Establishing a nomogram of the risk of in-hospital mortality in children with suspected sepsis and validating the applicability of the Phoenix

Abstract

Objective

This study aimed to develop a nomogram for assessing the risk of in-hospital mortality in children with suspected sepsis admitted to pediatric intensive care unit (PICU). Additionally, it sought to evaluate the applicability of the most recent international diagnostic criteria for pediatric sepsis, the Phoenix diagnostic criteria.

Methods

Data were collected from electronic medical records of pediatric sepsis patients admitted to the PICU at West China Second Hospital, Sichuan University, between January 1, 2021, and August 31, 2023, within the first 24 h of admission. A total of 539 children were randomly assigned to two groups: a training set (n = 458, 85%) and a validation set (n = 80, 15%). The LASSO regression was employed for variable selection. Multivariable logistic regression analysis was used to construct a predictive model incorporating the predictors selected through LASSO regression. A nomogram was constructed based on the selected variables for visual representation. Model validation was performed using a calibration plot, ROC curve, and DCA. The AUC of this model were compared with those of the SIRS, PELODS-2, pSOFA, PMODS, and Phoenix criteria. Subsequently, the pSOFA, SIRS, and Phoenix criteria were applied to data collected within the first 24 h of PICU admission to compare mortality rates based on these criteria in patients with suspected sepsis.

Results

Among the 42 variables examined, LASSO regression analysis identified six significant predictors: urea, P/F ratio, mechanical ventilation, HCO₃⁻ levels, vasoactive medications, and neurological. The model, incorporating these six predictors, demonstrated excellent predictive performance, with an AUC of 0.983 in the training set and 0.979 in the validation set. And it was confirmed by the DCA that the model has some clinical utility in terms of the risk of in-hospital mortality. ROC-AUC values and 95% CI of Phoenix, pSOFA, PMODS, SIRS and PELODS-2 scoring systems are 0.958 (0.941–0.976), 0.941 (0.916–0.966), 0.839 (0.787–0.891), 0.662 (0.597–0.726) and 0.918 (0.878–0.958), respectively.

Conclusion

(1) The inclusion of P/F ratio, mechanical ventilation, HCO₃⁻ levels, vasoactive medications, neurological, and urea in the risk nomogram enhanced its ability to predict mortality risk in sepsis patients. (2) The Phoenix diagnostic criteria exhibited superior discriminative power compared to pSOFA and SIRS.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12887-026-06791-w.

Introduction

Sepsis is still one of the primary causes of death among critically ill pediatric patients worldwide. Children are particularly vulnerable to sepsis, with approximately 25 million cases reported globally in 2017, resulting in over 3 million deaths [1]. This burden is especially pronounced in regions with limited economic resources, where sepsis is the leading cause of death among children under 5 years of age in Africa and parts of Asia, accounting for approximately 36.7% of all such deaths. Moreover, many survivors of pediatric sepsis experience long-term consequences, including cognitive, physical, and emotional impairments, which adversely affect growth and development and place a significant economic burden on both society and families [2–4]. As a result, the early identification of risk factors for poor outcomes in children with sepsis is critical for improving prognosis and reducing the long-term impact of this condition.

Numerous biomarkers, including body temperature, white blood cell count, creatinine levels, and the PaO₂/FiO₂ (P/F) ratio, have been utilized to develop various diagnostic criteria and prognostic models for sepsis. Further studies into the pathophysiology of sepsis revealed that the systemic inflammatory response syndrome (SIRS) criteria are overly sensitive, yet lack the specificity required to adequately address potentially fatal organ failure. To address these limitations, the criteria for diagnosing sepsis in adults were updated during the third international consensus conference on sepsis and septic shock, with the introduction of a sepsis-related organ failure assessment (SOFA) score of ≥ 2 as a diagnostic threshold [5]. Correlational studies have shown a significant association between adult sepsis inpatient mortality and SOFA scores, with higher scores correlating with increased mortality rates [6–8]. However, the adult SOFA score cannot be directly applied to pediatric patients, as physiological markers such as blood pressure and creatinine levels vary with age in children [9–11] Consequently, two pediatric SOFA scoring systems (pediatric SOFA, or pSOFA) were proposed by Schlapbach et al. and Matatics et al. Researchs suggest that the pSOFA score has superior predictive efficacy compared to other scoring systems, including SIRS, pediatric logistic organ dysfunction-2 (PELOD-2), pediatric multiple organ dysfunction score (P-MODS), pediatric risk of mortality (PRISM), and pediatric index of mortality (PIM) [12, 13]. However, the use of pSOFA in clinical practice remains limited due to discrepancies in the maximum score between the two pSOFA systems, their more complex scoring criteria, and the lack of multicenter validation. In January 2024, the American society of critical care medicine proposed an updated international consensus on pediatric sepsis, based on a multicenter retrospective analysis of over 3 million children from 10 medical centers across four countries. This consensus recommended the use of the Phoenix score (≥ 2) as the diagnostic criterion for pediatric sepsis [14]. However, significant regional variations exist in the diagnostic efficacy of the Phoenix score. In less economically developed regions, children with a Phoenix score ≥ 2 have an in-hospital mortality rate of 28.5%, compared to 7.1% in more economically developed regions [14].

To identify the risk factors for in-hospital mortality in children with sepsis and develop a reliable prognostic prediction model, this study conducted a retrospective analysis of children with suspected sepsis in our medical unit. Additionally, the study compared the clinical utility of this model with that of the Phoenix score, pSOFA [12], SIRS, PELOD-2 [15], and P-MODS [16], and assessed the applicability of the most recent international consensus criteria for pediatric sepsis.

Methods

Patient characteristics

The medical records of individuals with suspected sepsis admitted to the PICU of West China Second Hospital, Sichuan University, between January 1, 2021, and August 31, 2023, were retrospectively reviewed. Inclusion criteria: (1) required intravenous antibiotic treatment within 24 h of admission for acute or suspected infection (positive cultures or PCR from sterile sites or clinical evidence of infection. The clinical evidence could include infection-related symptoms, positive findings on physical examination, imaging results [such as pneumonia confirmed by chest X-ray], laboratory abnormalities [such as abnormal complete blood count, elevated CRP, or PCT], or characteristic clinical syndromes [like scarlet fever or staphylococcal scalded skin syndrome]); (2) age between 28 days and 18 years; (3) admission to the PICU at least 24 h. Exclusion criteria: (1) presence of a major organic disease, such as malignancy, hereditary metabolic disorders, congenital cardiac conditions, hematologic malignancies, or a history of major surgery; (2) PICU stay of less than 24 h. A flow diagram illustrating the study design is presented in Fig. 1. This study was approved by the medical ethics committee of the West China Second Hospital, Sichuan University (approval number:2017-46-4). This study received ethical approval from the Medical Ethics Committee of West China Second Hospital, Sichuan University (Approval No. 2017-046-4). As a retrospective observational analysis, clinical trial registration was not required according to international clinical trial regulations. The ethics committee granted a waiver for obtaining informed consent from participants in accordance with national regulations governing retrospective studies. All study procedures were conducted in accordance with the ethical principles outlined in the Declaration of Helsinki.

Fig. 1.

Flow diagram of study design

Data colletion

The research population was organized by a properly trained medical team through a review of hospital case files, applying the inclusion and exclusion criteria. Information on patient demographics, clinical and physiological characteristics, and prognosis was retrospectively collected. To standardize data collection, the most critical physiological and clinical conditions observed within the first 24 h of admission were recorded. The independent variables in the mortality risk model included patient demographics (age and sex), vital signs (body temperature, respiratory rate, heart rate, mean arterial pressure [MAP]), infection markers (leukocyte and platelet counts, procalcitonin [PCT], C-reactive protein [CRP] levels), organ dysfunction indicators (bilirubin, creatinine, platelet count [PLT], fibrinogen [FIB], international normalized ratio [INR], urea, D-dimer levels), blood gas parameters (partial pressure of carbon dioxide [PaCO₂], bicarbonate [HCO₃⁻], lactate [LAC], partial pressure of oxygen [PaO₂]), primary. infection site, use of mechanical ventilation, administration of one or more vasoactive medications, in-hospital prognosis, and neurological dysfunction (based on the Phoenix sepsis criteria: score of 0 for glasgow coma scale [GCS] > 10; 1 for GCS ≤ 10; 2 for bilaterally fixed pupils.

Statistical analyses

Statistical analysis was performed using R software (version 3.6.1; R Foundation for Statistical Computing, Vienna, Austria). First, We adopted a stratified random sampling strategy to divide our dataset into a 85% training dataset and 15% testing dataset stratified by outcome (28-day mortality). The glmnet package was employed to perform least absolute shrinkage and selection operator (LASSO) regression, a method for both shrinkage and variable selection in linear regression models. In this analysis, LASSO minimizes prediction error for a continuous response variable by applying a penalty to the model parameters, causing the regression coefficients of some variables to shrink toward zero. Variables with a regression coefficient of zero are excluded from the model, while those with non-zero coefficients are considered most strongly associated with the outcome variable. The family parameter was set to ‘binomial’ to accommodate the binary dependent variable, where the outcome of interest was the occurrence of death. The type measure parameter was set to “deviance,” which corresponds to -2 log-likelihood, as a measure of model fit. The LASSO regression procedure in R software performed 10-fold cross-validation for the centralization and normalization of the included variables, subsequently selecting the optimal value for the regularization parameter (λ). The model with the lowest λ that exhibited good performance, while retaining the fewest independent variables, was identified. Using the LASSO method, we analyzed the training set to identify the optimal predictors of mortality risk, including sex, age, disease duration, body temperature, respiratory rate, heart rate, mean arterial pressure, leukocyte and platelet counts, PCT, CRP, bilirubin, creatinine, PLT, FIB, INR, urea, D-dimer, PaCO₂, HCO₃⁻, LAC, PaO₂, neurological, use of mechanical ventilation, administration of vasoactive medications, and in-hospital prognosis. These variables were initially screened based on the aforementioned inclusion criteria.

We then employed the rms package in R to perform multivariable logistic regression and construct a predictive model by incorporating the features selected through LASSO regression. These features were evaluated based on their odds ratios and P-values, with a 95% confidence interval. All statistical tests were two-sided. After introducing all the selected features, we assessed their statistical significance and applied the significant predictors to develop a predictive model for in-hospital mortality risk in children with sepsis and suspected sepsis. Six features, which were statistically significant, were included in the final model and used to develop the nomogram prediction tool. The rms package was also utilized to create the nomogram diagram. Additionally, multiple validation methods were applied to evaluate the accuracy of the risk prediction model, using data from both the training and validation sets. The pROC package in R was used to generate the receiver operating characteristic (ROC) curve, and the area under the curve (AUC) was computed to assess the discrimination power of the nomogram in distinguishing true positives from false positives. The model’s performance was further compared with existing models, including Phoenix, pSOFA, pMODS, PELOD-2, and SIRS. To evaluate the calibration of the risk nomogram for hospital mortality, we used the rms package to plot and calculate calibration curves, accompanied by the Hosmer-Lemeshow test. Finally, the nricens package was used for decision curve analysis (DCA), which assesses the clinical utility of the nomogram based on the net benefit at different threshold probabilities in the sepsis cohort. The performance of our model was again compared with Phoenix, pSOFA, pMODS, PELOD-2, and SIRS.

Results

Characteristics of the study cohort

Among the 539 children with suspected sepsis included in our analysis, 235 were female, and 304 were male. In-hospital mortality occurred in 73 patients, accounting for 13.7% of the cohort. The patients were randomly assigned to either the training or validation sets in an 85:15 ratio, resulting in 485 patients in the training set and 81 patients in the validation set. No statistically significant differences were found between the two groups with respect to sex, age, vital signs, blood gas parameters, infection-related markers, indicators of organ dysfunction, mechanical ventilation within 24 h, neurological scores, use of vasoactive medications, etiology of sepsis, or mortality outcomes. Table 1 presents the basic characteristics of the patients in both sets.

Table 1.

Characteristics of the 539 infected patients enrolled in the study

Variables	Total(N = 539)	Test(N = 80)	Trian (N = 459)	z/χ2	p
Temperature [M (P25, P75)]	36.9(36.6, 37.7)	37.0(36.7, 37.8)	36.9(36.6, 37.7)	z:1.473	0.141
Heart rate [M (P25, P75)]	159(140, 171)	160(140, 172)	159(14, 171)	z:0.156	0.876
mean arterial pressure [M (P25, P75)]	65(58, 71)	64(59, 70)	65(58, 71)	z:0.102	0.919
Respiratory rate [M (P25, P75)]	50.00(40, 59)	52(40, 58)	50(40, 59)	z:0.066	0.947
Age [no. of cases (%)]				χ2:0.145	0.716
1–12 months	313 (58.1%) a	266(60.0%)	47(58.75%)
12–36 months	114 (21.2%)	96(20.9%)	18(22.5%)
36–84 months	68 (12.6%)	58(12.6%)	10(12.5%)
84–144 months	30 (5.6%)	26(5.7%)	4(5.0%)
> 144 month	14 (2.6%)	13(2.8%)	1(1.25%)
Gender [no. of cases (%)]				χ2:2.063	0.151
Male	304 (56.4%)	51 (63.7%)	253 (55.1%)
Female	235 (43.6%)	29 (36.2%)	206 (44.9%)
P/F [M (P25, P75)]	232(193, 300)	244(215, 317)	230(186, 290)	z:1.845	0.062
PaCO2 [M (P25, P75)]	36.6(30.7, 44.5)	36.10(32.0, 43.0)	36.6(30.5, 44.7)	z:0.307	0.759
HCO3− [M (P25, P75)]	22.2(19.1, 25.2)	22.2(19.6, 24.4)	22.4(18.9, 25.3)	z:0.174	0.862
Vasoactive medications [no. of cases (%)]				χ2: 0.103	0.724
No	479 (88.9%)	73(91.2%)	406(88.5%)
one	22 (4.1%)	3(3.8%)	19(4.1%)
two	38 (7.1%)	4(5.0%)	34(7.4%)
Bilirubin [M (P25, P75)]	7.5(5.0, 15.0)	8.4(4.8, 15.4)	7.3(5.1, 14.8)	z:0.586	0.558
Creatinine [M (P25, P75)]	21.0(18.0, 29.0)	20.5(17.0, 27.7)	22.0(18.0, 29.0)	z:0.699	0.485
WBC [M (P25, P75)]	9.1(6.4, 13.4)	8.7(5.9, 14.1)	9.2(6.5, 13.3)	z:0.524	0.601
PLT [M (P25, P75)]	344(227, 457)	352(217, 446)	344(22, 457)	z:0.254	0.799
Neurological	0(0, 0)	0(0, 0)	0(0, 0)	z:0.171	0.865
LAC [M (P25, P75)]	1.71(1.20,2.86)	1.45(1.19, 2.40)	1.77(1.20, 2.99)	z:1.791	0.073
UREA [M (P25, P75)]	0(0, 0)	0(0, 0)	0(0, 0)	z:0.962	0.336
FIB [M (P25, P75)]	239(178, 344)	246(189, 342)	238(177, 344)	z:0.441	0.660
INR [M (P25, P75)]	1.13(1.04, 1.27)	1.12(1.04, 1.27)	1.13(1.05, 1.27)	z:0.109	0.913
D-dimer [M (P25, P75)]	0.78(0.45, 1.68)	0.76(0.42, 1.32)	0.78(0.45, 1.73)	z:1.035	0.301
CRP [M (P25, P75)]	4.2(0.8, 24.0)	4.0(1.0, 30.4)	4.2(0.8, 23.2)	z:0.486	0.627
PCT [M (P25, P75)]	0.30(0.10, 1.50)	0.34(0.13, 1.00)	0.26(0.10, 1.67)	z:0.361	0.718
In-hospital death [no. of cases (%)]				χ2:0.120	0.729
No	465 (86.3%)	70 (87.5%)	395 (86.1%)
Yes	74 (13.7%)	10 (12.5%)	64 (13.9%)
Infection [no. of cases (%)]				χ2:0.189	0.709
Pulmonary	403(74.8%)	342(74.5%)	61(76.2%)
intracranial	70(13%)	58(12.6%)	12(15%)
Abdominal	45((8.3%)	39(8.5%)	6(7.5%)
Skin	7(1.3%)	7(1.5%)	0(0%)
Hematologic	12(2.2%)	11(2.4%)	1(1.3%)
Urinary	2(0.4%)	2(0.4%)	0(0%)

Abbreviations: P/F PaO₂/FiO_2,CRP C-reactive protein, PCT Procalcitonin, PLT Platelet, FIB Fibrinogen, INR International normalized ratio

Predictive model construction in the training group

The predictive model was developed using multivariate logistic regression after selecting the variables listed in Table 1 through LASSO regression analysis. Six of the initial 42 variables, including P/F ratio, mechanical ventilation, HCO₃⁻, vasoactive medications, neurological, and urea, were incorporated into the model as predictors. The LASSO regression analysis retained 13 variables with non-zero coefficients (Fig. 2). Among them, six indicators demonstrated significant statistical differences and were considered independent of one another (Table 2). These six variables were subsequently included in the final predictive model to create an in-hospital mortality risk nomogram (Fig. 3).

Fig. 2.

Variable selection by the LASSO binary logistic regression model. A coefficient profile plot was constructed against the log(lambda) sequence. a Ten variables with nonzero coefficients were selected by deriving the optimal lambda. b Following verification of the optimal parameter (lambda) in the LASSO model, we plotted the partial likelihood deviance (binomial deviance) curve versus log(lambda) and drew dotted vertical lines based on 1 standard error criteria. LASSO Least absolute shrinkage and selection operator

Table 2.

Logistic regression analysis of the predictors for the risk of mortality in suspicious sepsis patients in PICU

Variables	estiamte	Standard error	z	OR (95%CI)	p
mechanical ventilation	1.983	0.668	2.971	0.982-0.994	0.002
vasoactive medicine	1.853	0.467	3.968	3.442-41.043	<0.001
P/F	-0.011	0.005	-2.348	0.773-0.926	0.018
HCO3-	-0.119	0.057	-2.103	3.32-15.854	0.003
Urea	0.735	0.260	2.826	5.401-44.64	0.004
Neurological	2.334	0.845	2.760	1.577-4.151	0.005

Abbreviations: P/F PaO₂/FiO₂

Fig. 3.

A nomogram predicting in-hospital mortality in children with suspected sepsis

Predictive model validation

The ROC curve was used to evaluate the discriminatory capacity of the predictive model. As shown in Fig. 4a and b, this predictive model demonstrated an excellent discriminatory ability. In the training set, AUC was 0.983. As shown in Fig. 5a, b and a calibration curve was constructed by plotting the actual probability of in-hospital mortality (y-axis) against the predicted occurrence rate (x-axis), demonstrating a favorable agreement between the predicted and observed probabilities in both the training and validation sets. The DCA was employed to assess the usefulness of the model in predicting the risk of in-hospital mortality, supporting the value of the model in identifying high-risk patients (Fig. 6).

Fig. 4.

Area under the receiver operating characteristic (ROC) curve of the training set (a) and validation set (b)

Fig. 5.

The calibration curve of prediction nomogram of in-hospital mortality risk in children with suspected sepsis in the training set (a) and validation set (b)

Fig. 6.

Decision curve for the predicted in-hospital mortality risk in children with suspected sepsis in the training set (a) and validation set (b) using the nomogram

Discussion

To develop a predictive model for in-hospital mortality risk among children with suspected sepsis in the PICU, we integrated several organ dysfunction scoring systems. The identified independent risk factors included mechanical ventilation, P/F, type of vasoactive medication, neurological score, HCO₃⁻, and urea levels. While most identified risk factors for pediatric sepsis prognosis are consistent with those reported in previous studies, our study differs methodologically, as previous research primarily used logistic regression analysis. Previous research often relied on single-factor analysis to validate multivariate findings or conducted stepwise and iterative logistic regression analyses to derive their conclusions. For example, Chen et al. employed multivariate logistic regression to identify risk factors for in-hospital mortality in children with sepsis, highlighting brain natriuretic peptide, albumin, total bilirubin, D-dimer, lactate levels, and mechanical ventilation within 24 h as significant predictors [17]. However, these approaches must address confounding variables and multicollinearity. In this study, we applied LASSO regression analysis to mitigate these challenges and produce more accurate results. Variables were first screened using LASSO regression, followed by traditional logistic regression analysis. To validate the model’s accuracy and stability, we constructed ROC curves and DCA for both the training and validation sets. Our study found that mechanical ventilation was associated with higher in-hospital mortality, despite being a critical support for organ function. Mechanical ventilation is associated with worsening illness severity in the short term [18]. Furthermore, Villar et al. demonstrated that mechanical ventilation can exacerbate the inflammatory response, and increased inflammation is associated with an elevated risk of cardiovascular events, potentially leading to higher mortality rates [19]. A study found that 51% of participants believed invasive mechanical ventilation worsened the condition of septic shock patients and increased their mortality [20]. When considering disease severity, studies have shown that intubation within 24 h of septic shock onset did not lead to higher in-hospital mortality. However, increased mortality may result from inappropriate timing of intubation [21].

The P/F ratio, a widely used clinical indicator of pulmonary oxygenation function, is essential in assessing organ function. A prospective multicenter analysis evaluating the prognosis of patients with bloodstream infections found that a P/F ratio ≤ 300 was an independent risk factor for mortality [22]. A predictive model for adult sepsis indicates that the P/F ratio is one of the most accurate predictors of 28-day mortality in septic patients [23]. Additionally, it is an independent risk factor for 30-day mortality in sepsis patients admitted to the emergency department, as well as for 28-day mortality in elderly septic patients [24, 25]. Although our final model included different predictors, our results correspond to comparable research on adult sepsis, with the P/F ratio remaining one of the most significant independent factors. Higher PaO₂ levels during the first three days in the intensive care unit were associated with decreased 28-day morbidity compared to the conservative oxygen partial pressure group, according to a nationwide cohort study [26]. Unfortunately, our study’s predictive model did not include PaO₂ as an independent risk factor. Further research is needed to determine whether children’s responses to high oxygen partial pressure are comparable to those of adults. Additionally, no study has provided definitive recommendations on the optimal range of arterial oxygen partial pressure for septic patients, regardless of age.

Early administration of norepinephrine during shock may be justified in patients with severe vasoplegia and/or at high risk of fluid overload [27]. However, infection exacerbation, insufficient endogenous catecholamine secretion, or down-regulation of β-receptors leads to reduced responsiveness to vasoactive medications [28]. Combination therapy with multiple vasoactive medications is often required. The duration, type, and dose of vasoactive drug administration require careful consideration after septic shock onset. Studies have shown that prolonged norepinephrine use significantly increases the risk of death in patients [29]. In a study on septic shock prognosis, Dubin et al. proposed the blood pressure response index to assess the response of septic shock patients to combination therapy with vasoactive drugs. The more vasoactive drugs used in combination and the higher the doses, the greater the mortality rate [30]. Increased mortality may also be linked to severe side effects of vasoactive drugs, with some studies indicating that patients with severe drug side effects have twice the mortality rate compared to those without such effects [31]. The use of high doses of vasoactive drugs can lead to vasoconstriction, and higher blood urea nitrogen levels are independently associated with the presence and severity of neonatal sepsis, as well as the corresponding increase in gastrointestinal damage, leading to a poor prognosis [32].

Elevated urea levels are considered one of the most important risk factors for early sepsis, consistent with our findings and previous studies. Interestingly, low urea levels (< 2.86 mmol/L) may serve as a “safety sign,” indicating a low probability of sepsis within 72 h of emergency admission [33]. Blood urea nitrogen is derived from protein metabolites and is primarily cleared by the kidneys [34]. Due to the interplay between nutritional status, energy metabolism, and gastrointestinal bleeding, elevated urea levels typically reflect a rapid protein response to severe systemic injury, such as sepsis. Urea levels significantly increase in response to elevated protein metabolism and acute kidney injury in patients with sepsis [35]. A prospective multicenter randomized controlled trial by Gunst et al. suggests that the duration of elevated urea levels is an important independent determinant of renal replacement therapy duration [36]. Another study on post-discharge survival in patients with sepsis found that elevated urea nitrogen levels at discharge were associated with higher mortality rates [37].

Neurologic function is often an important indicator for assessing disease prognosis. As the only subjective measure in our scoring system, we found a significant difference between GCS scores greater than 10 and those below 10 in predicting inpatient mortality in patients with sepsis. We therefore followed the established Phoenix Sepsis Criteria: 0 points for GCS > 10, 1 point for GCS ≤ 10, and 2 points for fixation/pupil dilation. A study by Lai et al. found a difference between a GCS score of 15 and scores below 15 in predicting 28-day mortality in patients with sepsis [23]. Another study indicated that even mild alterations in consciousness, with a GCS score of 13–14, were independently associated with an increased risk of mortality [38]. Sepsis-associated encephalopathy (SAE) is considered one of the most common causes of impaired consciousness. The MIMIC-IV and eICU databases revealed that the incidence of SAE can be as high as 68% [39]. In contrast, a French multicenter cohort study reported an incidence of 53% [38]. It is well established that the occurrence of SAE is independently associated with short-term mortality. The severity of SAE is also correlated with mortality, with patients having GCS scores between 3 and 8 having the worst prognosis [38].

A decrease in HCO₃⁻ suggests the possibility of metabolic acidosis, which is associated with a negative prognosis. According to a large observational study using a binational ICU database from Australia and New Zealand, approximately 1.5% of ICU patients had severe metabolic acidosis. ICU mortality and in-hospital mortality rates were as high as 43.5% and 48.3%, respectively, exceeding the mortality rates for sepsis patients in the same database [40]. A recent international observational study conducted in 18 ICUs reported that 14% of critically ill patients had moderate-to-severe metabolic acidosis, with a median incidence of 172.5 cases per year. This suggests that metabolic acidosis is a significant issue among ICU patients [41]. A machine learning approach used to validate multiple bio-marker scoring systems for assessing sepsis severity in critically ill children identified angiopoietin-1, angiopoietin-2, and bicarbonate as having the strongest correlation with sepsis severity. Their concentrations were also predictive of the time-dependent nature of sepsis progression in children [42].

While the prediction model cannot definitively determine whether death occurs in a particular child, we hope that the predicted probability of in-hospital death in children with suspected sepsis can serve as one factor in clinical decision-making. Additionally, three commonly used organ function scoring systems-pSOFA, PELODS-2, PMODS and the newly introduced Phoenix criteria were selected for comparison, based on their practical application in this unit. Both PELODS-2 and pSOFA scores align with the core principles of the Sepsis 3.0 definition, which is widely utilized in pediatric sepsis prognostic studies. PMODS is also widely employed in pediatric critical care. The Phoenix criteria similarly emphasize that organ dysfunction is central to the definition of pediatric sepsis, though no additional relevant clinical studies have been reported. Our study demonstrated that PMODS has moderate predictive efficacy, with an ROC-AUC of 0.780 (95% CI: 0.718–0.843). The Phoenix, pSOFA, and PELODS-2 scoring systems exhibited high predictive efficacy, with ROC-AUC values of 0.943 (95% CI: 0.914–0.972), 0.932 (95% CI: 0.905–0.958), and 0.918 (95% CI: 0.878–0.958), respectively (additional Fig. 1). However, these models exhibited lower predictive efficacy compared to the prediction models developed in this study, which included fewer indicators and avoided age-stratified assessments of the renal and cardiovascular systems.

Additionally, our study is the first to examine the diagnostic effectiveness of the Phoenix criteria. Critically ill children from provinces of Sichuan, Tibet, Qinghai, and Gansu are treated at the Second West China Hospital of Sichuan University, a regional pediatric medical center in western China. As such, these patients are highly representative of pediatric patients with severe illnesses. Our study found that the mortality rate among children with suspected sepsis admitted to the PICU was 13.7%. However, in practice, different regions may employ varying diagnostic criteria, with the 2005 version of international guidelines being the most widely used, followed by pSOFA. pSOFA remains a key focus of current research in the field. Hence, we compared these three diagnostic criteria (additional Table 1). In the study cohort, 256 patients showed an increase of 2 or more points in the Phoenix score from baseline; 286 patients met 2 or more SIRS criteria; and 401 patients had a pSOFA score of 2 or more. The distributions of each score and their relationship to in-hospital mortality are presented in additional Fig. 2.

No deaths occurred in children with a Phoenix score or pSOFA score less than 2 (additional Fig. 2c, additional Fig. 2e), while 23 (9.1%) children died who met fewer than 2 SIRS criteria (additional Fig. 2a). The SIRS diagnostic criteria remain overly sensitive but lack specificity. Regarding in-hospital mortality, pSOFA demonstrated higher specificity but lower predictive accuracy compared to the Phoenix criteria. No significant positive correlation was found between the pSOFA score and mortality when the score ranged from 2 to 6 (additional Fig. 2d); however, a significant positive correlation was observed between the Phoenix score and mortality when the score was ≥ 2 (additional Fig. 2f). This suggests that, among children admitted to the intensive care unit with suspected infection, defining sepsis based on an increase in the Phoenix score offers superior prognostic accuracy for in-hospital mortality compared to either SIRS criteria or pSOFA.

Limitations

It’s also important to recognize the limitations of the study. First, the initial 24 h of PICU data were used to create and validate our model. This limited duration is the limitation since it is unable to capture later clinical deterioration, even while it allows for the risk of in-hospital mortality evaluation. Second, the model’s direct use in the general ward may be restricted due to its reliance on characteristics unique to the critical care unit (mechanical ventilation, vasopressin delivery, etc.). Furthermore, our study included only children with suspected sepsis who received antibiotics, which may have missed cases diagnosed with bacterial sepsis but did not receive antibiotics in a timely manner, cases with similar presentations but did not receive antibiotics, and cases of viral sepsis, which may have affected the generalizability of the model. Finally, since this was a single-center retrospective study, more prospective research and multi-center validation are required.

Conclusion

Our study identified six independent risk factors for in-hospital mortality in PICU patients with suspected sepsis: urea levels, P/F ratio, mechanical ventilation, HCO3⁻ levels, vasoactive medications, and neurological status. The nomogram based on these factors showed good predictive performance and clinical utility, serving as a practical tool for risk assessment and clinical decision-making in PICUs. Notably, the Phoenix diagnostic criteria demonstrated superior discriminative power compared to both pSOFA and SIRS criteria.

Abbreviations

PICU

Pediatric intensive care unit

SIRS

Systemic inflammatory response syndrome

PELODS-2

Pediatric logistic organ dysfunction-2

SOFA

Sepsis-related organ failure assessment

pSOFA

Pediatric sepsis-related organ failure assessment

PMODS

Pediatric multiple organ dysfunction score

P/F

PaO₂/FiO₂

PRISM

Pediatric risk of mortality

PIM

Pediatric index of mortality

MAP

Mean arterial pressure

PCT

Procalcitonin

CRP

C-reactive protein

PLT

Platelet

FIB

Fibrinogen

INR

International normalized ratio

PaCO₂

Partial pressure of carbon dioxide

HCO₃⁻

Bicarbonate

LAC

Lactate

PaO₂

Partial pressure of oxygen

GCS

Glasgow coma scale

LASSO

Least absolute shrinkage and selection operator

ROC

Receiver operating characteristic

DCA

Decision curve analysis

AUC

Area under the curve

SAE

Sepsis-associated encephalopathy

Authors’ contributions

LL and XL designed the study concept. And SH wrote the manuscript. SH and QL collected the data. XL and SH analyzed and interpreted the data. All authors were involved in drafting and revising the paper.

Funding

This study was funded by the Sichuan Science and Technology Program (2023ZYD0124).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This study was approved by the medical ethics committee of the West China Second Hospital, Sichuan University (approval number: 2017-46-4). Due to the retrospective nature of this study, the informed consent of patients was waived by the Ethics Committee of the West China Second Hospital, Sichuan University according to the national regulations. The procedures of this research adhered to the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.