Interpretable machine learning for identifying ICU readmission risk in subgroups with probabilistic rules

Area of Science:

• Medical Informatics
• Clinical Decision Support Systems
• Machine Learning in Healthcare

Background:

• Estimating intensive care unit (ICU) patient readmission risk is crucial for resource management and preventing early discharges.
• Current machine learning models often lack interpretability or fail to identify distinct patient subgroups with varying readmission risks.

Purpose of the Study:

• Introduce a novel rule-based model, truly unordered rule sets (TURS), to address the limitations of existing methods.
• Reveal heterogeneous readmission risks and subgroup-level patient characteristics within the ICU population.

Main Methods:

• Trained the TURS model on ICU admissions data from January 2011 to January 2020 at Leiden University Medical Center.
• Analyzed patient characteristics and feature variable influence on readmission risk for each identified subgroup.

Main Results:

• TURS identified patient subgroups with heterogeneous feature distributions and importance, offering actionable insights for ICU discharge planning.
• Achieved a predictive performance (ROC-AUC 70.5%) superior to other rule-based models, with a concise model complexity (5 rules, average length 2).

Conclusions:

• Subgroup analysis revealed significant patient heterogeneity, with distinct feature implications and varying impacts on readmission risk for each group.
• The TURS model provides a concise summary of patient subgroups, supporting ICU discharge decisions and enhancing knowledge discovery.