The Price of Explainability in Machine Learning Models for 100-Day Readmission Prediction in Heart Failure:...

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Clinical Decision Support

Background:

• Machine learning (ML) models can aid clinicians in managing heart failure (HF) patients post-discharge.
• Identifying factors for high readmission risk is crucial for effective HF patient management.

Purpose of the Study:

• Compare deep learning (DL) and traditional ML models for predicting 100-day HF readmissions.
• Assess the trade-offs between DL and traditional ML in terms of predictive performance and explainability.
• Provide global and local explanations for model predictions to highlight key risk factors.

Main Methods:

• Retrospective cohort study using data from Region Halland, Sweden (2017-2019).
• Developed and validated decision tree (traditional ML) and recurrent neural network (DL) models to predict 100-day HF readmissions.
• Utilized an ML explainer for model interpretability and compared performance against existing risk assessment tools.

Main Results:

• The study included 15,612 admissions, with a 35.85% readmission rate.
• A traditional, explainable ML model performed comparably to the DL model (AUC 68% vs. 66%) and outperformed conventional scoring methods.
• The explainable model provided actionable insights for enhanced care planning.

Conclusions:

• Explainable ML models can achieve performance comparable to deep learning models in predicting HF readmissions.
• Model transparency does not necessarily compromise predictive performance.
• Explainable models have the potential to facilitate clinical adoption and improve patient care planning.