Interpretable machine learning models for hospital readmission prediction: a two-step extracted regression tree...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Advanced machine learning (ML) models offer high accuracy in medical decision-making but lack interpretability, hindering clinical adoption.
• Interpretable ML tools can extract insights from complex models, yet their application to hospital readmission prediction is underexplored.

Purpose of the Study:

• To develop an ML algorithm for predicting 30- and 90-day hospital readmissions with high accuracy.
• To ensure the ML algorithm provides medically interpretable insights into readmission risk factors.

Main Methods:

• A two-step Extracted Regression Tree approach was employed.
• A black box prediction algorithm was trained first, followed by extracting an interpretable regression tree from its output.
• Data from a large Asian teaching hospital was used for model development and verification.

Main Results:

• The two-step method achieved prediction performance comparable to black box models (e.g., Neural Networks) in terms of accuracy, AUC, and AUPRC.
• Key readmission risk factors identified by the model align with established medical literature, confirming interpretability and validity.

Conclusions:

• The developed two-step approach provides accurate and interpretable predictions for hospital readmissions.
• This method offers a promising way to increase clinician trust and adoption of ML models in healthcare settings for readmission prediction.