Disentangling Blood-Based Markers of Multiple Sclerosis Through Machine Learning: An Evaluation Study

In the search for markers to aid early diagnosis, sustainable monitoring, and accurate prognosis of Multiple Sclerosis (MS), researchers have turned to blood-based markers. These provide rich information on a persons health while being easier to acquire than magnetic resonance images. To analyse blood data, researchers have used machine learning (ML) to support evaluation at scale, but because many different analytics pipelines exist, it is unclear how different ML methods compare and influence experimental outcomes. Therefore, this ML evaluation study compared in different configurations the performance of five ML algorithms, two methods to select their features, and approaches to evaluate them. The aim was to first assess how the ML methods influenced classifying people with MS and controls, and then disentangle the blood-based markers selected for the best performing classifiers. The results indicated that Logistic Regression with Random Forests for feature selection and 10-fold cross-validation produced the best results, that feature selection depended on the feature selection methods, and that data splits for training, validation, and testing were heterogeneous. This suggests experimental setups influence both the classification performance and disentangled markers, meaning that evaluation rigor matters when using ML to support discovery processes and knowledge creation in medical research.