Blood-based inflammatory parameters and machine learning models for mortality prediction in critically ill pediatric pneumonia: a retrospective cohort study

Abstract

Background

Pneumonia remains a leading cause of morbidity and mortality in critically ill children, yet comprehensive prognostic biomarkers specifically validated in pediatric intensive care settings remain limited. Identifying readily available blood parameters that predict mortality risk could enhance early risk stratification and guide therapeutic interventions. This study aimed to evaluate the prognostic value of blood parameters for hospital mortality in critically ill pediatric pneumonia patients and develop machine learning (ML)-based predictive models.

Methods

This retrospective cohort study analyzed pediatric pneumonia patients (aged 3 months to 18 years) admitted to intensive care units for ≥24 hours from the Pediatric Intensive Care (PIC) database version 1.1. Multivariate Cox regression analyses assessed associations between blood parameters and hospital mortality, adjusting for age, gender, bacterial infection status, and pathogen detection. Restricted cubic spline analyses examined dose-response relationships. Six ML algorithms incorporating significant blood parameters and clinical covariates were developed and validated using 7:3 train-test split.

Results

A total of 606 children were included, with an overall hospital mortality of 14.4% (87/606). Multivariate Cox regression identified three significant prognostic factors: neutrophil percentage adjusted hazard ratio (HR): 1.017, 95% confidence interval (CI): 1.004–1.031, P=0.01], lymphocyte percentage (adjusted HR: 0.985, 95% CI: 0.970–0.999, P=0.04), and platelet-to-neutrophil ratio (PNR) (adjusted HR: 0.995, 95% CI: 0.989–1.000, P=0.04). Restricted cubic spline analyses revealed predominantly linear or U-shaped dose-response relationships. Among ML models, random forest demonstrated superior predictive performance with test area under the curve (AUC) of 0.877 (95% CI: 0.860–0.895), sensitivity of 80.9%, specificity of 79.1%, and accuracy of 79.5%. The model significantly outperformed individual blood parameters (neutrophil percentage AUC: 0.712; lymphocyte percentage AUC: 0.660; PNR AUC: 0.689), and decision curve analysis confirmed positive net benefit across threshold probabilities of 5–50%.

Conclusions

Neutrophil percentage, lymphocyte percentage, and PNR are independent prognostic indicators for mortality in critically ill pediatric pneumonia. ML-based models incorporating these parameters show promise for early mortality risk stratification in pediatric intensive care settings.

Highlight box.

Key findings

• Neutrophil percentage, lymphocyte percentage, and platelet-to-neutrophil ratio (PNR) are independent prognostic indicators for hospital mortality in critically ill pediatric pneumonia patients.

• A random forest model incorporating these blood parameters achieved superior predictive performance with an area under the curve of 0.877, significantly outperforming individual biomarkers.

What is known and what is new?

• Inflammatory biomarkers such as the neutrophil-to-lymphocyte ratio have been established as prognostic markers in adult pneumonia patients. However, the prognostic value of composite inflammatory indices specifically in pediatric intensive care populations remains incompletely characterized, with limited validation across different age groups and disease severities.

• This study systematically evaluated twelve composite inflammatory indices in critically ill pediatric pneumonia patients and identified three significant prognostic blood parameters. We developed and validated machine learning models that demonstrated superior discrimination compared to single biomarkers, with decision curve analysis confirming clinical utility across relevant threshold probabilities.

What is the implication, and what should change now?

• Routine assessment of neutrophil percentage, lymphocyte percentage, and PNR should be considered for early risk stratification in pediatric pneumonia patients admitted to intensive care units. The validated random forest model may serve as a practical tool for identifying high-risk children who may benefit from intensive monitoring and aggressive therapeutic interventions.

Introduction

Pneumonia remains one of the leading causes of morbidity and mortality in children worldwide, with particularly devastating impacts in developing countries and among critically ill pediatric populations (1,2). Despite significant advances in vaccination and treatment strategies during the “Millennium Development Goal” period, pneumonia continues to account for substantial pediatric hospitalizations and deaths. Early childhood respiratory infections, including pneumonia, remain an important global public health issue, with more than 40 million annual cases resulting in approximately 650,000 deaths (3). A recent multicenter study from China reported 72,905 hospitalized cases of viral community-acquired pneumonia among 1,791,343 pediatric pneumonia cases between 2016 and 2022, with respiratory syncytial virus being the leading cause (57.21%) and infants under 1 year representing 59.84% of hospitalizations (4). Notably, the coronavirus disease 2019 (COVID-19) pandemic has significantly impacted the epidemiology of pediatric pneumonia. Studies have shown that Mycoplasma pneumoniae prevalence decreased during and immediately after the pandemic, followed by a sharp resurgence in 2024, with incidence rates reaching 12.5 per 1,000 hospitalizations compared to 2.1 during 2018–2023 (5). Similarly, a multicenter study from China demonstrated that after epidemic control measures were lifted, the number and proportion of children with Mycoplasma pneumoniae pneumonia increased sharply, with a significant increase in the proportion of children aged more than 7 years (6). The burden is especially pronounced in intensive care settings, where pneumonia represents a major cause of respiratory failure and carries mortality rates ranging from 14% to 28% depending on disease severity and patient characteristics (7-9). Moreover, early childhood respiratory infections and pneumonia may have long-term consequences, as a systematic review demonstrated that lower respiratory tract infections in children aged ≤5 years are associated with restrictive spirometry patterns in later childhood and adulthood, with eight of fourteen studies reporting significant reductions in forced expiratory volume (3).

Over the past two decades, substantial progress has been made in understanding the pathophysiology and prognostic factors of severe pediatric pneumonia. Multiple studies have demonstrated that inflammatory biomarkers, particularly the neutrophil-to-lymphocyte ratio (NLR), serve as valuable predictors of disease severity and mortality in pneumonia patients (10,11). The NLR has shown consistent prognostic value across different populations, with higher ratios associated with increased mortality risk, prolonged hospitalization, and development of complications such as necrotizing pneumonia (10,12). Recent investigations have also identified other blood-based parameters including platelet counts, lymphocyte percentages, and composite inflammatory indices as potential prognostic markers, with studies reporting area under the curve (AUC) values ranging from 0.70 to 0.88 for predicting adverse outcomes (13,14). Notably, Liu et al. demonstrated that erythrocyte sedimentation rate, lactate dehydrogenase, interleukin-6, C-reactive protein, neutrophil percentage, and NLR are valuable predictors for early identification of severe Mycoplasma pneumoniae pneumonia in children, with a combined AUC of 0.732 (15). Furthermore, recent studies have shown that the systemic immune-inflammation index (SII) and lymphocyte-to-monocyte ratio are valuable inflammatory biomarkers for predicting extrapulmonary complications in pediatric Mycoplasma pneumoniae pneumonia, with the combination of these markers achieving an AUC of 0.858 (16). Furthermore, machine learning (ML) approaches have emerged as promising tools for mortality prediction, with models incorporating multiple clinical and laboratory parameters achieving sensitivities of 79–88% and specificities of 76–83% for identifying high-risk patients (17,18).

However, significant knowledge gaps persist in optimizing risk stratification for pediatric pneumonia patients in intensive care settings. Most existing studies have focused on single biomarkers or limited combinations, potentially missing the complex interactions between multiple inflammatory and clinical parameters (19,20). The prognostic value of composite inflammatory indices specifically in pediatric populations remains incompletely characterized, with conflicting results regarding optimal cutoff values and limited validation across different age groups and disease severities (21,22). Additionally, while ML models have shown promise in adult populations, their application in pediatric pneumonia has been limited, and existing models often lack the integration of comprehensive blood parameters with clinical covariates necessary for accurate mortality prediction (23). A recent multinational prospective cohort study from the Pediatric Emergency Research Network developed severity prediction models for pediatric community-acquired pneumonia, achieving a c-statistic of 0.82 for discriminating between mild and moderate-to-severe cases (24). However, these models primarily focused on clinical signs and symptoms rather than comprehensive inflammatory biomarkers, and their applicability in intensive care settings remains to be established. The dose-response relationships between inflammatory markers and mortality risk, as well as potential effect modifications by age, gender, and infection type, require further elucidation to develop more precise risk assessment tools.

Therefore, this study aimed to comprehensively evaluate the prognostic value of blood-based inflammatory parameters and develop optimized prediction models for hospital mortality in pediatric pneumonia patients using data from the Pediatric Intensive Care (PIC) database. We systematically assessed twelve composite inflammatory indices alongside traditional blood markers, examined their dose-response relationships with mortality using restricted cubic splines, and evaluated effect modifications through comprehensive subgroup analyses. Furthermore, we developed and validated multiple ML models incorporating significant blood parameters with clinical covariates to establish an optimal risk stratification tool for early identification of high-risk pediatric pneumonia patients who may benefit from intensive monitoring and aggressive therapeutic interventions. We present this article in accordance with the TRIPOD reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-2025-1-846/rc).

Methods

Study design and data source

This retrospective cohort study utilized the PIC database version 1.1.0, containing comprehensive clinical data from patients admitted to intensive care units (ICUs) at Children’s Hospital, Zhejiang University School of Medicine [2010–2018] (25). This 1,900-bed tertiary pediatric center serves as Zhejiang Province’s largest pediatric facility and the Chinese National Clinical Research Centre of Child Health, managing over 3 million annual visits with 119 critical care beds across five specialized ICUs. The PIC database project was approved by the Institutional Review Board of the Children’s Hospital, Zhejiang University School of Medicine (Hangzhou, China) (approval No. 20190114-54). The requirement for individual patient consent was waived because the project did not impact clinical care, and all protected health information was de-identified. The PIC database is publicly available to researchers who have completed the required Collaborative Institutional Training Initiative (CITI) program. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Patient selection and inclusion criteria

Patients with pneumonia were identified using International Classification of Diseases, Tenth Revision (ICD-10) codes J18.900 and J18.901. Inclusion criteria: (I) first hospital admission with pneumonia diagnosis; (II) age 3 months–18 years; (III) ICU stay ≥24 hours. Exclusion criteria: (I) patients with incomplete essential demographic data (age or gender); (II) patients discharged within 24 hours of ICU admission; (III) patients with missing primary outcome data (hospital mortality status); (IV) readmissions (only the first ICU stay was analyzed for patients with multiple admissions). For multiple ICU admissions, only the first stay was analyzed.

Data extraction and variable collection

Clinical data were extracted using SQL queries through PostgreSQL. Variables collected within 24 hours included: demographics (age, gender); laboratory parameters (complete blood count, blood gases, biochemistry, infection markers, electrolytes, organ function tests, coagulation profile, cardiac biomarkers); vital signs; microbiological cultures (bacterial, viral, fungal, mycoplasma, chlamydia); comorbidities by organ system; and outcomes (hospital mortality, length of stay). Prolonged hospital stay was defined as total hospital length of stay exceeding 7 days, and prolonged ICU stay was defined as ICU length of stay exceeding 3 days.

Composite inflammatory indices calculation

Twelve blood-based indices were calculated: NLR, platelet-to-lymphocyte ratio (PLR), SII, neutrophil-to-lymphocyte-platelet ratio (NLPR), hemoglobin-to-red cell distribution width ratio (HRR), red cell distribution width-to-platelet ratio (RPR), red cell distribution width-to-red blood cell ratio (RDW-RBC), mean corpuscular hemoglobin concentration-to-red cell distribution width ratio (MCHC-RDW), mean corpuscular volume-to-red cell distribution width ratio (MCV-RDW), platelet-to-white blood cell ratio (PWR), platelet-to-red blood cell ratio (PRR), and platelet-to-neutrophil ratio (PNR).

Variables with >30% missing data were excluded to ensure data quality and reliability. For variables with ≤30% missing data, we employed multiple imputation using the random forest algorithm (missForest package), which has been demonstrated to outperform traditional imputation methods for mixed-type data in clinical settings. The imputation was performed with 10 iterations and 100 trees, repeated five times with different random seeds to ensure stability. However, imputation of variables with >30% missingness may introduce substantial bias and was therefore avoided in this study (26).

Statistical analysis

Continuous variables were presented as mean [standard deviation (SD)] or median [interquartile range (IQR)]; categorical variables as frequencies (percentages). Between-group comparisons used t-test/Mann-Whitney U test and Chi-squared/Fisher’s exact test. Univariate and multivariate Cox regression evaluated associations with hospital mortality, adjusting for age, gender, bacterial infection, and pathogen detection. Hazard ratios (HRs) with 95% confidence intervals (CIs) were reported.

Non-linear associations

Restricted cubic splines examined non-linear relationships using 3–6 knots determined by Akaike Information Criterion (AIC). Subgroup analyses were stratified by age (0–3, 4–6, 7–12, 13–18 years), gender, pathogen detection, infection types, and comorbidities. Interaction terms tested effect modification (P<0.05 indicating significance).

ML development

Six algorithms were implemented: logistic regression, random forest, XGBoost, Support Vector Machine (SVM), Least Absolute Shrinkage and Selection Operator (LASSO), and gradient boosting machine with optimized hyperparameters. Given the imbalanced nature of the outcome (14.4% mortality rate), we employed stratified sampling to ensure proportional representation of mortality cases in both training and test sets. Models used 70% training data with 5-fold stratified cross-validation repeated thrice. To address class imbalance, we implemented the Synthetic Minority Over-sampling Technique (SMOTE) in the training set, which generates synthetic samples for the minority class to balance the class distribution. Additionally, class weights were applied in algorithms that support this feature (logistic regression, random forest, and SVM) to penalize misclassification of the minority class more heavily. Performance metrics included AUC, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy on 20% test set. Variable importance was assessed for tree-based models. To enhance model interpretability, SHapley Additive exPlanations (SHAP) analysis was performed for the best-performing model. SHAP values were calculated to quantify the contribution of each feature to individual predictions, providing both global feature importance rankings and local explanations for individual patient risk assessments. SHAP summary plots and dependence plots were generated to visualize the direction and magnitude of feature effects on mortality prediction. All analyses were performed using R software (version 4.2.0) with the “shapr” and “SHAPforxgboost” packages. Two-sided P<0.05 was considered statistically significant.

Results

Study population and baseline characteristics

A total of 1,001 pediatric patients with pneumonia (ICD-10 codes: J18.900 or J18.901) were initially identified from the PIC database version 1.1. After applying sequential exclusion criteria, 606 patients comprised the final analytical cohort. Specifically, 79 patients were excluded due to non-first admissions, 247 patients were excluded for age outside the 3-month to 18-year range, and 69 patients were excluded for ICU stays less than 24 hours (Figure 1). The overall hospital mortality rate was 14.4% (87/606). The median age was 0.87 years (IQR: 0.44–2.61 years), with no significant difference between survivors and non-survivors (P=0.15). Children aged <1 year constituted the largest proportion of hospitalizations (53.0%), followed by 1–3 years (23.8%), 3–6 years (12.4%), and ≥6 years (10.9%). Male patients comprised 59.4% of the population (P>0.99). Most patients (547/606, 90.3%) had no documented comorbidities. Notably, all patients in the mortality group had no comorbidities (87/87, 100%), while 59 survivors (59/519, 11.4%) had at least one comorbidity, including 2 (0.4%) with one, 3 (0.6%) with two, 4 (0.8%) with three, and 50 (9.6%) with four or more comorbidities (P=0.006).

Figure 1.

Study population flowchart. Sequential exclusion criteria applied to identify the final analytical cohort from the PIC database. Numbers indicate patients excluded at each step with corresponding reasons. ICU, intensive care unit; PIC, Pediatric Intensive Care.

Regarding pathogen distribution, 208 patients (34.3%) had no pathogen detected, 193 (31.8%) had single pathogen infection, and 205 (33.8%) had mixed infection. Bacterial infection predominated (63.4%), while mycoplasma infection was rare (0.8%). Pathogen detection was significantly lower in non-survivors (50.6% vs. 68.2%, P=0.002). Laboratory parameters showed non-survivors had lower pH (7.34 vs. 7.36, P=0.048), higher blood urea nitrogen (P=0.02), and elevated alanine aminotransferase (P=0.002). Median hospital stay was shorter in non-survivors (12.00 vs. 14.00 days, P=0.04) (Table 1). Subgroup analysis by pathogen type (Table S1) revealed that patients without pathogen detection had higher mortality (20.7% vs. 13.0% vs. 9.3%, P=0.004), higher neutrophil percentage (P=0.04), and lower lymphocyte percentage (P=0.045) compared to single and mixed infection groups.

Table 1. Baseline characteristics of pediatric pneumonia patients stratified by hospital mortality.

Characteristics	Overall (n=606)	Survived (n=519)	Mortality (n=87)	P value	SMD
Age (years)	0.87 [0.44, 2.61]	0.93 [0.43, 2.76]	0.77 [0.44, 1.18]	0.15	0.125
Age group (years)
<1	321 (53.0)	267 (51.4)	54 (62.1)	0.23	0.260
1–<3	144 (23.8)	126 (24.3)	18 (20.7)
3–<6	75 (12.4)	69 (13.3)	6 (6.9)
≥6	66 (10.9)	57 (11.0)	9 (10.3)
Gender				>0.99	0.009
Male	360 (59.4)	308 (59.3)	52 (59.8)
Female	246 (40.6)	211 (40.7)	35 (40.2)
Pathogen detected				0.002**	0.365
No	208 (34.3)	165 (31.8)	43 (49.4)
Yes	398 (65.7)	354 (68.2)	44 (50.6)
Bacterial infection				0.002**	0.362
No	222 (36.6)	177 (34.1)	45 (51.7)
Yes	384 (63.4)	342 (65.9)	42 (48.3)
Viral infection
No	606 (100.0)	519 (100.0)	87 (100.0)		<0.001
Mycoplasma infection				>0.99	0.139
No	601 (99.2)	514 (99.0)	87 (100.0)
Yes	5 (0.8)	5 (1.0)	0 (0.0)
Mixed infection				0.01*	0.313
No	401 (66.2)	333 (64.2)	68 (78.2)
Yes	205 (33.8)	186 (35.8)	19 (21.8)
Number of pathogens	1.50 [0.00, 3.00]	2.00 [0.00, 4.00]	1.00 [0.00, 2.00]		0.277
Cardiovascular disease				0.02*	0.356
No	575 (94.9)	488 (94.0)	87 (100.0)
Yes	31 (5.1)	31 (6.0)	0 (0.0)
Respiratory disease				0.003**	0.397
No	568 (93.7)	481 (92.7)	87 (100.0)
Yes	38 (6.3)	38 (7.3)	0 (0.0)
Neurological disease				0.057	0.290
No	585 (96.5)	498 (96.0)	87 (100.0)
Yes	21 (3.5)	21 (4.0)	0 (0.0)
Congenital disease				0.09	0.268
No	588 (97.0)	501 (96.5)	87 (100.0)
Yes	18 (3.0)	18 (3.5)	0 (0.0)
Number of comorbidities
0	547 (90.3)	460 (88.6)	87 (100.0)	0.006**	0.506
1	2 (0.3)	2 (0.4)	0 (0.0)
2	3 (0.5)	3 (0.6)	0 (0.0)
3	4 (0.7)	4 (0.8)	0 (0.0)
≥4	50 (8.3)	50 (9.6)	0 (0.0)
Hospital length of stay (days)	13.00 [7.00, 23.00]	14.00 [8.00, 23.50]	12.00 [5.00, 20.00]	0.04*	0.167
Prolonged hospital stay				0.02*	0.262
No	148 (24.4)	118 (22.7)	30 (34.5)
Yes	458 (75.6)	401 (77.3)	57 (65.5)
Prolonged ICU stay				>0.99	0.008
No	96 (15.8)	82 (15.8)	14 (16.1)
Yes	510 (84.2)	437 (84.2)	73 (83.9)
White blood cell count (×109/L)	9.88 [7.33, 12.47]	9.80 [7.22, 12.34]	10.19 [8.80, 15.16]	0.053	0.156
Neutrophil percentage (%)	59.16 [18.85]	58.78 [19.20]	61.39 [16.57]	0.23	0.146
Lymphocyte percentage (%)	32.99 [17.25]	33.27 [17.46]	31.36 [15.93]	0.34	0.114
Hemoglobin (g/dL)	109.56 [17.04]	109.70 [17.58]	108.75 [13.46]	0.63	0.061
Platelet count (×109/L)	328.94 [239.25, 383.23]	328.00 [241.00, 382.87]	340.00 [219.50, 387.75]	0.82	0.017
pH	7.36 [0.08]	7.36 [0.08]	7.34 [0.09]	0.048*	0.220
PO2 (mmHg)	96.20 [52.02]	97.76 [51.86]	86.87 [52.25]	0.07	0.209
PCO2 (mmHg)	44.87 [16.08]	44.96 [16.70]	44.34 [11.80]	0.74	0.043
Lactate (mmol/L)	1.90 [1.42, 2.50]	1.90 [1.40, 2.50]	1.90 [1.70, 2.70]	0.12	0.227
Creatinine (mg/dL)	39.71 [36.00, 45.00]	40.00 [36.00, 45.00]	38.03 [36.52, 44.00]	0.29	0.077
Blood urea nitrogen (mg/dL)	3.31 [2.76, 4.29]	3.29 [2.70, 4.26]	3.48 [3.15, 4.42]	0.02*	0.122
Alanine aminotransferase (U/L)	30.00 [19.00, 48.00]	29.00 [18.00, 46.53]	35.00 [26.00, 50.72]	0.002**	0.079

Demographics, pathogen characteristics, comorbidities, clinical outcomes, and laboratory findings within the first 24 hours of admission are presented. Data are shown as n (%) for categorical variables, mean [SD] for normally distributed continuous variables, or median [IQR] for non-normally distributed continuous variables. P values were calculated using Chi-squared test or Fisher’s exact test for categorical variables, and Student’s t-test or Mann-Whitney U test for continuous variables. P<0.05 indicates statistical significance. *, P<0.05; **, P<0.01. ICU, intensive care unit; IQR, interquartile range; SD, standard deviation; SMD, standardized mean difference.

Prognostic value of blood parameters for hospital mortality

Multivariate Cox regression analysis was performed to evaluate the association between blood parameters and hospital mortality, adjusting for age, gender, bacterial infection status, and pathogen detection. Among single blood indicators, neutrophil percentage demonstrated a significant positive association with mortality risk (adjusted HR: 1.017, 95% CI: 1.004–1.031, P=0.01), indicating that each percentage point increase in neutrophil levels was associated with a 1.7% increased risk of death. Conversely, lymphocyte percentage showed a significant protective effect (adjusted HR: 0.985, 95% CI: 0.970–0.999, P=0.04), with each percentage point increase associated with a 1.5% reduced mortality risk. Other single blood parameters, including white blood cell count (P=0.27), hemoglobin (P=0.75), and platelet count (P=0.51), did not show significant associations with mortality (Figure 2).

Figure 2.

Forest plot of multivariate Cox regression analysis for blood parameters and hospital mortality. Adjusted HR with 95% CIs for single blood indicators and composite inflammatory indices. Models were adjusted for age, gender, bacterial infection status, and pathogen detection. The vertical dashed line indicates HR =1.0. Points to the right indicate increased mortality risk, while points to the left indicate protective effects. *, P<0.05 indicates statistical significance. CI, confidence interval; HR, hazard ratio; HRR, hemoglobin-to-red cell distribution width ratio; MCH, melanin-concentrating hormone; MCHC, mean corpuscular hemoglobin concentration; MCV, mean corpuscular volume; NLPR, neutrophil-to-lymphocyte-platelet ratio; NLR, neutrophil-to-lymphocyte ratio; PLR, platelet-to-lymphocyte ratio; PNR, platelet-to-neutrophil ratio; PRR, platelet-to-red blood cell ratio; PWR, platelet-to-white blood cell ratio; RBC, red blood cell; RDW, red cell distribution width; RPR, red cell distribution width-to-platelet ratio; SII, systemic immune-inflammation index; WBC, white blood count.

Among the twelve composite inflammatory indices evaluated, only the PNR demonstrated a statistically significant association with mortality (adjusted HR: 0.995, 95% CI: 0.989–1.000, P=0.04), suggesting a modest protective effect. The NLR showed a trend toward increased mortality risk but did not reach statistical significance (adjusted HR: 1.394, 95% CI: 0.978–1.988, P=0.07). Other composite indices, including the SII (P=0.19), PLR (P=0.90), and various red cell distribution width-based ratios, failed to show significant prognostic value after multivariate adjustment.

Dose-response relationships between blood parameters and mortality

Restricted cubic spline analyses examined potential non-linear associations between the three significant blood parameters and hospital mortality (Figure 3). For neutrophil percentage, the Cox regression model revealed an inverted U-shaped relationship with mortality risk, though neither the overall association (P=0.08) nor the non-linear component (P=0.12) reached statistical significance. The HR increased progressively above the median value of 59.31%, while the predicted mortality rate showed the lowest risk occurring at 18.70% (predicted mortality: 6.77%). Lymphocyte percentage demonstrated an inverse, nearly linear relationship with mortality risk. The HR showed a consistent protective effect as lymphocyte percentage increased, though the overall association was not statistically significant (P=0.27, non-linear P=0.37). The predicted mortality curve exhibited a gentle downward trend, with the lowest mortality risk observed at 72.91% (predicted mortality: 8.60%). The PNR displayed a predominantly linear inverse association with mortality, with minimal evidence of non-linearity (P=0.94) and the lowest risk at 248.19 (predicted mortality: 6.53%).

Figure 3.

Restricted cubic spline analysis of significant blood parameters. Non-linear associations between blood parameters and mortality risk. (A-C) HRs from Cox regression models for neutrophil percentage, lymphocyte percentage, and PNR, respectively. (D-F) Predicted mortality rates from logistic regression models for the corresponding parameters. The solid red line represents the point estimate with 95% CIs shown in blue shading. The horizontal dashed line indicates the reference value (HR =1.0 or mean mortality rate). Vertical dotted lines indicate median values or lowest risk points. Rug plots show the distribution of observations. P values for overall and non-linear associations are displayed. CI, confidence interval; HR, hazard ratio; PNR, platelet-to-neutrophil ratio.

Subgroup analyses and interaction effects

Stratified analyses evaluated the consistency of associations across various patient subgroups (Figures S1-S3). For neutrophil percentage, significant associations with mortality were observed in the youngest age group (0–3 years: HR 1.016, 95% CI: 1.003–1.030, P=0.02) and among male patients (HR 1.015, 95% CI: 1.001–1.030, P=0.04). However, formal interaction tests revealed no significant effect modification by any subgroup variable (all P>0.35). Lymphocyte percentage showed uniformly protective trends across all subgroups, with the most pronounced effects in the 0–3 years age group (HR 0.987, 95% CI: 0.972–1.001, P=0.07) and male patients (HR 0.986, 95% CI: 0.970–1.002, P=0.08). The PNR demonstrated significant protective associations in the 0-3 years age group (HR 0.995, 95% CI: 0.990–1.000, P=0.04), male patients (HR 0.993, 95% CI: 0.986–0.999, P=0.03), and patients with mixed infections (HR 0.987, 95% CI: 0.974–1.000, P=0.04).

Predictive model development and validation

To develop predictive models for hospital mortality, we incorporated the three blood parameters that demonstrated significant associations in multivariate analysis (neutrophil percentage, lymphocyte percentage, and PNR) along with the clinical covariates used in Cox regression adjustment (age, gender, bacterial infection status, and pathogen detection). Six ML algorithms were implemented incorporating these variables (Figure 4A,4B).

Figure 4.

Performance evaluation of machine learning models and clinical predictors. (A) ROC curves for six machine learning algorithms in the training set. (B) ROC curves for the corresponding models in the test set (internal validation). (C) Comparison of ROC curves between the optimal model and three blood parameters (neutrophil percentage, lymphocyte percentage, and PNR) that demonstrated significant associations in multivariate analysis in the training set. (D) Corresponding comparison in the test set. (E) DCA for the optimal model in the training set. (F) DCA for the optimal model in the test set. AUC, area under the curve; CI, confidence interval; DCA, decision curve analysis; PNR, platelet-to-neutrophil ratio; ROC, receiver operating characteristic.

Among all algorithms evaluated, random forest achieved the highest discrimination performance on both training and test sets. In the test set (internal validation), random forest attained an AUC of 0.877 (95% CI: 0.860–0.895), with an accuracy of 79.5%, sensitivity of 80.9%, specificity of 79.1%, and F1 score of 0.645. Gradient boosting demonstrated the second-best performance with a test AUC of 0.798 (95% CI: 0.776–0.820), followed by XGBoost with an AUC of 0.790 (95% CI: 0.767–0.812). LASSO regression achieved a test AUC of 0.742 (95% CI: 0.716–0.768), while support vector machine (AUC: 0.720, 95% CI: 0.694–0.746) and logistic regression (AUC: 0.715, 95% CI: 0.688–0.741) showed relatively inferior performance. Notably, the random forest model demonstrated robust generalization from the training set (AUC: 0.883) to the test set (AUC: 0.877), indicating minimal overfitting and stable predictive capability on unseen data.

We further compared the predictive performance of the optimal random forest model against the three individual blood parameters identified in multivariate analysis (Figure 4C,4D). The random forest model (test AUC: 0.877, 95% CI: 0.860–0.895) significantly outperformed neutrophil percentage (AUC: 0.712, 95% CI: 0.681–0.749), PNR (AUC: 0.689, 95% CI: 0.652–0.722), and lymphocyte percentage (AUC: 0.660, 95% CI: 0.625–0.697), demonstrating that the integrated ML approach provided superior discrimination compared to any single biomarker.

Decision curve analysis was performed to evaluate the clinical utility of the random forest model across a range of threshold probabilities (Figure 4E,4F). In both training and test sets, the model demonstrated positive net benefit compared to the “treat all” and “treat none” strategies across threshold probabilities ranging from 5% to 50%. In the test set, at a threshold probability of 20%, the model achieved a net benefit of 0.046, substantially exceeding the “treat all” strategy (net benefit: −0.038). This positive net benefit was maintained across clinically relevant threshold ranges, suggesting that the model could provide meaningful decision support in clinical practice for identifying high-risk patients with severe pneumonia. To enhance model interpretability, SHAP analysis was performed for the random forest model (Figure 5). The global feature importance ranking revealed that age was the most influential predictor (mean |SHAP value|: 0.550), followed by PNR (0.312), bacterial infection status (0.215), pathogen detection (0.190), lymphocyte percentage (0.135), neutrophil percentage (0.101), and gender (0.041). The SHAP summary plot demonstrated that higher PNR values showed predominantly negative SHAP values, confirming its protective effect against mortality.

Figure 5.

SHAP analysis for model interpretability. Global feature importance and SHAP summary plot for the random forest model. The left panel displays the mean absolute SHAP values for each feature, indicating their relative contribution to mortality prediction. PNR, platelet-to-neutrophil ratio; SHAP, SHapley Additive exPlanations.

Discussion

In this retrospective cohort study of 606 pediatric pneumonia patients from the PIC database, we identified significant associations between blood-based inflammatory parameters and hospital mortality. The overall mortality rate was 14.4%, with neutrophil percentage (adjusted HR: 1.017, P=0.01) and lymphocyte percentage (adjusted HR: 0.985, P=0.04) emerging as independent predictors of mortality, while the PNR demonstrated marginal significance (adjusted HR: 0.995, P=0.04). Restricted cubic spline analyses revealed predominantly linear relationships between these parameters and mortality risk, with neutrophil percentage showing a J-shaped pattern and lymphocyte percentage demonstrating an inverse association. Subgroup analyses identified consistent effects across most patient characteristics without significant interactions, and the random forest algorithm achieved the highest predictive performance (AUC: 0.877) among six ML models for mortality prediction.

Our finding of elevated neutrophil percentage as a risk factor for mortality aligns with extensive literature demonstrating the prognostic value of neutrophil-mediated inflammation in pediatric pneumonia (10,11). Previous studies have reported similar associations, with Li et al. finding that NLR >1.9 was associated with increased risk of necrotizing pneumonia and refractory disease in children with Mycoplasma pneumoniae pneumonia (10,27). The protective effect of higher lymphocyte percentage observed in our study (HR: 0.985 per percentage increase) is consistent with the established role of lymphocytes in mounting effective immune responses against respiratory pathogens (28). Notably, our mortality rate of 14.4% falls within the range reported in contemporary pediatric intensive care studies, which have documented rates between 10% and 28% depending on disease severity and population characteristics (7,8,29). The absence of significant associations for most composite inflammatory indices, except PNR, contrasts with adult studies where indices like NLR and PLR have shown robust prognostic value, suggesting potential age-related differences in inflammatory responses to pneumonia (13,14).

The dose-response relationships identified through restricted cubic spline analyses provide novel insights into optimal therapeutic targets for pediatric pneumonia management. The U-shaped mortality curve for neutrophil percentage, with the lowest risk at 18.7% (predicted mortality: 6.77%), suggests both insufficient and excessive neutrophil responses may be detrimental (21,29). This finding parallels observations in adult COVID-19 pneumonia, where Udompongpaiboon et al. reported non-linear associations between inflammatory markers and mortality, emphasizing the importance of balanced immune responses (30). The predominantly linear protective effect of lymphocyte percentage up to 72.9% aligns with studies demonstrating that lymphopenia is associated with poor outcomes in pediatric infections, with preserved lymphocyte counts indicating effective adaptive immunity (31,32). These patterns differ from adult pneumonia studies where more pronounced non-linear relationships have been reported, potentially reflecting the distinct immunological characteristics of pediatric populations (33).

Our subgroup analyses revealed important age-specific variations in the prognostic value of blood parameters, with the strongest associations observed in children aged 0–3 years. This finding is particularly relevant given that this age group experiences the highest pneumonia burden globally, accounting for the majority of the 138 million annual cases in children under 5 years (1). The significant effects of neutrophil percentage (HR: 1.016, P=0.02) and PNR (HR: 0.995, P=0.04) in this youngest cohort may reflect the immature immune system’s heightened vulnerability to dysregulated inflammatory responses (33). Gender differences were also notable, with male patients showing stronger associations for both neutrophil percentage and PNR, consistent with established sex-based differences in immune responses and pneumonia outcomes reported in pediatric literature (34). The absence of mortality events in patients with comorbidities was unexpected and likely reflects selection bias or successful risk-adapted management strategies in this high-risk population.

The ML models developed in our study demonstrated varying performance levels, with the random forest algorithm achieving superior discrimination (AUC: 0.877, 95% CI: 0.860–0.895) compared to traditional approaches such as logistic regression (AUC: 0.715) and support vector machine (AUC: 0.720). Notably, the random forest model significantly outperformed individual blood parameters including neutrophil percentage (AUC: 0.712), PNR (AUC: 0.689), and lymphocyte percentage (AUC: 0.660), demonstrating the added value of integrating multiple predictors through ML. Decision curve analysis further confirmed the clinical utility of the model, showing positive net benefit across threshold probabilities ranging from 5% to 50% in both training and validation sets, indicating that the model could provide meaningful decision support for identifying high-risk patients in clinical practice. This performance is comparable to recent pediatric pneumonia prediction models, such as the study by Ye et al. using CatBoost for severe Mycoplasma pneumoniae pneumonia prediction (AUC: 0.934), though their model focused on disease severity rather than mortality (18). Our finding that ensemble methods outperformed traditional statistical approaches aligns with growing evidence supporting ML applications in pediatric critical care, where complex non-linear interactions between clinical and laboratory parameters require sophisticated modeling approaches (17,23). The relatively modest performance of our models compared to some published studies may reflect the heterogeneous nature of our pneumonia cohort, which included various etiologies, whereas studies focusing on specific pathogens often achieve higher predictive accuracy (8,35).

The integration of multiple blood parameters with clinical covariates in our predictive models addresses a critical gap in pediatric pneumonia risk stratification. Previous systematic reviews have highlighted the limitations of single biomarkers, with Gunaratnam et al. reporting that even well-studied markers like CRP and procalcitonin achieve only moderate diagnostic accuracy (sensitivity: 70%, specificity: 64%) when used in isolation (19). Our comprehensive approach incorporating twelve composite inflammatory indices represents an advancement over existing tools, though the absence of significant associations for most composite indices suggests they may not offer substantial advantages over simple blood count percentages in pediatric populations. This finding contrasts with adult studies where composite indices have shown superior prognostic value, potentially reflecting differences in immune system maturity and inflammatory response patterns between children and adults (14,36).

This study has several important strengths that enhance its contribution to the pediatric pneumonia literature. The use of the PIC database, containing comprehensive clinical data from a major tertiary pediatric center with over 1,900 beds and 119 ICU beds, provides a robust foundation for our analyses (25,37). Our systematic evaluation of twelve composite inflammatory indices alongside traditional markers represents the most comprehensive assessment of blood-based prognostic factors in pediatric pneumonia to date. The application of advanced statistical techniques, including restricted cubic splines for dose-response modeling and multiple ML algorithms for prediction, provides methodological rigor often lacking in pediatric pneumonia studies. Furthermore, our focus on readily available blood parameters enhances the clinical applicability of our findings, particularly in resource-limited settings where advanced diagnostic tools may be unavailable. However, several limitations warrant consideration. The retrospective single-center design limits generalizability, as the study population was selected from a tertiary-level pediatric hospital, which means that the patients may differ from those treated in other pediatric hospitals. The absence of a control group of non-critically ill pneumonia patients limits our ability to determine whether the identified biomarkers are specific to mortality prediction or simply reflect disease severity. Additionally, the exclusion of patients with missing data may have introduced selection bias. The relatively small number of mortality events (n=87) constrained statistical power for subgroup analyses and may have contributed to wide CIs for some estimates. Additionally, the lack of detailed microbiological data prevented pathogen-specific analyses, which could have provided additional insights given the established differences in inflammatory responses to bacterial versus viral pneumonia (38-40).

The clinical implications of our findings are substantial for improving pediatric pneumonia management and outcomes. The identification of neutrophil and lymphocyte percentages as simple yet effective prognostic markers provides clinicians with immediately available tools for risk stratification without requiring specialized testing or additional costs. The optimal cutoff values identified through our analyses (neutrophil percentage >59.3%, lymphocyte percentage <33.0%) can guide clinical decision-making regarding intensive monitoring and aggressive therapeutic interventions. Our findings support the implementation of risk-adapted management strategies, particularly for children aged 0–3 years who demonstrated the strongest associations between inflammatory markers and mortality. Future research should focus on validating these findings in multicenter prospective cohorts, investigating the utility of serial measurements for dynamic risk assessment, and developing pathogen-specific prediction models that account for the distinct inflammatory profiles of bacterial, viral, and atypical pneumonia (21,38). Additionally, integration of emerging biomarkers such as cell-free DNA, microRNA profiles, and proteomics signatures may further enhance predictive accuracy (38). The development of clinical decision support systems incorporating our ML models could facilitate real-time risk assessment and personalized treatment strategies, ultimately improving outcomes for this vulnerable population.

Conclusions

This comprehensive analysis of 606 pediatric pneumonia patients demonstrates that simple blood count parameters, particularly neutrophil and lymphocyte percentages, provide valuable prognostic information for hospital mortality risk stratification. The predominantly linear dose-response relationships and age-specific variations identified highlight the importance of considering developmental factors in pediatric pneumonia management. While composite inflammatory indices showed limited additional value over traditional blood counts, the random forest algorithm successfully integrated multiple parameters to achieve good predictive accuracy for mortality (AUC: 0.877), with decision curve analysis confirming its clinical utility across a wide range of threshold probabilities. These findings support the implementation of readily available blood-based markers for early identification of high-risk pediatric pneumonia patients, particularly in the vulnerable 0–3 years age group, and provide a foundation for developing more sophisticated risk prediction tools to guide clinical decision-making and improve outcomes in pediatric intensive care settings.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The PIC database project was approved by the Institutional Review Board of the Children’s Hospital, Zhejiang University School of Medicine (Hangzhou, China) (approval No. 20190114-54). The requirement for individual patient consent was waived because the project did not impact clinical care, and all protected health information was de-identified.