strong discriminatory ability. These findings reinforce the importance of pediatric-specific severity scoring systems and highlight the limitations of applying adult-derived indices to pediatric populations without recalibration. Our study provides regional evidence from a South Asian PICU and underscores the need for locally validated prognostic tools. Future multicenter studies are warranted to refine existing models and develop context-appropriate risk prediction systems for pediatric critical care.

None declared.

None.

This study was approved by the Institutional Review Board (IRB) of Lahore General Hospital (LGH), Lahore, Pakistan. All procedures performed in this study involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards.