Abstract
Machine learning (ML) models based on clinical and non-contrast CT (NCCT) radiomics characteristics can be used for early risk evaluation of hemorrhagic transformation (HT) development in acute ischemic stroke (AIS) patients. This study demonstrates great potential for clinical precision in treatment and prognostic assessment.
Introduction
Acute Ischemic Stroke (AIS) remains a leading cause of disability and death worldwide. A critical complication in its management is Hemorrhagic Transformation (HT)—the leakage of blood into injured brain tissue following the initial stroke. HT significantly worsens patient prognosis and complicates decisions regarding reperfusion therapies like intravenous thrombolysis or mechanical thrombectomy.
While clinical tools such as the National Institutes of Health Stroke Scale (NIHSS) and standard non-contrast CT (NCCT) scans are used for risk assessment, they often fail to detect subtle, microscopic changes in brain tissue. This study, published in Frontiers in Neurology, investigates whether machine learning can leverage "radiomics"—the extraction of high-dimensional quantitative features from medical images—to provide an early warning system for HT using standard NCCT scans.
Methods and Materials
Patient Cohort and Study Design
This retrospective study analyzed 180 AIS patients from two medical centers in China between January 2019 and October 2023. The cohort was divided into a training set (n=144) and a validation set (n=36). Of the total patients, 104 developed HT.
Radiomics Pipeline
The researchers employed a multi-step radiomics workflow:
- Image Segmentation: Radiologists manually delineated the 3D volume of interest (ROI) representing the infarcted brain tissue on NCCT scans.
- Feature Extraction: 851 quantitative features were extracted using 3D-Slicer, including first-order statistics, shape features, and complex texture matrices (e.g., GLDM, GLRLM).
- Feature Selection: LASSO regression was used to identify the six most predictive features, minimizing overfitting.
- Model Development: Five algorithms (KNN, SVM, Decision Tree, XGB, and Logistic Regression) were compared. Three model types were constructed: Clinical, Radiomics, and a combined Clinical-Radiomics model.
Results
The combined Clinical-Radiomics model emerged as the superior predictor, achieving an Area Under the Curve (AUC) of 0.957 in the validation cohort with an accuracy of 86.1%. Notably, the model demonstrated high sensitivity (0.950), which is vital for clinical screening to ensure high-risk patients are not missed.
The primary predictors of HT identified were:
- Neutrophil-to-Lymphocyte Ratio (NLR): A systemic marker of inflammation.
- Baseline NIHSS Score: Reflecting stroke severity.
- Infarct Volume: Larger lesions correlate with higher bleeding risk.
- ASPECTS: A lower score indicates more extensive early ischemic changes.
- Rad-score: The calculated signature from the selected radiomics features.
Discussion
A key finding of this study is the "invisible" signal captured by radiomics. Features like the Gray-Level Dependence Matrix (GLDM) reflect microscopic blood-brain barrier disruptions that are not visible to the naked eye on standard CT. Furthermore, the study found that Logistic Regression—a simpler, more interpretable model—outperformed complex algorithms like XGBoost, suggesting robust linear relationships in this clinical context.
The synergy between systemic clinical markers (like NLR) and local tissue characteristics (radiomics) provides a holistic risk profile. Importantly, the model's performance remained consistent regardless of whether patients received reperfusion therapy, suggesting broad clinical applicability.
Conclusion
This study demonstrates that a combined clinical-radiomics model using standard NCCT scans can effectively predict HT in AIS patients. While promising, the study is limited by its retrospective nature and manual segmentation process. Future work should focus on prospective validation and the integration of automated AI-driven segmentation to facilitate real-time clinical use in emergency settings.
Summary updated