VERTIcal Grid lOgistic regression (VERTIGO).

Area of Science:

• Computational statistics
• Machine learning
• Health informatics

Background:

• Federated data analysis enables collaborative model training without data sharing.
• Vertically partitioned data, where different features are stored across institutions, presents unique challenges for analysis.
• Accurate logistic regression (LR) is crucial for many medical classification tasks.

Purpose of the Study:

• To develop an accurate logistic regression (LR) algorithm for federated analysis of vertically partitioned data.
• To address the technical challenges of scaling federated LR algorithms.
• To enable the creation of global predictive models from distributed datasets.

Main Methods:

• Proposed a novel dual optimization technique for binary logistic regression.
• Introduced the VERTIcal Grid lOgistic regression (VERTIGO) algorithm.
• Evaluated VERTIGO on artificial and real-world medical datasets for classification accuracy, calibration, and computational complexity.

Main Results:

• VERTIGO yields global model parameters identical to classical LR.
• The algorithm demonstrates comparable discrimination and calibration to traditional methods on simulated and real data.
• VERTIGO's computational cost is practical for real-world applications.

Conclusions:

• The VERTIGO algorithm provides an accurate solution for federated data analysis with vertically partitioned data.
• This method supports the development of robust global models in distributed healthcare settings.
• VERTIGO overcomes key scalability challenges in federated logistic regression.