Optimization of Cerebral Vasospasm Detection Using a Machine Learning Model with Clinical Features

Introduction: Cerebral vasospasm is a form of delayed cerebral ischemia that can occur after subarachnoid hemorrhage from aneurysm rupture. Given its high morbidity and mortality, it is essential that rapid and accurate diagnosis is made for management to be initiated promptly and appropriately. We know that vasospasm diagnosis requires various inputs including history, physical examination, and imaging findings. CTA is presently the most frequently employed imaging modality to determine vessel patency and caliber. The gold standard however is digital subtraction angiography (DSA) which is resource intensive, time-consuming, and sometimes inaccessible.

Methods: Our group developed an automatic segmentation model using 3D U-net (a convolutional neural network) to detect cerebral vasospasm using CTA input and DSA as the ground truth. However, automatic segmentations can be noisy. In the current work, we propose a second stage clinical model which integrates associative clinical and radiographical features (neurological exam, volume of hematoma, distance of aneurysm etc) on the output predicted segmentation to optimize diagnostic accuracy. We design a model using decision tree-based classifier and test its robustness to noise in the segmentations and feature ablation using manual segmentations.

Results: Our model improves vasospasm detection in each of the above instances. When we add noise to 403 manual segmentations such that the DICE coefficient is 0.70, we achieve optimized F1 scores of 0.93-0.97 and recall of 0.94-0.96. Human vasospasm diagnosis based on literature review is 79.6% sensitivity and 93.1% specificity. When we ablated on clinical features and added noise, we achieve F1 scores of 0.82 - 0.97.

Conclusion : We design a robust clinical model that optimizes our automatic segmentation model. Furthermore, using a decision tree we explain how each feature contributes to the classification network and each of their corresponding contributory weights. By integrating clinically relevant features and leveraging machine learning methods, we improve cerebral vasospasm diagnosis in our cohort.