Constrained Independent Component Analysis Based on Entropy Bound Minimization for Subgroup Identification from Multi-subject fMRI Data

Hanlu Yang, Fateme Ghayem, Ben Gabrielson, Mohammad A. B. S. Akhonda, Vince D. Calhoun, Tülay AdaliICASSP 2023 · DOI

Introduction

Neuroimaging faces a significant challenge in precision medicine: identifying meaningful patient "subtypes" within complex brain disorders without requiring prohibitive computational resources. Traditional methods like Group ICA (GICA) are efficient but often overlook individual variability, while Independent Vector Analysis (IVA) preserves individual nuances at the cost of extreme computational complexity. This paper introduces c-EBM (constrained Entropy Bound Minimization), a novel algorithm designed to bridge this gap by identifying functional brain network subgroups with high precision and significantly improved speed.

Methodology

The c-EBM algorithm is built upon three core technical pillars:

  1. Entropy Bound Minimization (EBM): Unlike standard ICA, which assumes specific statistical distributions, EBM estimates an upper bound of entropy using flexible measuring functions. This allows for more accurate separation of complex, non-Gaussian brain signals.
  2. Decoupling: By avoiding the strict requirement for orthogonal demixing matrices, c-EBM solves for each brain source individually. This decoupling provides a more natural fit for overlapping brain networks.
  3. Spatial Constraints: To maintain consistency across subjects, the algorithm utilizes Pearson correlation with reference maps. This ensures that identified sources correspond to known functional networks while still permitting subject-specific variations.

Performance Evaluation

The researchers validated c-EBM using both synthetic data and the COBRE dataset, which includes resting-state fMRI from 88 patients with schizophrenia.

Quantitative performance of c-EBM compared to MOO-ICAR, showing higher correlation and lower joint ISI across varying noise levels.

As shown in Figure 1, c-EBM demonstrates superior fidelity in separating brain sources. It achieves higher correlation with true sources and lower joint inter-symbol interference (joint-ISI) compared to existing constrained methods like MOO-ICAR, particularly when reference maps are less than perfect.

Key Results and Biological Validation

The most significant breakthrough is the computational efficiency. While IVA required 74 hours to process 50 individuals, c-EBM analyzed 88 individuals in just 5 hours—a nearly 15-fold speedup.

Beyond speed, the method successfully identified three distinct subgroups within the schizophrenia cohort. These subgroups exhibited significant differences in the visual, primary sensory, and motor networks, replicating findings from more computationally intensive methods.

Workflow for subgroup identification and resulting brain network differences across identified clusters.

Conclusion

c-EBM offers a scalable solution for large-scale clinical fMRI studies. By combining flexible entropy modeling with spatial constraints, it preserves the individual variability necessary for subgroup identification while remaining computationally feasible for modern datasets. Future work will focus on adaptive parameter tuning and extending the method to other heterogeneous conditions such as autism and depression.

Summary updated

BibTeX
@inproceedings{yang2023constrained,
  author = {Hanlu Yang and Fateme Ghayem and Ben Gabrielson and Mohammad A. B. S. Akhonda and Vince D. Calhoun and Tülay Adali},
  title = {{Constrained Independent Component Analysis Based on Entropy Bound Minimization for Subgroup Identification from Multi-subject fMRI Data}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  doi = {10.1109/icassp49357.2023.10095816},
}
Figure