Abstract TMP79: Deep Learning Model For Prediction Of Supratentorial Intracerebral Hemorrhage (ich) Expansion

Purpose: Hematoma expansion (HE) is associated with early clinical deterioration, worse long-term outcome, and higher mortality in ICH. Identification of patients at risk of HE may allow targeted anti-expansion therapies in future trials. We aimed to train and validate a deep learning model for prediction of HE based on admission non-contrast head CT. Methods: We utilized the clinical and imaging information from the Antihypertensive Treatment of Acute Cerebral Hemorrhage II (ATACH-2) trial and Yale ICH registry. All patients with spontaneous supratentorial ICH who had baseline head CT (28 were included. The HE was defined as an increase volume >33% or 6 ml. We trained a DenseNet121-based 3D convolutional neural network for prediction of HE (pipeline summarized in the Figure). Results: A total of 749 patients (610 from ATACH-2 and 183 from Yale) were randomly split (4:1) into 634 training/cross-validation (368 (58%) male, mean age of 62.8±13.1 years, median NIHSS 10, and 144 (23%) with HE); and 159 independent test set (92 (58%) male, mean age of 62.5±14.6 years, median NIHSS 12, and 36 (23%) with HE). The average of validation AUC in five-fold training/cross-validation was 0.62, with the best performing model achieving an AUC of 0.71 in validation fold. We then tested this model in the independent validation cohort isolated from training process, achieving an AUC of 0.67 in predicting HE. The attention heatmaps (Figure) confirm that deep learning model predictions were based on head CT regions containing ICH and perilesional edema. Conclusion: We could train, and independently validate a deep learning model for identification of supratentorial ICH patients at risk of impending HE. Compared to visual markers (including CTA spot sign), such model can provide an automated objective tool for future trial enrollment based on readily available non-contrast CT