Abstract Body: Background: Hemorrhagic transformation (HT) of ischemic stroke after intravenous thrombolytics is the most feared complication of treatment, occurring in 2-7% of patients. Patients with the most severe hemorrhage have an 18-fold increased risk of 24-hour deterioration and an 11-fold increase in 3-month mortality. Predicting those at higher risk of HT would allow more individualized care, potentially reducing the risk of harm. Utilizing only non-contrast computed tomography (NCCT) of the head, we sought to create a deep learning model to predict those at higher risk of HT from within a large, population-based study.

Methods: Utilizing the Greater Cincinnati/Northern Kentucky Stroke Study, we identified patients in the 2015 study epoch who presented with acute ischemic stroke and received intravenous thrombolytics. Images were obtained, visually checked for quality, brain extracted, aligned to a template, windowed to 10-100 HU, and cropped to maximize the ratio of brain-specific voxels in the image. Data were input into a 101-layer 3D Convolutional Neural Network with Residual Connections (ResNet-101) with a binary output (HT vs no HT) trained on weighted focal binary cross-entropy loss. Stratified 5-fold cross-validation was used to evaluate model performance. Smoothed salience mapping was used to interpret model output.

Results: 194 patients were initially evaluated, with 24 having imaging diagnosed HT. After quality checks, 121 images were used for analysis,17 having HT. Table 1 shows the demographics and clinical data which would be typically available at the time of thrombolytic administration. Cross-validation AUC varied from 0.73 to 0.87, with the full results displayed in Figure 2. Saliency map differences between correctly and incorrectly identified HT suggested the network focused on deep, subcortical structures and ventricles when predicting HT accurately (Figure 1).

Conclusions: We report pilot data of an accurate artificial neural network that correctly identifies HT after thrombolytics utilizing only baseline NCCT. Our results are consistent with a previous study suggesting a higher risk of HT after mechanical thrombectomy in patients with subcortical infarcts. More data is required to refine the algorithm, especially to understand why the ventricles were an area of interest to the model. Further work and improvement of this model has the potential to estimate risk of HT in a more individualized manner and provide decision support.