Predicting Modified Rankin Scale Scores of Ischemic Stroke Patients Using Radiomics Features and Machine Learning

Stroke affects millions of people worldwide. Because the symptoms of stroke are highly variable, it is not easy to predict clinical outcome. This limits doctors' ability to plan for personalized interventions to enhance recovery. Apparent diffusion coefficient (ADC) maps derived from diffusion-weighted imaging provide valuable information on ischemic lesions and have been shown to correlate with the modified Rankin Scale (mRS) score, a functional outcome score widely used in clinical settings. Here, we aim for developing an expert system to predict mRS scores from ADC maps. We used radiomics analysis to extract salient ADC map features. These features were then used to train a simple binary Naive Bayes classifier which grouped patients into two categories: favorable outcome (mRS < 2), and unfavorable outcome (mRS >= 2). The performance of the system was evaluated using ISLES 2017 dataset including brain scans from 43 ischemic stroke patients. We found that the highest performance was achieved using only 10 radiomics features out of 1132. The performance of the Naive Bayes classifier was comparable to more complex machine learning classifiers such as Support Vector Machine. In addition, we found that the performance of the Naive Bayes classifier dropped by a small margin when the input was reduced to include only the two most prominent features (original_shape_LeastAxisLength - shape feature and wavelet-HHL_firstorder_Skewness - intensity feature with wavelet filtering). These results are encouraging towards building a parsimonious and transparent mRS prediction tool that can be used as a clinical decision support system.