Comparing decentralized machine learning and AI clinical models to local and centralized alternatives: a systematic...

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Data Privacy in Medical AI

Background:

• Decentralized learning (DL) approaches like federated learning are emerging alternatives to traditional centralized models in healthcare.
• Healthcare applications face increasing demands for data privacy and regulatory compliance (e.g., GDPR, AI Act).

Purpose of the Study:

• To systematically review and compare the performance of decentralized learning (DL) models against traditional centralized learning (CL) and local learning (LL) in healthcare.
• To evaluate the clinical viability and trade-offs of DL in various medical domains.

Main Methods:

• Systematic review of 160 articles (01/2012-03/2024) from eight databases, analyzing 710 DL models and 8149 performance comparisons.
• Independent reviewer screening of 165,010 studies, focusing on oncology, COVID-19, and neurological diagnostics.
• Paired comparisons and clinical threshold analysis (≥0.80 performance) of DL, CL, and LL metrics.

Main Results:

• Centralized learning (CL) showed advantages in threshold-dependent metrics (accuracy, Dice score), while DL performed comparably in ranking metrics (AUROC).
• Decentralized learning (DL) consistently outperformed local learning (LL) in precision and accuracy.
• CL improved DL viability in 18% of cases; when both were viable, DL offered 'excellent' vs. 'acceptable' performance gains.

Conclusions:

• Decentralized learning (DL) provides clinically acceptable performance for privacy-constrained healthcare applications, balancing marginal trade-offs against regulations.
• DL significantly enhances local learning (LL) viability, offering substantial performance improvements.
• Standardized reporting of privacy and performance metrics is crucial for future decentralized learning research in medicine.