Explainable machine learning models for predicting 30-day readmission in pediatric pulmonary hypertension: A

Minjie Duan1,2, Tingting Shu3, Binyi Zhao4

  • 1College of Medical Informatics, Chongqing Medical University, Chongqing, China.

Insights

A new CatBoost model effectively predicts 30-day readmission risk in pediatric pulmonary hypertension (PH) patients. Key predictors include age, length of stay, congenital heart surgery, and discharge type, aiding personalized care for children with PH.

Area of Science:

  • Pediatric Cardiology
  • Medical Informatics
  • Machine Learning in Healthcare

Background:

  • Pediatric pulmonary hypertension (PH) readmissions pose significant burdens.
  • Current tools for predicting individualized readmission risk are insufficient.
  • Developing predictive models is crucial for managing pediatric PH patients.

Purpose of the Study:

  • To develop and validate machine learning models for predicting 30-day unplanned readmission in pediatric PH patients.
  • To identify key clinical factors influencing readmission risk.

Main Methods:

  • Utilized data from 5,913 pediatric PH inpatients (2012-2019).
  • Employed least absolute shrinkage and selection operator for variable selection.
  • Evaluated 15 machine learning algorithms, selecting the best performing model based on AUC.
  • Interpreted model outcomes using SHapley Additive exPlanations (SHAP).

Main Results:

  • The CatBoost model achieved an AUC of 0.81, with 74% accuracy, 78% sensitivity, and 74% specificity.
  • Age, length of stay (LOS), congenital heart surgery, and nonmedical order discharge were identified as significant predictors.
  • The CatBoost model demonstrated superior performance compared to traditional logistic regression.

Conclusions:

  • A validated CatBoost model can accurately predict 30-day readmission risk in pediatric PH.
  • Identifying high-risk patients allows for targeted interventions and improved outcomes.
  • Clinical factors like age, LOS, and specific surgical/discharge types are critical for risk stratification.
Abstract