Machine Learning Methods to Predict the Length of Stay for Acute Stroke: A Scoping Review and Meta-Analysis

Reliable hospital length of stay (LOS) predictions improve patient outcomes and optimize clinical pathways in stroke care. There is limited evidence systematically investigating benefits and limitations of existing machine learning (ML) models for determining factors associated with LOS. We conducted a scoping review to: (1) investigate factors affecting LOS in stroke to identify key predictors for enhancing ML models; and (2) assess the performance of predictive models (ML and traditional models) to generate insights for improving LOS predictions. This review was undertaken in Ovid MEDLINE, Ovid Embase, and Scopus, searching English-language records from 2014 to 2024 using the search terms "stroke" and "length of stay". We included studies investigating associated factors or assessing modeling approaches for predicting LOS in stroke. A meta-analysis using Bayesian methods was conducted to compare predictive performance (e.g., pooled C-statistics) of ML and logistic regression. Of 38 included studies, we identified 38 factors consistently extending LOS (patient characteristics [n = 20], health outcomes [n = 9], social circumstances [n = 2], and clinical care processes [n = 7]). Nine studies (24%) reported predictive models for predicting LOS (continuous or binary outcomes), with stroke severity (particularly the National Institutes of Health Stroke Scale at admission) being the most important predictor across seven studies. For binary outcomes, pooled C-statistic with 95% credible intervals was 0.764 (0.715-0.812) for best-performing ML models and 0.743 (0.692-0.795) for logistic regression across studies (no statistical difference; Bayes factor = 0.783). ML models did not demonstrate superior predictive performance, clinical reliability, and clinical utility over traditional models. Future ML models should prioritize a data-driven selection and transformation of informative inputs, enhance model generalizability and calibration, improve clinical interpretability, and implement a comprehensive model evaluation (including discrimination, calibration, and reclassification metrics). These methodological improvements enable more clinically meaningful LOS predictions to optimize personalized clinical pathways and support precise clinical decision-making at an early stage of stroke care.