Combining Transcranial Doppler Ultrasonography and Machine Learning for Early Detection of Cerebral Vasospasm in Aneurysmal Subarachnoid Hemorrhage

Introduction: Cerebral vasospasm (CV) is a serious complication following aneurysmal subarachnoid hemorrhage (aSAH), typically occurring 3-14 days post-hemorrhage. Current diagnostic methods like CT and MR angiography are costly, involve radiation exposure, and present logistical challenges. Transcranial Doppler ultrasonography (TCD) offers a non-invasive alternative, though it is limited by interrater variability and anatomical constraints. This study explores the combination of TCD with machine learning (ML) to enhance CV detection and diagnostic accuracy.

Methods: TCD data from 25 aSAH patients (3,287 cleaned samples) were examined. Key features, including middle cerebral artery (MCA) velocities and inter-arterial velocity differences, were extracted. Machine learning models were trained to predict CV, utilizing class balancing techniques to address data imbalance. Model performance was compared against clinical data.

Results: Initial analysis of TCD data from 25 aSAH patients yielded 9,123 measurements, with 70 indicating vasospasm, revealing a significant class imbalance (0.77% vasospasm cases). After cleaning, 3,287 samples remained for model training. Feature importance analysis showed that the velocity difference between the MCA and posterior cerebral artery (PCA) was the most predictive feature, followed by MCA velocity, PCA velocity, and the rolling average of internal carotid artery (ICA) velocities, indicating that multi-vessel flow dynamics are critical for predicting vasospasm. The ML model achieved high accuracy with perfect classification; however, this may reflect overfitting due to the class imbalance. Principal component analysis (PCA) demonstrated distinct clustering patterns for vasospasm risk, confirming the relevance of the identified features.

Conclusion : These preliminary findings suggest that integrating the full TCD dataset with ML could enhance non-invasive CV monitoring in aSAH patients, potentially decreasing reliance on invasive imaging. Further research is required to address data imbalance and validate results, but this approach may provide a scalable, efficient tool for CV detection, facilitating timely interventions in neurocritical care.