Multidataset Independent Subspace Analysis With Application to Multimodal Fusion

Area of Science:

• Machine Learning
• Signal Processing
• Data Science

Background:

• Unsupervised latent variable models, particularly blind source separation (BSS), are valued for interpretability.
• Existing BSS models typically analyze single datasets, limiting the potential for joint insights from multiple sources.
• Multidataset analysis offers richer solutions but lacks unified, principled integration methods.

Purpose of the Study:

• To develop a direct and principled method for combining information from multiple datasets within a BSS framework.
• To extend BSS models to effectively capture both shared and unique variability across and within datasets.
• To create a unified model capable of leveraging joint information from heterogeneous datasets flexibly and synergistically.

Main Methods:

• Introduced Multidataset Independent Subspace Analysis (MISA) for principled multidataset combination.
• Utilized the Kotz distribution for advanced subspace modeling.
• Employed novel combinatorial optimization to avoid local minima in model fitting.

Main Results:

• MISA provides a robust generalization of Independent Component Analysis (ICA), Independent Vector Analysis (IVA), and Independent Subspace Analysis (ISA).
• The method effectively captures underlying modes of shared and unique variability across datasets.
• Demonstrated utility in multimodal information fusion, even in challenging conditions like small sample sizes (N=600) and low signal-to-noise ratios.

Conclusions:

• MISA offers a unified approach to multidataset BSS, enhancing interpretability and joint solution discovery.
• The method facilitates flexible and synergistic information fusion from heterogeneous data sources.
• MISA shows promise for novel applications in unimodal and multimodal brain imaging and other data-intensive fields.