Evaluating Trade-offs In Iva of Multimodal Neuroimaging Using Cross-Platform Multidataset Independent Subspace Analysis

Multidataset independent subspace analysis (MISA) is a unified framework incorporating multiple linear blind source separation methods to analyze joint and unique information across multiple datasets. MISA can jointly analyze large multimodal neuroimaging datasets to advance our understanding of the brain from multiple perspectives. However, a systematic evaluation of the trade-offs between problem size/scale and total number of training samples is still absent in the literature. Aiming to support flexibility and replicability of deep latent variable modeling, and equip practitioners with crucial tools and usage guidelines, we developed a MISA PyTorch module incorporating the linked multi-network architecture and loss function of the original MISA MATLAB. We then investigated critical performance trade-offs between sample and latent space sizes in independent vector analysis (IVA) problems. Both platforms were highly similar in hundreds of simulation settings, demonstrating successful replication of the original framework and flexibility to evaluate multiple configurations. We observed that a larger sample size, fewer datasets and fewer sources can lead to better IVA model performance. We then performed an IVA experiment on a large multimodal neuroimaging dataset and observed high cross-modal correlation linkage among the identified sources in both MATLAB and Python platforms, supporting MISA's effectiveness for replicable multimodal linkage detection.