In-Hospital Mortality Prediction among Intensive Care Unit Patients with Acute Ischemic Stroke: A Machine Learning
Jack A Cummins1, Ben S Gerber2, Mayuko Ito Fukunaga3,4,5
1Manchester Essex Regional High School, Manchester, MA 01944, USA.
Health Data Science
|March 18, 2025
Summary
Machine learning models effectively predict in-hospital mortality for acute ischemic stroke patients in the ICU, identifying key risk factors like Glasgow Coma Scale scores. Further ethical considerations are needed to ensure equitable application across diverse patient groups.
Area of Science:
- Computational Medicine
- Artificial Intelligence in Healthcare
- Stroke Research
Background:
- Acute ischemic stroke is a significant cause of mortality in the United States.
- Accurate identification of high-risk patients is essential for timely intervention and resource allocation.
- Predicting in-hospital mortality in intensive care unit (ICU) patients with acute ischemic stroke is critical.
Purpose of the Study:
- To develop and validate machine learning models for predicting in-hospital mortality risk in ICU patients with acute ischemic stroke.
- To identify key factors associated with in-hospital mortality in this patient population.
- To evaluate model performance and fairness across demographic groups.
Main Methods:
- Utilized data from 3,489 acute ischemic stroke ICU admissions from the MIMIC-IV database.
- Extracted demographic, hospitalization, procedure, medication, intake, laboratory, vital signs, and clinical assessment data (e.g., Glasgow Coma Scale Scores) within the initial 48 hours.
- Developed and tuned three machine learning models (random forests, logistic regression, XGBoost) using Bayesian optimization, focusing on predicting mortality after 48 hours.
Main Results:
- The XGBoost model achieved the highest performance with an Area Under the Receiver Operating Characteristic Curve (AUC ROC) of 0.86 and an F1 score of 0.52.
- Key predictors of mortality included Glasgow Coma Scale Scores, blood urea nitrogen, and Richmond Agitation-Sedation Scale scores.
- The developed models demonstrated good fairness across male/female and various racial/ethnic groups.
Conclusions:
- Machine learning models show significant potential for predicting in-hospital mortality in ICU patients with acute ischemic stroke.
- The study highlights the importance of clinical assessment scores and laboratory values in mortality prediction.
- Ethical considerations are crucial to prevent exacerbation of health disparities and ensure equitable outcomes for all patient populations.
