Forecasting the information computational efficiency for multiple sclerosis using a machine learning-based approach on a massive multi-focal magnetic resonance imaging dataset

The Symbol Digit Modalities Test (SDMT) has been suggested as a reliable screening tool for information processing speed (IPS) deficiencies, which are common in MS patients. Understanding the mechanisms underlying cognitive problems in multiple sclerosis has significantly improved because to magnetic resonance imaging (MRI). It is yet unknown, nevertheless, which structural MRI signals have the most correlation with cognitive function. We extracted multimodal data (demographic, clinical, neuropsychological, and structural MRIs) from 540 MS patients using the multicenter 3T-MRI data set of the Italian Neuroimaging Network Initiative. Our objective was to evaluate, using machine learning techniques, how brain MRI structural volumes, in conjunction with clinical and demographic characteristics, contribute to the prediction of IPS abnormalities. In order to achieve dependable generalization performance, we trained and evaluated the eXtreme Gradient Boosting (XGBoost) model using a strict validation scheme. Based on SDMT scores, we performed a regression and classification exercise, feeding each model with various feature combinations. The model trained with the volumes of the thalamus, cortical gray matter, hippocampus, and lesions obtained an area under the receiver operating characteristic curve of 0.74 for the classification task.   A mean absolute error of 0.95 was attained by the model trained with cortical gray matter and thalamus volumes, EDSS, nucleus accumbens, lesions, and putamen volumes, as well as age, for the regression task. Our findings demonstrated that one of the most significant predictors of cognitive function in multiple sclerosis is damage to cortical gray matter and other deep and ancient gray matter regions, including the thalamus and hippocampus.