Abstract 101: Predicting Hematoma Expansion Using An Artificial Neural Network: An Exploratory Analysis Of The Atach 2 Trial

Abstract

Introduction: Hematoma expansion (HE) in patients with intracerebral hemorrhage (ICH) is a key predictor of poor prognosis and potentially amenable to treatment. Current clinical scoring models, radiological markers, and labs have limited ability to predict HE. This study aimed to create a more robust classification model to predict HE in patients with ICH using a deep learning algorithm, an artificial neural network (ANN). Methods: Data from the ATACH 2 trial (Antihypertensive Treatment of Acute Cerebral Hemorrhage) was utilized. Variables included in the models were chosen as per literature consensus on salient factors associated with HE and clinical expertise. HE was defined as increase in either >33% or 6mL in hematoma volume in the first 24 hours. A multi-layer feedforward ANN was trained using the back-propagation method to minimize the loss function with 5-fold cross validation. 80% of patients were used for training and 20% for testing. The ANN was compared to a logistic regression model (GLM). Given the imbalance in patients with HE, AUPRC (area under the precision-recall curve), recall and precision were calculated for the respective models. Results: Of the 963 patients in the study (mean age 62±13.0, 38.5% female), 31% had hematoma expansion. The median [interquartile range (IQR)] initial hematoma volume was 10.5mm 3 [5.16 - 20.25], median admission SBP [IQR] was 200 [184-217], median platelet count [IQR] was 213 [178-256], median INR [IQR] was 1.0 [0.9-1.0], and median admission GCS [IQR] was 15 [13-15]. The GLM model had AUPRC of 0.38 with recall of 0.60 and precision of 0.38 in the testing cohort. The ANN training model demonstrated AUPRC of 0.92 with recall of 0.72 with precision of 0.39 during testing. Initial ICH volume, platelets, INR, and GCS had the highest feature importance. Conclusion: We developed an ANN to predict HE in ICH patients performing with improved sensitivity and similar positive predictive value when compared to GLM. We support that ANN can capture complex associations between variables not attainable in classic regression models. This model may help identify patients at risk for HE who warrant careful monitoring and aggressive treatment upfront including those suitable for clinical trials for treatments.