Interpretable machine learning approach for optimizing hospice care predictions using health assessment data.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems

Background:

• Selecting appropriate end-of-life (EOL) care models (hospice home care, inpatient hospice care, shared hospice care) is complex and time-sensitive.
• No automated systems currently exist for identifying optimal hospice care models for EOL patients.
• This study addresses the need for data-driven tools to assist in hospice care model selection.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting the most suitable hospice care model for EOL patients.
• To utilize patient health assessment data for personalized hospice care recommendations.
• To create an interpretable ML model for enhanced clinical understanding and trust.

Main Methods:

• High-performance ML algorithms were employed to build predictive models using patient health assessment data.
• Knowledge distillation was used to transfer insights from a top-performing ML model to a decision tree for interpretability.
• Models were trained and validated on a dataset of 3,468 hospice patients.

Main Results:

• The developed ML models demonstrated high predictive performance, achieving a macro-F1 score of 0.88 and an AUPRC of 0.95.
• An interpretable decision tree model was successfully generated, maintaining high accuracy.
• The decision tree provides clear, visualizable pathways for selecting the best hospice care model.

Conclusions:

• ML models utilizing health assessment data show significant potential for guiding hospice care service selection for EOL patients.
• The findings support a data-driven approach to enhance personalized clinical decision-making in hospice care.
• This study represents a foundational proof of concept for integrating ML into hospice care planning.