IEEE Signal Processing Magazine paper on interpretable pattern discovery from complex brain signals
Published:
SimulaMet's Chief Research Scientist, Evrim Acar, has co-authored a survey article in the prestigious IEEE Signal Processing Magazine. The paper is featured in Accelerating Brain Discovery Through Data Science and Neurotechnology special issue, and highlights a powerful, but often underappreciated, set of tools that could help solve “the reproducibility crisis” in neuroscience.
The paper is co-authored together with Prof. Morten Mørup of the Technical University of Denmark and Prof. Tulay Adali, the Editor-in-Chief of the magazine. It is titled "Tensor and Coupled Decompositions: Interpretable Pattern Discovery in Multiset and Multimodal Functional Neuroimaging Data," and addresses an important challenge in neuroscience: joint analysis of neuroimaging signals from multiple subjects, multiple conditions, and/or multiple neuroimaging techniques.
With a consistently high impact factor, the IEEE Signal Processing Magazine is a leading journal in Electrical and Electronics Engineering and ranks among the top IEEE publications (signalprocessingsociety.org).
Read the full paper in the latest issue of the IEEE Signal Processing Magazine.
The reproducibility challenge
Functional neuroimaging tools, like functional Magnetic Resonance Imaging (fMRI), Electroencephalography (EEG), and Magnetoencephalography (MEG), are a “central window into our working mind”, observing the brain in action. These measurements are complex, and while combining them provides a more complete picture, it presents challenges.
The techniques record brain activity at different speeds and scales, and often measure different biological signals, such as blood flow in fMRI vs. electrical potentials in EEG (multimodal data). At the same time, the data varies from person to person and are analyzed as a collection (multiset data).
This complexity leads to a challenge in terms of the analysis of such data sets and capturing reproducible insights, with many studies relying on "fragile statistics and results that do not necessarily generalize".
The solution: tensor and coupled decompositions
The paper argues that tensor and coupled decompositions provide a powerful solution. These are advanced mathematical methods for data analysis that preserve the multi-dimensional structure of neuroimaging measurements. This allows them to discover interpretable patterns and provide "compressed and robust representations" of brain activity.
Unlike many conventional analyses, these decompositions are designed to find a unique solution. The paper emphasizes that this uniqueness is a "prerequisite for reliable and reproducible pattern discovery".
The survey goes beyond previous reviews by focusing on more advanced and flexible models, such as IVA (Independent Vector Analysis), PARAFAC2, and CGDs (Coupled Generator Decompositions). These methods are specifically chosen for their ability to handle the high variability in space and time that is characteristic of brain data.
The algorithm used for fitting the PARAFAC2 model was developed by SimulaMet PhD students Marie Roald and Carla Schenker.
Results: from theory to practice
To demonstrate the power of this framework, the authors applied these advanced methods to a real-world, public dataset of 16 subjects who had both EEG and MEG scans taken while viewing images of faces (famous, unfamiliar, and scrambled).
The team tested the models not only for reproducibility, getting consistent results across multiple runs, but also for replicability, whether the patterns hold up when the model is trained on a completely separate half of the data.
The paper demonstrates that the models were able to:
- Find consistent patterns: All three approaches (PARAFAC2, MSAA, and IVA) successfully extracted patterns that were both reproducible and replicable, demonstrating their robustness.
- Handle variability: The methods flexibly accounted for differences across subjects while still identifying shared, underlying brain activity.
- Provide interpretable insights: The discovered patterns corresponded to known brain functions, such as activation in regions known to be involved in face-recognition tasks.
By providing a comprehensive guide to these robust methods and demonstrating their practical success, the paper offers a clear path forward for researchers to "further our understanding of brain function in health and disease".