aNy-way Independent Component Analysis.

Area of Science:

• Multimodal data analysis
• Machine learning
• Signal processing

Background:

• Multimodal data fusion aims to uncover patterns across diverse datasets.
• Existing methods like jICA and mCCA have limitations in handling multiple modalities, shared parameters, or feature types.
• Techniques like DS-ICA and parallel ICA are restricted to two modalities or limited optimization.

Purpose of the Study:

• To develop a flexible and robust N-way data fusion model.
• To overcome the limitations of existing multimodal data fusion techniques.
• To enable the analysis of common and distinct features across any number of modalities with varying source numbers.

Main Methods:

• Proposed aNy-way ICA, combining infomax ICA and Gaussian independent vector analysis (IVA-G).
• Utilized a shared weight matrix model without orthogonality constraints.
• Validated through simulations assessing source and loading recovery accuracy.

Main Results:

• aNy-way ICA accurately recovers sources, loadings, and covariance/linkage patterns.
• The method performs well regardless of whether modalities have the same or different numbers of sources.
• Outperformed mCCA and mCCA+jICA in source and loading recovery, especially under noisy conditions.

Conclusions:

• aNy-way ICA provides a powerful N-way data fusion model for any number of modalities and linkage pairs.
• The model accommodates different numbers of non-orthogonal sources across modalities.
• Offers a significant advancement for complex multimodal data analysis in various scientific domains.