Implementing evolutionary optimization to model resting state functional connectivity (2019)
Computational models are crucial in understanding brain function. One particularly interested class of models is designed to replicate known brain structures using structural connectivity (SC) measured with diffusion spectrum imaging, and the behavior that emerges is then compared to empirical functional connectivity (FC), measured using functional magnetic resonance imaging. As the models become more accurate and more complex with more parameters, they can explain more of the observed phenomena, and may eventually be used for diagnosis and design of treatments of brain disorders. However, those parameters need to be carefully optimized for the models to work best, which becomes intractable to do manually or with traditional computational techniques like gradient descent as the models grow. Evolutionary computation techniques use adaptation over generations in nature as inspiration to optimize functions that are otherwise difficult and have succeeded in solving optimizations in similarly difficult search spaces (Miikkulainen, 2019). In this work, the Covariance Matrix Adaptation Evolutionary Strategy (CMA-ES) was configured to optimize continuous parameters of a large-scale biophysical network model that stimulates neural activity from structural connectivity of 66 cortical parcellations and computes FC (Deco et al., 2013). Empirical SC and FC data were collected and aggregated from 24 right-handed healthy young volunteers (Deco et al., 2013). The application of CMA-ES on this data resulted in a significantly better fit to empirical FC data than manually selected parameters in all four trials run of the CMA-ES algorithm so far. The best parameter set from the best trial run generates a functional connectivity matrix with a correlation of 0.5205 +- 0.0102 with empirical FC, compared to the previous best set of manually selected parameters that yield a correlation of 0.4647 +- 0.0114. This work provides a basis for utilizing evolutionary computation to optimize neural activity models. Statistically significant improvement has already been achieved with a relatively simple approach. To further increase the optimization potential, this approach will be combined with other evolutionary computation techniques to optimize other parameters, and these techniques will be scaled up to more detailed and patient-specific SC and FC data and more complex computational models. Optimizing neural activity models will further our understanding of the human brain and bring the field of neurology closer to personalized medicine.
Citation:
In Society for Neuroscience Abstracts, 2019. Society for Neuroscience.
Bibtex:

Presentation:
Slides (PDF)
Kaitlin Maile Former Ph.D. Student kmaile [at] cs utexas edu
Risto Miikkulainen Faculty risto [at] cs utexas edu
Manish Saggar Ph.D. Alumni saggar [at] stanford edu