Predictive performance of machine learning models in acute ischemic stroke: a systematic review and meta-analysis

 I'd fire anybody doing useless prediction research! This doesn't get any survivor recovered, does it?

Predictive performance of machine learning models in acute ischemic stroke: a systematic review and meta-analysis

• 

  Uzma Khanum 1

• 

  Vasudeva Guddattu 2

• S

  Shasthara Paneyala 3

• 

  Asha Srinivasan 4

• 

  Chaithra Nagaraju 1*

• 1. Division of Medical Statistics, School of Life Sciences, JSS Academy of Higher Education and Research, Mysuru, India

• 2. Department of Applied Statistics and Data Science, Manipal Academy of Higher Education, Udupi, India

Abstract

Introduction: 

Acute ischemic stroke (AIS) is a leading cause of global mortality and disability worldwide. Machine learning (ML) models enhance prognostic accuracy by analysing complex, multidimensional clinical data. The aim of this systematic review and meta-analysis is to identify the gaps in the current ML models, along with methodological and performance outcomes in AIS. Further, the study objective was to identify the most frequently used algorithms and compare their relative effectiveness, thereby supporting future research to develop novel ML-based predictive models for stroke care management.

Methods: 

The systematic review followed PRISMA guidelines with PROSPERO registration. A comprehensive search was performed in PubMed, Scopus, and Web of Science using MeSH keywords. Data extraction captured study characteristics, ML algorithms, and outcome metrics. We used the PROBAST and TRIPOD-AI to assess the qualities and bias of included studies. Meta-analysis of AUC values across ML models were conducted to synthesize model performance used a random-effects model to summarize and analyse the data and assessed heterogeneity (I2) statistic using SPSS-29 and R-Studio-4.2.0.

Results: 

A total of 14 studies were included in the systematic review, with 12 eligible for meta-analysis. The pooled AUC of ML models was 0.87 (95% CI, 0.83–0.91), demonstrating strong predictive performance despite substantial heterogeneity (I2 = 99%). Random forest (RF) (AUC = 0.85) and SVM (AUC = 0.82) outperformed logistic regression (LR) (AUC = 0.75), while XGBoost showed stable performance (AUC = 0.82); heterogeneity was mainly driven by study design, publication year, and algorithm type (p < 0.001).

Conclusion: 

ML-based models show potential for improving prognostic assessment in AIS; however, substantial heterogeneity and methodological limitations across studies limit the generalizability of pooled performance estimates.

Systematic review registration: 

https://www.crd.york.ac.uk/PROSPERO/view/CRD420251033217, (Registration number: CRD420251033217).