Critically ill children admitted to pediatric intensive care units (PICUs) represent a heterogeneous population with varying diagnoses, physiological disturbances, disease severity, and risks of death. Early and objective assessment of illness severity is essential for identifying high-risk patients, guiding clinical prioritization, counselling families, evaluating quality of care, and comparing outcomes across intensive care units. However, clinical judgement alone may be subjective and influenced by the experience of the treating physician. Standardized mortality-prediction scores provide an objective means of quantifying severity and estimating the probability of death using routinely available clinical and laboratory variables. The Pediatric Risk of Mortality-III (PRISM-III) score is a physiology-based scoring system derived from the most abnormal physiological and laboratory values recorded during the initial hours of PICU admission. It incorporates cardiovascular, neurological, respiratory, biochemical, and hematological parameters, with higher scores indicating greater physiological instability and mortality risk. PRISM-III has demonstrated useful discrimination for mortality across different pediatric critical-care populations; however, it requires multiple variables and may be influenced by treatment administered during the data-collection period [1,2]. The Pediatric Index of Mortality-3 (PIM-3) is an updated admission-based mortality-prediction model that uses information obtained at or around the time of the first face-to-face contact between the child and the PICU team. Its variables include systolic blood pressure, pupillary reaction, oxygenation indices, mechanical ventilation status, base excess, elective admission, recovery following a procedure, and specified high- and low-risk diagnoses. PIM-3 is relatively simple and allows risk estimation before extensive PICU treatment alters the patient’s physiological condition [3]. External validation studies have shown that PIM-3 generally provides good discrimination, although calibration may differ according to geographical location, case mix, referral pattern, and available healthcare resources [4]. PIM-3 and PRISM-III differ in their number of variables, timing of assessment, computational complexity, and potential susceptibility to treatment-related changes. Studies comparing pediatric mortality models have reported variable results, and the score with the best performance in one setting may not necessarily perform equally well in another [5]. Consequently, local validation is essential before a scoring system is routinely used for prognostication or evaluation of PICU performance. The present study compared the discrimination, calibration, and overall predictive performance of PIM-3 and PRISM-III for mortality among children admitted to a PICU. AIM To compare the performance of PIM-3 and PRISM-III scores in predicting mortality among patients admitted to the pediatric intensive care unit. OBJECTIVES 1. To calculate the PIM-3 and PRISM-III scores among eligible children admitted to the PICU. 2. To determine the association of PIM-3 and PRISM-III scores with PICU mortality and other clinical outcomes. 3. To compare the discrimination, calibration, sensitivity, specificity, and overall predictive accuracy of PIM-3 and PRISM-III for PICU mortality.

Source of Data Data were obtained from children admitted to the PICU during the study period. Information was collected from bedside clinical assessments, case records, nursing charts, laboratory reports, arterial blood-gas reports, treatment records, and the hospital information system. The parents or legally authorized representatives of eligible children were informed about the study, and written informed consent was obtained where required by the Institutional Ethics Committee. Study Design This was a hospital-based prospective observational comparative study. PIM-3 and PRISM-III were simultaneously assessed in the same participants; therefore, the study involved a paired comparison of the prognostic performance of the two scoring systems. No alteration was made to the diagnostic investigations or treatment provided to the participants. All children were managed according to the PICU’s prevailing standard treatment protocols. Study Location The study was conducted in the Pediatric Intensive Care Unit of the Department of Pediatrics, a tertiary-care teaching and referral hospital. The PICU received medical, surgical, emergency, and postoperative pediatric patients referred from the emergency department, pediatric wards, operating rooms, and surrounding healthcare centres. Study Duration The study was conducted over 18 months. This period included participant enrolment, clinical follow-up until PICU outcome, data verification, statistical analysis, and preparation of the study report. Sample Size A total of 120 eligible PICU patients were included. The sample size was determined on the basis of the anticipated discriminatory ability of the two mortality scores, the expected PICU mortality rate, a 95% confidence level, and 80% statistical power. Eligible participants were recruited consecutively until the required sample size was reached. Because the same children were assessed using both scores, the comparison of predictive accuracy was treated as a paired analysis. Study Population The study population consisted of critically ill children admitted to the PICU during the defined study period who met the eligibility criteria. Inclusion Criteria 1. Children aged 1 month to 18 years who were admitted to the PICU. 2. Children who remained in the PICU for a sufficient period to complete the required assessment. 3. Patients for whom the clinical and laboratory variables required to calculate both PIM-3 and PRISM-III were available. 4. Children whose parents or legally authorized representatives provided written informed consent, wherever applicable. Exclusion Criteria 1. Neonates younger than one month. 2. Patients who died, were discharged, or were transferred before the required score variables could be recorded. 3. Patients readmitted to the PICU during the same hospitalization; only the first admission was considered. 4. Children admitted exclusively for routine monitoring without critical illness, where specified by the scoring guidelines. 5. Patients with substantial missing clinical or laboratory information that prevented calculation of either score. 6. Children discharged or transferred against medical advice before the final PICU outcome could be established. 7. Parents or legally authorized representatives who declined participation, where consent was required. Sampling Technique A consecutive sampling method was used. Every eligible child admitted during the study period was screened and included until the predetermined sample size of 120 participants was achieved. Procedure and Methodology After admission to the PICU, each patient was screened according to the inclusion and exclusion criteria. A unique study identification number was assigned to maintain confidentiality. Demographic details, including age, sex, weight, source of admission, primary diagnosis, comorbidities, operative status, and reason for PICU admission, were recorded. PIM-3 assessment The PIM-3 variables were documented at the time of the first face-to-face contact with the PICU team or within the period specified by the original PIM-3 model. The variables included: • Elective or emergency admission; • Recovery following a procedure; • Cardiac bypass status; • Mechanical ventilation status; • Systolic blood pressure; • Pupillary reaction to bright light; • Base excess in arterial or capillary blood; • Fraction of inspired oxygen; • Partial pressure of arterial oxygen; • Presence of specified low-risk diagnoses; • Presence of specified high-risk diagnoses; and • Presence of specified very-high-risk diagnoses. Values obtained after death or cardiopulmonary resuscitation were not used unless explicitly permitted by the scoring rules. The PIM-3 logit was calculated using the original model equation, and the predicted probability of death was obtained as: PRISM-III assessment The PRISM-III score was calculated from the most abnormal eligible physiological and laboratory values recorded during the specified early PICU assessment period. A single abnormal value was considered sufficient when it satisfied the score definition. The assessed variables included: • Systolic blood pressure; • Heart rate; • Body temperature; • Mental status or Glasgow Coma Scale; • Pupillary responses; • Arterial pH; • Total carbon dioxide; • Arterial partial pressure of oxygen; • Arterial partial pressure of carbon dioxide; • Blood glucose; • Serum potassium; • Blood urea nitrogen; • Serum creatinine; • White blood-cell count; • Platelet count; and • Prothrombin time or activated partial thromboplastin time. Age-specific thresholds were applied according to the PRISM-III scoring instructions. The points assigned to abnormal variables were added to obtain the total score. Higher PRISM-III scores indicated greater physiological instability and a higher risk of mortality. To reduce observer-related variation, the investigators used predetermined definitions and standardized score sheets. Where uncertainty existed, the original case record and laboratory report were reviewed before finalizing the score. Both scores were calculated without influencing the clinical management of the child. Outcome assessment Participants were followed from PICU admission until discharge from the PICU, transfer to another unit, or death. The primary outcome was all-cause PICU mortality, categorized as survivor or non-survivor. Secondary outcomes included: • Requirement for invasive mechanical ventilation; • Requirement for vasoactive or inotropic support; • Development of multiple-organ dysfunction; • Duration of mechanical ventilation; • Length of PICU stay; and • Hospital outcome, where available. Sample Processing No additional blood sample was collected solely for the research. Blood samples required for routine clinical management were collected under aseptic precautions and processed in the hospital’s accredited laboratory. Arterial or capillary blood-gas analysis was performed promptly using a calibrated blood-gas analyser. Samples for complete blood count were collected in EDTA tubes and analysed using an automated hematology analyser. Samples for serum electrolytes, glucose, blood urea nitrogen, and creatinine were collected in appropriate tubes and analysed using the hospital’s automated biochemistry analyser. Coagulation samples were collected in sodium-citrate tubes and processed using an automated coagulation analyser. Laboratory values falling within the score-specific assessment periods were retrieved. When more than one eligible result was available, the most abnormal value was selected according to the PRISM-III scoring rules. Quality-control procedures were followed as per laboratory protocol. Data Collection Data were collected using a predesigned and pretested case-record form comprising the following sections: 1. Demographic and baseline characteristics; 2. Source and nature of PICU admission; 3. Primary diagnosis and comorbidities; 4. PIM-3 clinical and laboratory variables; 5. PRISM-III physiological and laboratory variables; 6. Respiratory, circulatory and organ-support requirements; 7. Duration of ventilation and PICU stay; and 8. Final PICU outcome. The collected forms were reviewed daily for completeness and consistency. Data were coded and entered into a password-protected electronic database. Patient names and hospital registration numbers were excluded from the analytical dataset. A proportion of the entries was cross-checked against the original records to minimize transcription errors. Statistical Methods Data were analysed using IBM SPSS Statistics 28.0. Continuous variables were assessed for normality using histograms, Q-Q plots, and the Shapiro-Wilk test. Normally distributed variables were expressed as mean and standard deviation, whereas skewed variables were presented as median and interquartile range. Categorical variables were summarized as frequencies and percentages. PIM-3 and PRISM-III scores were compared between survivors and non-survivors using the independent-samples t test or Mann-Whitney U test, as appropriate. Associations between categorical variables and mortality were evaluated using the chi-square test or Fisher’s exact test. The predictive performance of each score was evaluated as follows: • Discrimination: Receiver operating characteristic curves were constructed, and the area under the ROC curve (AUC) was reported with a 95% confidence interval. • Comparison of AUCs: As both scores were obtained from the same participants, correlated AUCs were compared using DeLong’s test. • Diagnostic accuracy: The optimal cut-off value was identified using the Youden index. Sensitivity, specificity, positive predictive value, negative predictive value, likelihood ratios, and overall accuracy were calculated with 95% confidence intervals. • Calibration: Agreement between predicted and observed mortality was assessed using calibration plots and the Hosmer-Lemeshow goodness-of-fit test. A non-significant result indicated no evidence of poor calibration, while interpretation was supported by the calibration plot. • Standardized mortality ratio: The number of observed deaths was divided by the sum of predicted deaths for each scoring system, and a 95% confidence interval was calculated. • Overall performance: The Brier score was calculated where individual predicted mortality probabilities were available; a lower Brier score represented better overall prediction. • Multivariable analysis: Binary logistic regression was used, where appropriate, to assess whether each score independently predicted mortality. Results were expressed as odds ratios with 95% confidence intervals. All statistical tests were two-tailed. A P value <0.05 was considered statistically significant. Ethical Considerations The study was initiated after approval from the Institutional Ethics Committee. Confidentiality of participant information was maintained throughout the study. The study involved no experimental intervention, and treatment decisions were made independently by the attending pediatric intensivists.

Table 1: Comparative performance of PIM-3 and PRISM-III in predicting PICU mortality (N=120) Performance parameter PIM-3 PRISM-III Effect estimate (95% CI) Test of significance P value Observed PICU mortality, n (%) 23 (19.2%) 23 (19.2%) 19.2% (13.1%-27.1%) Mean predicted mortality, Mean (SD), % 16.8 (19.7) 18.7 (21.4) Paired mean difference: 1.9% (0.3%-3.5%) Paired t=2.36 0.020* Median predicted mortality, median (IQR), % 8.9 (3.7-20.6) 10.8 (4.6-24.9) Median paired difference: 1.6% (0.4%-3.1%) Wilcoxon Z=2.18 0.029* Area under ROC curve 0.866 0.914 AUC difference: 0.048 (0.002-0.094) DeLong Z=2.04 0.041* Sensitivity at optimal cut-off, % 82.6% 91.3% Difference: 8.7% (-6.2%-23.6%) McNemar χ²=1.33 0.248 Specificity at optimal cut-off, % 85.6% 88.7% Difference: 3.1% (-5.2%-11.4%) McNemar χ²=0.50 0.480 Overall accuracy, n (%) 102 (85.0%) 107 (89.2%) Difference: 4.2% (-2.3%-10.6%) McNemar χ²=2.08 0.149 Brier score 0.112 0.084 Mean difference: -0.028 (-0.050 to -0.006) Paired t=-2.54 0.012* Predicted number of deaths 20.2 22.4 Difference: 2.2 deaths Standardized mortality ratio 1.14 1.03 Difference: -0.11 (-0.31-0.09) Wald Z=-1.08 0.281 Hosmer-Lemeshow goodness-of-fit χ²=9.18, df=8 χ²=5.36, df=8 Hosmer-Lemeshow test 0.327 / 0.718 *Statistically significant at P<0.05. A larger AUC and lower Brier score indicate better predictive performance. A non-significant Hosmer-Lemeshow test indicates no evidence of poor calibration. Among the 120 children admitted to the PICU, 23 (19.2%; 95% CI: 13.1%-27.1%) died. The mean predicted mortality was 16.8% (SD 19.7) with PIM-3 and 18.7% (SD 21.4) with PRISM-III. PRISM-III generated a significantly higher mean predicted mortality than PIM-3, with a paired mean difference of 1.9% (95% CI: 0.3%-3.5%; paired t=2.36, P=0.020). Similarly, the median predicted mortality was significantly higher with PRISM-III than with PIM-3 10.8% versus 8.9%, respectively (median paired difference 1.6%; 95% CI: 0.4%-3.1%; Wilcoxon Z=2.18, P=0.029). Both models demonstrated good discriminatory ability; however, the AUC was significantly greater for PRISM-III than for PIM-3 (0.914 versus 0.866), with an AUC difference of 0.048 (95% CI: 0.002-0.094; DeLong Z=2.04, P=0.041). PRISM-III also showed numerically higher sensitivity (91.3% versus 82.6%), specificity (88.7% versus 85.6%), and overall accuracy (89.2% versus 85.0%), although these differences were not statistically significant. The Brier score was significantly lower for PRISM-III than for PIM-3 (0.084 versus 0.112; mean difference -0.028, 95% CI: -0.050 to -0.006; P=0.012), indicating better overall predictive performance. PIM-3 predicted 20.2 deaths and produced a standardized mortality ratio of 1.14, whereas PRISM-III predicted 22.4 deaths and produced a ratio of 1.03, which was closer to the observed mortality; however, the difference was not significant (P=0.281). The non-significant Hosmer-Lemeshow results for PIM-3 (P=0.327) and PRISM-III (P=0.718) indicated acceptable calibration for both models. Table 2: Distribution of PIM-3 and PRISM-III scores among eligible PICU patients (N=120) Score parameter Overall (N=120) Survivors (n=97) Non-survivors (n=23) Effect estimate (95% CI) Test of significance P value PIM-3 predicted mortality, Mean (SD), % 16.8 (19.7) 9.2 (10.6) 48.9 (22.8) Mean difference: 39.7% (31.4%-48.0%) Welch t=9.45 <0.001* PIM-3 predicted mortality, median (IQR), % 8.9 (3.7-20.6) 6.1 (2.8-11.7) 47.3 (31.8-66.4) Hodges-Lehmann difference: 38.2% (29.6%-47.8%) Mann-Whitney Z=6.24 <0.001* PIM-3 risk <5%, n (%) 54 (45.0%) 53 (54.6%) 1 (4.3%) Reference PIM-3 risk 5%-<15%, n (%) 38 (31.7%) 33 (34.0%) 5 (21.7%) OR=8.03 (0.91-70.74) χ² for trend=57.82 <0.001* PIM-3 risk 15%-<30%, n (%) 17 (14.2%) 9 (9.3%) 8 (34.8%) OR=47.11 (5.43-408.52) PIM-3 risk ≥30%, n (%) 11 (9.2%) 2 (2.1%) 9 (39.1%) OR=238.50 (20.52-2772.25) PRISM-III score, Mean (SD) 10.9 (8.7) 7.6 (5.4) 24.8 (7.9) Mean difference: 17.2 (13.7-20.7) Welch t=9.80 <0.001* PRISM-III score, median (IQR) 8.0 (4.0-15.0) 7.0 (3.0-10.0) 24.0 (19.0-30.0) Hodges-Lehmann difference: 16.0 (13.0-19.0) Mann-Whitney Z=6.68 <0.001* PRISM-III score 0-5, n (%) 41 (34.2%) 41 (42.3%) 0 (0.0%) Fisher-Freeman-Halton test <0.001* PRISM-III score 6-10, n (%) 36 (30.0%) 34 (35.1%) 2 (8.7%) PRISM-III score 11-20, n (%) 27 (22.5%) 19 (19.6%) 8 (34.8%) PRISM-III score >20, n (%) 16 (13.3%) 3 (3.1%) 13 (56.5%) *Statistically significant at P<0.05. Percentages in the survivor and non-survivor columns were calculated within the respective outcome groups. The overall mean PIM-3 predicted mortality was 16.8% (SD 19.7), but it was substantially higher among non-survivors than survivors 48.9% versus 9.2%, respectively. The mean difference of 39.7% was statistically significant (95% CI: 31.4%-48.0%; Welch t=9.45, P<0.001). The median predicted mortality was also significantly higher among non-survivors than survivors (47.3% versus 6.1%), with a Hodges-Lehmann difference of 38.2% (95% CI: 29.6%-47.8%; P<0.001). Of the 54 children in the PIM-3 risk category below 5%, only one died. In comparison, mortality was observed in 5 of 38 children with a predicted risk of 5%-<15%, 8 of 17 with a risk of 15%-<30%, and 9 of 11 with a risk of at least 30%. Relative to the lowest-risk group, the odds of mortality increased markedly across successive PIM-3 categories, reaching an OR of 238.50 in the ≥30% category. This increasing trend was statistically significant (χ² for trend=57.82, P<0.001). The overall mean PRISM-III score was 10.9 (SD 8.7). Non-survivors had a significantly higher mean score than survivors (24.8 versus 7.6), with a mean difference of 17.2 points (95% CI: 13.7-20.7; Welch t=9.80, P<0.001). No mortality occurred among the 41 children with PRISM-III scores of 0-5, whereas 2 of 36 patients with scores of 6-10, 8 of 27 with scores of 11-20, and 13 of 16 with scores above 20 died. The distribution of PRISM-III categories differed significantly between survivors and non-survivors (P<0.001), demonstrating a strong progressive relationship between increasing illness-severity scores and mortality. Table 3: Association of PIM-3 and PRISM-III scores with mortality and other clinical outcomes (N=120) Clinical outcome Outcome present, n (%) PIM-3, Mean (SD), % PIM-3 effect estimate (95% CI) PIM-3 test; P value PRISM-III, Mean (SD) PRISM-III effect estimate (95% CI) PRISM-III test; P value PICU mortality: Yes 23 (19.2%) 48.9 (22.8) MD=39.7% (31.4%-48.0%) t=9.45; <0.001* 24.8 (7.9) MD=17.2 (13.7-20.7) t=9.80; <0.001* PICU mortality: No 97 (80.8%) 9.2 (10.6) Reference 7.6 (5.4) Reference Invasive ventilation: Yes 43 (35.8%) 31.6 (24.7) MD=23.1% (16.7%-29.5%) t=7.15; <0.001* 17.8 (9.3) MD=10.7 (7.9-13.5) t=7.55; <0.001* Invasive ventilation: No 77 (64.2%) 8.5 (9.8) Reference 7.1 (4.9) Reference Vasoactive support: Yes 31 (25.8%) 37.9 (25.6) MD=28.5% (20.6%-36.4%) t=7.19; <0.001* 20.3 (9.1) MD=12.7 (9.4-16.0) t=7.60; <0.001* Vasoactive support: No 89 (74.2%) 9.4 (10.8) Reference 7.6 (5.2) Reference Multiple-organ dysfunction: Yes 19 (15.8%) 52.3 (24.1) MD=42.2% (32.4%-52.0%) t=8.55; <0.001* 26.1 (7.6) MD=18.1 (14.3-21.9) t=9.46; <0.001* Multiple-organ dysfunction: No 101 (84.2%) 10.1 (11.9) Reference 8.0 (5.7) Reference PICU stay >7 days: Yes 37 (30.8%) 27.8 (24.2) MD=15.9% (8.8%-23.0%) t=4.44; <0.001* 16.4 (9.8) MD=7.9 (4.8-11.0) t=5.09; <0.001* PICU stay ≤7 days 83 (69.2%) 11.9 (15.1) Reference 8.5 (6.4) Reference Mechanical ventilation duration† 43 (35.8%) ρ=0.46 (0.19-0.66) Spearman ρ=0.46; 0.002* ρ=0.53 (0.28-0.70) Spearman ρ=0.53; <0.001* PICU length of stay 120 (100%) ρ=0.32 (0.15-0.47) Spearman ρ=0.32; <0.001* ρ=0.39 (0.23-0.53) Spearman ρ=0.39; <0.001* †Correlation with ventilation duration was calculated among the 43 mechanically ventilated patients. MD=mean difference; ρ=Spearman correlation coefficient. *Statistically significant at P<0.05. Higher PIM-3 and PRISM-III values were significantly associated with mortality and adverse PICU outcomes. Non-survivors had a mean PIM-3 predicted mortality of 48.9%, compared with 9.2% among survivors, giving a mean difference of 39.7% (95% CI: 31.4%-48.0%; P<0.001). Their corresponding mean PRISM-III scores were 24.8 and 7.6, respectively, with a mean difference of 17.2 points (95% CI: 13.7-20.7%; P<0.001). Among the 43 children requiring invasive mechanical ventilation, the mean PIM-3 predicted mortality was 31.6%, compared with 8.5% among those who did not require ventilation (mean difference 23.1%; P<0.001); corresponding PRISM-III scores were 17.8 and 7.1 (mean difference 10.7; P<0.001). Children requiring vasoactive support also had significantly higher PIM-3 predictions (37.9% versus 9.4%) and PRISM-III scores (20.3 versus 7.6), with both comparisons yielding P<0.001. The largest score differences were observed among children with multiple-organ dysfunction: affected children had mean PIM-3 and PRISM-III values of 52.3% and 26.1, compared with 10.1% and 8.0 among those without multiple-organ dysfunction (both P<0.001). Patients staying in the PICU for more than seven days also had significantly higher PIM-3 and PRISM-III values than those with shorter stays (P<0.001). Both scores demonstrated significant positive correlations with the duration of mechanical ventilation and PICU length of stay. PRISM-III showed slightly stronger correlations with ventilation duration (ρ=0.53 versus 0.46) and PICU stay (ρ=0.39 versus 0.32), although both scoring systems were consistently associated with greater resource requirements and poorer clinical outcomes. Table 4: Discrimination, calibration and diagnostic accuracy of PIM-3 and PRISM-III for PICU mortality (N=120) Predictive-performance measure PIM-3 PRISM-III Comparative estimate (95% CI) Test of significance P value Optimal mortality cut-off ≥12% predicted risk Score ≥14 Youden index True positives, n 19 21 Difference=2 McNemar exact test 0.500 False negatives, n 4 2 Difference=-2 McNemar exact test 0.500 True negatives, n 83 86 Difference=3 McNemar exact test 0.453 False positives, n 14 11 Difference=-3 McNemar exact test 0.453 Sensitivity 82.6% (62.9%-93.0%) 91.3% (73.2%-97.6%) Difference=8.7% (-6.2%-23.6%) McNemar χ²=1.33 0.248 Specificity 85.6% (77.2%-91.2%) 88.7% (80.8%-93.5%) Difference=3.1% (-5.2%-11.4%) McNemar χ²=0.50 0.480 Positive predictive value 57.6% (40.8%-72.8%) 65.6% (48.3%-79.6%) Difference=8.0% (-14.6%-30.6%) Wald Z=0.69 0.490 Negative predictive value 95.4% (88.8%-98.2%) 97.7% (92.1%-99.4%) Difference=2.3% (-2.6%-7.2%) Wald Z=0.92 0.357 Positive likelihood ratio 5.73 (3.46-9.48) 8.05 (4.63-14.01) Ratio of LR+=1.40 (0.72-2.73) Wald Z=1.00 0.318 Negative likelihood ratio 0.20 (0.08-0.50) 0.10 (0.03-0.37) Ratio of LR-=0.49 (0.09-2.61) Wald Z=-0.84 0.401 Youden index 0.682 0.800 Difference=0.118 (-0.034-0.270) Bootstrap Z=1.52 0.128 Overall accuracy 85.0% (77.5%-90.3%) 89.2% (82.3%-93.6%) Difference=4.2% (-2.3%-10.6%) McNemar χ²=2.08 0.149 AUC 0.866 (0.787-0.945) 0.914 (0.852-0.976) Difference=0.048 (0.002-0.094) DeLong Z=2.04 0.041* Hosmer-Lemeshow χ², df=8 9.18 5.36 Calibration test 0.327 / 0.718 Calibration slope 0.84 (0.63-1.05) 0.96 (0.76-1.16) Difference=0.12 (-0.17-0.41) Wald Z=0.81 0.419 Calibration intercept 0.18 (-0.21-0.57) 0.06 (-0.28-0.40) Difference=-0.12 (-0.64-0.40) Wald Z=-0.45 0.651 Brier score 0.112 0.084 Difference=-0.028 (-0.050 to -0.006) Paired t=-2.54 0.012* AUC interpretation: 0.70-0.79=fair, 0.80-0.89=good, and ≥0.90=excellent discrimination. *Statistically significant at P<0.05. Using the optimal cut-offs of ≥12% predicted mortality for PIM-3 and ≥14 points for PRISM-III, PIM-3 correctly identified 19 of the 23 deaths, whereas PRISM-III correctly identified 21 deaths. PIM-3 produced four false-negative and 14 false-positive classifications, compared with two false negatives and 11 false positives with PRISM-III. Accordingly, PRISM-III showed higher sensitivity (91.3% versus 82.6%), specificity (88.7% versus 85.6%), positive predictive value (65.6% versus 57.6%), negative predictive value (97.7% versus 95.4%), and overall accuracy (89.2% versus 85.0%). Nevertheless, the confidence intervals for these differences included the null value, and none of the individual classification measures differed significantly between the models. The positive likelihood ratio was higher for PRISM-III than PIM-3 (8.05 versus 5.73), while its negative likelihood ratio was lower (0.10 versus 0.20), indicating numerically better rule-in and rule-out performance. PRISM-III also produced a higher Youden index than PIM-3 (0.800 versus 0.682), although this difference was not significant (P=0.128). PIM-3 demonstrated good discrimination, with an AUC of 0.866 (95% CI: 0.787-0.945), whereas PRISM-III demonstrated excellent discrimination, with an AUC of 0.914 (95% CI: 0.852-0.976). The AUC difference of 0.048 was statistically significant (P=0.041). Both models demonstrated acceptable calibration according to the Hosmer-Lemeshow test, although PRISM-III had a calibration slope closer to one and an intercept closer to zero. Finally, its significantly lower Brier score (0.084 versus 0.112; P=0.012) confirmed that PRISM-III provided better overall mortality prediction than PIM-3.

Comparative performance of PIM-3 and PRISM-III In the present study, PICU mortality was 19.2%, while PIM-3 and PRISM-III predicted mortality rates of 16.8% and 18.7%, respectively. Thus, both models slightly underestimated observed mortality, although PRISM-III was more closely aligned with the actual outcome. The standardized mortality ratio was consequently closer to unity for PRISM-III than for PIM-3 (1.03 versus 1.14). Differences between observed and predicted mortality across settings may reflect variation in case mix, referral delays, pre-PICU stabilization, availability of organ support, diagnostic composition, and model-development populations. Pollack et al. (2016)[1], while updating the PRISM model, emphasized that mortality models require periodic revision because the relationship between physiological abnormalities and mortality changes as intensive-care practices improve. Tyagi et al. (2018)[2] compared PRISM-III, PIM-2, and PIM-3 in an Indian PICU and found that all were useful, although PIM-3 demonstrated better discrimination and was easier to calculate. In contrast, the present study showed significantly better discrimination with PRISM-III. Such disagreement highlights that superiority is not universal and is influenced by local patient characteristics and data-collection practices. The PIM-3 AUC of 0.866 in the present study indicated good discrimination. This was higher than the c-index of 0.76 reported by Lee et al. (2017)[3] in a Korean PICU and the AUC of 0.826 reported by Jung et al. (2018)[4]. Malhotra et al. (2020)[5] also reported satisfactory PIM-3 performance in a Dubai PICU, supporting its value as a practical admission-based prognostic score. The present PRISM-III AUC of 0.914 indicated excellent discrimination and was consistent with the strong performance reported by Mirza et al. (2020)[6] and Kaur et al. (2020)[7], both of whom demonstrated a significant association between increasing PRISM-III scores and mortality. Nasser et al. (2020)[8] directly examined the reliability of PRISM-III and PIM-3 and found both scores useful for mortality assessment, although their performance varied across risk groups. A meta-analysis by Shen et al. (2021)[9] reported summary ROC values of 0.84 for PRISM-III/IV and 0.82 for PIM-3, with respective pooled sensitivities of 78% and 75%. These pooled results support the present finding that both models were effective, while PRISM-III demonstrated a modest discriminatory advantage. The Brier score was significantly lower for PRISM-III than PIM-3 in the present study (0.084 versus 0.112, P=0.012), indicating better combined discrimination and probability accuracy. Rahmatinejad et al. (2022)[10] similarly found that models based on PRISM-III performed better than those based on PIM-3 for both PICU and hospital mortality, although they stressed the potential need for local recalibration. Conversely, Ekinci et al. (2022)[11] reported slightly higher AUCs for PIM-3 than PRISM in a multicentre Turkish cohort 0.934 and 0.917, respectively. However, PIM-3 had poor calibration in their population despite excellent discrimination. This distinction is important: a model may correctly rank patients from low to high risk while still systematically overestimating or underestimating absolute mortality. Both scores showed acceptable calibration in the present study, with Hosmer-Lemeshow P values of 0.327 for PIM-3 and 0.718 for PRISM-III. PRISM-III had a calibration slope closer to one and an intercept closer to zero. These findings agreed with Zhang et al. (2021)[12], who found good calibration between PRISM-III-predicted and observed mortality in two Chinese PICUs. Conversely, the poor PIM-3 calibration reported by Ekinci et al.[11] illustrates why external validation should examine both AUC and calibration rather than relying solely on discrimination. Distribution of scores among survivors and non-survivors The mean PIM-3 predicted mortality was markedly higher among non-survivors than survivors (48.9% versus 9.2%, P<0.001). The mean PRISM-III score was also considerably higher among non-survivors (24.8 versus 7.6, P<0.001). The narrow confidence intervals around these differences and the highly significant test results showed strong separation between outcome groups. A clear dose-response relationship was evident across PIM-3 categories. Only one death occurred in the <5% group, whereas 9 of 11 patients in the ≥30% group died. The odds of death rose progressively across the risk strata, and the trend was significant (P<0.001). Jung et al. (2018)[4] similarly demonstrated progressive increases in observed mortality with increasing PIM-3-predicted risk. Toteja et al. (2024)[13] also found significantly higher PIM-3 values among non-survivors and supported its use for mortality stratification in an Indian tertiary-care PICU. The PRISM-III categories demonstrated an equally strong gradient. No deaths occurred among patients scoring 0-5, compared with 13 deaths among the 16 children scoring >20. Zhang et al. (2021)[12] reported that median PRISM-III scores were substantially higher among non-survivors than survivors, which closely supports the present findings. Mirza et al. (2020)[6] and Kaur et al. (2020)[7] likewise observed that mortality increased significantly as PRISM-III scores rose. The strong gradient across both models confirms that these scores are useful for risk stratification. Nevertheless, the extremely large odds ratios and wide confidence intervals in the highest PIM-3 categories should be interpreted carefully. They resulted from the small number of patients and sparse outcome counts in individual categories. The estimates demonstrate a strong association but do not necessarily provide a precise measure of its magnitude. Genu et al. (2023)[14], in a Brazilian multicentre validation involving more than 41,000 admissions, found close agreement between PIM-3-predicted and observed mortality, with an SMR of 1.00. Their findings demonstrate that PIM-3 can perform well at the population level when applied to large, systematically collected datasets. However, performance in a large national registry may not be directly transferable to a smaller single-centre PICU with higher mortality and a different diagnostic spectrum. Association with mortality and other clinical outcomes Both scoring systems were strongly associated not only with mortality but also with invasive ventilation, vasoactive support, multiple-organ dysfunction, prolonged PICU stay, and duration of mechanical ventilation. Patients requiring invasive ventilation had a mean PIM-3 predicted mortality of 31.6%, compared with 8.5% among patients not requiring ventilation. Their mean PRISM-III score was also substantially higher (17.8 versus 7.1). These findings were clinically plausible because respiratory failure and mechanical ventilation indicate greater physiological instability and advanced critical illness. Patients receiving vasoactive support had significantly higher PIM-3 and PRISM-III values than those not receiving such support. Kumar et al. (2023)[15] similarly found that higher pediatric severity scores were associated with mortality, invasive ventilation, and greater therapeutic intensity. In critically ill pediatric hematology and oncology populations, ventilatory and inotropic support have also been repeatedly associated with poor outcome, indicating that the score-outcome relationship persists even in high-risk diagnostic subgroups. The largest score differences were found among patients with multiple-organ dysfunction. Their mean PIM-3 predicted mortality was 52.3%, compared with 10.1% among children without multiple-organ dysfunction, while the corresponding PRISM-III scores were 26.1 and 8.0. This finding was consistent with the physiological basis of PRISM-III, which integrates abnormalities across neurological, cardiovascular, respiratory, biochemical, and hematological systems. Zhang et al. (2021)[12] found that PRISM-III and organ-dysfunction scores were both significantly higher among non-survivors, reinforcing the close relationship between physiological instability, organ failure, and mortality. Agrwal et al. (2024)[16] compared pSOFA with PRISM-III and PIM-2 and found that all three measures were significantly associated with pediatric critical-care outcomes. Their findings support the interpretation that mortality scores may also provide information about the overall burden of organ dysfunction. Nevertheless, PIM-3 and PRISM-III were developed primarily to estimate mortality at the group level and should not replace dedicated organ-dysfunction scores when sequential assessment of organ failure is required. Patients with PICU stays exceeding seven days had significantly higher scores than those with shorter admissions. Both models were positively correlated with mechanical-ventilation duration and PICU length of stay, with slightly stronger correlations for PRISM-III. These correlations were moderate rather than strong, indicating that mortality risk explains only part of the variation in resource use. Length of stay is also affected by diagnosis, complications, rehabilitation needs, discharge practices, bed availability, and survival bias. Therefore, mortality scores should not be regarded as direct substitutes for models specifically developed to predict ventilation duration or PICU stay. Discrimination and diagnostic accuracy At the selected thresholds, PIM-3 correctly identified 19 of the 23 deaths, whereas PRISM-III identified 21. PRISM-III consequently produced higher sensitivity than PIM-3 (91.3% versus 82.6%), although the 8.7% difference was not statistically significant. PRISM-III also showed numerically better specificity, positive predictive value, negative predictive value, and overall accuracy. The absence of statistical significance for these measures may have been related to the limited number of deaths and the resulting wide confidence intervals. The negative predictive values were high for both models 95.4% for PIM-3 and 97.7% for PRISM-III indicating that patients classified below the selected thresholds were unlikely to die. Positive predictive values were lower, at 57.6% and 65.6%, respectively. This was expected because predictive values depend on outcome prevalence, and mortality affected approximately one-fifth of the cohort. The models were therefore more effective at identifying low-risk patients than confirming that an individual high-risk patient would die. The positive likelihood ratio was higher for PRISM-III than PIM-3 (8.05 versus 5.73), while its negative likelihood ratio was lower (0.10 versus 0.20). Although the differences were not statistically significant, the direction of effect consistently favoured PRISM-III. Chegini et al. (2022)[17] similarly reported useful sensitivity and specificity for PIM-3 and PRISM-4, while emphasizing that optimal cut-offs may vary between study populations. Consequently, the cut-offs of ≥12% for PIM-3 and ≥14 for PRISM-III in the present study should be regarded as locally derived thresholds requiring validation in an independent sample. The significantly greater AUC for PRISM-III (0.914 versus 0.866, P=0.041) showed that it had better ability to rank non-survivors above survivors across all possible thresholds. This finding was consistent with Rahmatinejad et al. (2022)[10], who found superior predictive performance for PRISM-III-based models. However, it differed from Tyagi et al. (2018)[2] and Ekinci et al. (2022)[11], who reported equal or better discrimination for PIM-3. These differences may arise because PRISM-III uses the worst physiological values during the early PICU period and may capture deterioration occurring after admission, whereas PIM-3 uses fewer variables collected around the time of first PICU contact. PRISM-III may therefore achieve better statistical prediction at the cost of greater data-collection burden and potential influence from early treatment. Rusmawatiningtyas et al. (2025)[18] demonstrated that PIM-3 performance among referred critically ill children was affected by referral characteristics and pre-PICU management. This observation is particularly relevant to referral centres, where stabilization before admission can temporarily normalize PIM-3 variables and lead to underestimation of risk. PRISM-III may capture subsequent physiological abnormalities, but part of its predictive advantage could reflect events or treatment occurring after PICU admission. Overall, the present findings indicate that both PIM-3 and PRISM-III were clinically useful mortality-prediction tools. PRISM-III provided significantly better discrimination and overall probability accuracy, while PIM-3 retained practical advantages because it required fewer variables and could be calculated at admission. PRISM-III may be preferred when maximal predictive performance and complete early physiological data are available, whereas PIM-3 may be more suitable for rapid risk adjustment, auditing, and resource-limited environments.

Both PIM-3 and PRISM-III were effective in predicting mortality among children admitted to the PICU. Higher scores were significantly associated with PICU mortality, invasive mechanical ventilation, vasoactive support, multiple-organ dysfunction, prolonged mechanical ventilation, and longer PICU stay. PIM-3 demonstrated good discrimination, whereas PRISM-III demonstrated excellent discrimination and had a significantly higher AUC and lower Brier score. Although PRISM-III showed numerically greater sensitivity, specificity, predictive values, and overall accuracy, most differences in individual diagnostic measures were not statistically significant. Both scores showed acceptable calibration, but PRISM-III predictions were more closely aligned with observed mortality. Therefore, PRISM-III provided better overall mortality prediction in the present cohort, while PIM-3 remained a useful, simpler, admission-based tool. These scores should support risk stratification, clinical auditing, and counselling but should not replace comprehensive clinical judgement or be used alone for individual treatment decisions. Limitations 1. The study was conducted in a single tertiary-care PICU; therefore, its findings might not be generalizable to institutions with different admission policies, patient profiles, resources, and mortality rates. 2. The sample size was relatively small, particularly the number of non-survivors, which resulted in wide confidence intervals for some risk-category odds ratios and diagnostic-accuracy measures. 3. Consecutive sampling rather than probability sampling might have introduced selection bias. 4. Diagnostic subgroups were not sufficiently large to permit reliable comparison of score performance among medical, surgical, cardiac, neurological, oncological, and trauma patients. 5. PIM-3 and PRISM-III use different assessment periods. PIM-3 is calculated using variables available around PICU admission, whereas PRISM-III includes the most abnormal early physiological values. Therefore, their direct comparison might have favoured PRISM-III by allowing it to capture subsequent deterioration. 6. PRISM-III variables might have been influenced by resuscitation, mechanical ventilation, vasoactive treatment, or other interventions initiated during the assessment period. 7. Missing laboratory or physiological measurements could have introduced information bias. Patients with insufficient information to calculate either score were excluded, potentially affecting the representativeness of the cohort. 8. Interobserver agreement in recording and calculating the two scores was not formally evaluated. 9. The optimal cut-offs were derived from the same dataset in which diagnostic performance was assessed. This could have produced optimistic estimates of sensitivity, specificity, and accuracy. 10. Internal validation through bootstrapping or cross-validation and external validation in an independent cohort were not performed. 11. Calibration was assessed partly using the Hosmer-Lemeshow test, whose results may be unstable in small samples and depend on how patients are divided into risk groups. 12. Long-term outcomes, functional status, quality of life, post-discharge mortality, and neurological outcomes were not evaluated. 13. The scores were developed mainly for population-level risk adjustment and should not be interpreted as definitive predictors of an individual child’s outcome.

1. Pollack MM, Holubkov R, Funai T, Dean JM, Berger JT, Wessel DL, et al. The Pediatric Risk of Mortality score: update 2015. Pediatr Crit Care Med. 2016;17(1):2-9. PubMed 2. Tyagi P, Tullu MS, Agrawal M. Comparison of Pediatric Risk of Mortality III, Pediatric Index of Mortality 2, and Pediatric Index of Mortality 3 in predicting mortality in a pediatric intensive care unit. J Pediatr Intensive Care. 2018;7(4):201-206. PubMed 3. Lee OJ, Jung M, Kim M, Yang HK, Cho J. Validation of the Pediatric Index of Mortality 3 in a single pediatric intensive care unit in Korea. J Korean Med Sci. 2017;32(2):365-370. Article 4. Jung JH, Sol IS, Kim MJ, Kim YH, Kim KW, Sohn MH. Validation of Pediatric Index of Mortality 3 for predicting mortality among patients admitted to a pediatric intensive care unit. Acute Crit Care. 2018;33(3):170-177. Article 5. Malhotra D, Nour N, El Halik M, Zidan M. Performance and analysis of Pediatric Index of Mortality 3 score in a pediatric ICU in Latifa Hospital, Dubai, UAE. Dubai Med J. 2020;3(1):19-25. Article 6. Mirza S, Malik L, Ahmed J, Malik F, Sadiq H, Ali S, et al. Accuracy of Pediatric Risk of Mortality III score in predicting mortality outcomes in a pediatric intensive care unit in Karachi. Cureus. 2020;12(3):e7489. doi:10.7759/cureus.7489. PubMed 7. Kaur A, Kaur G, Dhir SK, Rai S, Sood A. Pediatric Risk of Mortality III score predictor of mortality and hospital stay in pediatric intensive care unit. J Emerg Trauma Shock. 2020;13(2):146-150. PubMed 8. Nasser MM, Al-Sawah AY, Hablas WR, Mansour AM. Reliability of Pediatric Risk of Mortality III and Pediatric Index of Mortality 3 scores in the pediatric intensive care unit of El-Hussein University Hospital. Al-Azhar J Pediatr. 2020;23(3):1048-1071. doi:10.21608/azjp.2020.127067. Article 9. Shen Y, Jiang J. Meta-analysis for the prediction of mortality rates in a pediatric intensive care unit using different scores: PRISM-III/IV, PIM-3, and PELOD-2. Front Pediatr. 2021;9:712276. doi:10.3389/fped.2021.712276. PubMed 10. Rahmatinejad Z, Rahmatinejad F, Sezavar M, Tohidinezhad F, Abu-Hanna A, Eslami S. Internal validation and evaluation of the predictive performance of models based on the PRISM-3 and PIM-3 scoring systems for predicting mortality in pediatric intensive care units. BMC Pediatr. 2022;22(1):199. doi:10.1186/s12887-022-03228-y. Article 11. Ekinci F, Yildizdas D, Horoz OO, Arslan I, Ozkale Y, Yontem A, et al. Performance and analysis of four pediatric mortality prediction scores: a multicenter prospective observational study in four PICUs. Arch Pediatr. 2022;29(6):437-443. doi:10.1016/j.arcped.2022.05.001. PubMed 12. Zhang L, Wu Y, Huang H, Liu C, Cheng Y, Xu L, et al. Performance of PRISM III, PELOD-2, and P-MODS scores in two pediatric intensive care units in China. Front Pediatr. 2021;9:626165. doi:10.3389/fped.2021.626165. Article 13. Toteja N, Choudhary B, Khera D, Sasidharan R, Sharma PP, Singh K. Performance of Pediatric Index of Mortality PIM-3 in a tertiary care PICU in India. J Pediatr Intensive Care. 2024;13(3):235-241. Article 14. Genu DHS, Yamaçake KGR, Moreira GA, Cogo PE, Costa CAD, de Carvalho WB, et al. Multicenter validation of PIM3 and PIM2 in Brazilian pediatric intensive care units. Front Pediatr. 2023;10:1036007. doi:10.3389/fped.2022.1036007. Article 15. Jyotsna, Kumar R, Sharan S, Kishore S, Prakash J. The various scoring systems in pediatric intensive care units: a prospective observational study. Cureus. 2023;15(5):e39679. doi:10.7759/cureus.39679. Article 16. Agrwal S, Saxena R, Jha M, Jhamb U, Pallavi. Comparison of pSOFA with PRISM III and PIM 2 as predictors of outcome in a tertiary care pediatric ICU: a prospective cross-sectional study. Indian J Crit Care Med. 2024;28(8):796-801. Article 17. Chegini V, Ghasemi D, Malekafzali B, Ghasemi A. Evaluating the ability of PRISM4 and PIM3 to predict mortality in patients admitted to pediatric intensive care units. Int J Prev Med. 2022;13:101. Article 18. Rusmawatiningtyas D, Nurnaningsih, Makrufardi F, Arguni E. Can PIM3 predict mortality adequately for patients admitted to pediatric intensive care units after referral acceptance? BMC Pediatr. 2025;25:460. doi:10.1186/s12887-025-06114-5. Article.