Imaging-based machine learning to evaluate the severity of ischemic stroke in the middle cerebral artery territory

Gang Xie, Jin Gao, Jian Liu, Xuwei Zhou, Zhengkai Zhao, Wuli Tang, Yue Zhang, Lingfeng Zhang, Kang LiBMC Medical Imaging 2025 · DOI

Introduction

In the high-stakes environment of acute ischemic stroke (AIS), rapid and accurate assessment of severity is critical. Currently, the National Institutes of Health Stroke Scale (NIHSS) is the gold standard for this purpose. However, the NIHSS can be subjective and may be confounded by pre-existing neurological conditions such as dementia or prior strokes, which can artificially inflate baseline scores. This study explores the potential of machine learning (ML) to provide a more objective measure of stroke severity by integrating features from initial CT imaging.

Methods and Materials

The researchers conducted a retrospective analysis of 173 patients with AIS in the middle cerebral artery (MCA) territory, the most common site for strokes. Key imaging features were extracted from non-contrast CT and CT Angiography (CTA), including infarction volume, Alberta Stroke Program Early CT Score (ASPECTS), infarction side, basal ganglia involvement, and the presence of a dense MCA sign.

Study workflow including patient selection and model development

The cohort was divided into a training set (n=114, Jan 2020 – Feb 2022) and a test set (n=59, Mar 2022 – Mar 2023) from two different medical centers to ensure generalizability. Seven ML algorithms were evaluated: XGBoost, Random Forest, Decision Tree, Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Gaussian Naive Bayesian (GNB), and Logistic Regression.

Results

While complex models like XGBoost performed well in training, Logistic Regression emerged as the most robust and generalizable model for the final clinical application. The final model achieved an Area Under the Curve (AUC) of 0.815 in the training set and 0.780 in the test set, with an overall accuracy of 81.4% in the test set.

Several key insights were identified during the analysis:

  • The "Left Brain" Penalty: The study confirmed that left-hemisphere strokes often result in higher NIHSS scores, likely due to the dominance of language centers in the left hemisphere for most individuals.
  • Location Over Volume: Basal ganglia involvement was a massive predictor of severity. Because this area is a hub for motor and sensory fibers, even a relatively small infarction here can cause more profound deficits than a larger stroke elsewhere.
  • Clinical Utility: To facilitate bedside use, the authors developed a nomogram that maps CT findings to the probability of a "severe" stroke (defined as NIHSS ≥ 15). Notably, the model's classification of severity was more strongly associated with hemorrhagic transformation—a dangerous complication—than the traditional NIHSS score itself.

Discussion and Limitations

This study demonstrates that ML models can effectively translate objective imaging data into clinical severity scores, potentially bypassing the subjectivity of traditional exams. However, several hurdles remain before widespread adoption. The sample size of 173 patients is relatively small for ML, and the reliance on semi-automatic segmentation for volume calculation is time-consuming. Future work should focus on larger multi-center trials, fully automated deep learning segmentation, and the inclusion of CT Perfusion (CTP) data to better account for the ischemic penumbra.

Summary updated

BibTeX
@article{xie2025imaging,
  author = {Gang Xie and Jin Gao and Jian Liu and Xuwei Zhou and Zhengkai Zhao and Wuli Tang and Yue Zhang and Lingfeng Zhang and Kang Li},
  title = {{Imaging-based machine learning to evaluate the severity of ischemic stroke in the middle cerebral artery territory}},
  journal = {BMC Medical Imaging},
  year = {2025},
  doi = {10.1186/s12880-025-01745-7},
}
Figure