Scalable and distributed individualized treatment rules for multicenter datasets.

Area of Science:

• Biostatistics
• Machine Learning
• Health Informatics

Background:

• Synthesizing multicenter data is vital for accurate individualized treatment rules (ITRs).
• Privacy concerns hinder the integration of sensitive patient data from multiple sources.
• Classical meta-learning methods can be suboptimal due to biases in local estimates.

Purpose of the Study:

• To propose a novel, privacy-preserving method for learning optimal individualized treatment rules (ITRs) from distributed data.
• To address the limitations of classical meta-learning in handling multicenter data with privacy constraints.

Main Methods:

• Developed a convolution-smoothed weighted support vector machine (SVM) for optimal ITR learning.
• Utilized a convex and smooth loss function for efficient distributed learning.
• Implemented a coordinate gradient descent algorithm for guaranteed linear convergence.

Main Results:

• The proposed distributed learning procedure ensures optimal statistical performance with minimal communication rounds.
• The method effectively preserves data privacy by sharing only summary statistics, not raw subject-level data.
• Extensive simulations and a sepsis treatment case study demonstrated the method's effectiveness.

Conclusions:

• The convolution-smoothed weighted SVM offers an effective and privacy-preserving solution for learning individualized treatment rules from distributed data.
• This approach overcomes the challenges of data integration in multicenter studies, improving the accuracy and applicability of ITRs.
• The method has significant implications for personalized medicine, particularly in critical care settings.