FLCMC: Federated Learning Approach for Chinese Medicinal Text Classification.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Medical Informatics

Background:

• Federated learning faces challenges with data heterogeneity and privacy in sensitive domains like medical text.
• Existing federated algorithms struggle to maintain performance and stability with diverse datasets.

Purpose of the Study:

• To propose novel federated learning algorithms (FedPA, FedPAP) for Chinese medical text classification.
• To address privacy protection and data sharing issues in Chinese medical text analysis.
• To improve the performance and convergence stability of federated learning on heterogeneous data.

Main Methods:

• Introduced FLCMC, a federated learning approach incorporating self-attention mechanisms.
• Developed perturbed federated learning algorithms (FedPA, FedPAP) with customized optimizers.
• Integrated self-attention into model aggregation and added perturbation terms to local updates.

Main Results:

• FedPA and FedPAP demonstrated superior accuracy and convergence stability on synthetic and real-world Chinese medical datasets (IMCS-V2).
• The proposed algorithms outperformed FedAvg, FedProx, and FedAtt, especially with heterogeneous data.
• Achieved strong generalization ability on deep learning models for Chinese medical texts.

Conclusions:

• The proposed FedPA and FedPAP algorithms effectively handle data heterogeneity in Chinese medical text classification.
• These methods offer improved privacy, data sharing, accuracy, and stability for medical NLP tasks.
• FLCMC provides a robust solution for analyzing sensitive Chinese medical text data.