Abstract
BACKGROUND: The rise of federated learning (FL) as a novel privacy-preserving technology offers the potential to create models collaboratively in a decentralized manner to address confidentiality issues, particularly regarding data privacy. However, there is a scarcity of clear and comprehensive evidence that compares the performance of FL with that of the established centralized machine learning (CML) in the clinical domain. OBJECTIVE: This study aimed to review the performance comparisons of FL-based and CML models for mortality prediction in clinical settings. METHODS: Experimental studies comparing the performance of FL and CML in predicting mortality were selected. Articles were excluded if they did not compare FL with CML or only compared the effectiveness of different FL baseline models. Two independent reviewers performed the screening, data extraction, and risk of bias assessment. The IEEE Xplore, PubMed, ScienceDirect, and Web of Science databases were searched for articles published up to June 2024. The risk of bias was assessed using CHARMS (Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modeling Studies) and PROBAST (Prediction Model Risk of Bias Assessment Tool). Meta-analyses of the pooled area under the receiver operating curve (AUROC)/area under the curve (AUC) were performed for within-group comparisons (before and after federation). RESULTS: Nine articles with heterogeneous framework design, scenario, and clinical context were included: 4 articles focused on specific case types; 3 articles were conducted in intensive care unit settings; and 2 articles in emergency departments, urgent centers, or trauma centers. Cohort datasets involving 1,412,973 participants were used in all of the included studies. These studies universally indicated that the predictive performance of FL models is comparable to that of CML. The pooled AUC for the FL and CML performances were 0.81 (95% CI 0.76-0.85; I2=78.36%) and 0.82 (95% CI 0.77-0.86; I2=72.33%), respectively. The Higgins I2 test indicated high heterogeneity between the included studies (I2≥50%). In total, 4 out of 9 (44%) of the developed models were identified as having a high risk of bias. CONCLUSIONS: This systematic review and meta-analysis demonstrate that FL can achieve similar performance to CML while conquering privacy risks in predicting mortality across various settings. Owing to the small number of studies and a moderate proportion of the high risk of bias, the effect estimates might be imprecise.