Abstract
The machine learning (ML) models based on multi-modal MRI radiomics are of high value for predicting clinical outcomes in acute stroke patients. This study demonstrates that combining quantitative features from multiple MRI sequences significantly enhances the prediction of long-term recovery.
Introduction
Predicting the recovery trajectory of ischemic stroke patients is a complex clinical challenge. Traditional clinical scales, such as the National Institutes of Health Stroke Scale (NIHSS), provide essential neurological assessments but often fail to capture the underlying physiological nuances and potential for neuroplasticity. Furthermore, stroke-related psychiatric comorbidities like depression and anxiety complicate the prognosis.
Standard radiological inspection may overlook subtle "texture" patterns in brain tissue that indicate the severity of damage. This research addresses these limitations by leveraging radiomics—the high-throughput extraction of quantitative features from medical images—to develop objective, data-driven predictive models for stroke prognosis.
Methods
The study utilized a retrospective analysis of 148 patients, following a rigorous "imaging-to-insight" pipeline.
Multi-Modal Imaging and Segmentation
The researchers integrated five distinct MRI modalities to provide a comprehensive view of the infarct:
- DWI (Diffusion-Weighted Imaging)
- ADC (Apparent Diffusion Coefficient)
- FLAIR (Fluid-Attenuated Inversion Recovery)
- SWI (Susceptibility-Weighted Imaging)
- T1w (T1-weighted structural imaging)
Lesions were manually segmented on DWI images, which then served as a mask for the other co-registered modalities.
Feature Extraction and Selection
Using the pyradiomics library, 4,744 features were extracted per patient, including first-order statistics and texture features (e.g., GLCM). To prevent overfitting, Recursive Feature Elimination (RFE) was applied, selecting the 16 most impactful variables (3 per modality plus 1 shape feature).
Machine Learning Pipeline
Five models were evaluated using five-fold cross-validation: Support Vector Machine (SVM), Random Forest (RF), LightGBM, CatBoost, and XGBoost.
Datasets
Data was sourced from Liangxiang Hospital in Beijing, China (Oct 2020 – May 2022).
- Cohort: 148 patients with anterior circulation artery occlusion.
- Outcome Measure: The Modified Rankin Scale (mRS) at follow-up. "Good Prognosis" was defined as mRS ≤ 2 (n=83), and "Poor Prognosis" as mRS > 2 (n=65).
Results
The LightGBM model demonstrated superior performance in binary classification:
- AUC: 0.902
- Accuracy: 0.831
- Specificity: 0.902
- Sensitivity: 0.739
Clinical analysis revealed that patients with a poor prognosis were typically older (mean age 70.8 vs. 59.2) and presented with higher initial NIHSS scores (9.75 vs. 2.33).
Discussion
The integration of multi-modal MRI sequences allows for a holistic "fingerprint" of the stroke, capturing physiological aspects like oxygen extraction (SWI) and acute infarct size (DWI). Many selected features were mathematical abstractions (e.g., wavelet transforms) that capture sub-visual patterns indicative of tissue health.
Limitations
The study is limited by its relatively small sample size (n=148) and the use of manual segmentation, which is time-intensive and potentially subjective. Future work should focus on multi-center validation and the implementation of automated AI segmentation to improve scalability and objectivity. Additionally, accounting for stroke etiology (e.g., cardioembolic vs. atherosclerotic) could further refine the predictive accuracy of radiomic features.
Summary updated