Federated AI Framework for Privacy-Preserving Healthcare Prediction with Shuffle-Split Attention Networks

• Sivakumar Dhakshinamoorthy,
• Karthikeyan Periyasami,

Vol. 20, No. 3, March 31, 2026
10.3837/tiis.2026.03.006, Download Paper (Free):

• Artificial intelligence
• Federated learning
• Quantile normalization
• Shuffle Attention Network
• Split-Attention Networks

Abstract

Federated Learning (FL) based Artificial Intelligence (AI) governance framework for privacy-preserving healthcare prediction allows decentralized model training across various healthcare institutions without sharing of sensitive patient information. This approach enhances data privacy and supports in safeguarding patient confidentiality. However, there remains challenges, like ensuring model convergence, managing communication overhead, and maintaining robust security against attacks. To overcome these hurdles, the Shuffle Split Attention Forward Taylor Network (S2AFTNet) is proposed for privacy preserved healthcare prediction using FL. Firstly, the local training based on local data is done at each local model. The model training includes data acquisition, normalization using Quantile normalization, selecting features by Multi-Subspace Randomization and Collaboration-based unsupervised Feature Selection (SRCFS), data augmentation using Adaptive Synthetic (ADASYN) approach, and healthcare prediction using S2AFTNet is formed by combining Split-Attention Networks (ResNeSt), Shuffle Attention Network (SA-Net), and Taylor concept. Then, weights from local model training data are aggregated and averaged. Further, global model is applied on each local model based on averaged weights. Moreover, S2AFTNet performance is analyzed with metrics like, loss function, Mean Squared Error (MSE), and Root Mean Squared Error (RMSE), that achieved superior values of 0.034, 7.23, and 13.88%.

Statistics

Cite this article