EP4665234A2 - Devices and methods for identifying brain stimulation targets - Google Patents

Abstract

Methods for determining target locations for brain stimulation and for stimulating brain regions are disclosed. For each of a plurality of target stimulation locations of the brain, distance to a nearest white-matter vertex of a plurality of white-matter vertices of the brain is determined. The plurality of target stimulation locations are located at a lateral temporal cortex of the brain. A first subset of target stimulation locations is selected that are near a respective nearest white-matter vertex based on the determined distance. A functional brain connectivity quantitative model is received for the plurality of target stimulation locations. The functional brain connectivity quantitative model is built based on recordings of electrical signals collected during performance of one or more memory tasks. A second subset of target stimulation location are selected from the first subset that exhibit theta-connectivity to other electrodes in a montage of the brain based on the functional brain connectivity quantitative model. Stimulation is applied to the brain to improve cognitive function.

Description

Devices and Methods for Identifying Brain Stimulation Targets

[0001] This invention was made with government support under HR0011-41-1-2301 and N66001-14-2-4032 awarded by the Department of Defense. The government has certain rights in the invention.

TECHNICAL FIELD

[0002] The present invention generally relates to a method and apparatus for improving cognitive performance through the application of brain stimulation.

BACKGROUND

[0003] Direct electrical stimulation of the human brain can manipulate circuits underlying perception, cognition, and action (Siddiqi et al. 2022; Scangos et al. 2021b). Such stimulation has been used to treat network syndromes of brain dysfunction, suggesting that stimulation influences a broader network of brain regions beyond the stimulated location (Mayberg et al. 2005; Scangos et al. 2021; Limousin et al. 1998; Bouthour et al. 2019; Deuschi et al. 2006; Geller et al. 2017; Jobst et al. 2017; Lozano and Lipsman 2013). Stimulation, whether closed-loop or open-loop electrical stimulation, can also modulate behaviors, such as learning and memory, that depend on the coordinated activity of a network of brain regions (Mankin and Fried 2020; Das and Menon 2021; Voytek and Knight 2015; Keerativittayayut et al. 2018; Staresina and Wimber 2019).

[0004] Recent studies indicate that direct electrical stimulation of the brain, when applied or optimized in a closed-loop manner, can lead to improvements in memory and mood. In these approaches, brain states control the timing of stimulation or the choice of parameters, such as amplitude or frequency, because stimulation’s neural and behavioral effects largely depend on brain state at the time of delivery. While electrical stimulation is increasingly used as a therapeutic and experimental tool, variability in outcomes poses a critical challenge, in part because stimulation’s mechanisms of action remain poorly understood. Theoretical accounts evolved from models of local disruption of pathological activity (Benabid et al. 2004) to modulation of the broader network of areas connected to the stimulated location (Ashkan et al. 2017; McIntyre and Hahn 2010). Although closed-loop algorithms provide greater control over stimulation’s physiological and behavioral effects, variability in outcomes due to the unique anatomical and functional network properties of individual stimulation targets remains a primary challenge for using stimulation for clinical and experimental neuromodulation. SUMMARY

[0005] Accordingly, there is a need to reduce variability in outcomes of brain stimulation and more reliably improve function in individuals who are subject to electrical brain stimulation. One reason for the variable effects of stimulation on memory, emotion and cognition is that these functions depend on the interactions among multiple brain areas, and variability in the location being stimulated produces variable effects on the network. These variable effects are challenging to predict based on a strictly anatomical consideration for a stimulation target selection. For example, anatomical (or structural) connectivity of the brain is an insufficient predictor for selecting brain stimulation targets that if stimulated would reliably improve brain function.

[0006] In some embodiments, closed-loop stimulation therapy is improved by using widespread brain recordings to determine the target locations most likely to produce desired network effects. In some embodiments, electrical signals are recorded from multiple targets throughout the brain to build quantitative models of the functional connectivity among these regions. In some embodiments, electrical signals are recorded both during rest periods and during performance of tasks that tap into the desired behavior (e.g., learning, memory, executive function, or emotional processing). In some embodiments, improved therapeutic effects are achieved when the stimulation target locations exhibit strong functional connectivity to the network of brain areas supporting respective behavioral states. In some embodiments, networks of 3-13 Hz theta frequency connectivity are mapped using signal processing measures such as phase coherence, synchrony, and time-lagged correlations. In some embodiments, this class of quantitative methods, when applied to electrical signals from the brain, demonstrate strong positive correlations that rise during successful performance across a variety of cognitive tasks.

[0007] In some embodiments, the disclosed methods and systems reduce the variability of effects of closed-loop stimulation therapy that aims to improve memory and learning and/or other brain functions.

[0008] In accordance with one embodiment, a method for determining target locations for brain stimulation is performed. The method includes, for each of a plurality of target stimulation locations of the brain, determining distance to a nearest white-matter vertex of a plurality of white-matter vertices of the brain. The plurality of target stimulation locations are located at a lateral temporal cortex of the brain. The method further includes, selecting a first subset of target stimulation locations from the plurality of target stimulation locations that are near a respective nearest whitematter vertex based on the determined distance. The method further includes, receiving a functional brain connectivity quantitative model for the plurality of target stimulation locations, wherein the functional brain connectivity quantitative model is built based on recordings of electrical signals collected during an encoding phase. The method further includes, selecting a second subset of target stimulation locations from the first subset that exhibit theta-connectivity to other electrodes in a montage of the brain.

[0009] In a further embodiment, the method includes stimulating the brain electrically at the selected second subset of target stimulation locations.

[0010] In a further embodiment, stimulating the brain electrically includes delivering a closed- loop stimulation at the selected second subset of target stimulation targets, wherein one or more multivariate classifiers are used to determine timing and/or parameters of the stimulation.

[0011] In a further embodiment, the recordings of electrical signals are collected during the encoding phase of a memory task.

[0012] In a further embodiment, the method includes, for each target stimulation location, calculating a node-strength measure based on coherence between each pair of bipolar channels in the montage.

[0013] In a further embodiment, coherence C_xy between a pair of signals at electrodes x and j' sxy corresponds to normalized cross-spectral density, where C_xy = and Sxyis the cross-spectral sxx^syy density between the signals at the electrodes x and y, and where S_xx corresponds to auto-spectral density at the electrode x and S_yy corresponds to auto-spectral density at the electrode y.

[0014] In a further embodiment, the one or more multivariate classifiers are trained based on whole-brain patterns of spectral activity to discriminate encoding activity associated with recalled words and words that are not recalled.

[0015] According to another aspect of the invention, a method for delivering stimulation to a brain to improve cognition and/or memory is provided, the method comprising selecting one or more white-matter tracts of the brain, and delivering stimulation to the one or more white-matter tracts. In a further embodiment, the stimulation is electrical stimulation.

[0016] In a further embodiment, the one or more white-matter tracts is located within a middle temporal gyrus. In still a further embodiment, the middle temporal gyrus is a left middle temporal gyrus of the brain. In a further embodiment, the one or more white-matter tracts include at least one of an inferior longitudinal fasciculus and a superior longitudinal fasciculus. [0017] In an embodiment, the stimulation delivered to the one or more white-matter tracts is closed-loop stimulation, wherein one or more multivariate classifiers are used to determine timing and/or the parameters of the stimulation (e.g., amplitude, frequency, pulse width).

[0018] In an embodiment, the stimulation delivered to the one or more white-matter tracts is open-loop stimulation, wherein one or more multivariate classifiers and/or cognitive testing are used to determine the parameters of the stimulation (e.g., amplitude, frequency, pulse width).

[0019] In an embodiment, the electrical stimulation is delivered to the brain at from about

0.1mA to about 5.0mA. In a further embodiment, the electrical stimulation is delivered to the brain at about 1.0mA.

[0020] In an embodiment, the electrical stimulation is delivered to the brain at from about 20 Hz to about 200 Hz. In a further embodiment, the electrical stimulation is delivered to the brain at about 200 Hz.

[0021] In an embodiment, the electrical stimulation is delivered to the brain at a pulse width from about 15 psec to about 500 psec. In a further embodiment, the electrical stimulation is delivered to the brain at a pulse width of about 300 psec.

[0022] Note that the various embodiments described above can be combined with any other embodiments described herein. The features and advantages described in the specification are not all inclusive and, in particular, many additional features and advantages will be apparent to one of ordinary skill in the art in view of the drawings, specification, and claims. Moreover, it should be noted that the language used in the specification has been principally selected for readability and instructional purposes, and may not have been selected to delineate or circumscribe the inventive subject matter.

BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS

[0023] The following detailed description of embodiments of the invention will be better understood when read in conjunction with the appended drawings of an exemplary embodiment. It should be understood, however, that the invention is not limited to the precise arrangements and instrumentalities shown.

[0024] In the drawings:

[0025] Figure 1 A illustrates an example of performing an encoding task and training one more multivariate classifiers, according to at least one embodiment. Participants performed at least three sessions of the free recall task while being monitored with intracranial EEG. Multivariate classifiers were trained on whole-brain patterns of spectral activity to predict subsequently recalled vs. not recalled words.

[0026] Figure IB illustrates examples of recording electrode locations in a left and right lateral temporal cortex for multiple participants, according to at least one embodiment. Recording electrode locations for all participants in the Closed-loop (blue) and Random (green) groups, rendered on the Freesurfer average brain.

[0027] Figure 1C depicts each participant’s multivariate classifier serving as their personalized model to trigger stimulation. Classifiers trained on record-only data generalized to the stimulation session(s) for the Closed-loop group (P = 6.14 * 10-7) and outperformed classifiers for the Random group (P = 2.73 x 10-5).

[0028] Figure ID illustrates an analysis of feature importance for classifiers from the Closed- loop group and shows that successful memory states were associated with decreases in low- frequency activity and increases in high-frequency activity.

[0029] Figure 2 is a conceptual diagram illustrating an exemplary system that can monitor brain signals and/or deliver stimulation to a user to improve brain functionality (e.g., memory) of the user according to at least one embodiment.

[0030] Figures 3A and3B illustrate a flow chart of a method for determining target locations for brain stimulation, according to at least one embodiment.

[0031] Figure 4A depicts stimulation target locations for the Closed loop (blue) and Random (green) groups.

[0032] Figure 4B illustrates the Closed-loop stimulation strategy, based on detecting poor memory encoding states and intercepting them with stimulation.

[0033] Figure 4C illustrates that Closed-loop LTC stimulation improved memory performance (P=0.01) while random stimulation did not. Error bars in C reflect standard error of the mean.

[0034] Figure 5A illustrates that for the Closed-loop group, the effect of stimulation on memory depended on the target distance from the nearest white matter [left, P=0.007], The correlation was not significant for the Random group [right, P = 0.52], Error regions reflect the standard error of the estimate.

[0035] Figure 5B shows that Closed-loop LTC stimulation improved memory performance for targets located nearest to white matter (P=0.005). There was no effect for the Random group (P=0.62). Error bars reflect standard error of the mean. [0036] Figure 6A shows the assignment of each patient’s record-only electrodes to two Regions of Interest (ROIs) based on whether the electrode was located in a region that showed a memory- related spectral tilt or not, to isolate the brain’s memory encoding network.

[0037] Figure 6B illustrates that low-frequency connectivity was higher between the stimulation target and electrodes in classifier-defined memory regions, compared to electrodes in other regions (P=0.02) and compared to high-frequency network connectivity (P=0.02). In contrast, there was no difference in stimulation target high-frequency network connectivity.

[0038] Figure 6C illustrates that for closed-loop targets nearest to white matter, there was a significant correlation between stimulation target low-frequency connectivity and stimulation’s effect on memory [ = 0.69, P = 2 x 10-5], There was no effect for high-frequency connectivity. Error bars reflect standard error of the mean. Error regions reflect the standard error of the estimate. [0039] Figure 7A depicts a schematic of the analysis of stimulation-evoked physiology.

[0040] Figure 7B shows that, for stimulation targets near white matter, low-frequency functional connectivity predicted the stimulation-evoked change in low-frequency power (P = 0.02) [left side]. Additionally, high-frequency network connectivity did not predict stimulation’s effect [right side], [0041] Figure 8A illustrates an example of performing an encoding task and training one more multivariate classifiers, according to at least one embodiment.

[0042] Figure 8B illustrates examples of recording electrode locations in a left and right lateral temporal cortex for multiple participants, according to at least one embodiment.

[0043] Figure 8C illustrates stimulation-related memory change. Closed-loop stimulation of LTC improved recall performance (P=0.02) while random stimulation did not.

[0044] Figure 9A shows stimulation target locations. Targets in the LTC were stimulated either in closed-loop (blue) or randomly (green).

[0045] Figure 9B depicts how distance to white matter shown in mm affects stimulation-related memory change.

[0046] Figure 9C shows how distance to white matter categorized as “near,” “middle,” or “farthest” affects stimulation-related memory change.

[0047] Figure 9D shows how stimulation target node strength affects stimulation-related memory change.

[0048] Figure 10A shows that stimulation of theta hubs reduced downstream theta (4-8 Hz) power [r(36) = -0.475, P = 0.003],

[0049] Figure 10B shows that stimulation improved memory when it reduced theta power [r(36) = -0.522, P = 0.0008], DETAILED DESCRIPTION

[0050] I. Overview

[0051] The human brain is composed of billions of neurons electrically interconnected and organized into various areas to perform a variety of functions. The electrical activation and/or deactivation of neurons or groups of neurons is largely responsible for the function of the brain and communication among the various areas of the brain along the networks. It is generally thought that the activation of numerous neurons, among other types of network function, may be necessary to carry out each brain function when performing a cognitive task (e.g., creating a memory, solving a puzzle, recall of earlier memorized information, etc.).

[0052] In some instances, brain stimulation can be therapeutically applied in order to prevent the onset of or treat an undesirable state, such as in the cases of epilepsy or tremors that may be associated with Parkinson’s Disease. Direct electrical stimulation of the human brain is a powerful method for manipulating neural circuits underlying perceptual, motor, and cognitive systems. Focal electrical stimulation has been used to treat non-focal syndromes of brain dysfunction, such as depression and epilepsy, suggesting that stimulation influences a broader network of brain regions beyond the stimulated location. In other instances, beyond these clinical syndromes, several studies show that direct electrical stimulation can also modulate the brain’s ability to learn and remember, which depends on the coordinated activity of several key brain regions.

[0053] Although direct electrical stimulation of the human brain is increasingly used as a therapeutic and experimental tool, variability in outcomes continues to be a critical challenge. For example, stimulating the brain during episodic memory has led to reports of both memory enhancement as well as memory disruption. One reason is that stimulation’s mechanisms of action remain poorly understood. Following the success of stimulation for Parkinson’s, theoretical accounts of stimulation’s effects evolved from models of local disruption of pathological activity to modulation of the broader network of areas linked in circuit with the stimulated location. Because stimulation’s effects can be interpreted at the network level, variability in individual network structure could explain the variability in physiological and behavioral outcomes observed in the literature.

[0054] Several studies have sought to reveal how stimulation affects broader brain networks. Anatomically, whether stimulation targets gray matter, the gray-white matter boundary, or specific white-matter fibers determines the spread of physiological effects through the network. Previous research has demonstrated different excitation thresholds for neural elements in white and gray matter (Nowak and Bullier 1998), which may explain variability in the spatial extent over which stimulation exerts its effects (Histed et al. 2009). Compared to gray matter stimulation, white-matter stimulation leads to more broadly distributed excitation in downstream areas. White-matter-defined anatomical connectivity also constrains how stimulation affects the brain’s functional state. Behaviorally, stimulation of white matter has produced remission in depression and improvement in episodic memory. However, previous research has yet to show that variability in stimulation’s downstream effects depends on white vs. gray matter targeting in a way that predictably modulates episodic memory performance.

[0055] Beyond anatomical targeting of white-matter pathways, functional architecture and connectivity has been shown to mediate direct brain stimulation’s effects on the broader brain network, including the spread and persistence of stimulation’s physiological effects (Keller et al. 2011; Fox et al. 2020, 2014; Keller et al. 2018). Previous research further suggests this relation to be frequency-specific. For example, stimulating targets in the medial temporal lobe has been demonstrated to lead to greater downstream changes in low-frequency (5-13 Hz) activity in brain regions that are strongly connected, at low frequencies, to the stimulated site (Solomon et al. 2018). There are a variety of cognitive functions, including episodic memory, that have been linked to modulation of low-frequency activity (Colgin 2013; Burke et al. 2013; Donoghue et al. 2020; Koster and Gruber 2022; Griffiths et al. 2021). Thus, stimulating targets with strong low-frequency network connectivity is expected to reliably modulate such behaviors. For example, stimulation targets that are network hubs, defined by low-frequency functional connectivity, produce greater downstream changes in low-frequency activity when stimulated. Functional connectivity also explains the persistence of stimulation-evoked changes in neural excitability. Although the preceding data suggest a role for anatomical and functional networks in mediating stimulation’s effects, previous work has yet to link stimulation of functional and anatomical networks, evoked physiology, and modulation of behavioral performance. A demonstration that such stimulation does affect broad low-frequency activity in a way that is related to behavior would be consistent with the notion that low-frequency activity coordinates function across a distributed neural network. Such coordination may be especially important for dynamic cognitive functions like episodic memory (Solomon et al. 2017; Watrous et al. 2013) and suggests that low-frequency activity may be a more effective target for modulation with stimulation than high-frequency activity (Fries 2009; Harris and Gordon 2015; Fell and Axmacher 2011). As set forth herein, in some embodiments, anatomical and functional characteristics of the stimulation target are used as variables that control the effect of stimulation on the brain’s memory network.Figure 1 A illustrates an example of performing an encoding task and training one or more multivariate classifiers, according to one embodiment. There are at least three stages associated with memory, including memory encoding, memory retrieval and vocalization. During the encoding phase, a participant studies a list of words. In some embodiments, the word lists that are studied can be of various lengths, e.g., 10, 11, 12, 13, 14, and/or other lengths. In some embodiments, the words are presented to a participant in their respective native language. In the standard free recall task, words are selected randomly from a pool of common nouns. In the categorized free recall task, the word pool is constructed from different semantic categories (e.g., fruit, furniture, office supplies, and/or other semantic categories). After words are in the presented lists are studied, participants perform a distractor task, e.g., to attenuate the recency effect in memory (e.g., recency effect may last 20 seconds). In some embodiments, the distractor task is performed immediately following the final word in a respective list of words to be studied. Various distractor tasks can be performed. For example, a distractor task can include a series of arithmetic problems of the form A+B+C=??, where A, B and C are randomly chosen integers ranging from 1 to 9. In some embodiments, following the distractor task, participants are given optionally 30 seconds to verbally recall as many words as possible from the list in any order. In some embodiments, vocal responses are digitally recorded and later manually scored for analysis. In some embodiment, a respective session of this encoding-di stractor-recall procedure is performed for a number of lists of words, such as 10, 15, 20, 25, or 30 lists of words.

[0056] In some embodiments, participants perform a number of sessions (e.g., at least three) of a free recall task (e.g., standard or categorized) while being monitored with intracranial electroencephalogram (EEG). In some embodiments, multivariate classifiers are trained using whole-brain patterns of spectral activity to discriminate encoding activity associated with recalled vs. not recalled words.

[0057] Figure IB illustrates examples of recording electrode locations in a left and right lateral temporal cortex for multiple participants rendered on the Freesurfer average brain, according to one embodiment. In some embodiments, the dark grey (or blue in color) electrode locations are associated with participants in a closed-loop group and light grey (or green in color) electrode locations are associated with participants in a random or open-loop group. In some embodiments, deep brain stimulation (DBS) can be classified into open-loop (also known as conventional) and closed-loop (also known as adaptive). Closed-loop DBS employs a sensor to record a signal linked to symptoms while open-loop DBS does not use a sensor for recording the brain condition; therefore, stimulation parameters including duration, amplitude, and frequency of the pulse train remain constant in open-loop DBS regardless of fluctuations in the disease state. The recorded signal is known as a biomarker and can have varying nature, e.g., bioelectric, physiologic, biochemical, etc. In the open-loop DBS, a specialist tracks the patient’s clinical state and manually programs the device in a trial-and-error based manner. Adjustments of stimulation parameters are not conducted in real-time based on the ongoing neurophysiological variations in the brain; therefore, adverse effects on the patient may be induced due to the brain overstimulation. On the other hand, in the closed-loop DBS, the stimulation pulses are delivered when the brain is in an abnormal state, or they are automatically and dynamically adjusted based on the variations in the recorded signal over the time.

[0058] In some embodiments, a method for identifying potential locations of stimulation is implemented (e.g., instead of focusing on a specific location/region in the brain, states of the brain, and/or additive or cumulative effects of location of stimulation targets and brain states during stimulation). In some embodiments, potential locations for brain stimulation are identified based on physical (anatomical) and functional (electrical) connectivity. In some embodiments, anatomical and functional network properties of the stimulation target are used to predict stimulation’s behavioral and physiological effects. In some embodiments, closed-loop stimulation is delivered via intracranially implanted electrodes as neurosurgical patients perform a free recall memory task. Multivariate classifiers, which were trained to predict momentary lapses in memory function, trigger stimulation of the lateral temporal cortex (LTC) during the encoding phase of the task. Activity in the LTC correlates with episodic memory performance (Burke et al. 2014; Kim 2011; Kragel et al. 2017; Ojemann et al. 1988), and stimulation studies targeting this area suggest it may be an effective node for modulating the memory network (Bickford et al. 1958; Boggio et al. 2009; Curot et al. 2017; Ezzyat et al. 2018; Fl del et al. 2008; Kucewicz et al. 2018; Moriarity et al. 2001; Perrine et al. 1994). Stimulating LTC locations with high functional network connectivity leads to greater physiological and behavioral effects of stimulation. Further, stimulation improves memory performance when delivered to targets near white-matter pathways. These data suggest that in order to use stimulation effectively as a therapy for memory dysfunction, structural and functional characteristics of the stimulation target can be used to predictably modulate physiology and behavior (Ezzyat and Suthana In Press).

[0059] II. System Overview

[0060] Figure 2 is a conceptual diagram illustrating an exemplary system 200 that can record brain signals and/or deliver stimulation to a user to improve brain functionality (e.g., memory) of the user according to at least one embodiment. [0061] In some embodiments, electrophysiological data are collected from electrodes such as electrode 202 that are implanted subdurally (e.g., grid/strip configurations) on a cortical surface and/or electrodes within the brain parenchyma. In some embodiments, the placement of the electrodes is determined based on the epileptogenic monitoring needs of a patient. In some embodiments, the electrophysiological data are recorded by recording EEG system 220. In some embodiments, different recording EEG systems 220 can be used, such as Nihon Kohden EEG- 1200, Natus XLTek EMU 128 or Grass Aura-LTM64. In some embodiments, recording EEG system 220 is connected to the electrodes wirelessly or through a wired communication link. In some embodiments, after EEG data is recorded and/or collected by the recording EEG system 220, the EEG data can be stored locally in the recording EEG system 220 or externally, e.g., at computer system 260 or other storage medium. In some embodiments, recorded EEG data is sampled at 500 Hz, 1000 Hz, or 1600 Hz.

[0062] In some embodiments, the collected EEG data is analyzed by the computer system 260. For example, record-only sessions (e.g., sessions that record brain activity during performance of memory task without sessions that also include brain stimulation) are used as input to one or more logistic regression classifiers trained to discriminate encoding-related activity predictive of whether a word was later recalled or not recalled. In some embodiments, each of the one or more logistic regression classifiers are trained on EEG data specific for each participant. For example, participantspecific multivariate classifiers are trained to discriminate patterns of neural activity during record- only sessions of free recall.

[0063] In some embodiments, after participant-specific multivariate classifiers are trained on record-only data, the participant-specific multivariate classifiers are used to determine when to trigger stimulation during subsequent (independent) sessions. In some embodiments, stimulation strategy intercepts and rescues periods of poor memory encoding based on the predictions of the trained participant-specific multivariate classifiers. In some embodiments, stimulation of respective participants is optionally performed. Further, EEG data is recorded and analyzed for sessions where stimulation is performed and for sessions where stimulation is not performed.

[0064] In some embodiments, the stimulation is performed by stimulation system 240, such as External Neural Stimulator (ENS) (e.g, from Medtronic, Inc.). In some embodiments, recording and stimulation equipment can be obtained from Blackrock Microsystems. In some embodiments, computer system 260 (e.g., a laptop, a desktop computer, or other computer) records behavioral responses (vocalizations, key presses, etc.), synchronized to the ENS-recorded EEG via transmitted network packets. In some embodiments, behavioral responses are recoded during testing sessions. [0065] FIG. 3 illustrates a flow chart of a method 300 for determining (or identifying) target locations for brain stimulation, according to at least one embodiment.

[0066] For each of a plurality of target stimulation locations of the brain, distance to a nearest white-matter vertex of a plurality of white-matter vertices of the brain is determined (302). The plurality of target stimulation locations are located at a lateral temporal cortex of the brain.

[0067] Calculation of stimulation target distance to white matter

[0068] In some embodiments, brain scan segmentations were used to determine white-matter vertex locations. For example, Freesurfer method was used to segment participants’ T1 MRI scan to identify white-matter vertex locations, and the distance between the stimulation location and the nearest white-matter vertex is calculated based on the identified white-matter vertex locations. In some embodiments, respective distances between a target stimulation location and a nearest whitematter vertex are split into thirds to categorize target stimulation locations or sites as “near” (i.e., less than about 1mm in distance), “middle” (i.e., between about 1mm and 2mm in distance), or “far” (i.e., about 2mm or greater in distance) relative to the nearest white matter.

[0069] In some embodiments, to determine distance to a nearest white-matter vertex from a target stimulation location, an anatomical localization of the brain is performed.

[0070] Anatomical localization

[0071] In some embodiments, cortical surface regions are delineated on pre-implant whole brain volumetric T1 -weighted MRI scans using Freesurfer according to the Desikan-Kiliany atlas. In some embodiments, whole brain and high-resolution medial temporal lobe volumetric segmentation is also performed using the Tl-weighted scan and a dedicated hippocampal coronal T2-weighted scan with Advanced Normalization Tools (ANTS) and Automatic Segmentation of Hippocampal Subfields (ASHS) multi-atlas segmentation methods. In some embodiments, coordinates of the radiodense electrode contacts are derived from a post-implant CT and then registered with the MRI scans using ANTS. Subdural electrode coordinates are further mapped to the cortical surfaces using an energy minimization algorithm. In some embodiments, one or more neuroradiologists can optionally review cross-sectional images and surface renderings to confirm the output of the automated localization pipeline. In some embodiments, stimulation targets that are localized to the left inferior, middle, or superior temporal gyri (left or right hemispheres) are classified as LTC.

[0072] In some embodiments, recorded EEG data recorded by recording EEG system 220 (Figure 2) are filtered and analyzed.

[0073] EEG data analysis [0074] For example, intracranial electrophysiological data are filtered to attenuate line noise (5 Hz band-stop fourth order Butterworth, centered on 60 Hz). In some embodiments, the data is then referenced using a bipolar montage by identifying all pairs of immediately adjacent contacts on every depth, strip and grid and taking the difference between the signals recorded in each pair. The resulting bipolar time series are treated as a virtual electrode and used in all subsequent analysis. For example, for the purposes of anatomical localization, the midpoint of the bipolar pair is used as the location for this virtual electrode. In some embodiments, the same midpoint approach is used to localize stimulation targets.

[0075] A first subset of target stimulation locations from the plurality of target stimulation locations that are near a respective nearest white-matter vertex based on the determined distance are selected (304) (e.g., the distance is determined relative to white-matter vertices that are classified as “near”).

[0076] A functional brain connectivity quantitative model for the plurality of target stimulation locations is received (306). The functional brain connectivity quantitative model is built based on recordings of electrical signals collected during performance of one or more memory tasks. In some embodiments, the functional brain connectivity quantitative model is built based on recordings of electrical signals collected during rest periods and during performance of tasks that tap into the desired behavior, such as learning, memory, executive function, or emotional processing.

[0077] A second subset of target stimulation locations from the first subset are selected (308) selecting that exhibit theta-connectivity to other electrodes in a montage of the brain are selected. In some embodiments, the other electrodes are located in different brain regions (e.g., regions in the brain other than the region where respective target stimulation locations are located). In some embodiments, the target stimulation locations correspond to virtual electrodes. In some embodiments, a pair of electrodes are used to stimulate a respective target stimulation location. For example, a respective bipolar pair of electrodes is used to stimulate at a location of a virtual electrode. In some embodiments, target stimulation locations that exhibit theta-connectivity are determined based on the functional brain connectivity quantitative model. In some embodiments, to build the functional brain connectivity quantitative model, the recorded EEG fata is analyzed. For example, spectral decomposition (8 frequencies from 3-180 Hz, logarithmically spaced; Morlet wavelets; wave number = 5) is performed for 1366 ms epochs from 0 to 1366 ms relative to word onset. Mirrored buffers (length 1365 ms) are included before and after the interval of interest to avoid convolution edge effects. The resulting time-frequency data are then log-transformed, averaged over time, and z-scored within session and frequency band across word presentation events.

[0078] In some embodiments, each stimulation target is characterized based on its proximity to the nearest white-matter pathway, as well as its resting-state functional connectivity with the rest of the electrodes in the brain’s montage. In some embodiments, electrodes can be arranged differently. Montages are specific arrangements of channels, where a channel is a pair of electrodes. In some embodiments, bipolar montage is used that includes channels with adjacent electrode pairs. In some embodiments, the second subset of target stimulation locations are selected from the first subset of target stimulation locations, which are classified as near white-matter vertices, where the second subset of target stimulation locations exhibit theta-connectivity above a predefined threshold. In some embodiments, weights are assigned to each target stimulation location based on both proximity to white-matter vertex locations and theta-connectivity, and those target stimulation locations that are above a predefined weight are selected for the stimulation of the brain electrically. In some embodiments, stimulating the brain at locations in the LTC locations with high functional network connectivity improve the physiological and behavioral effects of stimulation. Further, in some embodiments, stimulation improves memory performance when delivered to targets near white- matter pathways. In some embodiments, when cognitive functions different from memory are addressed, the target stimulation locations are selected to be located in different regions of the brain depending on the corresponding function (e.g., regions other than LTC).

[0079] The brain is (310) optionally electrically stimulated at the selected second subset of target stimulation locations. In some embodiments, closed-loop stimulation is delivered via intracranially implanted electrodes as neurosurgical participants (e.g., patients) perform a free recall memory task. In some embodiments, to improve memory and/or learning, multivariate classifiers are used to determine timing of the brain stimulation. In some embodiments, multivariate classifiers are trained to predict momentary lapses in memory function, and stimulation of the LTC is triggered at times when such memory lapses are expected to occur based on the predictions of the multivariate classifiers. In some embodiments, in addition to the timing of stimulation, brain states as detected or predicted by the multivatiate classifier also control the choice of parameters, such as amplitude or frequency.

[0080] In some embodiments, a method for delivering stimulation to a brain to improve cognition and/or memory includes selecting one or more white-matter tracts of the brain and delivering stimulation to the one or more white-matter tracts. In some embodiments, the one or more white-matter tracts is located within a middle temporal gyrus (MTG). In particular embodiments, the middle temporal gyrus is a left middle temporal gyrus of the brain. In still further embodiments, the one or more white-matter tracts include at least one of an inferior longitudinal fasciculus and a superior longitudinal fasciculus. These two white-matter tracts are the two major white-matter vertices of the MTG, which leave the MTG to connect to other areas of the brain, and have been found to play an important role in language processing and comprehension. Briggs, et al. 2021.

[0081] In some embodiments, the brain is optionally electrically stimulated. In some embodiments, the stimulation delivered to the one or more white-matter tracts is closed-loop stimulation, and one or more multivariate classifiers are used to determine timing and/or the parameters (e.g., amplitude, frequency, pulse width) of the stimulation. In some embodiments, the stimulation delivered to the one or more white-matter tracts is open-loop stimulation, and one or more multivariate classifiers and/or cognitive testing are used to determine the parameters (e.g., amplitude, frequency, pulse width) of the stimulation.

[0082] Stimulation methods

[0083] In some embodiments, at the start of each session, the safe amplitude for stimulation of the selected target (e.g., middle temporal gyrus, the inferior longitudinal fasciculus, the superior longitudinal fasciculus) is determined using a mapping procedure in which stimulation is applied at 0.5 mA while a neurologist monitors for after discharges. This procedure is repeated, incrementing the amplitude in steps of 0.5 mA, up to a maximum of 1.5 mA for depth contacts and 3.5 mA for cortical surface contacts. The electrical stimulation applied to the brain may be within a range of about 0.1mA to about 5.0mA. In some embodiments, the electrical stimulation is applied to the brain within a range of 0.1mA to about 0.5mA, about 0.5 mA to about 1 .0mA, about 1 .0mA to about 1.5mA, about 1.5 mA to about 2.0mA, about 2.0mA to about 2.5mA, about 2.5 mA to about 3.0mA, about 3.0mA to about 3.5mA, about 3.5 mA to about 4.0mA, about 4.0mA to about 4.5mA, about 4.5 mA to about 5.0mA, about 0.5mA to about 1.5mA, about 1.5mA to about 2.5mA, about 2.5mA to about 3.5mA, about 3.5mA to about 4.5mA, about 0.5mA, about 1.0mA, about 1.5mA, about 2.0mA about 2.5mA, about 3.0mA, about 3.5mA, about 4.0mA, about 4.5mA, or about 5.0mA.

These maximum amplitudes are chosen to be below the after-discharge threshold and below accepted safety limits for charge density. For each stimulation session, electrical current is passed through a single pair of adjacent electrode contacts. In some embodiments, the locations of implanted electrodes are determined by the monitoring needs of the clinicians (e.g., recording sites depicted in Figure 4B). Accordingly, a combination of anatomical and functional information is used to select stimulation sites (e.g., locations depicted in Figure 4 related to an experiment discuss below).

[0084] In some embodiments, stimulation is delivered to the selected target (e.g., middle temporal gyrus, the inferior longitudinal fasciculus, the superior longitudinal fasciculus) using charge-balanced biphasic rectangular pulses (e.g., pulse width = 300 ps) at either 50 Hz, 100 Hz or 200 Hz frequency (a single frequency was chosen for each subject), and is applied for 500 ms in response to classifier-detected poor memory states. In some embodiments, the electrical stimulation applied to the brain may have a frequency within a range of about 20 Hz to about 200 Hz, about 20 Hz to about 50 Hz, about 50Hz to about 100 Hz, about 100 Hz to about 150 Hz, about 150 Hz to about 200 Hz, about 20 Hz, about 30 Hz, about 40Hz, about 50 Hz, about 60 Hz, about 70 Hz, about 80 Hz, about 90 Hz, about 100 Hz, 110 Hz, 120 Hz, about 130 Hz, about 140Hz, about 150 Hz, about 160 Hz, about 170 Hz, about 180 Hz, about 190 Hz, or about 200 Hz.

[0085] In some embodiments, stimulation is delivered to the selected target (e.g., middle temporal gyrus, the inferior longitudinal fasciculus, the superior longitudinal fasciculus) at a particular pulse width. In some embodiments, the electrical stimulation applied to the brain may have a pulse width within a range of about 15 psec to about 500 psec, about 15 psec to about 50 psec, about 50 psec to about 100 psec, about 150 psec to about 200 psec, about 200 psec to about 250 psec, about 250 psec to about 300 psec, about 300 psec to about 350 psec, about 350 psec to about 400 psec, about 400 psec to about 450 psec, about 450 psec to about 500 psec, about 15 psec to about 150 psec, about 150 psec to about 250 psec, about 250 psec to about 350 psec, about 350 psec to about 450 psec, about 15 psec, about 50 psec, about 100 psec, about 150 psec, about 200 psec, about 250 psec, about 300 psec, about 350 psec, about 400 psec, about 450 psec, or about 500 psec.

[0086] In the experiment described below in section III, participants perform one practice list followed by 25 task lists: lists 1-3 were used as a baseline for normalizing the classifier; lists 4-25 consisted of 11 lists each of Stim and NoStim conditions, randomly interleaved. On NoStim lists, stimulation was not triggered in response to output of participant-specific multivariate classifier. In some embodiments, “Stim list” corresponds to a session where stimulation is delivered (e.g., when predicted by the multivariate classifiers that stimulation is needed) and “NoStim list” corresponds to a session where stimulation has not been delivered.

[0087] In some embodiments, stimulating the brain electrically includes (312) delivering a closed-loop stimulation at the selected second subset of target stimulation targets, wherein one or more multivariate classifiers are used to determine timing and/or parameters of the stimulation. In some embodiments, the performance of a respective memory task includes (314) an encoding phase, a distracting phase, and a recall phase.

[0088] In some embodiments, for each target stimulation location, a node-strength measure is calculated (316) based on coherence between each pair of bipolar channels in the montage.

[0089] Calculation of stimulation target node strength

[0090] In some embodiments, the node strength of stimulation targets is calculated using the MNE-Python software package. In some embodiments, data is extracted from record-only sessions of each patient and the data is used to calculate the coherence between each pair of bipolar channels in the patient’s montage. In some embodiments, coherence C_xy between a pair of signals at sxy electrodes x andy corresponds (318) to normalized cross-spectral density, where C_xy = and sxx^syy

Sxy is the cross-spectral density between the signals at the electrodes x andy and S_xx corresponds to auto-spectral density at the electrode x and S_yy corresponds to auto-spectral density at the electrode y. In some embodiments, the multitaper method is used to estimate spectral density (“On the performance advantage of multitaper spectral analysis,” Bronez, 1992). In some embodiments, a time-bandwidth product of 4 and a maximum of 8 tapers (tapers with spectral energy < 0.9 were removed) is used, and coherence for frequencies between 5 Hz and 13 Hz is computed. In some embodiments, inter-electrode coherence within non-overlapping 1-s windows of data collected during a 10-second baseline (countdown) period that occurred at the start of each word list is computed. In some embodiments, the resulting coherence values between each pair of electrodes are then regressed on the Euclidean distance between each pair of electrodes, to account for the correlation between inter-electrode coherence and distance. This distance-residualized measure of coherence was then used in the node-strength calculation.

[0091] In some embodiments, the one or more multivariate classifiers are trained (320) based on whole-brain patterns of spectral activity to discriminate encoding activity associated with words that recalled and words that are not recalled.

[0092] Multivariate classification

[0093] In some embodiments, to train the one or more multivariate classifiers data collected in record-only sessions is selected as input to a logistic regression classifier trained to discriminate encoding-related activity predictive of whether a word was later recalled or not recalled. Spectral power averaged across the time dimension for each word-encoding epoch (0-1366 ms relative to word onset) as the input data are used. Accordingly, in some embodiments, the features for each individual word-encoding observation are the average power across time, at each of the 8 analyzed frequencies ^x N electrodes. In some embodiments, L2 -penalization (Hastie et al., 2001) is used and the penalty parameter is set to C = 2.4* 10 ⁴ for all subjects (e.g., participants or patients). In some embodiments, the penalty parameter is based on an analysis of a large pre-existing (independent) dataset of iEEG patients performing free recall, in which this penalty parameter maximized crossvalidated classifier performance across subjects. Each patient’s penalty parameter is weighted in inverse proportion to the number of recalled and not recalled words . This ensures the classifier learned equally from both recalled and not recalled items when there was an unequal number of exemplars in each class.

[0094] Analysis of physiological effects of stimulation

[0095] To assess the effect of LTC stimulation on neural activity, recording channels (i.e., those that were not stimulated) were analyzed and spectral power in the theta band (4-8 Hz) was compared on Stim lists vs. NoStim lists. In some embodiments, “Stim list” corresponds to a session where stimulation is delivered (e.g., when predicted by the multivariate classifiers that stimulation is needed) and “NoStim list” corresponds to a session where stimulation has not been delivered. In some embodiments, electrodes exhibiting non-physiological post-stimulation artifacts (such as amplifier saturation/relaxation) are first excluded using three different measures of the EEG time series before and after stimulation. Intervals before and after stimulation are compared for changes in variance using an F-test and for changes in signal amplitude using a /-test. Additionally, a polynomial function is fit to the time series before and after each stimulation event and used a /-test to compare the resulting betas for the quadratic term. These three measures are calculated using the signal from -400 ms to -100 ms relative to stimulation onset and 100 ms to 400 ms relative to stimulation offset. In order to select statistical thresholds for each measure, the same analysis is conducted on each participant’s record-only data. Next, p-value thresholds associated with a 5% detection rate in the record-only data (i.e., false positives) are selected. Any channel that is significant on any of the three measures is excluded from analysis. To measure stimulation’s effect on theta power, spectral power from -1100 ms to -100 ms relative to stimulation onset and 100 ms to 1100 ms relative to stimulation offset is extracted. Morlet wavelets (wave number = 5) are used to estimate spectral power for 45 logarithmically spaced frequencies from 4 Hz to 200Hz. Mirrored buffers are used to eliminate edge artifacts. The resulting spectral power estimates are then z-scored within each frequency, separately for each session. Next, the power within the 4-8 Hz band across the time dimension for each 1000 ms pre-stimulation period and for each matched 1000 ms poststimulation period is averaged. Next, the pre-stimulation data is subtracted from the post-stimulation data to yield a distribution of change in 4-8 Hz power. The distribution of power changes for stimulation events are compared to the analogous power changes from NoStim lists. To do so, 4-8 Hz power is extracted and calculated using identical parameters. However, because there are no actual stimulation events in NoStim lists, a synthetic distribution of onset times is generated by extracting the lag (in ms) between each word onset and stimulation event in Stim lists, and sampling randomly from that distribution of onset times to determine when to extract data relative to word onset events in NoStim lists. This was done for all NoStim lists, excluding the first three (see Analysis of memory performance). Finally, an independent samples /-test were used to compare the distribution of Stim list power differences to the distribution of NoStim list power differences. The resulting distribution of t-statistics was then averaged across all electrodes to estimate stimulation- evoked 4-8 Hz power (Figure 3).

[0096] III. Experimental Procedures and Results

Examples

[0097] Example 1 :

[0098] Participants

Forty-seven patients undergoing intracranial electroencephalographic monitoring as part of clinical treatment for drug-resistant epilepsy were recruited to participate in a study. In total, N= 57 brain locations were stimulated: 38 patients were stimulated in one location, 8 patients were stimulated in two separate locations, and 1 patient was stimulated in three separate locations. Only one location was stimulated per session. Of this dataset, data from 14 patients are included in an earlier publication. Data from all other patients have not been previously published, and all of the presently reported analyses and results are novel.

[0099] The experiment assessed the effects of electrical stimulation on memory -related brain function. Electrophysiological data were collected from electrodes implanted subdurally (grid/strip configurations) on the cortical surface as well and/or electrodes within the brain parenchyma. The clinical team determined the placement of the electrodes based on the epileptogenic monitoring needs of the patient.

[00100] Anatomical localization

[00101] Cortical surface regions were delineated on pre-implant whole brain volumetric Tl- weighted MRI scans using Freesurfer (Fischl et al. 2004) according to the Desikan-Kiliany atlas (Desikan et al. 2006). Whole brain and high resolution medial temporal lobe volumetric segmentation was also performed using the Tl-weighted scan and a dedicated hippocampal coronal T2-weighted scan with Advanced Normalization Tools (ANTS) (Avants et al. 2008) and Automatic Segmentation of Hippocampal Subfields (ASHS) multi-atlas segmentation methods (Yushkevich et al. 2015). Coordinates of the radiodense electrode contacts were derived from a post -implant CT and then registered with the MRI scans using ANTS. Subdural electrode coordinates were further mapped to the cortical surfaces using an energy minimization algorithm (Dykstra et al. 2012). Two neuro-radiologists reviewed cross-sectional images and surface renderings to confirm the output of the automated localization pipeline. Stimulation targets localized to the inferior, middle, or superior temporal gyri (left or right hemispheres) were classified as LTC. For region of interest analyses, electrodes were assigned to regions using Freesurfer atlas labels (IFG: inferior frontal gyrus; MFG: middle frontal gyrus; SFG: superior frontal gyrus; MTLC: medial temporal lobe cortex; HIPP: hippocampus; ITG: inferior temporal gyrus; MTG: middle temporal gyrus; STG: superior temporal gyrus; IPC: inferior parietal cortex; SPC: superior parietal cortex; OC: occipital lobe).

[00102] Verbal memory task

[00103] Across participants, data were collected from two behavioral tasks: standard delayed free recall and categorized delayed free recall. In both tasks, participants were instructed to study lists of words for a later memory test; no explicit encoding task was used. Lists were composed of 12 words (https://memory.psych.upenn.edu/Word_Pools) presented in either English or Spanish, depending on the participant’s native language. In the standard free recall task, words were selected randomly from a pool of common nouns. In the categorized free recall task, the word pool was constructed from 25 semantic categories (e.g., fruit, furniture, office supplies). Each list of 12 items in the categorized version of the task consisted of four words drawn from each of three categories. Overall, N= 19 participated in standard free recall only; N= 26 participated in categorized free recall only; and N= 2 participated in both free and categorized recall (in separate sessions).

[00104] Immediately following the final word in each list, participants performed a distractor task (to attenuate the recency effect in memory, length = 20 seconds) consisting of a series of arithmetic problems of the form A+B+C=??, where A, B and C were randomly chosen integers ranging from 1- 9. Following the distractor task participants were given 30 seconds to verbally recall as many words as possible from the list in any order; vocal responses were digitally recorded and later manually scored for analysis. Each session consisted of 25 lists of this encoding-di stractor-recall procedure. [00105] EEG recording and analysis

[00106] Electrophysiological recording and stimulation was conducted using a variety of systems. Recording and stimulation equipment included clinical EEG systems (Nihon Kohden EEG-1200, Natus XLTek EMU 128 or Grass Aura-LTM64), equipment from Blackrock Microsystems, as well as the External Neural Stimulator (ENS) (Medtronic, Inc.). Data were sampled at 500, 1000, or 1600 Hz (depending on the clinical site). During the sessions, a computer recorded behavioral responses (vocalizations, key presses), synchronized to the recorded EEG via transmitted network packets. [00107] Intracranial electrophysiological data were filtered to attenuate line noise (5 Hz bandstop fourth order Butterworth, centered on 60 Hz). The data was referenced using a bipolar montage (Burke et al. 2013) by identifying all pairs of immediately adjacent contacts on every depth, strip and grid and taking the difference between the signals recorded in each pair. The resulting bipolar timeseries was treated as a virtual electrode and used in all subsequent analysis. For the purposes of anatomical localization, the midpoint of the bipolar pair was used as the location for this virtual electrode. The same midpoint approach was used to localize stimulation targets and to measure stimulation target distance to white matter.

[00108] Multivariate classification of memory

[00109] Spectral decomposition was performed (8 frequencies from 6-175 Hz, logarithmically- spaced; Morlet wavelets; wave number = 5) for 1366 ms epochs from 0 to to 1366 ms relative to word onset. Mirrored buffers (length = 1365 ms) were included before and after the interval of interest to avoid convolution edge effects. The resulting time-frequency data were then log- transformed, averaged over time, and z-scored within session and frequency band across word presentation events. For a subset of participants, the same spectral decomposition procedure was also performed on record-only data from the memory recall phase of each list. These data were then used in addition to the encoding data to train the classifier (Kragel et al. 2017). To do so, spectral power was computed for the 500 ms interval preceding a response vocalization, as well as during unsuccessful periods of memory search (the first 500 ms of any 2500 ms interval in which no recall response was made). For both trial types (correct vocalizations and unsuccessful search periods), it was further stipulated that no vocalization onsets occurred in the preceding 2000 ms.

[00110] This closed-loop stimulation approach was based on using individualized memory classifiers to control the timing of stimulation in response to brain activity. Thus, after collecting at least three record-only sessions from an individual patient, the data was then used as input to a logistic regression classifier that would trigger closed-loop stimulation during the later stimulation session(s). To build the classifier, patterns of brain activity collected during record-only sessions were used and trained the classifier to discriminate words that were recalled vs. not recalled. The input features were spectral power at the eight analyzed frequencies x N electrodes (Fig 1 A). L2- penalization was used to prevent overfitting (Hastie et al. 2001) and set the penalty parameter (C) to 2.4 x io-4 based on the optimal penalty parameter calculated across the large pre-existing dataset of free-recall participants (Kragel et al. 2017; Ezzyat et al. 2018). The penalty parameter was weighted separately for each participant in inverse proportion to their number of recalled and not recalled words; this was done so that the model would learn equally from both classes (Hastie et al. 2001). Classification analyses were programmed using either the Matlab implementation of the LIBLINEAR library (Fan et al. 2008) or the Python library scikit-leam (Pedregosa et al. 2011). [00111] For the Closed-loop group (34 participants, N = 40 stimulation targets), classifiers were trained using the true mapping of features (spectral power x electrodes) to recall outcomes. In contrast, for the Random group (13 participants, N = 17 stimulation targets), a technical error in labeling features during classifier training led to classifiers that were trained on permuted data, eliminating the true mapping between neural activity on each trial and recall outcomes. This provided a natural experiment for testing whether the closed-loop nature of stimulation enhanced the efficacy of LTC stimulation.

[00112] To assess the importance of individual features to the classifier’s performance, we calculated a forward model (Haufe et al. 2014): where Sx is the data covariance matrix, w is the vector of feature weights from the trained classifier, and ^s 'y is the variance of the logit-transformed classifier outputs for all recalled/ not recalled events y. Positive values in A suggest a positive relation between power for a given feature and successful memory recall. The A value was computed separately for each participant (averaging features within anatomical regions of interest based on the Freesurfer labels derived from anatomical localization of electrodes) before conducting across-participant statistical tests (Fig ID).

[00113] Calculation of stimulation parameters

[00114] At the start of each stimulation session, the safe amplitude for stimulation was determined using a mapping procedure in which stimulation was applied at 0.5 mA while a neurologist monitored for after discharges. This procedure was repeated, incrementing the amplitude in steps of 0.5 mA, up to a maximum of 1.5 mA for depth contacts and 3.5 mA for cortical surface contacts. These maximum amplitudes were chosen to be below the after discharge threshold and below accepted safety limits for charge density (Shannon 1992). For each stimulation session, electrical current was passed through a single pair of adjacent electrode contacts. The locations of implanted electrodes were determined strictly by the monitoring needs of the clinicians (recording sites depicted in Figure IB). A combination of anatomical and functional information was therefore used to select stimulation sites, prioritizing (if available) targets in the middle temporal gyrus (stimulation targets depicted in Figure 5A). This choice was guided by prior work identifying the middle temporal gyrus as an effective target for modulating memory with stimulation (Ezzyat et al. 2018; Kucewicz et al. 2018). Stimulation was delivered using charge-balanced biphasic rectangular pulses (pulse width = 300 ps) at either 50, 100 or 200 Hz frequency (a single frequency was chosen for each subject), and was applied for 500 ms in response to classifier-detected poor memory states. Participants performed one practice list followed by 25 task lists: lists 1-3 were used as a baseline for normalizing the classifier; lists 4-25 consisted of 11 lists each of Stim and NoStim conditions, randomly interleaved. NoStim lists were identically structured to Stim lists, except that stimulation was never delivered in response to classifier output.

[00115] To determine (in actuality) how well the classifier predicted recalled and forgotten words in a given participant’s stimulation session, area under the receiver operating characteristic curve (AUC) was used. The true classifier outputs and true recall outcomes from the NoStim lists were used to calculate the classifier generalization AUC for the stimulation sessions. To generate the corresponding receiver operating characteristic curves for visualization (Figure 1C), the classifier outputs for recalled and not recalled words were modeled using signal detection theory (Wixted 2007). This was done by using the classifier outputs to estimate the mean and variance of hypothetical (normal) distributions of memory strength for recalled and not recalled words. A curve relating true and false positive rates was then generated by varying the assumed decision criterion (Wixted 2007).

[00116] Analysis of memory performance

[00117] All participants completed at least three sessions of the record-only task (for purposes of classifier training) and at least one session of the stimulation task. For the stimulation session(s) stimulation’s effect on recall performance was calculated as follows: 4 = x 100, where Rs is the average recall for stimulated lists and RNS is the average recall for non-stimulated lists. Because the first three lists of every stimulation session were always non-stimulated (used for normalization of the classifier input features for that session), these lists were excluded from the calculation of RNS to avoid introducing a temporal order confound. All participants were required to demonstrate a minimum RNS = 8.33% (1 out of 12 words per list) for inclusion in the sample.

[00118] Calculation of stimulation target distance to white matter

[00119] Using Freesurfer to segment patients’ T1 MRI scan, white-matter vertex locations were identified, then the distance between the stimulation location (midpoint of the bipolar pair) and the nearest white matter vertex was calculated. These distances were then split into thirds in order to categorize stimulation sites as Near, Mid, or Far relative to the nearest white matter (Solomon et al. 2018; Mohan et al. 2020).

[00120] Calculation of stimulation target node strength

[00121] A previously reported method for calculating the resting-state functional connectivity between channels using the MNE-Python software package was adopted (Gramfort et al. 2014; Solomon et al. 276 2018). Data from non-task periods of the record-only sessions of each patient were extracted and used to calculate the coherence between each pair of bipolar channels in the patient’s montage. The coherence (Cxy) between two signals is the normalized cross- spectral density. This measure reflects the consistency of phase differences between signals at two electrodes, weighted by the correlated change in spectral power at both sites:

[00122] where Sxy is the cross-spectral density between signals at electrodes x and y; Sxx and S yy are the auto-spectral densities at each electrode. The multitaper method was used to estimate spectral density (Bronez 1992). A time-bandwidth product of 4 and a maximum of 8 tapers was used (tapers with spectral energy < 0.9 were removed), computing coherence for frequencies between 285 5 and 13 Hz. Inter-electrode coherence within non-overlapping 1-s windows of data collected during a 10-second baseline (countdown) period that occurred at the start of each word list was computed. The resulting coherence values between each pair of electrodes were then regressed on the Euclidean distance between each pair of electrodes, to account for the correlation between interelectrode coherence and distance (Solomon et al. 2018). This distance-residualized measure of coherence was then used in the node-strength calculation. This entire procedure was then repeated for calculating high-frequency functional connectivity in the 45-90 Hz range.

[00123] Analysis of physiological effects of stimulation

[00124] To assess the effect of LTC stimulation on neural activity, recording channels (i.e., those that were not stimulated) were analyzed and spectral power in the theta band (4-8 Hz) was compared on Stim lists vs. NoStim lists. First electrodes exhibiting non-physiological post-stimulation artifacts (such as amplifier saturation/relaxation) were excluded using three different measures of the EEG time series before and after stimulation. Intervals before and after stimulation were compared for changes in variance using an /’-test and for changes in signal amplitude using a Etest. Additionally, a polynomial function was fit to the time series before and after each stimulation event and used a t- test to compare the resulting betas for the quadratic term. These three measures were calculated using the signal from -400 ms to -100 ms relative to stimulation onset and 100 ms to 400 ms relative to stimulation offset. In order to select statistical thresholds for each measure, the same analysis was conducted on each participant’s record-only data. Then p-value thresholds associated with a 5% detection rate in the record-only data (i.e., false positives) were selected. Any channel that was significant on any of the three measures was excluded from analysis.

[00125] To measure stimulation’s effect on theta power, spectral power was extracted from -1100 ms to -100 ms relative to stimulation onset and 100 ms to 1100 ms relative to stimulation offset. Morlet wavelets (wave number = 5) were used to estimate spectral power for 45 logarithmically spaced frequencies from 4 to 200Hz. Mirrored buffers were used to eliminate edge artifacts. The resulting spectral power estimates were then z-scored within each frequency, separately for each session. Then power within the 4-8 Hz band was averaged across the time dimension for each 1000 ms pre-stimulation period and for each matched 1000 ms post-stimulation period. Then the prestimulation data was subtracted from the post-stimulation data to yield a distribution of change in 4- 8 Hz power.

[00126] The distribution of power changes for stimulation events were compared to the analogous power changes from NoStim lists. To do so, 4-8 Hz power was extracted and calculated using identical parameters. However, because there were no actual stimulation events in NoStim lists, a synthetic distribution of onset times was generated by extracting the lag (in ms) between each word onset and stimulation event in Stim lists, and sampling randomly from that distribution of onset times to determine when to extract data relative to word onset events in NoStim lists. This was done for all NoStim lists, excluding the first three (see Analysis of memory performance above). Finally, an independent samples /-test was used to compare the distribution of Stim list power differences to the distribution of NoStim list power differences. The resulting distribution of /-statistics was then averaged across all electrodes to estimate stimulation-evoked 4-8 Hz power (Figure 6).

[00127] To measure stimulation’s effect on low-frequency power, spectral power was extracted from -600 ms to -100 ms relative to stimulation onset and 100 ms to 600 ms relative to stimulation offset. Morlet wavelets (wave number = 5) were used to estimate spectral power for the same set of frequencies used to train the classifier with buffers to eliminate edge artifacts. The resulting spectral power estimates were then z-scored within each frequency, separately for each session. Then power within each frequency was averaged across the time dimension for each pre-stimulation period and for each matched post-stimulation period. Then the pre-stimulation data was subtracted from the post-stimulation data to yield a distribution of change in spectral power for each electrode.

[00128] The distribution of power changes for stimulation events were compared to the analogous power changes from NoStim lists. To do so, spectral power was extracted and calculated using identical parameters. However, because there were no actual stimulation events in NoStim lists, a synthetic distribution of onset times was generated by extracting the lag (in ms) between each word onset and stimulation event in Stim lists, and sampling randomly from that distribution of onset times to determine when to extract data relative to word onset events in NoStim lists.

[00129] Finally, an independent samples /-test was used to compare the distribution of Stim list power differences to the distribution of NoStim list power differences within each electrode. The resulting distribution of /-statistics was then averaged across all electrodes to estimate stimulation- evoked change in power (Figure 7A). These values were then averaged separately within clusters of low and high frequencies that significantly predicted memory performance (based on classifier feature importance, Figure ID).

[00130] Statistics

[00131] Data are presented as mean ± standard error of the mean; scatterplots show the standard error of the estimate. All statistical comparisons were conducted as two-tailed tests. Spearman rank correlation was used for non-normally distributed variables (e.g., white-matter distance, Figure 5B); other correlations were conducted as Pearson correlation. To account for the fact that some participants were stimulated at more than one target (always in separate sessions), linear mixed effects models were used to assess the effect of stimulation on memory and differences between the Closed-loop and Random groups. The models assumed separate intercepts and slopes for each participant. Linear mixed effects models were also used in analyzing the effects of white matter distance on memory; low- and high-frequency memory network node strength; and the effect of node strength on stimulation-evoked physiology. Data distributions were either visually inspected or assumed to be normal for parametric tests.

[00132] Results

[00133] Multivariate classifiers identify memory lapses

[00134] The simulation strategy sought to intercept and rescue periods of poor memory encoding. To do so, participant-specific multivariate classifiers were trained to discriminate patterns of neural activity during record-only sessions of free recall (Figure 1A). For the Closed-loop group (N = 40), classifiers were trained using the true mapping of features (spectral power x electrodes) to recall performance; for the Random group (N = 17), due to a technical error in labeling features (see Methods), classifiers were trained on permuted features. The recording electrode locations for the Closed-loop and Random groups appear as spheres in Figure IB. After training the classifiers on record-only data, they were used in later (independent) sessions to identify poor memory states for targeting with stimulation.

[00135] The first question was how well the classifiers predicted memory outcomes during the stimulation sessions (i.e. out-of-sample generalization). To answer this question, data from NoStim lists was used in which classifier predictions were obtained about the probability of recall for each word, but these predictions were not used to trigger stimulation (see Methods). Using area under the receiver operating characteristic curve as an index of classification accuracy, it was found that classifiers for the Closed-loop reliably exceeded chance performance [Mean AUC = 0.62 (chance AUC = 0.50), Wilcoxon signed rank test P = 5.73 x 10-7], Closed-loop classifiers also outperformed classifiers for the Random group [Mann-Whitney U = 586.0, P = 1.85 x 10-5], As expected, Random classifiers did not exceed chance [Mean AUC = 0.49, Wilcoxon signed rank test P = 0.55; Figure 1C],

[00136] To understand what features the classifier used to discriminate good vs. poor memory encoding states, a forward model was used for each participant to derive importance estimates for each feature (Haufe et al. 2014). The feature importance values were averaged within a set of regions of interest (ROIs) separately for each classifier frequency. Across participants, classifiers predicted successful memory encoding based on increased high-frequency activity (especially in frontal, lateral temporal, and medial temporal lobe areas) and decreased low-frequency activity across much of the recorded cortex and subcortex (Figure ID). This pattern, which is referred to as the spectral tilt, has been observed in previous studies to be a biomarker of successful episodic memory encoding and retrieval (Ezzyat et al. 2017; Burke et al. 2014; Long et al. 2014).

[00137] Closed-loop LTC stimulation improves memory

[00138] Having established that classifiers in the Closed-loop group reliably discriminate memory encoding states, it was next asked if memory performance could be increased via stimulation of the LTC (Figure 4A). The stimulation strategy was based on detecting poor memory encoding states and intercepting them with stimulation (Figure 4B). For the Closed-loop and Random groups, recall performance was compared for lists in which stimulation (Stim lists) was delivered vs. identically structured lists which stimulation was not delivered (NoStim lists, as described above). In the Closed-loop group, recall was higher on Stim lists compared to NoStim lists [A = 10.4% ± 4.2; z = 2.5, P = 0.01, Figure 4C], suggesting that intercepting poor memory encoding states with LTC stimulation enhanced recall. In contrast, there was no difference in memory performance for the Random group [A = -3.4% ± 6.6; z = -0.52, P = 0.60], There was a trend for greater memory enhancement for the Closed-loop compared to the Random group [z = 1.77, P = 0.08]. These findings are the first to compare closed-loop LTC stimulation with a random/open-loop stimulation control and are consistent with previous studies showing memory enhancement via LTC stimulation (Ezzyat et al. 387 2018; Kucewicz et al. 2018; Kahana et al. 2023).

[00139] White matter proximity mediates stimulation’s effect on memory

[00140] Motivated by physiological studies of electrical stimulation’s effects on downstream targets (Mohan et al. 2020; Solomon et al. 2018; Keller et al. 2018), it was asked whether stimulating close to white-matter tracts would produce greater positive or negative effects on memory. If so, this would suggest that the brain’s anatomical network structure plays a key role in determining how effectively stimulation can modulate cognitive function (Stiso et al. 2019; Crocker et al. 2021). To answer this question, it was examined how stimulation’s effect on memory performance varied as a function of the stimulation target’s proximity to white matter. For the Closed-loop group, lower distance to white matter predicted greater stimulation-related memory improvement [z = 3.51, P = 0.01; Figure 5A], In the random stimulation group, there was neither expected nor observed a correlation between white matter distance and the memory effect (P = 0.66). There was no difference between the distances to white matter for the Closed-loop and Random groups [P = 0.65] and the median distance was in fact numerically greater for the Closed loop (1.58 mm) compared to the Random group (1.39 mm). This suggests that distance to white matter alone does not explain the finding of improved memory in the Closed-loop group. Instead, proximity to white matter appears to enhance the effectiveness of closed-loop stimulation.

[00141] To further test this idea, the stimulation targets were divided into terciles and asked whether stimulation near white matter was particularly effective in modulating memory performance Figure 5A. Indeed, Closed-loop stimulation targets near white matter enhanced memory performance on Stim lists compared to NoStim lists [Near: M = 28.3% ± 7.0%, z = 4.06, P = 5 * 10-5], This memory improvement was larger than for Closed-loop stimulation targets further away from white matter [Mid: M = 1.5% ± 7.4%, z = 2.59, P = 0.01; Far: M = 0.6% ± 7.4%, z = 2.61, P = 0.009], Closed-loop stimulation near white matter also significantly outperformed the Random stimulation near white matter group [Random M = -7.9% ± 13.3%, z = 2.31, P = 0.02, Figure 5B], As expected, the Random group did not show improved memory (Stim vs. NoStim within- participant) in any white matter distance bin (all P > 0.32). These data suggest that stimulating near white matter leads to greater modulation of memory, and extend previous work that linked white matter proximity to stimulation’s effect on electrophysiology (Mohan et al. 2020; Solomon et al. 2018; Keller et al. 2018; 416 Stiso et al. 2019; Crocker et al. 2021).

Stimulation target functional connectivity predicts the change in memory

[00142] It was then asked why closed-loop stimulation delivered near white matter reliably modulated memory function. One possibility is that stimulating near white matter allows more reliable and direct access to the broader memory network connected to the stimulated location (Khambhati et al. 2019; Stiso et al. 2019; Solomon et al. 2018; Mohan et al. 2020). The functional connectivity between the brain’s memory encoding network and the stimulation targets located near white matter was then measured. Critically, separate measurements of connectivity at low (5-13 Hz) and high frequencies (45-90 Hz) were constructed by calculating coherence using participantspecific resting-state data (see Methods). Then, to isolate the brain’s memory encoding network, all electrodes that were in brain regions that showed a spectral tilt that predicted memory success during the task were identified, assessed using classifier feature importance Figure 6A. Stimulation target connectivity to electrodes In vs. Out of the memory network were compared, for both low and high-frequency coherence (referred to as Node Strength). Stimulation targets showed stronger low- frequency connectivity to electrodes in the memory network than to electrodes outside of the memory network [z = 2.31, P = 0.02, Figure 6B], For memory network electrodes, low-frequency connectivity was also higher than high-frequency connectivity [z = 2.39, P = 0.02], In contrast, stimulation targets showed equivalent high-frequency connectivity In vs. Out of the memory network [P = 0.49, Figure 6B],

[00143] Although stimulation targets near white matter showed greater overall low-frequency connectivity with memory -predicting brain areas, this finding leaves open the question of whether variability in connectivity strength with the memory network predicts variability stimulation’s effect on memory. To answer this question, low-frequency node strength with stimulation-related memory change was correlated. It was found that low-frequency node strength predicted closed-loop stimulation’s effect on memory [ = 0.69 ± 0.16, P = 2 x 10-5, Figure 6C] while high-frequency node strength did not (P = 0.56). The difference in correlation for low vs. high-frequency node strength was also significant (two-tailed permutation test P = 0.03). For all other targets that were further from white matter, there was no relation between node strength and stimulation-related memory change (all P > 0.19).

Functional connectivity mediates stimulation’s effect on downstream physiology

[00144] The preceding results indicate that low-frequency functional connectivity to the memory network predicts stimulation effects on memory. The final question was whether low-frequency connectivity also predicts stimulation’s physiological effects across the memory network. To test this prediction the Closed-loop stimulation targets near white matter was again examined and correlated each stimulation target’s connectivity to the memory network with the stimulation-evoked spectral power in this network (Figure 7A). Two participants’ data were excluded due to excessive stimulation artifact on the recording channels. In the remaining participants, it was found that stimulation -target functional connectivity predicted stimulation-related changes in low-frequency power [P = -0.72 ± 0.23, P = 0.001, Figure 7B). The correlation was not significant when using high-frequency connectivity and evoked power (P = 0.83), Figure 7B).

[00145] Discussion

[00146] Direct electrical stimulation has emerged as a powerful tool for manipulating neural activity. The hypothesis that network properties of a stimulated brain location predict stimulation’s effects on both memory and network physiology was evaluated herein. Prior studies suggest that white matter pathways mediate stimulation’s network-level physiological effects (Paulk et al.

2022; Solomon et al. 2018; Mohan et al. 2020; Khambhati et al. 2019; Stiso et al. 2019). Other studies demonstrate that measures of structural and functional connectivity predict stimulation’s effects on downstream targets (Keller et al. 2011; Fox et al. 2020; Solomon et al. 2018). However, none have simultaneously linked structural/functional connectivity with both (1) a reliable improvement over baseline cognitive functioning and (2) concomitant changes in neurophysiology that explain the behavioral effect. To directly address these questions, it was asked whether whitematter proximity and functional connectivity underlie the degree to which stimulation of LTC produces improvements or impairments of memory, alongside changes in oscillatory signatures of mnemonic function.

[00147] It was found that closed-loop stimulation of LTC reliably improved memory on stimulated vs. non-stimulated lists. Consistent with the hypothesis that white-matter pathways convey the effects of stimulation to the broader memory network, the benefits of closed-loop LTC stimulation were found to arise principally from stimulating in, or near, white matter pathways. For the electrodes nearest to white matter, stimulation yielded a 28% increase in recall performance, whereas there was a failure to observe any reliable increase when delivering stimulation far from these pathways (1%). In a subgroup of subjects who received randomly timed stimulation in LTC targets there was a failure to observe any improvement in memory performance.

[00148] To evaluate how stimulation— target functional connectivity mediates stimulation’s behavioral and physiological effects, participant-specific large-scale neural recordings obtained during prior record-only sessions were analyzed. Prior studies have shown that brain networks become coherent at low-frequencies during successful memory encoding and retrieval (Solomon et al. 2017; Kragel 481 et al. 2021a), so low-frequency coherence was used to measure the network node strength of each stimulation target. It was then asked if greater node strength between LTC stimulation targets and downstream memory-predicting areas resulted in greater effects of stimulation on memory performance. Consistent with this hypothesis, there was found a strong positive correlation (P = 0.69, see Figure 6C) between low-frequency connectivity and stimulation- related memory improvement. Finally, LTC stimulation engaged low-frequency activity across a broader brain network in a way that matched the network position of the stimulated location (Figure 7).

[00149] The data highlight how precise targeting improves stimulation efficacy by showing that delivering stimulation near LTC white-matter leads to greater stimulation-related memory gains (Figure 5). By linking low-frequency network connectivity with physiological and behavioral outcomes, the results also point to a neural mechanism for modulating memory with stimulation. [00150] This result extends earlier work that demonstrated the potential to modulate episodic memory by targeting LTC with stimulation (Ezzyat et al. 2018; Kucewicz et al. 2018). Directly comparing closed-loop and open-loop stimulation strategies in the same study helps to establish a causal role for the closed-loop approach (Hampson et al. 2018; Ezzyat and Rizzuto 2018). Finally, the data from 57 stimulation targets (across 47 patients) also represents a substantial increase compared to sample sizes described in related prior studies (Ezzyat et al. 2018; Hampson et al. 2018).

[00151] Prior work has linked successful memory function with theta power and coherence (Burke et al. 2013; Solomon et al. 2017; Herweg et al. 2020; Griffiths et al. 2019; Kragel et al. 2021b; Ter Wai et al. 2021; Osipova et al. 2006; Guderian and Diizel 2005; Klimesch et al. 1997; Staudigl and Hanslmayr 2013). Here, this physiological correlate of memory function was investigated by testing how memory-modulating LTC stimulation affects low-frequency physiology. It was found that stimulation’s effect on low-frequency activity depends on the low-frequency functional connectivity of the stimulation target. This suggests that identifying strong functional connections can produce stronger modulation of low-frequency activity within the memory network. Furthermore, it was found that stimulation that modulated low-frequency activity also modulated memory performance.

[00152] Several prior studies found potential therapeutic benefits of closed-loop stimulation triggered by decoding of intracranial brain recordings (Ezzyat et al. 2018; Scangos et al. 2021a; Hampson et al. 2018; Kahana et al. 2023). However, with some important exceptions (Hampson et al. 2018), this work has lacked an open-loop or random stimulation control condition, leaving open the question of what specific role the closed-loop nature of stimulation played in its therapeutic effects. Here, the effects of closed-loop stimulation with a random stimulation condition were compared. Closed-loop participants received stimulation only for those items predicted to be forgotten. Participants in the Random group followed the same protocol, but using classifiers trained on permuted data, resulting in stimulation being applied without regard to predicted memory success. This led to reliable memory improvement for the Closed-loop group and none for the Random group, despite following an otherwise identical protocol (Figure 1C).

[00153] It was found that closed-loop stimulation improved memory the most when it was delivered to LTC targets in or near white matter. This finding builds on a growing literature that indicates that stimulation is most effective when it is delivered in or near white matter pathways (Khambhati et al. 2019; Stiso et al. 2019; Mohan et al. 2020; Solomon et al. 2018; Paulk et al. 2022). One explanation for this phenomenon is that only stimulation of white matter pathways successfully engages broader brain networks, perhaps via oscillatory synchronization. In contrast, gray matter stimulation tends to cause more local effects (Mohan et al. 2020; Paulk et al. 2022). Though purely local effects may sometimes be desirable, the key cognitive and pathophysiological processes of greatest interest to neuroscientists tend to involve multiple interconnected brain regions.

[00154J Among its many applications for modulating cognition and behavior (Siddiqi et al. 2022; Fox et al. 2020; Sreekumar et al. 2017), a number of recent studies have evaluated stimulation’s potential for enhancing episodic memory (Mankin and Fried 2020; Suthana and Fried 2014; Curot et al. 2017; Lee et al. 2013; Sankar et al. 2014). While this investigated explored numerous stimulation targets within the LTC, future work should compare stimulation of this region to other brain areas within the broader episodic memory network. Recent work suggests that stimulating white matter pathways in the medial temporal lobe, for example, can also improve memory (Titiz et al. 2017; Mankin et al. 2021; Suthana et al. 2012). However, these previous studies used visual and/or spatial memoranda, while the present study focused on encoding and retrieval of verbal material. Thus, future research should compare stimulation to the lateral and medial temporal lobes, to determine whether stimulation target location interacts with the modality of the to-be-remembered information. This could contribute to other work that has used stimulation to study the component processes that contribute to successful episodic memory (El-Kalliny et al. 2019).

[00155] Stimulation was delivered using macroelectrodes, consistent with its clinical applications (Krauss et al. 2021; Morrell 2011; Sun et al. 2008). Macroelectrode stimulation alters local activity at the spatial scale of the distance between the anode and cathode (approximately 1 cm), but can also alter more distant regions. Because memory relies on a broad network of cortical and subcortical regions, including the hippocampus (Kim 201 1 ; Keerativittayayut et al. 2018), stimulating a broader network may be necessary to impact cognitive function. On the other hand, memory also relies on the recapitulation of specific patterns of neuronal activity, especially within the hippocampus (Foster 2017; Staresina and Wimber 2019). Thus, other work has stimulated through microelectrodes to mimic and reinstate memory-related hippocampal activity using a model -based closed loop approach (Hampson et al. 2018, 2013; Deadwyler et al. 2017). An avenue for future work could use macroelectrode stimulation in a similar vein, by triggering stimulation at multiple macroelectrode contacts in order to synchronize a particular spatiotemporal pattern of activity across key memory- related regions (Kim et al. 2016, 2018).

[00156] In relating low-frequency network connectivity, physiology, and behavior, the study contributes to methodological development for invasive stimulation (Krauss et al. 2021; Cagnan et al. 2019) that illuminates the critical role of low-frequency networks in cognition (Voytek and Knight 557 2015). In addition, the present study also suggests that other methods that manipulate low-frequency activity could be leveraged to modulate neural and cognitive function. Several recent studies using non-invasive methods have leveraged low-frequency theta-patterned stimulation to modulate episodic and working memory (Nilakantan et al. 2017; Hermiller et al. 2020; Tambini et al. 2018; Warren et al. 2019; Grover et al. 2022). Such low-frequency stimulation modulates electrophysiology perhaps by entraining low-frequency oscillations that are associated with cognitive function (Solomon et al. 2021; Reinhart and Nguyen 2019; Reinhart et al. 2017; Hanslmayr et al. 2019).

[00157] There are some limitations to the current work that emerge from trade-offs associated with this approach. An attempt to construct the largest existing dataset examining invasive closed- loop stimulation for memory modulation was made. In so doing, although most participants were stimulated at a single frequency (200 Hz), included in this analysis were participants who underwent stimulation at other stimulation frequencies. A critical question for future research will be to directly compare stimulation at different frequencies within-participant (Mohan et al. 2020). In addition, the comparisons of (1) closed-loop vs. random stimulation and (2) white matter distance as between- participants tests were carried out. Future work should therefore address the effectiveness of closed- loop stimulation and white matter targeting within-participant, by focusing resources on collecting data within-participant. In the case of invasive stimulation, however, this approach presents its own challenges in terms of patients’ clinical priorities. [00158] In summation, this demonstration of improved memory with closed-loop stimulation supports the idea that memory function is dynamic, and that closed-loop algorithms that account for moment-to-moment variability in the brain’s memory state can selectively deliver stimulation only when it is needed. The present study also links closed-loop stimulation efficacy to white matter targeting, brain-wide evoked physiology, and changes in episodic memory performance. The findings suggest strategies for using the functional and anatomical network profile of putative stimulation targets to optimize downstream changes in oscillatory activity and cognition.

[00159] Example 2

[00160] Figure 8 illustrates stimulation strategy and classifier performance. Figure 8A illustrates performance of memory tasks and multivariate classifier training. Participants performed at least three sessions of the free recall task while being monitored with intracranial EEG. Using wholebrain patterns of spectral activity, multivariate classifiers were trained to discriminate encoding activity associated with recalled vs. not recalled words. Figure 8B illustrates recording electrode locations for all participants in the Closed-loop (blue) and Random (green) groups, rendered on the Freesurfer average brain. Figure 8C illustrates stimulation-related memory change. Closed-loop stimulation of LTC improved recall performance (P = 0.02) while random stimulation did not.

[00161] The stimulation strategy sought to intercept and rescue periods of poor memory encoding. To do so, participant-specific multivariate classifiers were trained to discriminate patterns of neural activity during record-only sessions of free recall (Figure 8A). For the Closed-loop group (N= 40), classifiers were trained using the true mapping of features to recall performance; in contrast, for the Random group (N= 17), a technical error in labeling features during classifier training led to classifiers that were trained on permuted data. This provided a natural experiment for testing whether the closed-loop nature of stimulation enhanced the therapeutic efficacy of LTC stimulation. The recording electrode locations for the Closed-loop (blue) and Random (green) groups appear as spheres in Figure 8B. After having trained these classifiers on record-only data, the classifiers were then used them to trigger stimulation during subsequent (independent) sessions.

[00162] This first question asked was how well the classifiers performed during the stimulation sessions in discriminating neural activity associated with recalled vs. not recalled words (i.e., out-of- sample classifier generalization). To answer this question, the area under the receiver operating characteristic curve was calculated as an index of classification accuracy (AUC), using only the NoStim lists. For the Closed-loop group (A= 40), it was found that the distribution of AUCs significantly exceeded chance [Mean AUC = 0.62 (chance AUC = 0.50), Wilcoxon signed rank test P = 5.73x 10 ⁷], As expected, classifier AUCs for the Random group did not exceed chance [(# = 17); Mean AUC = 0.49, Wilcoxon signed rank test L = 0.55], Closed-loop AUCs were significantly higher than Random AUCs [Mann-Whitney U= 586.0, P = 1.85 X 10 ⁵], confirming the difference between groups (Figure 1C).

[00163] Having established that the Closed-loop and Random groups differed in terms of their classifier generalization performance, next it was asked whether stimulating LTC in closed-loop increased memory performance, as suggested by prior work. For the Closed-loop group, it was found that recall was higher for Stim lists compared to NoStim lists (A = 10.6% ± 4.3; /(39) = 2.45, P = 0.02). In contrast, memory performance on Stim and NoStim lists did not differ for the Random group (A = -3.2% ± 5.5; /(16) = -0.59, P = 0.57). There was a trend for greater memory enhancement for the Closed-loop compared to the Random group (Z(55) = 1.83, P = 0.07).

[00164] Whereas previous work suggested that stimulating LTC could improve memory performance, these findings are the first to directly compare closed-loop LTC stimulation with a random/open-loop stimulation control. In addition, these data also demonstrate the robustness of these earlier findings in a larger replication sample.

[00165] Figure 9 illustrates experimental results showing that white-matter proximity and functional connectivity predict stimulation-related memory change. Figure 9A show stimulation target locations. Targets in the LTC were stimulated either in closed-loop (blue) or randomly (green). Figure 9B shows how distance to white matter shown in mm affects stimulation-related memory change. Stimulation’s effect on memory were analyzed as a function of the distance between the nearest white matter and the stimulation target [p(38) = -0.420, P = 0.007], Figure 9C shows how distance to white matter categorized as “near,” “middle,” or “farthest” affects stimulation-related memory change. Closed-loop LTC stimulation improved memory performance for targets located nearest to white matter. Figure 9B shows how stimulation target node strength affects stimulation-related memory change. For closed-loop LTC stimulation, there was a significant negative correlation between the stimulation target distance to white matter and stimulation’s effect on memory [r(12) = 0.595, P = 0.025], Error bars in C reflect standard error of the mean Error regions in B and D reflect the standard error of the estimate.

[00166] The first of our two main questions, motivated by physiological studies of electrical stimulation’s effect on downstream targets, was whether stimulating close to white-matter tracts produced greater positive or negative effects on memory. To answer this question, it was examined how stimulation’s effect on memory performance varied as a function of the stimulation target’s proximity to white matter (see Figure 9A). These analyses revealed that for closed-loop stimulation, increased proximity to white matter predicted greater stimulation-related memory improvement [Spearman p(38) = -0.42, P = 0.007, Figure 9B],

[00167] In contrast, distance to white matter did not correlate with the memory effect when LTC stimulation was randomly timed (Spearman p(l 3) = -0.13, P = 0.65). There was no difference between the distances to white matter for the Closed-loop and Random groups [Mann-Whitney U = 331.0, P = 0.44] and the median distance was in fact numerically greater for the Closed-loop group (1.58 mm) compared to the Random group (1.39 mm). This suggests that distance to white matter does not on its own explain the finding of improved memory in the closed-loop group.

[00168] To further evaluate this apparent distinction between closed-loop and random stimulation, stimulation targets were divided into terciles, repeating these analyses separately for targets in the nearest, middle or father third of the distribution of white-matter distances. Consistent with the preceding analyses, closed-loop stimulation improved memory when delivered to electrodes nearest to white matter [Closed-loop M = 28.25% ± 8.14%, Random M = -7.89% ± 8.40%, f(17) = 2.36P = 0.03, Figure 9C], Within the Closed-loop group, stimulating electrodes nearest to white matter significantly improved memory compared to electrodes in the middle [M= 1.55% ± 4.22%, 7(24) = 2.67, P = 0.01] and farthest \M= 0.64% ± 5.87%, 7(26) = 2.65, P = 0.01] thirds of the distance distribution. Finally, closed-loop stimulation near white matter significantly improved memory compared to NoStim lists within-participant [/-test vs. 0: 7(13) = 3.23, P = 0.005], In contrast, there was no effect of stimulation in any white matter distance bin in the Random group (all > 0.38).

[00169] Next, it was asked why stimulation delivered near white matter reliably modulated memory function. One possibility is that stimulating near white matter allows more reliable and direct access to the broader neural network connected to the stimulated location. This could then allow closed-loop stimulation to more reliably modulate the brain’s memory network. To determine whether stimulation target connectivity predicted memory modulation, functional connectivity between the stimulation target and all other electrodes in the patient’s montage was measured. Using target node strength as measure of functional connectivity, correlated node strength with the stimulation-related change in memory performance was measured. Consistent with the idea that greater connectivity leads to more reliable memory modulation, it was found that functional node strength reliably predicted closed-loop stimulation’s effect on memory [r(12) = 0.595, P = 0.025, Figure 9D],

[00170] Figure 10 illustrates experimental results showing that theta-connectivity predicts stimulation’s effect on physiology and behavior, according to at least one embodiment. Figure 10A shows that stimulating theta hubs reduced downstream theta (4-8 Hz) power [r(36) = -0.475, P = 0.003], Figure 10B shows that stimulation improved memory when it reduced theta power [r(36) = -0.522, P = 0.0008],

[00171] The preceding results indicate that white-matter proximity and functional connectivity to the rest of the brain predict stimulation’s effects on memory performance. Our closed-loop system entrained the timing of stimulation to the waxing and waning of successful memory states as determined by the classifier. Therefore, it was hypothesized that stimulation’s effects on memory would be further related to changes in memory-related physiological activity across a broader brain network. To address this prediction, each stimulation target’s connectivity was calculated to the rest of the brain, now defined by connectivity associated with successful recall. Two participants’ data were excluded due to excessive stimulation artifact on the recording channels. In the remaining participants, it was found that stimulation-target functional connectivity predicted stimulation- related changes in (4-8 Hz) theta power [r(36) = -0.475, P = 0.003] (Figure 10A).

[00172] Next, it was asked whether the stimulation-related change in 4-8 Hz power correlated with the stimulation-related change in memory performance. If stimulation affects physiology across the brain, and thereby causally manipulates distributed memory states, then stimulating should lead to physiological changes that reflect changes in memory. Consistent with this hypothesis, stimulation improved memory when it reduced theta power [r(36) = -0.522, P = 0.0008] (Figure 10B).

[00173] In at least one embodiment, there is included one or more computers having one or more processors and memory (e.g., one or more nonvolatile storage devices). In some embodiments, memory or computer readable storage medium of memory stores programs, modules and data structures, or a subset thereof for a processor to control and run the various systems and methods disclosed herein. In one embodiment, a non-transitory computer readable storage medium having stored thereon computer-executable instructions which, when executed by a processor, perform one or more of the methods disclosed herein.

[00174] It will be appreciated by those skilled in the art that changes could be made to the exemplary embodiments shown and described above without departing from the broad inventive concept thereof. It is understood, therefore, that this invention is not limited to the exemplary embodiments shown and described, but it is intended to cover modifications within the spirit and scope of the present invention as defined by the claims. For example, specific features of the exemplary embodiments may or may not be part of the claimed invention, and features of the disclosed embodiments may be combined. Unless specifically set forth herein, the terms “a,” “an” and “the” are not limited to one element but instead should be read as meaning “at least one.” [00175] It is to be understood that at least some of the figures and descriptions of the invention have been simplified to focus on elements that are relevant for a clear understanding of the invention, while eliminating, for purposes of clarity, other elements that those of ordinary skill in the art will appreciate may also comprise a portion of the invention. However, because such elements are well known in the art, and because they do not necessarily facilitate a better understanding of the invention, a description of such elements is not provided herein.

[00176] Further, to the extent that the method does not rely on the particular order of steps set forth herein, the particular order of the steps should not be construed as limitation on the claims.

The claims directed to the method of the present invention should not be limited to the performance of their steps in the order written, and one skilled in the art can readily appreciate that the steps may be varied and still remain within the spirit and scope of the present invention.

Claims

CLAIMS What is claimed is:

1. A method for determining target locations for brain stimulation, the method comprising: for each of a plurality of target stimulation locations of the brain, determining distance to a nearest white-matter vertex of a plurality of white-matter vertices of the brain, wherein the plurality of target stimulation locations are located at a lateral temporal cortex of the brain; selecting a first subset of target stimulation locations from the plurality of target stimulation locations that are near a respective nearest white-matter vertex based on the determined distance; receiving a functional brain connectivity quantitative model for the plurality of target stimulation locations, wherein the functional brain connectivity quantitative model is built based on recordings of electrical signals collected during performance of one or more memory tasks; and selecting, based on the functional brain connectivity quantitative model, a second subset of target stimulation locations from the first subset that exhibit theta-connectivity to other electrodes in a montage of the brain.

2. The method of claim 1, further comprising: stimulating the brain electrically at the selected second subset of target stimulation locations.

3. The method of claim 1 or claim 2, wherein stimulating the brain electrically includes: delivering a closed-loop stimulation at the selected second subset of target stimulation targets, wherein one or more multivariate classifiers are used to determine timing and/or parameters of the stimulation.

4. The method according to any one of the preceding claims, wherein performance of a respective memory task of the one or more memory tasks includes an encoding phase, a distracting phase, and a recall phase.

5. The method according to any one of the preceding claims, further comprising: for each target stimulation location, calculating a node-strength measure based on coherence between each pair of bipolar channels in the montage.

6. The method according to any one of the preceding claims, wherein: coherence C_xy between a pair of signals at electrodes x and y correspond to normalized cross- spectral density, where C_xy = and Sx>- is the cross-spectral density between the signals at the electrodes x and y, and where S_xx corresponds to auto-spectral density at the electrode x and S_yy corresponds to auto-spectral density at the electrode y.

7. The method according to any one of the preceding claims, wherein the one or more multivariate classifiers are trained based on whole-brain patterns of spectral activity to discriminate encoding activity associated with words that recalled and words that are not recalled.

8. A computer system comprising: one or more processors; and memory storing one or more programs configured to be executed by the one or more processors, the one or more programs include instructions for: for each of a plurality of target stimulation locations of the brain, determining distance to a nearest white-matter vertex of a plurality of white-matter vertices of the brain, wherein the plurality of target stimulation locations are located at a lateral temporal cortex of the brain; selecting a first subset of target stimulation locations from the plurality of target stimulation locations that are near a respective nearest white-matter vertex based on the determined distance; receiving a functional brain connectivity quantitative model for the plurality of target stimulation locations, wherein the functional brain connectivity quantitative model is built based on recordings of electrical signals collected during performance of one or more memory tasks; and selecting, based on the functional brain connectivity quantitative model, a second subset of target stimulation locations from the first subset that exhibit theta-connectivity to other electrodes in a montage of the brain.

9. The computer system of claim 8, wherein the one or more programs include instructions for: stimulating the brain electrically at the selected second subset of target stimulation locations.

10. The computer system of claim 8 or claim 9, wherein the one or more programs include instructions for: delivering a closed-loop stimulation at the selected second subset of target stimulation targets, wherein one or more multivariate classifiers are used to determine timing and/or parameters of the stimulation.

11. The computer system according to any one of claims 8 through 10, wherein performance of a respective memory task of the one or more memory tasks includes an encoding phase, a distracting phase, and a recall phase.

12. The computer system according to any one of claims 8 through 11, wherein the one or more programs include instructions for: for each target stimulation location, calculating a node-strength measure based on coherence between each pair of bipolar channels in the montage.

13. The computer system according to any one of claims 8 through 12, wherein: coherence C_xy between a pair of signals at electrodes x and y correspond to normalized crosssxy spectral density, where C_xy and Sxy is the cross-spectral density between the signals at the sxx^syy electrodes x and y, and where S_xx corresponds to auto-spectral density at the electrode x and S_yy corresponds to auto-spectral density at the electrode y .

14. The computer system according to any one of claims 8 through 13, wherein the one or more multivariate classifiers are trained based on whole-brain patterns of spectral activity to discriminate encoding activity associated with words that recalled and words that are not recalled.

15. A non-transitory computer-readable storage medium storing one or more programs configured to be executed by one or more processors of a computer system that is in communication with a display generation component and one or more input devices, the one or more programs include instructions for: for each of a plurality of target stimulation locations of the brain, determining distance to a nearest white-matter vertex of a plurality of white-matter vertices of the brain, wherein the plurality of target stimulation locations are located at a lateral temporal cortex of the brain; selecting a first subset of target stimulation locations from the plurality of target stimulation locations that are near a respective nearest white-matter vertex based on the determined distance; receiving a functional brain connectivity quantitative model for the plurality of target stimulation locations, wherein the functional brain connectivity quantitative model is built based on recordings of electrical signals collected during performance of one or more memory tasks; and selecting, based on the functional brain connectivity quantitative model, a second subset of target stimulation locations from the first subset that exhibit theta-connectivity to other electrodes in a montage of the brain.

16. The non-transitory computer-readable storage medium of claim 15, wherein the one or more programs including instructions for: stimulating the brain electrically at the selected second subset of target stimulation locations.

17. The non-transitory computer-readable storage medium of claim 15 or claim 16, wherein the one or more programs including instructions for: delivering a closed-loop stimulation at the selected second subset of target stimulation targets, wherein one or more multivariate classifiers are used to determine timing and/or parameters of the stimulation.

18. The non-transitory computer-readable storage medium according to any one of claims 15 through 17, wherein performance of a respective memory task of the one or more memory tasks includes an encoding phase, a distracting phase, and a recall phase.

19. The non-transitory computer-readable storage medium according to any one of claims 15 through 18, wherein the one or more programs including instructions for: for each target stimulation location, calculating a node-strength measure based on coherence between each pair of bipolar channels in the montage.

20. The non-transitory computer-readable storage medium according to any one of claims 15 through 19, wherein: coherence C_xy between a pair of signals at electrodes x and y correspond to normalized crosssxy spectral density, where C_xy = and Sxy is the cross-spectral density between the signals at the sxx^syy electrodes x and y, and where S_xx corresponds to auto-spectral density at the electrode x and S_yy corresponds to auto-spectral density at the electrode y.

21. The non-transitory computer-readable storage medium according to any one of claims 15 through 20, wherein the one or more multivariate classifiers are trained based on whole-brain patterns of spectral activity to discriminate encoding activity associated with words that recalled and words that are not recalled.

22. A method for delivering stimulation to a brain to improve cognition and/or memory, the method comprising selecting one or more white-matter tracts of the brain, and delivering stimulation to the one or more white-matter tracts.

23. The method according to claim 22, wherein the stimulation is electrical stimulation.

24. The method according to claim 22 or claim 23, wherein the one or more white-matter tracts is located within a middle temporal gyrus.

25. The method according to any one of claims 22 through 24, wherein the middle temporal gyrus is a left middle temporal gyrus of the brain.

26. The method according to any one of claims 22 through 25, wherein the one or more whitematter tracts include at least one of an inferior longitudinal fasciculus and a superior longitudinal fasciculus.

27. The method according to any one of claims 22 through 26, wherein the stimulation delivered to the one or more white-matter tracts is closed-loop stimulation, wherein one or more multivariate classifiers are used to determine timing and/or parameters of the stimulation.

28. The method according to any one of claims 22 through 26, wherein the stimulation delivered to the one or more white-matter tracts is open-loop stimulation, wherein one or more multivariate classifiers and/or cognitive testing are used to determine parameters of the stimulation.

29. The method according to any one of claims 22 through 28, wherein the electrical stimulation is delivered to the brain at from about 0.1mA to about 5.0mA.

30. The method according to any one of claims 22 through 29, wherein the electrical stimulation is delivered to the brain at about 1.0mA.

31. The method according to any one of claims 22 through 30, wherein the electrical stimulation is delivered to the brain at from about 20 Hz to about 200 Hz.

32. The method according to any one of claims 22 through 31, wherein the electrical stimulation is delivered to the brain at about 200 Hz.

33. The method according to any one of claims 22 through 32, wherein the electrical stimulation is delivered to the brain at a pulse width from about 15 psec to about 500 psec.

34. The method according to any one of claims 22 through 33, wherein the electrical stimulation is delivered to the brain at a pulse width of about 300 psec.