Network Analysis-Driven Machine Learning Model for Identifying High-Cost Stroke Inpatients Using Hospital Discharge...

Area of Science:

• Medical informatics
• Health services research
• Computational epidemiology

Background:

• Stroke imposes a severe and escalating medical burden.
• A small subset of high-cost patients accounts for a disproportionate share of expenditures.
• Effective cost-risk stratification is essential for optimizing care and resource allocation.

Purpose of the Study:

• To construct a comorbidity network for stroke patients using electronic health record data.
• To extract network features reflecting complex disease interactions.
• To develop machine learning models for identifying patients at high risk of high hospitalization costs.

Main Methods:

• Retrospective study utilizing inpatient stroke data (2021-2023).
• Comorbidity network construction and feature extraction from near-discharge data.
• Development and comparison of 5 machine learning models for high-cost patient identification.
• Application of Shapley Additive Explanations for model interpretability.

Main Results:

• Integrating network features significantly enhanced model performance, with Extreme Gradient Boosting showing the best results.
• Network features constituted a major proportion of global feature importance.
• Shapley Additive Explanations revealed potential phased changes in patient resource consumption.
• Early identification models using only admission data exhibited limitations.

Conclusions:

• An integrated framework of comorbidity network analysis and machine learning accurately identifies stroke patients at high risk of excessive hospitalization costs.
• The core model shows promise for risk stratification near discharge, supporting risk management and resource optimization.
• This study provides a foundation for developing more precise early identification models for high-cost stroke patients.