Toward explainable machine learning methods for stroke patient outcomes in Tennessee

Stroke is one of the leading causes of long-term disability and death in the United States. Stroke patients often face severe health consequences, significantly impacting their lives and placing a substantial financial burden on their families and the wider healthcare system. Therefore, reliable predictions of various patient outcomes, such as early hospital readmission, length of stay (LOS) in the hospital, and risk of mortality, can help patients and healthcare providers in various aspects. Furthermore, successful modeling of such phenomena can help identify the influential factors affecting the patient outcomes, and, by this, improve the quality of care for patients. In this research, we have combined statistical analysis and machine learning (ML) algorithms to enhance the prediction of three patient outcomes — i.e. 30-day readmission, LOS, and mortality — for stroke patients in Tennessee. Since typically such a dataset is imbalanced, due to a small fraction of those events, various ML algorithms, suitable for imbalanced data, such as XGBoost, LightGBM, and CatBoost, were employed in this work. To further improve the performance of the models, various data-level approaches were used to overcome the imbalanced nature of the data. These methods include cluster centroids, NearMiss, and Instant Hardness Threshold. It was shown that such a combination of data modification, especially with under-sampling methods, and suitable ML algorithms can lead to high model performance, measured in terms of Recall and other metrics. Furthermore, based on the features of the data available in our work, using SHAP explainable ML method, the influential factors affecting these outcomes were identified; higher age and mostly the vital signs at the time of admission play an important role in LOS. For 30-day readmission peripheral artery disease, sleep disorders, as well as prescribed medicine such as anticoagulant and antibiotic agents were among the most influential features. For mortality, static patient health conditions were the most influential factors. A simple Graphical User Interface (GUI) was also developed for one of the LOS outcomes, which can be extended to other outcomes, to demonstrate the capability of this work for practical applications.

Grantor dc:publisher

University of Tennessee at Chattanooga

Year dc:date.available

2026

Author dc:creator

• Rahmati, Monireh

Contributors dc:contributor

• Sartipi, Mina
• Fell, Nancy; Gao, Lani; Cho, Jin
• College of Engineering and Computer Science

• Cerebrovascular disease--Patients--Tennessee--Data processing
• Medical informatics
• Medical statistics

Licence

• http://rightsstatements.org/vocab/InC/1.0/

Language dc:language

English, eng

Repository record dc:identifier

https://scholar.utc.edu/theses/1014

OAI identifier oai:identifier

oai:scholar.utc.edu:theses-2198