Validation of generative artificial intelligence techniques for synthetic data generation in multiple sclerosis research: A comparison with real-world evidence from the Italian multiple sclerosis registry

Background: The use of generative artificial intelligence to create synthetic data is an emerging strategy to enhance real-world evidence. Objective: To assess the validity of artificial intelligence-generated synthetic data in replicating data collected in the Italian Multiple Sclerosis and Related Disorders Register for the evaluation of progression independent of relapse activity risk under early intensive treatment and escalation treatment strategies. Methods: Artificial intelligence generative models were t…

By the Truvace deskOctober 15, 20261 min read1 sourcePeer-reviewedTRV-2026-1393· cite & verify

Validation of generative artificial intelligence techniques for synthetic data generation in multiple sclerosis research: A comparison with real-world evidence from the Italian multiple sclerosis registry — "Axial PSIR MRI of a brain with multiple sclerosis lesions" by Article authors: Alice Favaretto, Davide Poggiali, Andrea Lazzarotto, Giuseppe Rolma, Francesco Causin, Paolo Gallo is licensed under CC BY 4.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/.

In brief

Researchers trained generative AI models on a sub-cohort of 1666 patients from the Italian Multiple Sclerosis and Related Disorders Register and produced a synthetic dataset of 4878 patients. They then compared treatment strategies, early intensive versus escalation, using both real and synthetic data.

The comparison matters because synthetic data could widen access to registry-style evidence without exposing real patient records. The authors support use only in clinically defined research settings, so how well the approach works for other questions or populations remains to be tested.

Main points

Generative models were trained on 1666 registry patients and produced 4878 synthetic patients.
Validation used the Synthetic vAlidation FramEwork, Clinical Synthetic Fidelity, and Nearest Neighbor Distance Ratio.
The outcome studied was the risk of progression independent of relapse activity.
Cox proportional hazards models on synthetic data preserved the robustness of the estimates for a first progression event.
Associations with radiological risk factors were broadly similar in real and synthetic datasets.

The gain

AI-generated synthetic patients reproduced treatment-effect estimates for escalation versus early intensive therapy in MS while preserving privacy, suggesting a way to extend real-world evidence.

The rundown

The study addressed a specific question: whether AI-generated synthetic data can replicate findings from the Italian MS register on progression independent of relapse activity under two treatment strategies.

Fidelity and privacy were assessed with named validation tools, and the synthetic data were reported to meet both. Treatment-effect estimates for escalation versus early intensive treatment were similar between real and synthetic datasets.

The authors conclude that the synthetic data reliably replicated treatment-effect outcomes and support their potential use in clinically defined research settings.