TRANSFORM's goal is closed-loop brain stimulation: the creation of devices that can interpret and respond to brain activity in real time. Those devices depend on robust internal models of brain activity, psychological states, and the connection between the two. Our mathematical and computational neuroscience efforts under this program follow two major tracks: neural decoding and simulation/modeling.
Neural decoding has a long history in the brain-computer interface community, and has achieved some remarkable results in human motor control. However, those analyses are often driven more by algorithmic convenience than by principled understanding of the underlying statistics. Moreover, TRANSFORM faces a more complex challenge, in that we must model and decode the complex trans-diagnostic psychological domains, rather than simple and well-characterized movements in three-dimensional space.
Our decoding approach consists of a dual state-space formulation, following the framework that has been developed by Drs. Emery Brown and Uri Eden and applied to a variety of problems in neuroscience. Our psychological domains such as cognitive flexibility, reward motivation, or risk aversion are modeled as latent state variables that are assumed to be encoded in subjects' performance on behavioral task. Using a variety of Bayesian filtering approaches, we can recover an a posteriori estimate of those state variables at each timestep of the task. That same latent variable can then be considered to be encoded in the local field potential (LFP) and the spiking activity of recorded neurons. Using a very similar optimal filtering process, combined with our prior results in modeling spiking neurons as point processes having state-conditional intensity functions, we can estimate the same hidden state from ongoing observations of neural signals. Those state estimations form the heart of the device's controller, which can then seek to steer the state to more optimal regions through application of electrical stimulations.
The goal of the brain simulation and modeling team is to use computational methods to derive a better understanding of the mechanisms underlying the observed brain activity evoked by the TRANSFORM task battery. To do so, we start with observations of the brain's dynamic activity, which consists of rich spatiotemporal patterns propagating within the complex brain anatomy. As part of the TRANSFORM-DBS project, large spatiotemporal data sets are being created that span many different recording modalities. To understand these data, modeling approaches that mathematically express scientific knowledge and rigorously connect the observed activity with complex patterns and hidden biological mechanisms are required. This models will assist the control engineers in designing optimal interventions and will help provide a biological understanding for the features observed in the decoding analyses.
To understand the brain activity associated with severe psychiatric illness, we implement two main categories of modeling: physical modeling, and statistical modeling. Physical models consist of mathematical formulations of biological processes. For example, the Hodgkin-Huxley equations are a system of differential equations that relate the opening and closing of ion channels in an individual neuron to the voltage dynamics of action potential generation. These models quickly become extraordinarily complex when the goal is an accurate representation of brain activity across multiple areas and types of neurons. Statistical models are often descriptive, without an immediate connection to biological processes. For example, neuronal spiking activity can be characterized as a process of biased “coin-flips”, where the goal is to predict the outcome (such as “heads”) at each moment in time. Both physical and statistical modeling have rich histories in neuroscience, and provide explicit, quantitative representations of our knowledge of neuronal activity.
With the complexity of neurological data recorded by the TRANSFORM team - and the wealth of scientific knowledge possessed by the team members – the brain simulation team implements an approach that combines both physical and statistical reasoning. Using this framework, we will seek new insights into the mechanisms that support complex cognitive tasks. That, in turn, will allow more rational design of the brain stimulation interventions used by the final device. Ultimately, we hope that this better mathematical understanding of the brain's activity will help improve the care of patients with severe psychiatric illness.