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

2026-02-14 · Source: Machine learning : nature.com subject feeds · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Advanced, medium

Summary

A systematic review published on February 14, 2026, analyzed 160 articles, encompassing 710 decentralized learning (DL) models and 8149 performance comparisons in healthcare, primarily in oncology, COVID-19, and neurological diagnostics. The review compared DL approaches, such as federated learning and swarm learning, against centralized learning (CL) and local learning (LL) models. Findings indicate that CL generally outperformed DL in threshold-dependent metrics like accuracy and Dice score, showing 78% favourability with large effect sizes. However, DL achieved comparable performance to CL in ranking metrics like AUROC (51% CL favourability with small effect size). Crucially, DL consistently surpassed LL across all metrics, with 86% favourability for precision and 83% for accuracy. While CL could "rescue" DL viability in up to 18% of comparisons, DL significantly improved LL viability by 7.6–27 percentage points, demonstrating its clinical acceptability in privacy-constrained healthcare settings.

Key takeaway

For healthcare data scientists and CTOs evaluating AI model deployment strategies, this review highlights that decentralized learning (DL) is a robust, clinically acceptable alternative to centralized models, particularly where data privacy regulations like GDPR and the AI Act are paramount. While centralized learning may offer slight performance advantages in some metrics, DL consistently outperforms local learning and provides a viable path for secure, multi-institutional collaboration. Prioritize DL for privacy-constrained clinical applications, carefully weighing its marginal performance trade-offs against regulatory compliance and implementation complexity.

Key insights

Decentralized learning offers clinically acceptable performance in healthcare, especially for privacy-constrained contexts, outperforming local learning.

Principles

• Centralized learning often yields marginal performance gains.
• DL consistently outperforms local learning.
• Privacy-performance trade-offs are key for DL implementation.

Method

A systematic review of 165,010 studies from eight databases (01/2012-03/2024) was conducted, analyzing 160 articles, 710 DL models, and 8149 performance comparisons using paired and clinical threshold analyses.

In practice

• Implement DL for privacy-sensitive healthcare data.
• Consider CL for marginal performance boosts.
• Balance performance with regulatory compliance (GDPR, AI Act).

Topics

• Decentralized Learning
• Federated Learning
• Healthcare AI
• Clinical Model Evaluation
• AI Regulation

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Researcher, AI Scientist, AI Data Scientist

Related on AIssential

• Federated Learning over Blockchain-Enabled Cloud Infrastructure — Integrating Federated Learning with blockchain in cloud-edge environments enhances data privacy and security.
• Decentralized Autoregressive Generation — Decentralized autoregressive generation, using expert flows, achieves performance parity with centralized MLLM training.
• A Data-Centric Framework for Intraoperative Fluorescence Lifetime Imaging for Glioma Surgical Guidance — A data-centric AI framework improves FLIm accuracy for glioma surgery by refining labels and enhancing model robustness.
• Representation learning to advance multi-institutional studies with electronic health record data from US and France — A graph-based framework harmonizes heterogeneous EHR data across institutions using representation learning and privacy-preserving methods.
• Complex & Intelligent Systems, Volume 12, Issue 4, April 2026 — The issue highlights diverse AI advancements across networking, computer vision, NLP, and secure systems.
• Patient privacy in the age of clinical AI: Scientists investigate memorization risk — AI models trained on de-identified EHRs can memorize and leak patient data, necessitating rigorous privacy evaluation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine learning : nature.com subject feeds.