COMMUTE: Communication-efficient transfer learning for multi-site risk prediction.

Area of Science:

• Computational biology
• Biostatistics
• Machine learning in healthcare

Background:

• Multi-site healthcare data presents challenges like population heterogeneity and data sharing constraints.
• Accurate risk prediction models are crucial for targeted interventions and improved patient outcomes.

Purpose of the Study:

• To develop a communication-efficient transfer learning approach (COMMUTE) for multi-site healthcare data.
• To address population heterogeneity and data sharing limitations in training risk prediction models.

Main Methods:

• COMMUTE trains population-specific source models locally, then learns calibration terms to adjust for data heterogeneity.
• It supports both centralized data pooling and federated learning by generating synthetic data when individual data is not shareable.

Main Results:

• COMMUTE significantly outperforms methods that ignore population heterogeneity or are trained on single populations in simulations.
• In predicting extreme obesity using eMERGE Network data, COMMUTE achieved an AUC of ~0.80, surpassing benchmarks (AUC 0.51-0.70).

Conclusions:

• COMMUTE improves risk prediction accuracy, especially for target populations with limited data.
• The method is communication-efficient in federated settings, requiring only single-round model parameter sharing.