Grid Binary LOgistic REgression (GLORE): building shared models without sharing data.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomic Data Analysis

Background:

• Classifying complex patterns in clinical and genomic data necessitates large, labeled patient datasets.
• Current methods often rely on centralized data, raising privacy concerns and limiting data sharing.
• Distributed modeling approaches for sensitive health data remain underexplored.

Purpose of the Study:

• To investigate the feasibility of developing binary logistic regression (LR) models in a distributed manner.
• To enable researchers to share predictive models without sharing sensitive patient data.
• To introduce and evaluate the Grid Binary Logistic Regression (GLORE) model.

Main Methods:

• The GLORE model brings computation to the data, avoiding the need for a central data repository.
• It integrates decomposable partial elements or non-privacy sensitive prediction values.
• Key outputs include model coefficients, variance-covariance matrix, goodness-of-fit, and ROC curve AUC.

Main Results:

• Experiments on simulated and clinical data demonstrated GLORE's accuracy, yielding results identical to traditional LR (to 10^-15 precision).
• GLORE was found to be computationally efficient compared to traditional LR methods.
• The study confirmed that GLORE allows data to remain protected at its original sites.

Conclusions:

• GLORE offers a viable distributed approach for building logistic regression models.
• It effectively addresses data privacy concerns by keeping patient data decentralized.
• Synchronization of coefficient gradients and consistent predictor formatting across sites are necessary considerations.