A Comparative Analysis of Centralized and Federated Learning for Multimodal ECG and PCG Classification.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Cardiology

Background:

• Cardiovascular disease detection relies on analyzing physiological signals like ECG and PCG.
• Existing machine learning models often require centralized data, posing privacy challenges.
• Federated learning (FL) offers a decentralized approach to model training, preserving patient data privacy.

Purpose of the Study:

• To evaluate federated learning approaches for cardiac abnormality detection using ECG and PCG data.
• To compare the performance of multimodal federated models against centralized single-modality models.
• To assess the impact of FL on data privacy and model performance metrics.

Main Methods:

• Analysis of ECG and PCG data from the PhysioNet 2016 challenge dataset.
• Implementation and testing of various federated learning strategies.
• Comparison of centralized and federated model performance with varying numbers of clients (2 and 4).

Main Results:

• Multimodal federated models (ECG + PCG) consistently outperformed centralized single-modality models.
• Performance gains from multimodal approaches compensated for potential losses from distributed learning.
• Federated models achieved performance comparable to centralized single-modality approaches.

Conclusions:

• Multimodal federated learning demonstrates significant potential for improving cardiovascular disease detection.
• FL offers decentralization benefits while maintaining high model performance.
• This approach facilitates optimized machine learning deployment in clinical settings with enhanced patient privacy.