Spatial modeling of multiple sclerosis for disease subtype prediction.

Area of Science:

• Neuroimaging
• Biostatistics
• Machine Learning

Background:

• Magnetic Resonance Imaging (MRI) is crucial for Multiple Sclerosis (MS) management.
• Current MS assessment relies on subjective clinical scores and image ratings.
• There is a need for objective and quantitative methods for MS subtype classification.

Purpose of the Study:

• To develop and compare objective methods for a 5-way classification of MS disease subtypes.
• To evaluate the performance of spatially informed models against machine learning approaches.
• To improve the accuracy of MS subtype classification beyond current subjective standards.

Main Methods:

• Proposed two spatially informed models: Bayesian Spatial Generalized Linear Mixed Model (BSGLMM) and Log Gaussian Cox Process (LGCP).
• BSGLMM accounts for voxel spatial dependence and lesion map binary nature.
• LGCP models random spatial variation in lesion location.
• Compared spatial models with a multi-class Support Vector Machine (SVM) using quantitative MRI features and clinical data.

Main Results:

• Spatially informed models demonstrated superior performance compared to standard approaches.
• Average prediction accuracies of up to 85% were achieved using the proposed spatial models.
• The BSGLMM and LGCP models effectively handle spatial dependencies ignored by mass univariate analyses.

Conclusions:

• Objective, spatially informed models offer a significant advancement in classifying MS subtypes.
• These methods provide a more accurate and reliable alternative to subjective assessments in MS diagnosis and management.
• The developed models have the potential to enhance clinical decision-making for patients with Multiple Sclerosis.