Federated Learning-Based Model for Predicting Mortality: Systematic Review and Meta-Analysis

Area of Science:

• Clinical informatics
• Machine learning in healthcare
• Privacy-preserving technologies

Background:

• Federated learning (FL) offers a privacy-preserving approach for collaborative model development in decentralized settings.
• Existing evidence comparing FL performance with centralized machine learning (CML) in clinical applications, particularly for mortality prediction, is limited.
• Addressing data privacy concerns is crucial in clinical machine learning.

Purpose of the Study:

• To systematically review and compare the performance of FL-based models against CML models for mortality prediction in clinical settings.
• To synthesize evidence on the effectiveness of FL in clinical mortality prediction through meta-analysis.

Main Methods:

• A systematic review and meta-analysis of experimental studies comparing FL and CML for mortality prediction.
• Searches conducted in IEEE Xplore, PubMed, ScienceDirect, and Web of Science up to June 2024.
• Risk of bias assessed using CHARMS and PROBAST; pooled area under the curve (AUC) calculated.

Main Results:

• Nine articles were included, covering diverse clinical settings and involving 1,412,973 participants.
• FL models showed comparable predictive performance to CML models, with pooled AUCs of 0.81 for FL and 0.82 for CML.
• High heterogeneity was observed across studies (I2≥50%), and 44% of models had a high risk of bias.

Conclusions:

• Federated learning achieves performance comparable to centralized machine learning for clinical mortality prediction while addressing privacy risks.
• The findings suggest FL is a viable alternative in clinical settings where data privacy is paramount.
• The precision of effect estimates may be limited by the small number of studies and the proportion of high-risk-of-bias models.