Vertical federated learning based on data subset representation for healthcare application.

Area of Science:

• Artificial Intelligence in Healthcare
• Machine Learning for Medical Diagnosis
• Federated Learning for Data Privacy

Background:

• Vertical Federated Learning (VFL) enables collaborative model training across hospitals while preserving patient privacy.
• Existing VFL methods struggle with healthcare data due to limited samples/labels and lack of cross-hospital data correlation.
• Healthcare AI demands robust disease classification and diagnosis models.

Purpose of the Study:

• To introduce FedRL, a novel representation-based VFL method for improved healthcare AI.
• To address challenges in VFL for low-sample, low-label healthcare scenarios.
• To leverage intrinsic data connections across hospitals for enhanced model performance.

Main Methods:

• FedRL utilizes aligned data for federated representation pretraining.
• The method involves splitting local data, exploiting subset relationships, and using a custom loss function.
• A collaborative representation model is trained across hospitals to capture global data latent representations.

Main Results:

• FedRL demonstrated superior performance on three healthcare datasets.
• The method achieved average improvements of 4.7% in accuracy, 5.6% in AUC, and 4.8% in F1-score.
• FedRL showed robustness and consistent performance with limited labeled samples.

Conclusions:

• FedRL is an effective VFL method for healthcare data analysis.
• The proposed approach enhances disease classification and clinical diagnosis.
• FedRL offers a promising solution for privacy-preserving AI in healthcare with limited data.