Machine Learning for Predicting Pediatric Mortality After Cardiopulmonary Bypass: Harnessing the Complex Predictive Power of Accessible Inflammatory Markers

Technical Summary: Machine Learning for Predicting Pediatric Mortality After Cardiopulmonary Bypass

Problem Statement
Children undergoing cardiac surgery with cardiopulmonary bypass (CPB) are at high risk for Systemic Inflammatory Response Syndrome (SIRS), which can progress to Multiple Organ Dysfunction Syndrome (MODS) and death. While specific inflammatory biomarkers (e.g., IL-6, IL-8, TNF-α) are known to predict adverse outcomes, they are often costly and require specialized laboratory facilities, limiting their utility for routine risk stratification. Conversely, standard Complete Blood Count (CBC) parameters and C-reactive protein (CRP) are readily available and low-cost but may not capture complex, non-linear interactions when analyzed via traditional statistical methods. The study aims to evaluate whether an Artificial Intelligence (AI)–based Machine Learning (ML) model can accurately predict 30-day postoperative mortality in pediatric patients (<15 years) using only these accessible inflammatory markers.

Methodology

• Study Design & Cohort: A retrospective observational study was conducted at the Instituto do Coração (University of São Paulo) involving 1,499 pediatric cardiac surgery procedures with CPB performed between March 2020 and August 2025. The cohort included patients under 15 years of age with procedures classifiable by the Risk Adjustment for Congenital Heart Surgery (RACHS-1) score.
• Data Collection: Laboratory data were collected preoperatively and at 12, 24, 36, 48, and 72 hours postoperatively. Features included standard CBC parameters (platelets, hemoglobin, hematocrit, leukocytes, neutrophils, monocytes, immature neutrophils), CRP levels, and derived inflammatory indices (NLR, PLR, MLR, SIRI, AISI, SII, NLPR, dNLR).
• Model Architecture: The authors developed an Ensemble Voting Classifier combining two algorithms:
  1. XGBoost: For learning non-linear interactions.
  2. Random Forest: To improve stability and generalization.
• Handling Class Imbalance: Given the low mortality rate (5.7%), the authors avoided synthetic data generation (e.g., SMOTE) to prevent noise. Instead, they employed cost-sensitive learning, weighting the models to penalize the misclassification of mortality events more heavily than survivors.
• Training & Validation: The dataset was split into an 80% training set and a 20% independent test set. Hyperparameters were optimized via 5-fold stratified cross-validation. Performance was evaluated using bootstrapping (1,000 iterations) to calculate 95% Confidence Intervals (CI) for metrics including ROC AUC, sensitivity, specificity, PPV, and NPV.
• Interpretability: SHapley Additive exPlanations (SHAP) were used to quantify feature contributions and visualize the physiological drivers of the model's predictions.

Key Results

• Predictive Performance: The ensemble model achieved a ROC AUC of 0.927 (95% CI: 0.882–0.966) on the independent test set, demonstrating robust discriminative ability. The model showed excellent calibration (Brier score = 0.040).
• Clinical Thresholds: Optimizing for high sensitivity (≥80%) to minimize missed lethal events, the model achieved:
  • Sensitivity: 79.9% (identifying nearly 80% of mortality events).
  • Negative Predictive Value (NPV): 98.7%, indicating high reliability in identifying low-risk patients who could potentially undergo de-escalation of surveillance.
  • Specificity: 84.5%.
  • Positive Predictive Value (PPV): 21.9%.
• Physiological Drivers (SHAP Analysis): Two distinct phenotypes were identified as primary predictors of mortality:
  1. Thrombocytopenia: Low minimum platelet counts were the strongest discriminator. Non-survivors had significantly lower median minimum platelets (63,000 vs. 112,500/µL; p < 0.001).
  2. "Immune Paralysis" Phenotype: Contrary to the expectation that hyperinflammation drives mortality, non-survivors exhibited significantly lower minimum CRP levels (30.1 vs. 46.9 mg/L; p = 0.027). This suggests that a failure of the acute-phase response (potentially due to hepatic dysfunction or immune exhaustion) is a critical risk factor.
• Feature Importance: While derived inflammatory indices (e.g., NLR) were included, individual hematological parameters—specifically platelet count and CRP kinetics—held greater predictive weight than the composite ratios.

Significance and Claims
The paper claims that machine learning models can accurately predict 30-day mortality in pediatric CPB patients using only low-cost, routinely available laboratory data. The study highlights that:

1. Accessibility: Systematic monitoring of standard inflammatory markers, when analyzed via ML, can serve as an effective screening tool without the need for expensive cytokine assays.
2. Biological Insight: The identification of the "immune paralysis" phenotype (low CRP associated with death) challenges the conventional focus solely on hyperinflammation, suggesting that an inadequate acute-phase response is a critical pathway to adverse outcomes.
3. Clinical Utility: The high NPV of the model supports its potential use in clinical settings to safely identify low-risk patients, potentially allowing for the de-escalation of invasive postoperative surveillance.
4. Methodological Advantage: The study demonstrates that AI can capture complex, non-linear interactions between inflammatory and hematologic variables that traditional biostatistical methods might miss, offering a step toward earlier recognition of postoperative deterioration.

The authors note limitations, including the single-center retrospective nature of the study and the exclusion of specific cytokine data (e.g., IL-6), but assert that their findings align with broader literature regarding the prognostic value of hematologic markers in cardiac surgery.