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

This systematic review evaluates how decentralized learning (DL) approaches-e.g., federated learning, swarm learning, ensemble-compare with traditional models in healthcare applications. We searched eight databases (01/2012 to 03/2024), screening 165,010 studies with two independent reviewers. Analysis included 160 articles comprising 710 DL models and 8149 performance comparisons across clinical domains, predominantly in oncology, COVID-19, and neurological diagnostics. In paired comparisons, centralized learning (CL) demonstrated advantages in threshold-dependent metrics (78% favourability for accuracy and Dice score with large effect sizes), while DL achieved comparable performance in ranking metrics (51% centralized favourability for AUROC with small effect size). DL consistently outperformed local learning (LL) across all metrics, particularly precision (86% favourability) and accuracy (83% favourability). Clinical threshold analysis (≥0.80 performance) revealed that CL rescued DL viability in up to 18% of comparisons, though when both achieved clinical viability, improvements typically represented "excellent versus acceptable" performance (median difference of 0.7-1.5pp) rather than "acceptable versus inadequate." DL rescued LL viability with substantial improvements (median difference of 7.6-27pp). These findings demonstrate DL offers clinically acceptable alternatives for privacy-constrained contexts, with implementation decisions balancing marginal performance trade-offs against regulation (e.g., GDPR, AI Act) and application. Future research requires standardized privacy-performance reporting.