# Extrapolating from trials to clinic: a predictive model defining the boundaries of benefit for multiple sclerosis therapies in real-world populations based on systematic review

## TL;DR

This study creates a model to predict which MS patients will benefit from therapies and which may face harm, using trial data and real-world evidence.

## Contribution

A predictive model is developed to guide real-world MS treatment decisions by analyzing trial and clinical data.

## Key findings

- Baseline characteristics predict treatment efficacy and risk of harm in MS patients.
- High-efficacy treatments initiated early with strategic de-escalation yield best long-term outcomes.
- Prescribing to trial-excluded patients is more likely to cause harm than benefit.

## Abstract

Clinical trials for multiple sclerosis (MS) disease-modifying treatments selectively enroll patients with favorable risk–benefit profiles. However, these therapies are often prescribed more broadly in clinical practice. We aimed to identify which patients are unlikely to benefit and may face substantial harm, and codify this into a data-driven framework for guiding real-world MS treatment decisions.

Systematic searches of PubMed and ClinicalTrials.gov identified 61 randomized, blinded phase 2b/3 trials with ≥ 100 adults per arm (all pediatric trials were included due to rarity), ≥ 48 weeks of treatment, and Expanded Disability Status Scale–based confirmed disability progression as an outcome. These trials enrolled 46,611 participants and contributed 91,787 patient-years. We extracted 80 baseline variables per trial arm and derived 30 additional features to reduce bias and train multivariable regression models. Model performance was validated using an independent, longitudinal real-world MS cohort. Infection-related mortality risk was estimated from national life tables and adjusted by treatment-specific hazard ratios.

Baseline characteristics predicted both untreated progression and treatment efficacy. Therapeutic benefit increased with higher relapse rates and presence of enhancing lesions and declined with age and disease duration. Relapse rates in placebo arms declined across trial periods, mirrored by waning treatment efficacy on disability progression, which was confirmed in real-world data. In contrast, treatment-related morbidity and mortality increased with age, disability, and comorbidities. These opposing trends were integrated into a web-based personalized risk–benefit estimator.

Interpretable models offer a unified view of MS evolution and treatment effects. They show that the therapeutic risk–benefit ratio is dynamic, shaped by individual characteristics and predictable over time. The models project that initiating high-efficacy treatments early, followed by strategic de-escalation yields the best long-term outcomes. Critically, they extrapolate, and real-world data confirm that prescribing disease-modifying treatments to patients who would have been excluded from pivotal trials is more likely to cause harm than benefit. By enabling individualized, evidence-based decisions, this estimator can help clinicians deliver safer, more effective MS care worldwide.

The online version contains supplementary material available at 10.1186/s12916-025-04603-z.