Assisting the diagnosis of multiple sclerosis using a set of regional brain volumes: A classification model for patients and healthy controls. (2021)

• Record Type: Journal Article
• Authors: Aoki, Sorama; Fujimori, Juichi; Mikami, Ryoko; Hoshi, Kenji; Kawakami, Junko; Sato, Kenichi; Nakashima, Ichiro
• Is Part Of: Informatics in medicine unlocked. Volume 26(2022)
• Publication Date: 2021
• Subjects: Multiple sclerosis -- Normalized brain volume -- Classification -- Bayesian regularized neural networks -- Support vector machine -- Self-organizing maps; Medical informatics -- Periodicals; 610.285
• DOI: 10.1016/j.imu.2021.100766
• Languages: English
• ISSNs: 2352-9148

Abstract

1. Abstract: Multiple sclerosis (MS) is an inflammatory disease of unknown etiology in the central nervous system characterized by dissemination in time and space. Difficulties are associated with definitively diagnosing MS early because there are no disease-specific symptoms or diagnostic markers; therefore, a comprehensive judgment to differentiate MS from other diseases is made based on the clinical features of repeated relapses and remission and characteristic magnetic resonance imaging (MRI) findings. In this study, we attempted to construct a predictive classification model using a machine learning method that accurately distinguishes healthy controls (HC) and MS patients based on a quantitative assessment of brain atrophy characteristics caused by MS. We used brain volumes from 55 segments of each brain region calculated from the MRI images of 72 MS patients and 21 HC. These data were input into supervised learning methods (Bayesian regularized neural networks (BRNN) and support vector machine (SVM)) for training on fluctuation patterns in brain atrophy. The obtained accuracy of the model was 77.8% for sensitivity and 95.2% for specificity with cross-validation. The MS prediction rate, calculated by this model, was correlated with Expanded Disability Status Scale (EDSS) scores (r = 0.413, p < 0.0001), which indicates the severity of physical disability due to MS. We also constructed a model using self-organizing maps (SOM) to confirm fluctuations in 15 segments of the brain regions that contributed the most to the classification based on the automatic relevance determination (ARD) function of BRNN. Confirmed fluctuations may be a characteristic of MS brain atrophy, consistent with previous findings. Based on the results obtained herein, we concluded that the model constructed to differentiate between healthy and MS might appropriately reflect the complex data structure of brain atrophy in MS patients. Furthermore, this model's MS predictive rate has potential as a quantifiable pseudo-marker in MS medical care. Although the present model has some limitations, such as the extent of age alterations, this method has the advantage of not including a subjective evaluation depending on individual doctors' experience. Therefore, it may provide an objective indicator that facilitates the diagnosis of MS, which is now mainly being evaluated by a qualitative approach. Highlights: Learning model using a set of regional brain volumes can predict multiple sclerosis. Especially 15 brain regions among all 55 regions were essential to prediction. The model's output significantly correlated with the degree of physical disability. There are no quantifiable markers to identify multiple sclerosis by MRI at present. The model's output has the potential to develop a quantifiable pseudo-biomarker.