Machine learning-based time-to-event survival analysis in pediatric patients with severe sepsis

Comprehensive Summary

The aim of this study was to use survival analysis machine learning programs to create a time-to-event survival prediction model for pediatric sepsis. The retrospective study evaluates five survival machine learning algorithms: CoxPHSurvivalAnalysis, HingeLossSurvivalSVM, GradientBoostingSurvivalAnalysis, RandomSurvivalForest, and ExtraSurvivalTrees. The data used in this study was obtained from the Paediatric Intensive Care database at Children’s Hospital ZheJiang University School of Medicine between 2010 to 2018; patients under 18 with primary or secondary sepsis and a complete set of data were examined. 223 patients were identified of which 200 survived and 23 did not. Ultimately out of the 233 patients, 67 were used to train the machine learning algorithms and 156 were used to test the algorithms, creating a 7:3 training-testing split. RandomSurvivalForest had the best performance with a mean td-AUC of 0.97, c-index of 0.85 (95% CI: 0.65-0.96), and the best calibration. Due to its high performance RandomSurvivalForest was selected and used in SHAP analysis that identified calcium total and RDW (red cell distribution width) as the two most influential factors regarding pediatric sepsis. The study establishes a U-shaped relationship for creatinine and lymphocytes with regards to the increased mortality risk. Calcium total <1.10 mmol/L, RDW >15.07%, sodium <131.68 mmol/L, and pH <7.32 were also associated with increased mortality risk in patients with pediatric sepsis. Using the model, an online prediction calculator was made which the authors note can assist clinicians with personalizing survival probabilities, improving risk assessments, and informing treatment planning for critically ill pediatric sepsis patients.

Outcomes and Implications

This study is important because sepsis is a significant source of complication and mortality in pediatric intensive care units and predicting its impact on survival, with limited reliability of current biomarkers and high variability of underlying conditions, remains very difficult. While the machine learning models in this study show great promise in aiding clinicians to predict critical pediatric patient prognoses, it is limited by a small sample size, even smaller number of patients who did not survive, and the exclusion of patients who may not have survived past the 24-hour threshold for study inclusion. Regardless, with external validation and larger, independent cohorts, the creation of an accessible and accurate online machine learning model to predict survivability in critically ill pediatric patients with sepsis may be an extremely useful tool for clinical practice in the near future.