Maximum-Likelihood Independent Component Analysis with Multi-View Data: A Comparative Study

In this article, we propose a new method called Maximum-Likelihood ICA for Multiview Data (MVICAD) to improve the signal-to-noise ratio of sources estimated by Independent Component Analysis (ICA). MVICAD uses multiple views to take advantage of the group structure in the data and increase the accuracy of source estimation. By comparing MVICAD with an exhaustive search method that includes a term from Equation 2 in delay optimization, we found that ignoring this term does not affect performance.
To evaluate MVICAD’s performance, we conducted simulations using a synthetic dataset with 5 subjects, 3 sources, and 700 samples. We simulated data according to our model (1) and observed that MVICAD provides more accurate source estimation than the exhaustive search method.
In real-world applications, doctors may have multiple views of the same patient’s data, such as an MRI scan and a CT scan. By using MVICAD on these multiple views, we can improve the quality of the estimated sources compared to using ICA on each view separately. Moreover, since the sources are assumed to be the same for all views, MVICAD takes advantage of the group structure in the data without assuming that the sources are independent.
MVICAD is written in Python and available on GitHub. Our simulations show that MVICAD provides more accurate source estimation than an exhaustive search method that includes a term from Equation 2 in delay optimization. By using multiple views, MVICAD can improve the signal-to-noise ratio of source estimation and provide more reliable results for real-world applications.