Related Experiment Videos
Machine Learning Methods to Predict the Length of Stay for Acute Stroke: A Scoping Review and Meta-Analysis
Zhenran Xu1,2, Monique F Kilkenny1,2, Tzu-Yung Kuo1,2
1Stroke and Ageing Research, Department of Medicine, School of Clinical Sciences at Monash Health Monash University Melbourne Australia.
Health Care Science
|July 25, 2026
Summary
Predicting hospital length of stay in stroke care is crucial. Machine learning models do not currently outperform traditional methods in predicting stroke length of stay, highlighting areas for future improvement.
Area of Science:
- Neurology
- Medical Informatics
- Health Services Research
Background:
- Accurate hospital length of stay (LOS) predictions are vital for optimizing stroke care pathways and patient outcomes.
- Limited systematic evidence exists on the benefits and limitations of machine learning (ML) models for predicting stroke LOS.
- Identifying key predictors of LOS is essential for enhancing predictive models in stroke management.
Purpose of the Study:
- To conduct a scoping review to identify factors influencing stroke length of stay (LOS).
- To investigate and identify key predictors for enhancing ML models for LOS prediction.
- To assess the performance of ML and traditional predictive models for stroke LOS to inform future improvements.
Main Methods:
- A comprehensive scoping review was performed across Ovid MEDLINE, Ovid Embase, and Scopus databases.
- Searched for English-language records from 2014-2024 using terms "stroke" and "length of stay".
- Included studies investigating LOS factors or predictive modeling approaches; meta-analysis using Bayesian methods compared ML and logistic regression performance.
Main Results:
- Identified 38 factors consistently extending LOS, categorized into patient characteristics, health outcomes, social circumstances, and clinical care processes.
- Stroke severity, particularly the National Institutes of Health Stroke Scale score, was the most significant predictor across studies.
- Pooled C-statistics showed no significant difference between best ML models (0.764) and logistic regression (0.743) for binary LOS prediction.
Conclusions:
- Machine learning models currently do not demonstrate superior predictive performance, clinical reliability, or utility over traditional models for stroke LOS.
- Future ML models require improved data-driven input selection, enhanced generalizability and calibration, and better clinical interpretability.
- Methodological advancements are needed to enable more clinically meaningful LOS predictions for personalized stroke care and decision-making.