Clinical Outcome Prediction Pipeline for Ischemic Stroke Patients Using Radiomics Features and Machine Learning

Ischemic stroke is a debilitating brain injury affecting millions of people, causing long-term disabilities. Immediately after stroke, it is not easy to predict the extent of the injury and its long-term effects, yet outcome prediction is desired to inform clinical decision-making processes. Apparent diffusion coefficient (ADC) maps, calculated from diffusion-weighted imaging, are widely used in clinics to diagnose and monitor ischemic stroke. Radiomics analysis is an emerging feature extraction method providing many quantitative imaging indicators from the ADC maps. Here, we have utilized these features to predict the clinical outcome of 43 ischemic stroke patients. We divided the clinical outcome into two groups (good and poor outcomes) based on the patients' modified Rankin Scale scores and trained a binary classifier to predict the correct outcome group. We compared various machine learning classifiers and feature selection and pre-processing techniques to create a parsimonious mRS score prediction pipeline. Our results showed that the best-performing classifier was a multi-layer perceptron classifier which used three radiomics features to achieve a classification accuracy of 0.94. This is a marked improvement compared to our previous results, where the classification accuracy was around 0.7 and matches the performance of previous studies reported in the literature. In the clinics, our pipeline can help doctors and stroke patients plan recovery and rehabilitation processes.