Federated Choquet Regression with LASSO for Outcome Prediction in Multisite Longitudinal Trial Data

Area of Science:

• Biomedical data analysis
• Machine learning in clinical research
• Decentralized modeling

Background:

• Privacy regulations hinder multi-site clinical data aggregation.
• Decentralized approaches are crucial for biomedical research.
• Existing methods may not capture complex feature interactions in longitudinal data.

Purpose of the Study:

• Introduce a federated nonlinear regression algorithm for outcome prediction in multisite longitudinal trials.
• Address privacy constraints and complex data characteristics.
• Develop a robust method for handling missing values and nonlinear relationships.

Main Methods:

• Developed the Multiple Imputation Choquet Integral Regression with LASSO (MIChoquet-LASSO) algorithm.
• Applied a federated variant (FEDMIChoquet-LASSO) for decentralized analysis.
• Evaluated performance using synthetic, public, and proprietary longitudinal randomized controlled trial data.
• Compared against Ordinary Least Squares (OLS) regression and Choquet OLS regression.

Main Results:

• MIChoquet-LASSO effectively handles nonlinear longitudinal data with missing values, outperforming benchmarks in overfitting scenarios.
• Federated Choquet OLS underperformed in federated settings.
• FEDMIChoquet-LASSO demonstrated consistently superior performance in federated contexts.
• Mean Squared Error (MSE) was the primary performance metric.

Conclusions:

• FEDMIChoquet-LASSO provides a reliable solution for outcome prediction in privacy-constrained, multisite longitudinal trials.
• The algorithm successfully addresses missing data and nonlinear relationships.
• This federated approach maintains strong predictive performance within privacy limitations.