Fed-MStacking: Heterogeneous Federated Learning With Stacking Misaligned Labels for Abnormal Heart Sound Detection

Area of Science:

• Artificial Intelligence
• Smart Healthcare
• Internet of Health Things (IoHT)

Background:

• Ubiquitous sensing in smart healthcare enables intelligent heart sound auscultation but raises user privacy concerns due to sensitive data on smart devices.
• Federated learning (FL) addresses privacy by enabling decentralized learning without data sharing in the Internet of Health Things (IoHT).
• Traditional FL lacks model heterogeneity and client personalization, as it requires uniform architectural models across clients and servers.

Purpose of the Study:

• To propose Fed-MStacking, a heterogeneous FL framework using stacking ensemble learning for personalized client models in medical institutions.
• To address inconsistent data labeling across local clients, where each client may only have one case type and data cannot be shared.
• To train a global multi-class classifier by aggregating missing class information and building meta-data for FL training via a meta-learner.

Main Methods:

• Developed Fed-MStacking, a heterogeneous FL framework incorporating stacking ensemble learning.
• Implemented a meta-learner to aggregate missing class information from clients with inconsistent labeling, creating meta-data for FL.
• Utilized random forests (RFs), feedforward neural networks (FNNs), and convolutional neural networks (CNNs) as base classifiers on a multi-institutional heart sound database.

Main Results:

• Fed-MStacking demonstrated superior performance compared to homogeneous stacking in multi-institutional heart sound analysis.
• The framework successfully supported clients in building personalized, heterogeneous models.
• Aggregation of missing class information enabled the training of a global multi-class classifier despite data inconsistencies.

Conclusions:

• Fed-MStacking offers an effective solution for privacy-preserving, personalized, and heterogeneous federated learning in smart healthcare applications like heart sound analysis.
• The proposed approach enhances model performance by leveraging stacking ensemble learning and addressing data heterogeneity and labeling inconsistencies.
• This framework advances the application of FL in the Internet of Health Things (IoHT) for multi-institutional medical data analysis.