ST-VTD framework improves fMRI analysis with LSTM-based tensor decomposition

A new research paper on arXiv (2607.22262) presents ST-VTD (spatiotemporal variational tensor decomposition), a framework designed to model shared and subject-specific structure in multisubject spatiotemporal data, particularly functional MRI (fMRI). Traditional matrix and tensor decompositions often rely on fixed multilinear structures that struggle with complex variability. ST-VTD addresses this by combining a tensor factorization generative model with structured priors: spatial factors are regularized via an LL1 low-rank decomposition, while temporal factors are modeled using a learned LSTM-based prior, enabling flexible and adaptive dynamics. The inference is performed through an amortized variational formulation that unrolls optimization iterations, creating an interpretable and parameter-efficient architecture. A key practical innovation is a warm-start strategy using group independent component analysis (group ICA), which notably improves optimization performance.

Experiments on realistic synthetic fMRI datasets demonstrate that ST-VTD significantly improves latent factor recovery compared to representative classical and probabilistic decomposition benchmarks. The method jointly represents spatial maps and temporal dynamics, capturing both shared patterns across subjects and individual variability. With only 6 pages and 2 figures, the paper focuses on the algorithmic framework and validation, highlighting potential applications in neuroimaging where understanding both group-level and subject-specific brain activity is critical. Future work could extend to real fMRI data and other domains like social media or climate modeling that require analyzing multiple time-series sources with heterogeneous patterns.