Federated Fuzzy Clustering for Decentralized Incomplete Longitudinal Behavioral Data

Area of Science:

• Health Informatics
• Machine Learning
• Data Privacy

Background:

• Medical data analysis is constrained by privacy regulations like HIPAA.
• Anonymization is often insufficient for sensitive health data.
• Traditional clustering methods struggle with longitudinal, incomplete behavioral health data.

Purpose of the Study:

• To develop a privacy-preserving, decentralized federated clustering algorithm for complex longitudinal behavioral health data.
• To address limitations of existing methods in handling missing data and varying time points across multisite trials.

Main Methods:

• A decentralized federated multiple imputation-based fuzzy clustering algorithm was developed.
• The algorithm uses federated learning to aggregate model parameters, preserving data privacy.
• It requires minimal communication rounds and accommodates clients with incomplete longitudinal data.

Main Results:

• The algorithm demonstrated rapid convergence and high performance on clustering metrics.
• Evaluations were conducted on real dietary health data and simulated datasets with varying parameters.
• The method effectively handles complex longitudinal data from multisite randomized controlled trials.

Conclusions:

• The proposed algorithm offers a robust solution for analyzing sensitive health data while ensuring patient privacy.
• It enables targeted treatments by identifying patient subgroups in behavioral health.
• Potential applications extend to the Internet of Medical Things for broader health data analysis.