Investigation of imaging biomarkers on Multiple Sclerosis (MS) disability by using a shallow artificial neural network model.

The severity of disability and its progression rate are variable in MS. The expanded disability status scale (EDSS) has been correlated with features on magnetic resonance imaging (MRI), magnetic resonance spectroscopy (MRS), magnetization transfer ratio (MTR), diffusion tensor imaging (DTI) and optical coherence tomography (OCT). A shallow artificial neural network (ANN) is a suitable application of artificial intelligence for modeling the interplay of imaging biomarkers and their effect on clinical disease severity.

Fifty MS patients on Fingolimod were enrolled in a two-year study. Imaging biomarkers were measured across three time-points at one year intervals including: MRI, MRS, MTR, DTI and OCT. A single hidden-layered feed-forward ANN is constructed with seventeen input and one output nodes. The ANN undergoes supervised learning using Bayesian regularization. The data consisted of 135 timepoints; 15% of this data was reserved for testing the network to ensure adequate generalization.

A feed-forward network with 5 neurons containing 96 parameters is designed. The mean squared error of testing (and training) is 1.213 (0.2021). The accuracy of the network is 77.9% (87.3%). T-test between the model's output and the target EDSS is found to be statistically similar with p-values equating to 0.87 (0.97). Additionally, the model correctly captured the trend in EDSS for two SPMS patients, and accurately predicted their EDSS despite not being trained on SPMS patients.

This work demonstrates a successful proof-of-concept application of a robust ANN to accurately predict the clinical disability based on multiple important imaging biomarkers. The model can now be applied to further investigate sensitivity-analyses of each imaging biomarker across a wide matrix that spans the ranges of age, disease duration and lesion-load that the model was trained on.