Applying Federated Learning to Block-Term Tensor Regression for Decentralised Data Analysis of Biomedical Data

Abstract

Block-Term Tensor Regression (BTTR) is a powerful method for modeling complex, high-dimensional data through multilinear relationships, making it well-suited for healthcare and neuroscience. However, its reliance on centralized datasets raises privacy concerns and limits collaboration. To address this, we propose Federated Block-Term Tensor Regression (FBTTR), an extension of BTTR for federated learning that enables decentralized model building while preserving privacy and regulatory compliance. We evaluate FBTTR in two case studies: finger movement decoding from Electrocorticography (ECoG) signals and disease risk prediction. On the BCI Competition IV dataset, FBTTR outperforms non-multilinear models and achieves higher accuracy than centralized BTTR (e.g., subject 3 thumb decoding: 0.76 +/- 0.05 vs. 0.71 +/- 0.05). On real-world clinical data, FBTTR surpasses both standard federated approaches and centralized BTTR (e.g., Fed-Heart-Disease Dataset AUC-ROC: 0.872 +/- 0.02 vs. 0.812 +/- 0.003; accuracy: 0.772 +/- 0.02 vs. 0.753 +/- 0.007). These results show that FBTTR is scalable, computationally efficient, and achieves predictive performance comparable to centralized models. Released as open-source software, it provides a practical and transparent framework for advancing federated analytics in healthcare and brain-computer interface applications.