Performance of PIM, PRISM, and PELOD scores in a resource-limited North Indian paediatric intensive care unit

Abstract

BACKGROUND

Pediatric mortality prediction models are widely used in high-income nations but remain under-validated in resource-limited settings with different disease profiles and care resources. This study evaluated the performance of pediatric risk of mortality score III (PRISM-III), pediatric index of mortality-3 (PIM-3), and pediatric logistic organ dysfunction score-2 (PELOD-2) in predicting mortality in a resource limited paediatric intensive care unit.

AIM

To determine the discriminative ability and calibration of PRISM III, PIM-3, and PELOD-2 for predicting 30-day mortality.

METHODS

This was an observational study over 15 months and included 175 children aged 1 month to 18 years. Physiological variables were recorded within the first hour (PIM-3) and the first 24 hours (PRISM III, PELOD-2). Performance was measured using the area under the receiver operating characteristic curve for discrimination and the Hosmer-Lemeshow goodness-of-fit test for calibration.

RESULTS

The observed mortality rate in the cohort was 13.7% (24 deaths). Mortality was significantly associated with sepsis and the early requirement for mechanical ventilation or inotropes. All scores showed high negative predictive values (> 95%) and excellent calibration (Hosmer-Lemeshow P values: PRISM-III = 0.952, PIM-3 = 0.692, PELOD-2 = 0.518), with standardized mortality ratio of 1.04, indicating accurate overall mortality prediction. PIM-3 achieved the highest discrimination (area under the receiver operating characteristic 0.951), followed closely by PELOD-2 (0.939), while PRISM III performed lowest (0.884). PELOD-2 offered the highest sensitivity (87.5%).

CONCLUSION

All three scores fit the observed risk distribution well, demonstrating clinically acceptable discrimination and calibration for predicting mortality in the local patient population. However, the small number of mortality events limits the precision of these estimates and precludes claims of broader applicability.

Key Words: Mortality prediction; Severity of illness score; Discrimination; Calibration; Risk stratification; Outcome prediction; Critical illness; Vasoactive; Inotropic support

Core Tip: Our findings confirm that paediatric intensive care unit mortality scores (pediatric index of mortality-3, pediatric risk of mortality score III, pediatric logistic organ dysfunction score-2) demonstrate acceptable discrimination and calibration for our low resource setting paediatric intensive care unit population (all Hosmer-Lemeshow P > 0.05; standardized mortality ratios 1.00-1.04). Pediatric index of mortality-3 achieved the highest discriminative ability (area under the receiver operating characteristic curve 0.951), outperforming pediatric logistic organ dysfunction score-2 (0.939) and pediatric risk of mortality score III (0.884), and coupled with a negative predictive value > 95%, represents the most practical choice for early risk stratification and efficient resource allocation within our center.

INTRODUCTION

Mortality in the paediatric intensive care unit (PICU) is driven primarily by severity of illness and patient-specific characteristics, but important additional determinants such as comorbidities, nutritional status, socioeconomic constraints, and delays in accessing care also influence outcomes[1]. Over the past three decades, mortality prediction scores have been developed to quantify physiologic derangement converting routinely available clinical and laboratory data into composite severity scores that estimate in-hospital death risk[2,3]. These tools also enable comparison between observed and predicted deaths, often summarized through the standardized mortality ratio (SMR). Their usefulness is based on the premise that physiological status at admission provides an objective estimate of baseline risk, enabling comparisons of performance across units when the models are accurate and well calibrated.

Among the most widely used PICU models are pediatric risk of mortality score (PRISM)[4], pediatric index of mortality (PIM)[5], and pediatric logistic organ dysfunction score (PELOD)[6]. PRISM evolved from earlier, complex indices into a more practical physiologic scoring system, while PIM was designed for rapid calculation using variables collected within the first hour of admission. PELOD was primarily descriptive of multi organ dysfunction over time, complements prognostic scores by quantifying organ failure trajectories that are strongly associated with mortality. However, all these models were validated largely in high-income patient settings and may not perform equivalently in resource-limited regions. In the hilly state of Himachal Pradesh, India, the PICU manages different patient profiles (higher infection burden, seasonal/tropical illnesses, frequent malnutrition, variable pre-hospital stabilization, delayed referrals, and inconsistent access to advanced technologies). These contextual factors may cause systematic under or overestimation of risk when international scores are applied without local validation or recalibration.

Therefore, this study was planned to evaluate the performance of established mortality prediction scores (PRISM-III, PIM-3, and PELOD-2) in the PICU of a resource-limited tertiary-care hospital by validating their discrimination and calibration in local patients, examining associations between score severity and outcomes (at discharge and 30-day mortality) to determine whether adjustment for local risk factors is required to strengthen clinical decision-making and future research planning.

MATERIALS AND METHODS

This was an observational study conducted in a 24-bedded PICU (8 beds in the intensive area and 16 beds in the step-down area) over a 15-month period (June 2024 to August 2025). The unit admits both medical and postsurgical patients requiring intensive care and is staffed by full-time paediatricians and staff nurses, with 3-4 nurses per 8-hour shift. The unit is equipped for continuous monitoring, mechanical ventilation, blood gas analysis, ultrasonography, and X-ray services. The study was initiated after approval from the Institutional Ethics Committee.

Consecutive sampling was used and admissions with a minimum PICU stay > 4 hours were enrolled after written informed consent was obtained from parents/guardians (and assent from older children where applicable). Confidentiality was maintained using anonymized study identifiers. Data was collected by the trained investigator; and cross-checked by both co-investigators, discrepancies were discussed and resolved. Those with insufficient data due to any reason, including non-functioning laboratory services or unrecorded parameters were excluded. Age, sex, admission diagnosis, and all physiological variables required to calculate PRISM-III, PIM-3, and PELOD-2 were documented. For PIM-3, data collection was strictly limited to the first hour following PICU admission, with no subsequent values incorporated into the final score. PRISM-III and PELOD-2 utilized the worst clinical values observed over the initial 24-hour period. The primary outcome of this study was all-cause mortality (survival vs death) at 30 days following hospital admission, along with the length of PICU stay and total hospital stay as secondary outcomes.

Statistical analysis

Data was summarized using mean ± SD or median [interquartile range (IQR)] for continuous variables and n (%) for categorical variables. Discrimination of PRISM-III, PIM-3, and PELOD-2 for mortality was assessed using the area under the receiver operating characteristic curve (AUROC) with 95% confidence intervals (CIs)[7]. Calibration was evaluated using calibration plots and the Hosmer-Lemeshow goodness-of-fit test[8]. All 3 models performance was also examined across age subgroups. Statistical analysis and visualization were conducted using R version 4.2.2 (R Foundation for Statistical Computing, Vienna, Austria).

RESULTS

Total 151 children survived out of 175 PICU admissions enrolled. Ninety-six were males (54.9%), among them 85 survived; there was no significant association between gender and mortality (P = 0.339). The demographic details are shown in Table 1. Patients were stratified into four age groups (1-12, 13-60, 61-120, and > 120 months), with the highest proportion of admissions in the 1-12-month group (n = 56, 32%). Survival proportions were comparable across age groups (P = 0.686).

Table 1 Demographic and clinical characteristics of patients, median (interquartile range)/n (%).

Characteristics	Total, n = 175	Survivors, n = 151	Non-survivors, n = 24	P value
Age in months	48 (12-126)	42 (12-132)	54 (4.7-102)	0.845
Age subgroups
1-12 months	56	49 (87.5)	7 (12.5)	0.686
13-60 months	46	40 (87)	6 (13)
61-120 months	29	23 (79.3)	6 (20.7)
> 120 months	44	39 (88.6)	5 (11.4)
Gender
Male	96	85 (88.5)	11 (11.5)	0.339
Female	79	66 (83.5)	13 (16.5)
Primary diagnosis
Respiratory disease	41	41 (100.0)	0	0.372
Renal disease	2	2 (100.0)	0
Sepsis	30	15 (50.0)	15 (50.0)
Neurological disease	43	41 (95.2)	2 (4.7)
Cardiac disease	10	10 (100.0)	0
Postoperative	1	1 (100.0)	0
Oncology	8	6 (75.0)	2 (25.0)
Trauma	2	1 (50.0)	1 (50.0)
Endocrine	4	0	4 (100.0)
Hepatic disease	6	4 (66.7)	2 (33.3)
Poisoning	9	9 (100.0)	0
Metabolic disease	1	0	1 (100.0)
Gastrointestinal disease	18	17 (94.4)	1 (5.6)
Mechanical ventilation in first 24 hours	25	9 (36)	16 (64)	< 0.001
Inotropes in first 24 hours	26	14 (53.8)	12 (46.2)	< 0.001
Duration of PICU stay (days)	3 (1-5)	3 (1-4)	4 (2-7)	0.031
Duration of total hospital stay (days)	5 (4-9)	6 (4-9)	4 (2-7)	0.011
Referred out	13	7 (53.8)	6 (46.2)	0.231

PICU: Paediatric intensive care unit.

Neurological (n = 43) and respiratory (n = 41) conditions were the most frequent admission diagnosis, followed by sepsis (n = 30) and gastrointestinal illnesses (n = 18). Overall outcome varied by diagnosis: Survival was 100% in respiratory, renal, cardiac, postoperative, and intoxication admissions, whereas mortality was higher in endocrine and metabolic diseases. Sepsis showed substantial risk with 50% mortality (15/30 deaths), intermediate mortality was observed in hepatic disease (2/6, 33.3%) and oncology (2/8, 25%), while neurological (2/43, 4.7%) and gastrointestinal diagnosis (1/18, 5.6%) were associated with low mortality.

The median PICU stay for the cohort was 3 days (IQR 1-5); non-survivors had a significantly longer median PICU stay than survivors [4 (IQR 2-7) days vs 3 (IQR 1-4) days; P = 0.031]. Total hospital stay also differed significantly, with survivors staying longer than non-survivors [6 (4-9) days vs 4 (2-7) days; P = 0.011]. At admission, 26 (14.9%) needed inotropic support and 25 (14.3%) needed mechanical ventilation. Patients on ventilators (16/25) and those on inotropes (12/26) had considerably higher mortality (P < 0.001; Table 1). Across the cohort, the mean and median scores were highest for PRISM-III followed by PIM-3 and PELOD-2 (Table 2). Across all scoring systems, non-survivors had significantly higher scores than survivors (all P < 0.001) indicating children who died had more severe disease.

Table 2 Comparison of pediatric risk of mortality score III, pediatric index of mortality-3, and pediatric logistic organ dysfunction score-2 mortality scores between survivors and non-survivors, mean ± SD/median (interquartile range).

	Survivors, n = 151	Non survivors, n = 24	Total, n = 175	P value	AUROC (95%CI)	Observed death (O)	Expected death (E)	SMR (O/E) 95%CI	Calibration HL-GOF
PRISM-III	4.35 ± 3.92	13.04 ± 7.68	5.54 ± 5.55	< 0.001	0.884 (0.827-0.927)	24	23	1.04 (0.63-0.46)	χ2 = 2.69, P = 0.952
	3 (2-6)	10.5 (6-18)	4 (2-7)
PIM-3	2.52 ± 1.75	18.79 ± 28.31	4.75 ± 1.84	< 0.001	0.951 (0.907-0.978)	24	24	1.0 (0.64-1.47)	χ2 = 5.592, P = 0.692
	2.1 (1.6-2.8)	8.6 (4-13.5)	2.3 (1.7-3.5)
PELOD-2	1.74 ± 2.08	7.08 ± 4.20	2.47 ± 3.07	< 0.001	0.939 (0.893-0.970)	24	23	1.04 (0.63-1.46)	χ2 = 7.36, P = 0.518
	2 (0-2)	7 (4-9)	2 (0.6-3)

HL-GOF: Hosmer-Lemeshow goodness of fit; AUROC: Area under the receiver operating characteristic curve; CI: Confidence interval; SMR: Standardized mortality ratio; PELOD-2: Pediatric logistic organ dysfunction score-2; PRISM-III: Pediatric risk of mortality score III; PIM-3: Pediatric index of mortality-3.

Discrimination

Discrimination was assessed using the AUROC with 95%CIs, which summarizes each model’s ability to distinguish survivors from non-survivors across the full range of predicted risks. An AUROC > 0.70 was considered acceptable, > 0.80 good, and > 0.90 excellent[7]. In this cohort, excellent discrimination was observed for two of the three scores: PIM-3 showed the highest discrimination (AUROC 0.951; 95%CI: 0.907-0.978), followed closely by PELOD-2 (AUROC 0.939; 95%CI: 0.893-0.970) and PRISM-III (AUROC 0.884; 95%CI: 0.827-0.927) in this population as illustrated in Figure 1 and detailed in Table 2.

Figure 1 Receiver operating characteristic curve. A: Pediatric logistic organ dysfunction score-2; B: Pediatric index of mortality-3; C: Pediatric risk of mortality score III. PELOD-2: Pediatric logistic organ dysfunction score-2; PRISM-III: Pediatric risk of mortality score III; PIM-3: Pediatric index of mortality-3; AUC: Area under the curve.

Calibration

Calibration was assessed using the Hosmer-Lemeshow goodness-of-fit test, which evaluates agreement between predicted and observed mortality across strata of predicted risk[8]. The χ² statistic ∑(O-E)2/E was used to compare the observed (O) and expected (E) deaths for each of the ten equal-sized groups (deciles) of patients ranked by predicted mortality. A corresponding P value was then calculated. Higher χ² values indicate greater mismatch between observed and predicted risk, while a P value > 0.05 was considered evidence of good calibration. In this study, Hosmer-Lemeshow goodness-of-fit results demonstrated no statistically significant lack of fit for all three scores, with a high P value across all models suggesting predicted risks closely matched observed mortality (Table 2).

SMR

The SMR compares observed deaths (O) with expected deaths (E) predicted by scoring system over a defined period (SMR = O/E) and is commonly used to evaluate PICU performance and the overall accuracy of mortality prediction models[9]. An SMR > 1 suggests underestimation of mortality implying higher-than-predicted mortality, an SMR < 1 indicates over prediction implying lower-than-predicted mortality, and an SMR of 1 reflects perfect agreement. In this cohort, the observed mortality closely matched predictions and yielded an SMR of 1.04 (95%CI: 0.63-1.46) for each scoring system, indicating only a slight, statistically non-significant under prediction, with the CI including 1.0 and therefore demonstrating good agreement at the population level (Table 2).

Performance characteristics across various age groups

In infants aged 1-12 months, all three models showed broadly similar central tendency and overall strong performance (Table 3): Median (IQR) score for PRISM-III was 3 (1.5-6), PIM-3 was 2.6 (1.9-3.1), and PELOD-2 was 2 (2-3). Discrimination and calibration was good across all scores, indicating close agreement between predicted and observed mortality within all age groups. In children aged 13-60 months, all three scores showed good-to-excellent calibration and discrimination, with PIM-3 and PELOD-2 demonstrating higher AUROC than PRISM-III (Table 3). There was overall agreement between predicted and observed mortality in this age group with SMR (1.0 for PIM-3/PELOD-2 and 1.2 for PRISM-III).

Table 3 Age-stratified predictive performance of pediatric risk of mortality score III, pediatric index of mortality-3, and pediatric logistic organ dysfunction score-2 mortality scores.

	PRISM-III	PIM-3	PELOD-2
Age group 1-12 months
mean ± SD	4.52 ± 5.48	4.85 ± 11.70	2.64 ± 2.75
median (interquartile range)	3 (1.5-6)	2.6 (1.9-3.1)	2 (2-3)
AUROC (95%CI)	0.856 (0.656-1.000)	0.845 (0.550-1.000)	0.883 (0.722-0.990)
Observed death (O)	7	7	7
Expected death (E)	7	6	7
SMR (O/E) 95%CI	1	1.2	1
HL χ2	1.55	7.07	1.72
P value	0.907	0.53	0.423
Age group 13-60 months
mean ± SD	5.61 ± 5.61	5.19 ± 13.74	2.26 ± 3.30
median (interquartile range)	4 (2.2-6.8)	2.3 (1.8-3.7)	1.5 (0-2)
AUROC (95%CI)	0.872 (0.701-1.000)	0.976 (0.925-1.000)	0.982 (0.938-1.000)
Observed death (O)	6	6	6
Expected death (E)	5	6	6
SMR (O/E) 95%CI	1.2	1	1
HL χ2	5.84	2.58	6.14
P value	0.441	0.958	0.105
Age group 61-120 months
mean ± SD	6.97 ± 6.20	2.87 ± 2.57	3.14 ± 3.69
median (interquartile range)	5 (2-10)	2.1 (1.4-3.7)	2 (0-6)
AUROC (95%CI)	0.812 (0.635-0.953)	0.884 (0.718-1.000)	0.783 (0.530-0.971)
Observed death (O)	6	6	6
Expected death (E)	6	5	6
SMR (O/E) 95%CI	1	1.2	1
HL χ2	7.65	12.06	4.17
P value	0.468	0.149	0.383
Age group > 120 months
mean ± SD	5.93 ± 5.06	5.41 ± 13.68	2.05 ± 2.77
median (interquartile range)	5 (2-7)	2.05 (1.6-3.55)	1 (0-3)
AUROC (95%CI)	0.849 (0.640-1.000)	0.764 (0.333-1.000)	0.879 (0.660-1.000)
Observed death (O)	5	5	5
Expected death (E)	5	4	5
SMR (O/E) 95%CI	1	1.3	1
HL χ2	10.06	8.99	4.02
P value	0.185	0.343	0.546

PELOD-2: Pediatric logistic organ dysfunction score-2; PRISM-III: Pediatric risk of mortality score III; PIM-3: Pediatric index of mortality-3; HL: Hosmer-Lemeshow; AUROC: Area under the receiver operating characteristic curve; CI: Confidence interval; SMR: Standardized mortality ratio.

In older children (61-120 months), all these scores demonstrated excellent calibration with good discrimination overall, with AUROC ranging from 0.783 (PELOD-2) to 0.884 (PIM-3), while PRISM-III (0.812) also performed acceptably (Table 3). Observed deaths and expected deaths were similar, resulting in SMR (1.0 for PRISM-III/PELOD-2 and 1.2 for PIM-3), indicating close agreement between predicted and observed mortality in this age group. In adolescents (> 120 months), no statistically significant lack of fit was observed by Hosmer-Lemeshow for all three.

Diagnostic performance of mortality prediction scores

Overall, all scores showed good diagnostic performance for mortality prediction at the chosen cut-offs, with high negative predictive values (NPV, 95.4%-97.6%) indicating strong ability to rule out death among low-risk patients (Table 4). PIM-3 (> 3.6) had balanced performance across all metrics. Its positive predictive value (PPV, 45.2%) and NPV (96.2%) were favourable, and the diagnostic accuracy reached 84.0%, making it a reliable predictor overall. PELOD-2 (> 3.8) had the highest sensitivity (87.5%) among all scores, along with strong specificity (80.8%). Although the PPV was relatively low (42.0%), its NPV was excellent (97.6%), indicating that patients with scores below the threshold were very unlikely to die. PRISM-III (> 6) showed comparatively lower diagnostic accuracy (77.1%), a notably low PPV (34%), however, it retained a high NPV (94.4%), suggesting its utility in identifying low-risk patients (Table 4).

Table 4 Diagnostic accuracy of pediatric risk of mortality score III, pediatric index of mortality-3, and pediatric logistic organ dysfunction score-2 mortality scores.

		Mortality		Total, 175	Sensitivity, %	Specificity, %	PPV, %	NPV, %	Diagnostic accuracy
		Yes, 24	No, 151
PRISM-III	> 6	17	33	50	70.8	78.1	34	95.4	77.1
	≤ 6	7	118	125
PIM-3	> 3.6	19	23	42	79.2	84.8	45.2	96.2	84
	≤ 3.6	5	128	133
PELOD-2	> 3.8	21	29	50	87.5	80.8	42	97.6	81.7
	≤ 3.8	3	122	125

PELOD-2: Pediatric logistic organ dysfunction score-2; PRISM-III: Pediatric risk of mortality score III; PIM-3: Pediatric index of mortality-3; PPV: Positive predictive value; NPV: Negative predictive value.

DISCUSSION

Reliable mortality prediction tools are essential in the PICU to support risk stratification, research comparability, and prognostication alongside clinical judgment. Although PRISM, PIM, and PELOD (and their updated versions) are widely used, their performance may differ when applied outside the high-resource settings, in which they were developed, hence the need for local validation. Therefore, this prospective study evaluated and compared the discrimination and calibration of PRISM-III, PIM-3, and PELOD-2 in a resource limited PICU.

In our cohort of 175 admissions, the observed mortality was 13.7% (24 deaths), which was higher than rates reported in earlier Indian studies (e.g., 5.1% by Raghavendra Bhupal et al[10]. and 6.3% reported by Srinivas et al[11]), possibly reflecting differences in illness severity and baseline morbidity, and other health-system factors. Mortality was not significantly associated with age or gender in our study, consistent with prior reports[12]. In contrast, certain diagnosis (notably sepsis and endocrine disorders)[13] and the need for early intensive therapies including mechanical ventilation[14] and inotropic support[15] were strongly associated with higher mortality (P < 0.001), underscoring their role as markers of severe physiological derangement.

Validating prognostic models requires assessment of both discrimination and calibration. Discrimination (i.e., the ability to distinguish survivors from non-survivors) was evaluated for each scoring system using AUROC[7]. In our cohort, discrimination (AUROC) was excellent for PIM-3 (0.951; 95%CI: 0.907-0.978), and PELOD-2 (0.939; 95%CI: 0.893-0.970), while PRISM-III also demonstrated good-to-excellent performance (0.884; 95%CI: 0.827-0.927). These AUROC values compare favourably with reports from other settings; for example, Ramazani and Hosseini[16] reported lower AUROC values for PIM-3 (0.824), PELOD-2 (0.803), and PRISM-III (0.773) in a mixed ICU population. Srinivas et al[11] reported good discrimination for both PRISM-III (AUROC = 0.888) and PIM-3 (AUROC = 0.934) in South India, with PIM-3 outperforming PRISM-III, similar to our cohort. Similarly, Uzunay Gündoğan et al[17] from Turkey found good discrimination for both PRISM-III and PIM-3. Together, these findings illustrate that mortality scores performance varies across settings, while underscoring the strong discriminatory ability of PRISM-III, PIM-3, and PELOD-2 in our population. The overlapping CIs among these models suggest broadly comparable ability to rank patients by mortality risk.

Calibration was assessed using the Hosmer-Lemeshow goodness-of-fit test, which evaluates agreement between predicted and observed outcomes across risk strata[8]. Our analysis showed calibration for all three scoring systems (PRISM-III, PIM-3, and PELOD-2), supported by uniformly non-significant Hosmer-Lemeshow P values ranging from 0.518 to 0.952, with minimal deviation between observed and expected deaths across risk strata. This is noteworthy because multiple studies from India[2,10,11], Türkiye[17], Egypt[18], Pakistan[19], Hong Kong[20], Portugal[21], and selected high-risk cohorts[22,23] have reported significant goodness-of-fit test (P < 0.05), implying lack of fit and the need for recalibration. In contrast, the consistently strong calibrations in our cohort suggested that all three mortality scores provided unbiased risk estimates and accurately reflect the mortality risk gradient in our PICU population, supporting their applicability in our setting.

We utilized the SMR to evaluate aggregate mortality prediction accuracy in our PICU[9]. In our cohort, SMR was close to the ideal value of 1.0 (overall 1.04); indicating strong agreement between predicted and observed mortality (23 expected vs 24 observed). This contrasts with reports from other settings showing under prediction (SMR > 1.0) or over prediction (SMR < 1.0), necessitating local recalibration[2,19,24]. Nevertheless, these scores were calculated from information around the time of PICU admission (a “snapshot”), so they estimate baseline risk but do not update as the child’s condition improves or deteriorates during the PICU stay. Change in patient condition may be better reflected by serial assessments over time (sequential organ dysfunction monitoring tools) rather than a single admission score.

We also evaluated their performance across four paediatric age strata; infants (1-12 months), young children (13-60 months), older children (61-120 months), and adolescents (> 120 months)[12]. A consistent finding from these subgroup analyses for all scores (PRISM-III, PIM-3, and PELOD-2) demonstrated calibration across each age group. This was supported by uniformly non-significant Hosmer-Lemeshow P values (range: 0.105-0.958) for all three mortality scores within each age category. This is particularly relevant because reduced calibration across age ranges has been reported in some settings; in contrast, the consistently no statistically significant lack of fit was observed across the paediatric age spectrum in our cohort suggests minimal age-related systematic bias in risk prediction using these standard models.

Using ROC analysis and the Youden index to improve clinical applicability, we identified locally exploratory cut-offs of > 6 for PRISM-III, > 3.6 for PIM-3, and > 3.8 for PELOD-2. At these thresholds, PELOD-2 showed the highest sensitivity (87.5%), making it useful for screening when maximizing detection of high-risk patients is prioritized. PIM-3 demonstrated a balanced profile (sensitivity 79.2%, specificity 84.8%) supporting its use for general risk assessment. All scores generated high NPV (> 95%), underscoring their utility in reliably identifying patients at low risk of death. In contrast, PPVs were modest (34%-45%), indicating that scores above the cut-off reflect increased risk but should be interpreted alongside the overall clinical context rather than only determinant of bad outcome. A fundamental prerequisite for any severity scoring system is that higher scores reflect greater underlying illness severity; our study confirmed this, as PRISM-III, PIM-3, and PELOD-2 were all significantly higher in non-survivors than survivors (P < 0.001 for all).

Key strengths of our study include its prospective design, the simultaneous comparison of three widely used prognostic scores within the same cohort, and a rigorous statistical evaluation of both discrimination and calibration (AUROC, Hosmer-Lemeshow, and SMR). However, it also had certain limitations. First, as a single-center study, the generalizability of our results may be restricted. Second, all these scores relied primarily on admission data; and did not capture dynamic clinical trajectories or non-mortality outcomes such as long-term morbidity and resource utilization. Importantly, the high NPV observed in this study are expected and are driven in part by low baseline mortality in cohort rather than model discrimination alone. While the predictive metrics for the overall cohort are robust, the age-stratified subgroup analyses are statistically underpowered due to a low number of mortality events (5-7 deaths per group). Although we report these subgroup metrics to maintain full transparency and align with prior literature, their wide CIs limit their stability. Therefore, they preclude definitive conclusions and necessitate cautious interpretation.

CONCLUSION

Although our results support the utility of these scores in our centre, extrapolating data (e.g., AUROC and cut-offs) to other PICU should be done cautiously. Future work should prioritise multi-centre validation in diverse settings to confirm generalizability and identify context-specific differences, and should build larger collaborative datasets to enable robust evaluation in high-risk subgroups (e.g., sepsis, congenital anomalies, therapy-specific cohorts). In addition, further research is needed to clarify how sequential dynamic scoring system can best complement admission-based scores to improve prognostic accuracy throughout the PICU stay.

ACKNOWLEDGEMENTS

We sincerely thank the paediatric intensive care unit nurses and residents, whose diligent and timely recording of clinical data was essential to the success of this study. We are also deeply grateful to the participating children and their families for their invaluable contribution to advancing pediatric critical care. We gratefully acknowledge Dr. Bhagat Ram Thakur for his early efforts during development of this research project.