Time-Dependent Deep Learning Prediction of Multiple Sclerosis Disability.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Neurology

Background:

• Current deep learning models often analyze single timepoints, unlike clinical practice which correlates changes over time.
• Multiple sclerosis (MS) research has focused on lesion segmentation, with limited analysis of long-term disability progression.
• Longitudinal analysis is crucial for understanding disease trajectory and informing patient management.

Purpose of the Study:

• To propose and evaluate time-dependent deep learning models for predicting long-term disability in MS.
• To benchmark a Video Vision Transformer (ViViT) against Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) and Vision Transformer-LSTM (ViT-LSTM) architectures.
• To assess the models' ability to predict disability progression using the Extended Disability Severity Score (EDSS).

Main Methods:

• A cohort of 703 patients with cervical spine MRIs (2002-2023) was utilized.
• A Video Vision Transformer (ViViT) was compared with a VGG-16 based CNN-LSTM and a ViT-LSTM.
• Ablation analysis was performed to determine the time-dependency of model performance.

Main Results:

• ViViT achieved a higher AUC (0.84) than VGG16-LSTM (0.74) for predicting trinary EDSS in 6 years (p < 0.001).
• VGG16-LSTM outperformed ViViT when analyzing shorter durations (2 years of MRI data).
• Exact EDSS classification (regression or classification) yielded collectively worse performance.

Conclusions:

• Time-dependent deep learning models can effectively predict MS disability using trinary stratification, mirroring clinical assessment.
• The choice of model architecture may depend on the available longitudinal data duration.
• Further validation in external cohorts and clinical trials is warranted to confirm these findings.