Predicting multiple sclerosis prognosis using AI and machine learning: integrating clinical, immunological, and radiological variables

TL;DR

This study uses machine learning to predict the progression of multiple sclerosis by combining immune data, clinical information, and MRI results.

Contribution

The novel integration of cytokine profiles with machine learning models to predict both disability and MRI lesion progression in MS patients.

Findings

Ensemble models, particularly Random Forest, showed high accuracy in predicting EDSS disability levels.

Random Subspace classifiers achieved 82.4% sensitivity and specificity in predicting new MRI lesions.

Including patient ID improved the performance of logistic regression models for EDSS prediction.

Abstract

Accurate prediction of disease progression in multiple sclerosis (MS) remains a critical challenge in clinical management. This study investigates the utility of supervised machine learning (ML) models in predicting clinical disability, as measured by the Expanded Disability Status Scale (EDSS), and radiological activity based on MRI lesion changes in patients with relapsing-remitting MS (RRMS).

Using peripheral cytokine profiles (IL-12, TNF-α, IFN-γ, IL-4, IL-10) along with patient metadata (e.g., sex, family history, relapse status), 43 ML classifiers were trained and evaluated for their ability to discriminate between mild and moderate disability (EDSS <1 vs >1, and <2.5 vs >2.5), and to predict new MRI lesions in 15 MS patients.

Ensemble models consistently outperformed simpler algorithms. For EDSS prediction, Random Forest achieved 90.1% sensitivity and 89.7% specificity, while Simple Logistic Regression reached 92.6% for both metrics when patient ID was included. In predicting new MRI lesions, Random Subspace classifiers performed best, with 82.4% sensitivity and specificity.

These findings suggest that combining cytokine-based immune signatures with machine learning strategies can provide clinically meaningful predictions of both functional disability and radiological progression. Such tools may support more proactive patient monitoring, informed therapeutic decision-making, and risk stratification in the care of RRMS. Further validation in prospective cohorts is warranted to support clinical implementation.

Figures