Client Participation per Round in Federated Learning for Multiple Sclerosis with Real-World Data

Area of Science:

• Artificial Intelligence
• Neuroscience
• Medical Informatics

Background:

• Federated learning (FL) enables collaborative machine learning without centralizing sensitive data.
• Predicting multiple sclerosis (MS) disability progression is crucial for patient management.
• Routine clinical data offers a valuable resource for developing predictive models.

Purpose of the Study:

• To investigate the impact of client participation rate (k) on federated learning performance and runtime.
• To evaluate the trade-offs between participation levels and model accuracy for predicting MS disability progression.

Main Methods:

• Utilized original study data, preprocessing, model architecture, and evaluation protocols.
• Varied the per-round client participation rate (k) to values of 1.0, 0.6, and 0.4.
• Assessed model performance using ROC-AUC and AUC-PR metrics and measured runtime.

Main Results:

• Lowering client participation (k) reduced computational wall time.
• A participation rate of k ≈ 0.6 maintained nearly all predictive performance (ROC-AUC and AUC-PR).
• Runtime was reduced by approximately one-third with k ≈ 0.6 compared to full participation (k=1.0).

Conclusions:

• Client participation rate is a critical parameter influencing federated learning efficiency.
• Optimizing k, such as to ≈ 0.6, offers a practical balance between runtime reduction and predictive accuracy in MS disability progression models.
• Federated learning with optimized participation is a viable approach for leveraging routine clinical data in neurological disease research.