Jennifer Stiso's Publications

A schematic of a model of sequence learning. Sequences can be generated from a latent space of items, and potential transitions between them. For people experiencing these sequences, a reconstruction of the original latent space can help make predictions about which items come next. The structure of these reconstructing varies based on how much temporal discounting is employed when learning the sequence. Neurophysiological evidence of cognitive map formation during sequence learning eNeuro, (2022). Humans are adept at learning the statistics of sequences. This ability is facilitated by learning a latent space of transition probabilities between items, or a cognitive map. However, work testing explicit theories of how these maps are built, vary across individuals, and are reflected in neural activity is sparse. We use a model that infers an individual’s cognitive map from sequential reaction times and intracranial encephalography (iEEG) recordings to address these gaps. We find that neural activity in the temporal lobe most often reflects the structure of maps and easily identifies task-relevant features of the latent space. We also identify features of individual learning strategies and latent spaces that influence how quickly maps are learned. These discoveries advance our understanding of humans’ highly generalizable ability to learn spaces.

A schematic processing pipeline for iEEG data. The pipeline includes IED detection, artifcat rejection, and functional connectivity calulation with different measures and bands. Functional connectivity is then averaged into different groups and assessed for time windows with, and without IEDs Fluctuations in functional connectivity associated with interictal epileptiform discharges (IEDs) in intracranial EEG bioaRxiv, (2022). We quantify how much functional connectivity changes during IEDs. We quantify changes in 5 functional connectivity measures (orthogonal amplitude envelope correlation, imaginary coherence, imaginary phase-locking value, autoregressive model fit, and cross correlation) in 5 differet frequencey bands in a large sample of 145 individuals undergoing iEEG monitoring. Epliepsy is an extremely heterogeneous disease, but is characterized by a propensity for periods of hyper synchronous activity known as seizures. In order to better understand this underlying propensity for synchrony, it would be useful to characterize how global measures of synchrony change during subseizure epileptic activity, or interictal epileptiform dishcarges (IEDs). However, it has been difficult to obtain large samples of intracranial neurophysiology data required ot ansewr this question directly. Here, we use a large sample of publicly available data demonstrate consistent increases in functional connectivity during IEDs. Our results demonstrate global changes to neural dynamics during focal IEDs and inform future modeling studies on basic properties of synchrony in epilepsy.

A schematic analysis pipeline. MEG data is segmented into 1s windows, and functional connectivity is calulated in each window. Each window is then pair with its correponding BCI performance measure. Non-negative matrix factorization is then used to decompose bothe brain connectivity and behavior into subgraphs with temporal expression. The method also gives each subgraph a quantitative association with behavior called the performance loading. Learning in brain-computer interface control evidenced by joint decomposition of brain and behavior Journal of Neural Engineering, (2020). We identify and characterize separate groups of interacting brain areas that predict learning in a brain computer interface (BCI) task and suggest a key role for networks that support learning in sustaining attention. We employ non-negative matrix factorization to obtain a single, data-driven decomposition of brain and behavioral data taken during BCI learning and use models from network control theory to test theories about their role. BCIs have marked potential as a therapy to restore neural function, but their application is limited by the difficulty, and heterogeneity of learning to reliably control a BCI. Previous work has demonstrated that neural activity in diverse cognitive systems predicts learning across individuals, but is unclear how these different systems interact through time to cohesively support the volitional control of neural activity. In 20 individuals learning to control a BCI, we investigate time varying interactions across the whole brain that predict learning, and find a separation of different neural process that work to selectively facilitate sustained attention. Our results inform avenues of research seeking to identify candidates for BCI based therapies based on their time varying brain activity.

A schematic of the network control theoretic model used to model stimulation. Data include network estimates of white matter connectivity and neural activity recordings from iEEG. The model them defines future brain activity (x(t+1)) as the current activity (x(t)) spread along white matter connections (A), and added to any external stimulation (Bu(t)). The formal model of activity is given as x(t+1) = Ax + Bu(t). White matter network architecture guides direct electrical stimulation through optimal state transitions Cell Reports, (2019). A mathematical model of stimulation spread across white matter tracts in the brain makes predictions about which regions, and patterns of brain activity will be eneficial for stimulation targetted to improve memory. We use electrocorticographic recordings of brain activity from a stimulation experiment conducted in epilepsy patients to show how important an individual’s patterns of white matter connections is for determining how their brain responds to stimulation and how they influence predictions of when and where to stimulate. Brain stimulation is an effective therapy for depression, epilepsy, Parkinson’s, and other neurological disorders. Personalizing and optimizing these therapies to treat other disorders is an active area of research, but progress is hampered by how difficult it is to predict how the entire brain will respond to stimulation at a given region. We used features of an individual’s brain activity, and pattern of connections between brain regions to model what happens to the whole brain when we stimulate a specific region. This work is an important, early step towards developing a fast, generalizable model of an individual’s response to a specific stimulation therapy.

Spatial embedding imposes constraints on neuronal network architectures Trends in Cognitive Science, (2018). Recent progress towards understanding circuit function has capitalized on tools from network science to parsimoniously describe the spatiotemporal architecture of neural systems. Such tools often address systems topology divorced from its physical instantiation. Nevertheless, for embedded systems such as the brain, physical laws directly constrain the processes of network growth, development, and function. We review here the rules imposed by the space and volume of the brain on the development of neuronal networks, and show that these rules give rise to a specific set of complex topologies. These rules also affect the repertoire of neural dynamics that can emerge from the system, and thereby inform our understanding of network dysfunction in disease. We close by discussing new tools and models to delineate the effects of spatial embedding.

PDFs are provided to ensure timely dissemination of academic work. They can be downloaded for noncommercial, information purposes only, and may not be reposted without permission from the respective copyright holders.