dGAMLSS: an exact, distributed algorithm to fit Generalized Additive Models for Location, Scale, and Shape for...

Area of Science:

• Biostatistics
• Computational Biology
• Data Science

Background:

• Estimating population reference ranges across age and sex is crucial for identifying atypical measurements.
• Generalized Additive Models for Location, Scale, and Shape (GAMLSS) are recommended for modeling non-linear growth and population heterogeneity.
• Challenges in data sharing for GAMLSS model fitting due to privacy and practical constraints hinder multi-site studies.

Purpose of the Study:

• To introduce a privacy-preserving distributed algorithm for fitting GAMLSS models across multiple sites.
• To enable the construction of population reference charts without sharing patient-level data.
• To address the need for federated learning algorithms applicable to GAMLSS.

Main Methods:

• Proposed distributed GAMLSS (dGAMLSS), a novel distributed algorithm for fitting GAMLSS models in a federated manner.
• Incorporated specific strategies for fitting smooth functions with varying communication efficiency.
• Developed an R package (dGAMLSS) for implementing the algorithm and managing site-specific parameters.

Main Results:

• Demonstrated the effectiveness of dGAMLSS in constructing population reference charts across clinical, genomics, and neuroimaging data.
• Showcased the ability of dGAMLSS to accurately reproduce pooled reference charts and statistical inference.
• Validated the algorithm's performance in scenarios requiring privacy preservation.

Conclusions:

• dGAMLSS successfully enables the creation of population reference charts using multi-site data while preserving patient privacy.
• The algorithm facilitates robust statistical modeling and inference in distributed settings where data cannot be pooled.
• dGAMLSS offers a practical solution for collaborative research requiring sensitive health data.