This cross-sectional diagnostic study was conducted at the PICU of Can Tho Children's Hospital, a tertiary referral center in the Mekong Delta of Vietnam, from May 2024 to May 2025.

A convenience sample of children aged 2-months to 15-years admitted to the PICU with a clinical diagnosis of sepsis was enrolled. Sepsis was defined according to the 2020 Surviving Sepsis Campaign guidelines, requiring suspected infection accompanied by cardiovascular dysfunction.8 Clinically, infection was indicated by fever (temperature > 38.5°C) or hypothermia (< 36°C), age-inappropriate tachycardia or tachypnea, or an identifiable infectious focus, along with signs of a systemic inflammatory response such as leukocytosis, leukopenia, or an increased proportion of band forms. Cardiovascular dysfunction was defined by the presence of at least one of the following criteria: persistent hypotension unresponsive to initial fluid resuscitation (≥ 40‒60 mL/kg in the first hour), the need for vasoactive support to maintain age-appropriate mean arterial pressure, or the presence of two or more clinical signs of inadequate perfusion (e.g., prolonged capillary refill, weak or absent peripheral pulses, a central-to-peripheral temperature gap > 3°C, elevated lactate, urine output < 0.5 mL/kg/h, or altered mental status) despite fluid and vasoactive therapies.

Patients were excluded if they had pre-existing chronic comorbidities, including viral-induced acquired immunodeficiency, thalassemia, nephrotic syndrome, cancer, complex congenital heart disease, or other immunodeficiencies. Further exclusion criteria included death before complete data could be captured, incomplete medical charts that prevented the calculation of PSS, pSOFA, or PRISM-III scores, or inter-hospital transfers with missing external medical records.

This study was a single-gate, cross-sectional investigation designed to assess diagnostic test accuracy. A convenience sampling method was employed. The sample size was determined using Buderer’s formula for specificity, targeting a 95 % confidence level (Zα/2​ = 1.96) with an allowable error margin (w) of 4 %.9 Based on an expected prevalence (p) of septic shock of 0.17, which was proxied from mortality rates in similar pediatric populations, and an anticipated high Specificity (Sp) of 0.97 for modern prognostic scores, the minimum required sample size was calculated to be 84 participants.10,11,12 A total of 86 consecutively eligible patients were ultimately enrolled.

Upon PICU admission, demographic data, clinical features, and laboratory results were collected. The primary infectious focus was determined from clinical, laboratory, and imaging findings. Septic shock classification was performed independently and single ‒ blinded to score calculations to minimize assessment bias. Prognostic scores were calculated as follows: The PSS was used to identify sepsis (score ≥2) and septic shock by evaluating four organ systems.1 The pSOFA score assessed six organ systems, utilizing the SpO2​/FiO2​ ratio when arterial blood gas data were unavailable.2 The PRISM III score was calculated using the most severe values of 17 physiological variables recorded within the first 24 hours of PICU admission to estimate mortality risk.5 For temporal alignment, PSS and pSOFA scores were determined using the earliest available data within the first six hours, while PRISM-III utilized the worst values over the first 24 hours.

Nutritional status was assessed using World Health Organization (WHO) z-scores: weight-for-length/height for children aged 0–59 months and BMI-for-age for those aged 5–15 years. For analysis, patients were categorized into a binary variable of either “Normal” or “Malnutrition/Overweight-Obesity” based on these z-scores.13,14 Tachypnea was defined using age-specific respiratory rate thresholds from WHO-Integrated Management of Childhood Illness (IMCI) guidelines for children under five years and American Heart Association Pediatric Advanced Life Support (PALS) guidelines for older children.15,16 The primary infectious focus was determined through a comprehensive review of clinical, laboratory, imaging, and microbiological data and was categorized accordingly. All data adjudication was independently performed by two investigators, with any discrepancies resolved through consensus.

All data were cleaned, coded, and subsequently analyzed using Statistical Package for the Social Sciences (SPSS) version 22.0 (IBM Corp., Armonk, NY) and MedCalc version 22.017 (MedCalc Software Ltd., Ostend, Belgium).

Categorical variables were summarized as frequencies and percentages, with group comparisons conducted using the Chi-Squared (χ2) test or Fisher’s exact test where appropriate. The distribution of continuous variables was assessed for normality using the Shapiro-Wilk test, supplemented by visual inspection of histograms and Q-Q plots. Normally distributed data are presented as mean ± Standard Deviation (SD) and were compared using two-sided independent t-tests. Data that were not normally distributed are reported as median and Interquartile Range (IQR) and were compared using the Mann-Whitney U test.

To identify independent predictors, variables with a p-value less than 0.10 in univariable analysis, along with clinically prespecified variables, were entered into a multivariable logistic regression model. Results are presented as Odds Ratios (OR) with their corresponding 95 % Confidence Intervals (95 % CI). The discriminative ability of the models was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUROC), with 95 % CIs. AUROCs were compared using the DeLong test. The optimal cut-off value for each model was determined using Youden’s index (J = sensitivity + specificity − 1). At these thresholds, we calculated sensitivity, specificity, Positive Predictive Value (PPV), Negative Predictive Value (NPV), and positive and negative likelihood ratios (LR+ and LR-).

Model calibration was assessed via the Hosmer-Lemeshow goodness-of-fit test, calibration-in-the-large (intercept), and the calibration slope, which are visually represented in Fig. 1. A two-sided p-value less than 0.05 was considered statistically significant for all analyses.

Fig. 1.

Calibration plot comparing predicted vs observed shock risk by deciles. Calibration plot comparing predicted versus observed septic shock risk by deciles for the exploratory multivariable model. Given the limited sample size, this calibration assessment is descriptive and should be interpreted cautiously.

Given the limited number of septic shock events, multivariable logistic regression was performed for exploratory purposes only. Model estimates were interpreted cautiously due to the potential risk of overfitting and limited events-per-variable.

The study protocol was approved by the Institutional Review Board of Can Tho University of Medicine and Pharmacy (Approval n° 24.045.SV/PCT-HĐĐĐ, dated May 24, 2024). Informed consent was obtained from the parents or legal guardians of all participants.

During the study period, a total of 86 children with sepsis were enrolled, of whom 23 (26.7 %) developed septic shock (Fig. 2).

Fig. 2.

Classification of pediatric sepsis.

The baseline characteristics of the cohort are presented in Table 1. Children older than 12-months had significantly higher odds of developing septic shock compared to those aged 2 to < 12-months (78.3 % vs. 21.7 %; OR = 3.07, 95 % CI: 1.01‒9.30, p = 0.041). No significant associations were found between septic shock and sex or nutritional status.

The distribution of primary infection foci differed between groups. Respiratory infections were the most common focus in both groups, but were more prevalent in the shock group (65.2 % vs. 42.9 %). Conversely, neurological and soft tissue foci were less common in the shock group (Fig. 3).

Fig. 3.

. Common infectious foci.

Fever at admission (≥ 38.5°C) was strongly associated with septic shock (91.3 % vs. 39.7 %; OR = 16.0, 95 % CI: 3.44‒74.13, p < 0.001), while the presence of tachypnea was not (78.3 % vs. 66.7 %; p = 0.41). As expected, the median scores for pSOFA (6 vs. 2, p < 0.001) and PSS (4 vs. 2, p < 0.001) were significantly higher in the shock group. The PRISM-III score was also higher in the shock group, but this difference did not reach statistical significance (9 vs. 7, p = 0.07) (Table 2).

In multivariable logistic regression analysis, both PSS (OR = 2.78, 95 % CI: 1.43–5.38, p = 0.002) and PRISM-III (OR = 1.23, 95 % CI: 1.02–1.48, p = 0.029) were independent predictors of septic shock, while pSOFA was not (p = 0.984) (Table 3).

The diagnostic performance of the scores at their optimal cut-offs is detailed in Table 4. The PSS, at a cut-off of 2.5, showed exceptional sensitivity (95.7 %) and negative predictive value (97.6 %), making it an excellent rule-out tool. In contrast, PRISM-III, at a cut-off of 12.5, exhibited high specificity (90.5 %) and a strong positive likelihood ratio (LR+ 6.41), rendering it a useful rule-in tool. The pSOFA score showed balanced but moderate sensitivity (69.6 %) and specificity (81.0 %).

The PSS demonstrated the highest discriminative ability (AUROC = 0.867, 95 % CI: 0.777‒0.931), followed by PRISM-III (AUROC = 0.826, 95 % CI: 0.729‒0.899) and pSOFA (AUROC = 0.791, 95 % CI: 0.690‒0.871) (Fig. 4). However, pairwise comparisons revealed no statistically significant differences between the AUROCs (all p > 0.05, Table 5).

Fig. 4.

ROC curves for pSOFA, PSS, and PRISM III. Receiver Operating Characteristic (ROC) curves for the pediatric Sequential Organ Failure Assessment (pSOFA), Phoenix Sepsis Score (PSS), and Pediatric Risk of Mortality-III (PRISM-III) scores for the prediction of septic shock for exploratory purposes. The optimal cut-off values (determined by Youden's index) and associated performance metrics should be interpreted as preliminary and require validation in larger cohorts.

The multivariable model showed good calibration, with a non-significant Hosmer-Lemeshow test (χ2 = 10.64; p = 0.155) and a calibration slope of 0.814, indicating minimal systematic overfitting or underfitting (Fig. 1).

Stratification by the PSS cut-off of 2.5 effectively identified high-risk patients. The shock rate was 48.9 % in the high-score group (≥ 2.5) compared to only 2.4 % in the low-score group (< 2.5), corresponding to a markedly increased odds of shock (OR = 38.3, 95 % CI: 4.84‒302.79, p < 0.001) (Table 6).

The single-center design and modest sample size (n = 86), with only 23 septic shock events, limit statistical power and increase the risk of unstable estimates due to a low events-per-variable ratio. Accordingly, multivariable regression results should be interpreted with caution and considered exploratory and hypothesis-generating rather than confirmatory. Calibration analyses were also descriptive and should be interpreted cautiously. Cut-offs derived using Youden’s index are optimistic and should not be interpreted as clinically actionable thresholds without external validation.

Convenience sampling may introduce selection bias and limit the generalizability of our findings to the broader Vietnamese pediatric population. In addition, differences in score timing and clinical management may introduce incorporation bias. Specifically, time-varying therapies administered between early scoring (Phoenix Sepsis Score and pSOFA within the first 6-hours) and 24-hour PRISM III assessment may influence both shock classification and score components. Furthermore, several variables used to define septic shock overlap with components of Phoenix and pSOFA, which may artificially inflate discrimination metrics and limit the interpretability of comparative performance; this is one of the main limitations of the study. Future multicenter studies with larger samples, preregistered thresholds, and external validation (with planned subgroup analyses by age/infectious focus) are warranted. Similar single-center constraints have been highlighted in a Mekong Delta PICU cohort evaluating biomarker-based prognostication, reinforcing the need for broader validation.30