Accelerating Prediction of Malignant Cerebral Edema after Ischemic Stroke with Automated Image Analysis and Explainable Neural Networks

BACKGROUND/OBJECTIVE: Malignant cerebral edema is a devastating complication of stroke, resulting in deterioration and death if hemicraniectomy is not performed prior to herniation. Current approaches for predicting this relatively rare complication often require advanced imaging and still suffer from suboptimal performance. We performed a pilot study to evaluate whether neural networks incorporating data extracted from routine CT imaging could enhance prediction of edema in a large diverse stroke cohort. METHODS: An automated imaging pipeline retrospectively extracted volumetric data, including CSF volumes and hemispheric CSF volume ratio, from baseline and 24-hour CTs performed in participants of an international stroke cohort study. Fully connected and long short-term memory (LSTM) neural networks were trained using serial clinical and imaging data to predict those who would require hemicraniectomy or die with midline shift. The performance of these models were tested, in comparison with regression models and the EDEMA score, using cross-validation to construct precision-recall curves. RESULTS: Twenty of 598 patients developed malignant edema (12 required surgery, 8 died). The regression model provided 95% recall but only 32% precision (area under precision-recall curve [AUPRC] 0.74), similar to the EDEMA score (precision 28%, AUPRC 0.66). The fully connected network did not perform better (precision 33%, AUPRC 0.71) but the LSTM model provided 100% recall, 87% precision (AUPRC of 0.97) in the overall cohort and the subgroup with NIHSS ≥ 8 (p=0.0001 vs. regression and fully connected models). Features providing the most predictive importance were the hemispheric CSF ratio and NIHSS score measured at 24-hours. CONCLUSION: A LSTM neural network incorporating volumetric data extracted from routine CTs identified all cases of malignant cerebral edema by 24-hours after stroke, with significantly fewer false positives than a fully connected neural network, regression model and the validated EDEMA score. This preliminary work requires prospective validation but provides proof-of-principle that a deep learning framework could assist in selecting patients for surgery, prior to deterioration.