Personalized federated learning for predicting disability progression in multiple sclerosis using real-world routine...

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Clinical Informatics

Background:

• Predicting multiple sclerosis (MS) disability progression is crucial for treatment but remains challenging.
• Existing federated learning (FL) methods face limitations due to data heterogeneity across institutions.
• Privacy-aware collaborative modeling is essential for leveraging multi-center patient data.

Purpose of the Study:

• To systematically evaluate personalized federated learning (PFL) for predicting 2-year MS disability progression.
• To assess the efficacy of PFL in overcoming data heterogeneity in multi-center real-world MS datasets.
• To compare PFL strategies against conventional FL and centralized approaches.

Main Methods:

• Utilized multi-center real-world data from over 26,000 multiple sclerosis patients.
• Implemented and evaluated two PFL strategies: AdaptiveDualBranchNet and personalized fine-tuning.
• Benchmarked PFL performance against baseline FL, centralized models, and client-specific models.

Main Results:

• Personalized federated learning significantly improved prediction performance compared to baseline FL.
• Personalized FedProx and FedAVG achieved high predictive accuracy (ROC-AUC ~0.84).
• PFL methods effectively addressed data heterogeneity challenges while maintaining privacy.

Conclusions:

• Personalization is critical for developing scalable, privacy-aware clinical prediction models in MS.
• PFL demonstrates significant potential for improving early intervention strategies in multiple sclerosis.
• This approach can be extended to other complex neurological disorders and clinical prediction tasks.