Differential Diagnostic Value of Machine Learning–Based Models for Embolic Stroke

Abstract

Cancer-associated thrombosis (CAT) and atrial fibrillation (AF)-related stroke are two subtypes of acute embolic stroke with distinct lesion patterns on diffusion weighted imaging (DWI). This pilot study aimed to evaluate the feasibility and performance of DWI-based machine learning models for differentiating between CAT and AF-related stroke. Patients with CAT and AF-related stroke were enrolled. In this pilot study with a small sample size, DWI images were augmented by flipping and/or contrast shifting to build convolutional neural network (CNN) predicative models. DWI images from 29 patients, including 9 patients with CAT and 20 with AF-related stroke, were analyzed. Training and testing accuracies of the DWI-based CNN model were 87.1% and 78.6%, respectively. Training and testing accuracies were 95.2% and 85.7%, respectively, for the second CNN model that combined DWI images with demographic/clinical characteristics. There were no significant differences in sensitivity, specificity, accuracy, and AUC between two CNN models (all P = n.s.). The DWI-based CNN model using data augmentation may be useful for differentiating CAT from AF-related stroke.