For Heart failure (HF) in pediatric populations represents a significant global health challenge, contributing substantially to mortality rates among children under five years of age worldwide (1). In young children, the most common underlying etiologies of HF include congenital heart disease and cardiomyopathy (2). Although the overall incidence of pediatric HF is relatively low—estimated between 0.9 and 7.4 cases per 100,000 children annually—the condition carries a markedly high morbidity and mortality burden. Reported in-hospital mortality rates among pediatric HF patients range from 7% to as high as 26%, particularly in younger children or those with complex comorbidities (3, 4). In the United States alone, more than 14,000 pediatric hospitalizations annually are attributed to heart failure, highlighting its substantial clinical impact relative to its low incidence (5). While the absolute burden of HF in children is lower than that observed in adult populations, affected pediatric patients often demonstrate greater severity of illness. Children with HF demonstrate significantly higher resource utilization—including ICU admissions, longer hospital stays, and mechanical circulatory support—than adults with HF. Moreover, mortality rates for children with HF in emergency departments and inpatient settings are frequently higher than those reported for adults, reflecting both the clinical complexity and fragility of the pediatric HF population (6).

These findings highlight the urgent need for improved risk stratification tools tailored to the pediatric HF context, particularly for children under five who are at heightened risk of rapid deterioration. Nonetheless, widely used pediatric risk scores such as PRISM-III and PIM-2 offer limited prognostic insight, as they are not specifically designed for HF (7, 8). This further reinforces the necessity for risk assessment tools specifically tailored to pediatric HF. In recent years, advancements in artificial intelligence have expanded the application of machine learning (ML) in clinical research. ML techniques excel at analyzing complex datasets, enhancing disease diagnosis, prognostication, and treatment outcome prediction. Unlike conventional statistical methods, ML algorithms can decipher intricate nonlinear relationships and uncover previously unrecognized associations, thereby identifying prognostic patterns that may be overlooked by traditional scoring systems (9).

In cardiovascular medicine, ML algorithms have demonstrated success in prognostic prediction for adult populations with HF and other cardiac conditions (10–12). However, research on ML-based risk modeling for pediatric HF remains scarce, with existing models often limited by poor interpretability and generalizability. To address this gap, this study aims to develop and validate an optimized ML-based mortality risk prediction model using a comprehensive dataset to predict in-hospital mortality among children under five years of age with HF. Furthermore, SHapley Additive exPlanations (SHAP) will be employed to improve interpretability by quantifying the contribution of individual clinical features to mortality risk. This dual approach aims to offer a robust, interpretable framework to support clinical decision-making and personalized care in this vulnerable population.

In this study, we developed an XGB machine learning model to predict in-hospital mortality among children under five years old with HF, and demonstrated robust performance across both internal and external validation cohorts. The model achieved an AUC of 0.916 in the training set and 0.851 in the internal validation set, significantly outperforming other commonly used models. Notably, performance remained stable in an external validation cohort (AUC = 0.846), indicating strong generalizability and the ability to effectively discriminate between survivors and non-survivors in this high-risk population. This level of accuracy is comparable to, or exceeds, that reported in previous pediatric mortality prediction studies. For instance, Du et al. reported an XGB-based model that achieved a sensitivity of 78.5% and specificity of 82.4% for predicting postoperative mortality in children with congenital heart disease. Our model demonstrated similar performance, suggesting that it is capable of identifying the majority of fatal cases while minimizing false positives among survivors (18). Such performance represents a notable improvement over traditional logistic regression-based tools (7) and highlights the added value of machine learning in integrating and analyzing complex pediatric datasets.

Beyond overall accuracy, an important finding lies in the model's identification of top predictive features: notably NT-proBNP, blood pH, PCT, LDH, WBC, creatinine, and PLT. These variables reflect well-established pathophysiological mechanisms associated with critical illness and offer insight into risk stratification in young HF patients. NT-proBNP emerged as the most influential predictor—a result aligned with extensive evidence of its prognostic utility in both adult and pediatric heart failure (19). The American Heart Association has emphasized NT-proBNP as a key biomarker for assessing severity and prognosis in pediatric HF (20). In a study by Chowdhury et al., an NT-proBNP level ≥520.2 pg/ml predicted moderate-to-severe HF (≥ class II) with 83% sensitivity and 91% specificity. Median NT-proBNP in non-survivors (11,681.01 pg/ml) was significantly higher than in survivors (839.4 pg/ml, p < 0.001) (21). A recent meta-analysis further confirmed the association between elevated NT-proBNP and increased mortality risk, reporting a hazard ratio of 1.65 (95% CI: 1.55–1.76) (22). The strong contribution of NT-proBNP in our model reinforces its clinical relevance. A child presenting with markedly elevated levels should be regarded as high-risk and considered for early intensive management or ICU monitoring. Likewise, blood pH was identified as a top feature, indicating the prognostic relevance of metabolic acidosis. A recent review highlighted compelling evidence linking critically low arterial pH (mean 6.15) to sudden infant death, underscoring its role in reflecting severe physiological instability and end-organ failure (23).

Inflammatory and tissue injury markers also featured prominently. PCT, a well-known biomarker of bacterial infection and sepsis, is frequently elevated during systemic inflammatory responses (24). Its inclusion in our model suggests that infectious complications or systemic inflammation may be common precipitants of decompensation in pediatric HF. Elevated PCT levels on admission may help identify patients in early sepsis or with infections exacerbating cardiac dysfunction. Prior studies have shown that high PCT is associated with increased risk of organ dysfunction and mortality in critically ill pediatric populations (25, 26). Similarly, LDH—a nonspecific enzyme released during tissue breakdown or hypoxia—was identified as a strong predictor. LDH often increases in conditions involving multi-organ injury, hemolysis, or hepatic congestion, all of which are common in advanced heart failure (27, 28). In a cohort of 4,343 critically ill children, Wang et al. found that LDH had the highest predictive accuracy for in-hospital mortality (AUC = 0.729) and remained independently associated with death after adjusting for age and organ dysfunction (OR = 2.45, 95% CI: 1.84–3.24) (29). Furthermore, LDH was significantly associated with 30-day mortality and ICU length of stay, surpassing traditional cardiac biomarkers in predictive performance. The combination of low pH and elevated LDH effectively captured the clinical profile of systemic shock and widespread cellular injury. This represents an especially ominous pattern in children with underlying HF.

Creatinine elevation reflects impaired renal function, which in HF patients may result from low cardiac output or venous congestion (30). A landmark meta-analysis of 16 studies involving over 80,000 HF patients reported a 15% increased risk of mortality per 0.5 mg/dl increase in creatinine, with more than a twofold mortality risk in patients with moderate-to-severe renal dysfunction (creatinine ≥1.5 mg/dl) (31). Our findings confirm that elevated creatinine in children is associated with increased mortality risk, likely reflecting severe hemodynamic compromise, renal hypoperfusion, or concurrent nephrotoxic injury.

Elevated WBC count is commonly linked to systemic inflammation or infection, which are frequent precipitants of decompensation in pediatric heart failure. In addition to WBC, platelet count has shown independent prognostic significance. A large retrospective analysis of the MIMIC-IV database revealed that thrombocytopenia was strongly associated with 28-day mortality in sepsis patients, with a nearly twofold increase in risk for those with platelet counts below 50 × 10⁹/L (32). Notably, WBC and platelet indices can be jointly interpreted as markers of the host hematologic response to critical illness. The platelet-to-white cell ratio has been shown to outperform other complete blood count–derived indices in predicting mortality in several acute inflammatory diseases, including acute heart failure. Lower platelet-to-white cell ratio, reflecting elevated WBC and/or thrombocytopenia, was significantly associated with short-term mortality and in some cohorts exceeded age in prognostic relevance (33). Collectively, these findings illustrate that our model not only identifies clinically relevant biomarkers associated with pediatric HF mortality but also mirrors established physiological mechanisms, reinforcing its potential as a clinically meaningful tool.

A key strength of our study was the use of SHAP analysis to enhance transparency and bridge the gap between algorithmic predictions and clinical reasoning. A common criticism of machine learning in clinical contexts is the “black box” nature of many algorithms, which can limit clinician trust and uptake (34). By applying SHAP, we were able to deconstruct each prediction into individual feature contributions, thereby improving transparency in decision-making (35). For example, elevated NT-proBNP and low pH may emerge as dominant contributors to a high-risk classification, whereas normal WBC count may offset risk. This level of interpretability ensures that the model's predictions are aligned with established clinical reasoning, increasing its potential for integration into real-world practice.

When embedded within hospital systems, interpretable ML models like ours could actively support real-time decision-making by alerting clinicians to high-risk pediatric HF patients and guiding timely intervention. In resource-limited environments, ML-powered triage systems could help allocate scarce critical care resources—such as ICU beds or surgical capacity—to those pediatric patients most likely to benefit (36, 37). At a broader level, such tools may support global health equity by addressing preventable HF-related mortality among children under five. Given that our model relies solely on routinely collected laboratory and clinical variables, it can be integrated into electronic health record (EHR) systems to generate automated risk scores at the point of admission. Such implementation could allow frontline providers to prioritize high-risk patients early and streamline escalation-of-care decisions. Our model thus functions not only as an early warning system, but also as a strategic tool for informing both clinical practice and policy development aimed at achieving Sustainable Development Goal 3.2 (38).

Although the model performed well, certain methodological limitations remain. Boruta may underrepresent weak predictors in imbalanced data, and early stopping in XGB, while helpful, does not fully eliminate overfitting risk given the low event rate.

Despite the promising results, our study has several important limitations. First, it was retrospective in nature, relying on observational medical record data. This introduces potential for missing data, documentation errors, and residual confounding, and limits causal inference. Second, although we performed external validation, the size of the external cohort was relatively small. Larger and more heterogeneous validation cohorts—across multiple institutions and geographic regions—are needed to confirm generalizability. Further multi-center validation studies are currently being planned. Third, the presence of unmeasured confounders cannot be excluded. Variables such as nutritional status, pre-admission medication history, and timing of hospital presentation may influence disease severity, treatment decisions, and outcomes, potentially affecting model performance. Lastly, only basic echocardiographic parameters (e.g., ejection fraction, chamber size) were consistently available, while advanced imaging metrics (e.g., strain imaging, tissue Doppler, cardiac MRI) were largely missing and thus excluded, which may limit the model's physiological depth.

This study developed and externally validated an XGB-based model to predict in-hospital mortality in children under five with heart failure, achieving high accuracy and generalizability. By identifying clinically meaningful predictors and incorporating SHAP analysis, the model offers both predictive performance and interpretability. These findings support its potential as a clinical decision-support tool for early risk stratification, guiding interventions and resource allocation. Further prospective studies are warranted to confirm its utility across diverse healthcare settings.