Modelling and prediction of the dynamic responses of large-scale brain networks during direct electrical stimulation

Our Nature Biomedical Engineering paper develops a technology to model and predict the dynamic response of multiregional brain networks during electrical brain stimulation. Our neurotechnology can enable personalized closed-loop deep brain stimulation for neurological and mental disorders.
Modelling and prediction of the dynamic responses of large-scale brain networks during direct electrical stimulation
Like

Our paper in Nature Biomedical Engineering1 is now published.

Millions of patients with neurological and mental disorders do not respond to current treatments – for example 20-30% in major depression2,3. As these disorders are linked to abnormal brain network activity, applying deep brain stimulation (DBS) has the potential to normalize this activity and provide an alternative therapy. Despite its promise, DBS has shown variable efficacy in mental disorders. The vision in my lab is to address this challenge by developing personalized closed-loop brain-machine interface (BMI) systems that tailor the DBS to a subject’s needs based on feedback of neural response to stimulation. My lab is in the Ming Hsieh Department of Electrical and Computer Engineering at the USC Viterbi School of Engineering and works at the interface of machine learning, control theory, and neuroscience to develop these closed-loop BMI systems2.

A major challenge for realizing such a system is that we don’t know how each patient’s brain activity and their symptoms would respond to stimulation. This makes it difficult to know how to adjust the dose of stimulation – that is, its amplitude or frequency – over time to tailor it to a patient’s needs. In our new Nature Biomedical Engineering1 paper, we resolve this standing challenge by developing novel electrical stimulation waveforms and new machine learning models of the effect of stimulation on brain networks.

I set out to develop these models over 6 years ago, back in 2014, when starting my own lab. The challenge at the time was that we had no mathematical framework for thinking about input-output modelling for the brain – with input being the stimulation and output being the brain network activity. So first, we developed a theoretical system identification framework for formulating this modelling. I recruited my PhD student Yuxiao Yang to start working on the problem. We hypothesized that ongoing stimulation would lead to a dynamic brain network response and thus designed dynamic input-output (IO) models. But a challenge was how to fit these models. We argued that obtaining informative datasets for model fitting requires delivering an input stimulation waveform that is “white-spectrum” in its amplitude and frequency so that it can sufficiently excite brain network activity. To achieve this, we designed a multi-level noise (MN)-modulated stimulation pulse train that stochastically switched its amplitude and frequency between discrete levels (Fig. 1a). We published an initial theoretical framework and numerical simulations in Journal of Neural Engineering4 in 2018 but of course this was just the beginning.

 

Fig. 1. a, We developed a novel stochastic MN stimulation waveform and machine learning algorithms to fit dynamic input-output models for the brain. b, Our models accurately predicted the single-trial neural response to stimulation across a distributed brain network.

To test and demonstrate our modeling and prediction framework in the brain, about 3 years ago, we started collaborating with Bijan Pesaran, Professor of Neural Sciences at NYU, who had developed a state-of-the-art brain network recording and stimulation hardware and thus performed the required new experiments with his postdoctoral scholar Shaoyu Qiao. We delivered continuous MN stimulation at a given brain site while simultaneously recording local field potential (LFP) activity across multiple brain regions. We then used the collected data to fit the dynamic IO models and evaluate them.

 Our results showed that the fitted dynamic IO models accurately predicted the single-trial dynamic neural response to stimulation for every stimulation region tried and outperformed any alternative model we built (Fig. 1b). Interestingly, even stimulating one site modulated the LFP activity in a distributed network of brain regions but with variable modulation strength. So we asked what explains this variability? We devised an at-rest functional controllability measure that computed the energy at the stimulation site to modulate each network site's LFP activity. Despite being based on at-rest data without stimulation, this measure explained the variability in modulation strength and IO prediction during stimulation. Finally, using extensive control experiments, we showed that our fitted dynamic IO models enable model-based closed-loop control of neural biomarkers.

 Together, the results in our paper establish the ability to predict how changes in stimulation would change multiregional brain network dynamics that underlie various mental states. But where do we go from here to realize the ultimate vision of a personalized closed-loop DBS system?

As I describe in my Perspective article in Nature Neuroscience2, realizing these closed-loop systems requires two major components: 1) A decoder that estimates mental states from large-scale brain activity in real time, 2) An input-output model to describe the effect of stimulation on this activity and thus mental states. We previously demonstrated the first component in a paper in Nature Biotechnology5 by developing machine learning algorithms in work that was spearheaded by my PhD students Omid Sani and Yuxiao Yang. Our algorithms achieved successful decoding of mood variations from intracranial human brain activity that was collected by Edward Chang’s lab at UCSF.  Now our Nature Biomedical Engineering1 paper demonstrates the second component. Together, these components can predict the effect of stimulation on mental state, and thus find the “right” stimulation parameters in real time to reach a target state2.

Of course, there is still much to be done. Our next steps are to validate these IO models in human patients – for which we already have promising results6 – and then to combine them with our decoder to develop a closed-loop personalized DBS system. We are motivated and hopeful that this research could lead to effective personalized therapies for tens of millions of patients with treatment-resistant mental disorders such as depression, chronic pain or addiction.

 References:

  1. Yang, Y. et al. Modelling and prediction of the dynamic responses of large-scale brain networks during direct electrical stimulation. Nature Biomedical Engineering 1–22 (2021) doi:10.1038/s41551-020-00666-w.
  2. Shanechi, M. M. Brain-machine interfaces from motor to mood. Nature Neuroscience 22, 1554–1564 (2019).
  3. Rush, A. J. et al. Acute and longer-term outcomes in depressed outpatients requiring one or several treatment steps: A STAR*D report. American Journal of Psychiatry 163, 1905–1917 (2006).
  4. Yang, Y., Connolly, A. T. & Shanechi, M. M. A control-theoretic system identification framework and a real-time closed-loop clinical simulation testbed for electrical brain stimulation. Journal of Neural Engineering 15, 066007 (2018).
  5. Sani, O. G. et al. Mood variations decoded from multi-site intracranial human brain activity. Nature Biotechnology 36, 954 (2018).
  6. Yang, Y. et al. A novel framework for dynamic modeling of brain-network response to electrical stimulation. in 2018 Computational and Systems Neuroscience (Cosyne) (2018).

 

Please sign in or register for FREE

If you are a registered user on Research Communities by Springer Nature, please sign in

Subscribe to the Topic

Biotechnology
Life Sciences > Biological Sciences > Biotechnology