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

Determining the appropriate end-of-life (EOL) care model within a short time frame is challenging and requires extensive experience. To the best of our knowledge, no studies have developed automatic systems for identifying the hospice care models: hospice home (HHC), inpatient (HIC), and shared care (HSC). This study aimed to determine the optimal hospice care model for EOL patients with machine learning (ML) methods based on health assessment data.

We employed high-performance ML methods to build prediction models that could predict the most appropriate hospice care service for each patient using their health assessment data. Furthermore, we employed the knowledge distillation technique to transfer knowledge from the best-performing ML model to a decision tree model for classification interpretation.

Experiments were conducted on a dataset of 3,468 hospice patients from National Cheng Kung University Hospital (2005-2020). ML models were built and validated, achieving high performance, with a macro-F1 score of 0.88 and an area under the precision-recall curve (AUPRC) of 0.95. In addition, an interpretable decision tree model was generated, which maintained high performance while providing clear, visualizable decision paths for the best hospice care model.

ML models were developed using health assessment data to explore their potential in guiding the selection of hospice care services for end-of-life patients. The findings demonstrate a data-driven approach that may support more informed and personalized clinical decisions, while representing an initial proof of concept for integrating ML into hospice care planning.