Early Prediction of Ischaemic Stroke Outcomes: A Novel Computational Approach

Abstract

Malignant stroke can lead to a death rate as high as 80%. Although early interventions can improve patient outcomes, they also lead to side effects. Therefore, the early prediction of stroke outcomes is crucial for clinical strategies. Imaging markers such as brain swelling volume and midline shift have been critical predictors in various stroke scoring systems. However, these markers can only become visible on brain images days after stroke onset, which delays clinical decisions. A primary challenge in predicting these markers is that brain swelling is a biomechanical process that relies on anatomical features, such as lesion size and location. To tackle this problem, we propose a novel computational pipeline to predict brain swelling after stroke. We first provide a mathematical model of the brain by using a five-compartment poroelastic theory. It allows us to generate high-quality stroke cases with varied 3D brain and lesion anatomy, which are then used to train and validate a deep neural network (DNN). Our in silico experimentation with 3,000 cases shows that anatomical features of stroke brains are well-learned by the DNN, with minimal errors in brain swelling prediction found in the hold-out testing cases. In addition, we used the DNN to process clinical imaging data of 60 stroke patients. The results show that the markers generated from the DNN can predict 3-month stroke outcomes with an AUC of around 0.7. It indicates that the proposed computational pipeline can potentially advance the time point for clinical decisions. Significance StatementStroke is the second leading cause of death in the world, and malignant stroke can lead to a death rate of 80%. Early interventions can improve patient outcomes but can also cause side effects. Therefore, it is crucial to predict stroke outcomes at an early stage. Radiological markers such as brain swelling volume and midline shift have been used in various stroke scoring systems. However, these markers can only become visible after days to stroke onset, which delays clinical decisions. To tackle this issue, we propose a novel computational pipeline to predict brain swelling after stroke onset. The proposed pipeline is found to predict brain swelling accurately and can potentially assist early clinical decision-making.