Blood pressure regulation systems and methods
The blood pressure regulation system addresses dysregulated blood pressure in spinal cord injuries by using a detector and stimulator to apply electrical stimuli based on brain tissue signals, effectively stabilizing blood pressure and reducing cardiovascular risks.
Patent Information
- Application Number
- PCT/CA2026/050674
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-08
- Filing Date
- 2026-05-08
- Publication Date
- 2026-11-12
AI Technical Summary
Individuals with spinal cord injuries above the T5 level experience dysregulated blood pressure, leading to orthostatic hypotension and increased risk of cardiovascular diseases due to disrupted sympathetic control over heart and blood vessels.
A blood pressure regulation system that includes a detector to identify electrical signals from brain tissue, a stimulator to apply electrical stimulus signals to spinal cord tissue, and a processor to control the stimulus based on detected biomarkers, using spectral signatures and neural firing patterns to regulate blood pressure.
Effectively stabilizes blood pressure by applying targeted electrical stimuli to the spinal cord, mitigating orthostatic hypotension and reducing the risk of cardiovascular complications.
Smart Images

Figure CA2026050674_12112026_PF_FP_ABST
Abstract
Description
BLOOD PRESSURE REGULATION SYSTEMS AND METHODSCROSS-REFERENCE TO RELATED APPLICATION
[0001] This application claims the benefit of U. S. Provisional Patent Application No.63 / 802,448, filed May 8, 2025, the entire contents of which are incorporated herein by reference.FIELD
[0002] The present disclosure relates to blood pressure regulation systems and methods.BACKGROUND
[0003] The following paragraph is provided by way of background to the present disclosure. They are not however an admission that anything discussed therein is prior art or part of the knowledge of a person of skill in the art.
[0004] Spinal cord injury is a debilitating injury that disrupts the transmission of neural information from the brain to the rest of the body. Injuries at, or above, the T5 spinal cord level (i.e., the upper thoracic spinal cord) may lead to dysregulated sympathetic control over the heart and blood vessels since sympathetic nerves exit the spinal cord between the T5 and T12 spinal levels. Consequently, a person living with a spinal cord injury may have trouble regulating their blood pressure. For example, a person living with a spinal cord injury may face significant decreases in blood pressure when transitioning from a supine to an upright position (known as orthostatic hypotension (OH) which occurs due to a loss of sympathetically mediated vasoconstriction). Episodes of orthostatic hypotension may, for example, result in dizziness, light-headedness or loss of consciousness. Episodes of orthostatic hypotension may occur several times per day (such as up to 28 times per day, for example). Over time, the risk of developing cardiovascular diseases increases fourfold among individuals with spinal cord injury when compared to the general population. As such, improved systems and methods for regulating blood pressure are desirable.SUMMARY
[0005] In one aspect, in accordance with the teachings herein, there is provided at least one embodiment of the invention where the at least one embodiment includes a blood pressure regulation system comprising: a detector operable to detect at least one electrical signal from brain tissue; a stimulator device operable to apply at least one electrical stimulus signal to spinal cord tissue to regulate blood pressure; and at least one processorconfigured to control application by the stimulator device of the at least one electrical stimulus signal to the spinal cord tissue based at least partially on a presence of at least one blood pressure correlated biomarker in the detected at least one electrical signal.
[0006] In at least one embodiment, the at least one processor is further configured to scan the at least one electrical signal for the presence of the at least one blood pressure correlated biomarker.
[0007] In at least one embodiment, the at least one processor is further configured to, in response to detecting the at least one blood pressure correlated biomarker in the detected at least one electrical signal, determine at least one parameter of the at least one electrical stimulus signal based at least partially on the at least one blood pressure correlated biomarker.
[0008] In at least one embodiment, the detector is operable to detect the at least one electrical signal from medullary neurons.
[0009] In at least one embodiment, the detector is operable to detect the at least one electrical signal from ventrolateral medulla neurons.
[0010] In at least one embodiment, the detector is operable to detect the at least one electrical signal from bulbospinal C1 neurons.
[0011] In at least one embodiment, the at least one blood pressure correlated biomarker comprises a spectral signature of a local field potential, wherein the spectral signature was a current band power or a time-lagged power.
[0012] In at least one embodiment, the spectral signature is between about 80±10% Hz to about 200±10% Hz.
[0013] In at least one embodiment, the spectral signature is between about 120±10% Hz to about 200±10% Hz.
[0014] In at least one embodiment, the spectral signature is between about 80±10% Hz to about 120±10% Hz.
[0015] In at least one embodiment, the spectral signature is between about 0.01 ±10% Hz to about 1 ±10% Hz.
[0016] In at least one embodiment, the at least one blood pressure correlated biomarker comprises a number of neural firings above a threshold value.
[0017] In at least one embodiment, the threshold value is four standard deviations greater than a baseline number of neural firings.
[0018] In at least one embodiment, the threshold value is five standard deviations greater than a baseline number of neural firings.
[0019] In at least one embodiment, when the at least one blood pressure correlated biomarker comprises a number of neural firings, the detected at least one electrical signal is from a local field potential (LFP).
[0020] In at least one embodiment, the at least one blood pressure correlated biomarker is determined at least partially by isolating neurons based on template matching.
[0021] In at least one embodiment, the detector is invasive.
[0022] In at least one embodiment, the detector comprises at least one electrode operable to detect the at least one electrical signal.
[0023] In at least one embodiment, the at least one electrode of the detector is implantable within the brain tissue.
[0024] In at least one embodiment, the at least one electrode of the detector is implantable within the rostral ventrolateral medulla.
[0025] In at least one embodiment, the at least one electrode of the detector is an electrocorticography electrode.
[0026] In at least one embodiment, the at least one electrode of the detector is an endovascular electroencephalography electrode.
[0027] In at least one embodiment, the detector is non-invasive.
[0028] In at least one embodiment, the detector comprises at least one electrode operable to detect the at least one electrical signal.
[0029] In at least one embodiment, the at least one electrode of the detector is an electroencephalography electrode.
[0030] In at least one embodiment, the stimulator device is invasive.
[0031] In at least one embodiment, the stimulator device is non-invasive.
[0032] In at least one embodiment, the stimulator device comprises at least one electrode operable to apply the at least one electrical stimulus signal to the spinal cord tissue.
[0033] In at least one embodiment, the stimulator device is operable to apply the at least one electrical stimulus signal to spinal cord tissue between a T10 spinal cord level and a T12 spinal cord level.
[0034] In at least one embodiment, the detector communicates wirelessly with the at least one processor.
[0035] In at least one embodiment, the stimulator device communicates wirelessly with the at least one processor.
[0036] In at least one embodiment, the system further comprises a feedback controller.
[0037] In at least one embodiment, the feedback controller is configured to receive as input a measure of blood pressure.
[0038] In at least one embodiment, the feedback controller is configured to receive as input the at least one electrical signal.
[0039] In at least one embodiment, the feedback controller is a Proportional-Integral-Derivative (PID) controller.
[0040] In at least one embodiment, the feedback controller is configured to use a weighted least squares linear model to control application by the stimulator device of the at least one electrical stimulus signal to the spinal cord tissue.
[0041] In at least one embodiment, the feedback controller is configured to implement a logistic regression model to control application by the stimulator device of the at least one electrical stimulus signal to the spinal cord tissue.
[0042] In at least one embodiment, the logistic regression model was trained on features including current and time-lagged band powers from high-gamma1, high-gamma2 and VFL spectral bands.
[0043] In at least one embodiment, the logistic regression model is paired with a weighted least squares (WLS) model or an exponentially weighted moving average (EWMA) model.
[0044] In at least one embodiment, an average over time of the detected at least one electrical signal was used to trigger stimulation to avoid stimulation in response to noise spikes.
[0045] In another aspect, in accordance with the teachings herein, there is provided at least one embodiment of the invention where the at least one embodiment includes a method of regulating blood pressure. The method may comprise: detecting presence of at least one blood pressure correlated biomarker in at least one electrical signal from brain tissue; and based on the detected presence of the at least one blood pressure correlated biomarker, applying at least one electrical stimulus signal to spinal cord tissue to regulate the blood pressure.
[0046] In at least one embodiment, the detecting is performed using a detector and at least one processor, and the at least one electrical stimulus signal is applied using a stimulator device controlled by the at least one processor.
[0047] In at least one embodiment, the at least one electrical signal is detected from rostral ventrolateral medulla neurons.
[0048] In at least one embodiment, the at least one electrical signal is detected from bulbospinal C1 neurons.
[0049] In at least one embodiment, the at least one electrical stimulus signal is applied between a T10 spinal cord level and a T12 spinal cord level.
[0050] In at least one embodiment, the method is performed using a system as described herein.
[0051] In another aspect, in accordance with the teachings herein, there is provided at least one embodiment of the invention where the at least one embodiment includes use of the system described herein as a blood pressure regulator.
[0052] In another aspect, in accordance with the teachings herein, there is provided at least one embodiment of the invention where the at least one embodiment includes a non-transitory computer readable medium storing program code instructions that, when executed by at least one processor, configure the at least one processor to perform a method described herein for regulating blood pressure.
[0053] In another aspect, in accordance with the teachings herein, there is provided at least one embodiment of the invention where the at least one embodiment includes a blood pressure regulation system comprising: a detector operable to detect at least one blood-pressure correlated signal from brain tissue; a stimulator device operable to apply at least one electrical stimulus signal to spinal cord tissue to regulate blood pressure; and at least one processor configured to control application by the stimulator device of the at least one electrical stimulus signal to the spinal cord tissue based at least partially on a presence of at least one blood pressure correlated biomarker in the detected at least one blood-pressure correlated signal.
[0054] It will be appreciated that the foregoing summary sets out representative aspects of embodiments to assist skilled readers in understanding the following detailed description. Other features and advantages of the present application will become apparent from the following detailed description taken together with the accompanying drawings. It should be understood, however, that the detailed description and the specific examples, while indicating preferred embodiments of the application, are given by way of illustration only, since various changes and modifications within the spirit and scope of the subject matter will become apparent to those skilled in the art from this detailed description.BRIEF DESCRIPTION OF THE DRAWINGS
[0055] For a better understanding of the various embodiments described herein, and to show more clearly how these various embodiments may be carried into effect, reference will be made, by way of example, to the accompanying drawings which show at least oneexample embodiment, and which are now described. The drawings are not intended to limit the scope of the teachings described herein.
[0056] FIG. 1 is a schematic view of a blood pressure regulation system according to one example embodiment;
[0057] FIG. 2A is a schematic view of a blood pressure regulation system according to one example embodiment;
[0058] FIG. 2B is a schematic view of a blood pressure regulation system according to one example embodiment;
[0059] FIG. 2C is a schematic view of a blood pressure regulation system according to one example embodiment;
[0060] FIG. 2D is a schematic view of a blood pressure regulation system according to one example embodiment.
[0061] FIG. 3 is a flowchart diagram of an example embodiment of a method implemented when following some example program codes of a program memory of the blood pressure regulation system of FIG. 1.
[0062] FIG. 4 depicts a reduction in neural firing rates in response to increases in blood pressure.
[0063] FIG. 5 depicts an increase in neural firing rates following the application of a vasodilator.
[0064] FIG. 6 depicts a decrease in neural firing rates following cyclical reductions in blood pressure
[0065] FIG. 7 depicts a correlation between neural firing rates and systolic blood pressure.
[0066] FIG. 8 depicts the integration of chemoreflex and baroreflex on firing rate.
[0067] FIG. 9 depicts local field potential response to increased blood pressure from vasopressor application.
[0068] FIG. 10 depicts correlation between high-gamma Local Field Potentials (LFP) band-powers and Multiunit Activity (MUA) firing rates.
[0069] FIG. 11 depicts LFP analysis on chemoreflex data.
[0070] FIG. 12 depicts increases in neural firing rates when blood pressure drops.
[0071] FIG. 13 depicts increases in high-gamma LFP band-power in response to blood pressure decreases.
[0072] FIGS. 14A-14B show a schematic and actual BCI implantation and electrode configuration, respectively.
[0073] FIGS. 15A-15B show the LBNP and chemoreflex experimental setup, respectively.
[0074] FIGS. 16A-16C show signals related to a data-driven beat detection algorithm for systolic and diastolic blood pressure.
[0075] FIGS. 17A-17F show data for hypotensive induction via SNP from a single animal.
[0076] FIGS. 18A-18F show data for hypotensive induction via LBNP from a single animal.
[0077] FIGS. 19A-19F show data for hypotensive induction via PE from a single animal.
[0078] FIGS. 20A-20C show LFP band power modulation from SNP, LBNP and LFP, respectively, across animals (n=5).
[0079] FIGS. 21A-21G show data for a hypoxia ramp from 11% to 21% O2 in a single animal.
[0080] FIG. 22 shows LFP band power modulation from hypoxia ramps across animals (n=5).
[0081] FIGS. 23A-23G show data for a hypoxia jump from 21% to 11% O2 in a single animal.
[0082] FIG. 24 shows LFP band power modulation from hypoxia jumps across animals (n=5).
[0083] FIGS. 25A-25G show data for a hypercapnia ramp from 0% to 10% CO2 in a single animal.
[0084] FIG. 26 shows LFP band power modulation from hypercapnia ramps across animals (n=5).
[0085] FIGS. 27A-27G show data for a hypercapnia jump from 0% to 10% CO2 in a single animal.
[0086] FIG. 28 shows LFP band power modulation from hypercapnia jumps across animals (n=5).
[0087] FIGS. 29A-29F show data for sympathetic activation via GLP-1 from a single animal.
[0088] FIG. 30 shows LFP band power modulation from GLP-1 across animals (n=5).
[0089] FIG. 31 shows an example embodiment of an architecture of the β-VAE used for latent feature extraction from LFP data.
[0090] FIG. 32 shows an example of a custom-built dashboard for online BSI experiments.
[0091] FIG. 33 is an illustration of an overview of the online BSI experimental method.
[0092] FIGS. 34A-34C show least-squares k-fold cross validation results for GLS and WLS models that were evaluated.
[0093] FIGS. 35A-35C show least-squares LODO cross validation results for GLS and WLS models that were evaluated.
[0094] FIGS. 36A-36E show LODO cross validation logistic regression results for an example BSI decoder.
[0095] FIGS. 37A-37F show data for BSI effects on BP following SNP injections (n=5).
[0096] FIGS. 38A-38F show data for BSI effects on BP following PE injections (n=5).
[0097] FIGS. 39A-39F show data for BSI effects on BP during 30 minutes of LBNP (n=10).
[0098] Further aspects and features of the example embodiments described herein will appear from the following description taken together with the accompanying drawings.DETAILED DESCRIPTION
[0099] The headings and Abstract of the Disclosure provided herein are for convenience only and do not interpret the scope or meaning of the embodiments.
[0100] Various embodiments in accordance with the teachings herein will be described below to provide examples of at least one embodiment of the claimed subject matter. No embodiment described herein limits any claimed subject matter. The claimed subject matter is not limited to devices, systems or methods having all of the features of any one of the devices, systems or methods described below or to features common to multiple or all of the devices, systems or methods described herein. It is possible that there may be a device, system or method described herein that is not an embodiment of any claimed subject matter. Any subject matter that is described herein that is not claimed in this document may be the subject matter of another protective instrument, for example, a continuing patent application, and the applicants, inventors or owners do not intend to abandon, disclaim or dedicate to the public any such subject matter by its disclosure in this document.
[0101] Furthermore, it will be appreciated that for simplicity and clarity of illustration, where considered appropriate, reference numerals may be repeated among the figures to indicate corresponding or analogous elements or steps. In addition, numerous specific details are set forth in order to provide a thorough understanding of the embodiments described herein. However, it will be understood by those of ordinary skill in the art that the embodiments described herein may be practiced without these specific details. In other instances, well-known methods, procedures and components have not been described in detail so as not to obscure the embodiments described herein. Also, the description is not to be considered as limiting the scope of the embodiments described herein.
[0102] It should also be noted that the terms “coupled” or “coupling” as used herein can have several different meanings depending on the context in which these terms are used. For example, the terms coupled or coupling can have a mechanical, electrical or communicative connotation. For example, as used herein, the terms coupled or coupling can indicate that two elements or devices can be directly connected to one another or connected to one another through one or more intermediate elements or devices via an electrical element, an electrical signal, a light signal or a mechanical element depending on the particular context.
[0103] Similarly, throughout this specification and the appended claims the term “communicative” as in “communicative pathway”, “communicative coupling”, and in variants such as “communicatively coupled” is generally used to refer to any engineered arrangement for transferring and / or exchanging information. Examples of communicative pathways include, but are not limited to, electrically conductive pathways (e.g., electrically conductive wires, physiological signal conduction), electromagnetically radiative pathways (e.g., radio waves, optical signals, etc.), or any combination thereof. Examples of communicative couplings include, but are not limited to, electrical couplings, magnetic couplings, radio couplings, optical couplings or any combination thereof.
[0104] When introducing elements of the present subject matter or the embodiments thereof, the articles “a,” “an,” “the,” and “said” are intended to mean that there are one or more of the elements. Unless the context requires otherwise, throughout the specification and claims which follow, the word “comprise”, “include” or “have” and variations thereof, such as, “comprises” and “comprising”, “includes” and “including” or “has” and “having” are to be construed in an open, inclusive sense, that is, as “including, but not limited to”.
[0105] It should also be noted that, as used herein, the wording “and / or” is intended to represent an inclusive-or. That is, “X and / or Y” is intended to mean X or Y or both X and Y, for example. As a further example, the phrases “X, Y, and / or Z”, “any operable combination of X, Y and Z”, “X, Y, Z or any combination thereof’ or “any combination of X, Y and Z” is intended to mean X, or Y, or Z, or X and Y, or X and Z, or Y and Z, or X, Y and Z.
[0106] It should be noted that terms of degree such as “substantially”, “about” and “approximately” as used herein mean a reasonable amount of deviation of the modified term such that the end result is not significantly changed. These terms of degree may also be construed as including a deviation of the modified term, such as by 1%, 2%, 5%, 10% or 15%, for example, if this deviation does not negate the meaning of the term it modifies.
[0107] Furthermore, the recitation of numerical ranges by endpoints herein includes all numbers and fractions subsumed within that range (e.g., 1 to 5 includes 1, 1.5, 2, 2.75, 3, 3.90, 4, and 5). It is also to be understood that all numbers and fractions thereof are presumed to be modified by the term “about” which means a variation of up to a certain amount of the number to which reference is being made if the end result is not significantly changed, such as 1%, 2%, 5%, 10% or 15%, for example.
[0108] FIG. 1 is a schematic illustration of an example embodiment of a blood pressure regulation system 10. The blood pressure regulation system 10 may detect and decode (or otherwise process) blood pressure related neural activity from brain tissue of a patient. In response to the decoded / detected blood pressure related neural activity, the blood pressure regulation system 10 may stimulate the patient’s spinal cord tissue to regulate the patient’s blood pressure based at least partially on the decoded blood pressure related neural activity. Stimulating a patient’s spinal cord tissue to regulate blood pressure may be particularly advantageous for patients who have suffered a spinal cord injury, one or more spinal cord tumors, Multiple System Atrophy, Parkinson’s Disease, Alzheimer’s Disease, or any condition which at least partially inhibits natural blood pressure regulation and / or one or more brainstem and / or spinal cord pathways. The patient is typically a human. However, the patient need not be a human and may instead be another animal (such as, for example, a non-human primate, a rat, etc.).
[0109] The blood pressure regulation system 10 comprises a detector 12. The detector 12 is operable to detect at least one electrical signal from the patient’s brain tissue. The at least one electrical signal may be generated by neural activity of the brain tissue. Based on the captured (or detected) neural activity, the at least one electrical signal may encode blood pressure regulation related information. The detector 12 may be an invasive device (i.e., at least a portion of the detector 12 is inserted into at least some tissue of the patient) or a non-invasive device (i.e., no portion of the detector 12 is inserted into any tissue of the patient). Although the detector 12 is described herein as being operable to detect at least one electrical signal from the patient’s brain tissue, the detector 12 may additionally, or alternatively, be operable to detect other signals which encode blood pressure regulation related information such as functional near infrared spectroscopy (fNIRS) signals which may be derived from blood oxygen levels.
[0110] The detector 12 may be positioned proximal to medullary neurons. In at least one embodiment, the detector 12 is positioned proximal to ventrolateral medulla neurons.
[0111] In at least one embodiment, the detector 12 is positioned proximal to the rostral ventrolateral medulla. In at least one embodiment, the detector 12 is positioned proximal to one or more bulbospinal C1 neurons of the rostral ventrolateral medulla.
[0112] In at least one embodiment, the detector 12 is positioned proximal to the caudal ventrolateral medulla. In at least one embodiment, the detector 12 is positioned proximal to one or more A1 neurons of the caudal ventrolateral medulla. In at least one embodiment, the detector 12 is positioned proximal to one or more gamma-aminobutyric acid (GABA)-ergic neurons of the caudal ventrolateral medulla.
[0113] In at least one embodiment, the detector 12 is positioned proximal to the nucleus of the solitary tract (which also may be referred to as the “nucleus solitarius tractus”).
[0114] In the illustrated embodiment, the detector 12 comprises at least one electrode 13. The at least one electrode 13 may be operable to detect the at least one electrical signal from the brain tissue. The at least one electrode 13 may be an invasive electrode (i.e., at least a portion of the at least one electrode 13 is inserted into at least some tissue of the patient) or a non-invasive electrode (i.e., no portion of the electrode 13 is inserted into any tissue of the patient). The at least one electrode 13 may be positioned proximal to the rostral ventrolateral medulla. In at least one embodiment, the at least one electrode 13 is positioned proximal to one or more bulbospinal C1 neurons of the rostral ventrolateral medulla.
[0115] As the level of invasiveness of the detector 12 increases, a spatiotemporal resolution of the detector 12 may become sufficient to capture individual action potentials of proximal neurons (such as neurons close to the at least one electrode 13, for example).
[0116] In some cases, at least a portion of the detector 12 (or the at least one electrode 13) is surgically placed (such as by way of craniotomy, for example) within the brain or peripheral nervous system.
[0117] The blood pressure regulation system 10 also comprises a stimulator device 14 (which may also be referred to as a neuromodulation device). The stimulator device 14 is operable to apply at least one electrical stimulus signal to the patient’s spinal cord tissue to regulate the patient’s blood pressure. The application of the at least one electrical stimulus signals may be referred to as neuromodulation therapy. The stimulator device 14 may be an invasive device or a non-invasive device. At least a portion of the stimulator device 14 may be positioned proximal to the patient’s spinal cord tissue. In at least one embodiment, at least a portion of the simulator device 14 is positioned proximal to caudal thoracic spinal cord tissue. In at least one embodiment, the stimulator device 14 is operable to apply the at least one electrical stimulus signal to the patient’s spinal cordtissue between the T10 and T12 spinal cord levels. In at least one embodiment, the primary target of the stimulation is to target the patient’s blood vessels within the splanchnic region (e.g. kidney, liver, and intestinal tissues) to contract these blood vessels rapidly thereby increasing blood pressure. Blood vessels in such abdominal tissue can quickly vary blood pressure based on an applied stimulus (or lack of stimulus).
[0118] In the illustrated embodiment, the stimulator device 14 comprises at least one electrode 15. The at least one electrode 15 is operable to apply the at least one electrical stimulus signal to the patient’s spinal cord tissue. In at least one embodiment, the at least one electrode 15 is operable to apply the at least one electrical stimulus signal to caudal thoracic spinal cord tissue. In at least one embodiment, the at least one electrode 15 is operable to apply the at least one electrical stimulus signal to the patient’s spinal cord tissue between the T10 and T12 spinal cord levels. The at least one electrode 15 may be an invasive electrode or a non-invasive electrode. The at least one electrode 15 may be positioned proximal to the patient’s spinal cord tissue. In at least one embodiment, the at least one electrode 15 is positioned proximal to the patient’s spinal cord tissue between the T10 and T 12 spinal cord levels.
[0119] The blood pressure regulation system 10 also comprises at least one processor 16. The at least one processor 16 may decode the at least one electrical signal detected by the detector 12. Based on the decoded signal, the at least one processor 16 may control application of the at least one electrical stimulus signal by the stimulator device 14. For example, the at least one processor 16 may be configured to scan or otherwise process the at least one electrical signal detected by the detector 12 for the presence of at least one blood pressure correlated biomarker. As described elsewhere herein, the at least one blood pressure correlated biomarker may be, or include, a spectral signature, a number of detected neural firings which is higher than a threshold amount, etc. The spectral signature may, for example, be, or include, a band-power between two frequencies (e.g. a band power between about 80 Hz and about 125 Hz, a band power between about 80 Hz and about 200 Hz, a band power between about 125 Hz and about 200 Hz, a band power between about 0.01 Hz and about 0.5 Hz, etc.). In at least one embodiment, the spectral signature is, or includes, a high-gamma band-power (such as the high-gamma band-power shown in FIGS. 9-11 and 13, for example). The presence of at least one blood pressure correlated biomarker in the detected at least one electrical signal may represent that an increase in blood pressure is desired. As such, in response to the detection of at least one blood pressure correlated biomarker, the at least one processor 16 may control the stimulator device 14 to apply the at least one electricalstimulus signal to the patient’s spinal cord tissue in order to increase the patient’s blood pressure.
[0120] In at least one embodiment, the at least one processor 16 decodes the at least one electrical signal detected by the detector 12 within about 5-100 ms of measuring the at least one electrical signal.
[0121] Blood pressure may, for example, be decreased by ceasing to, or not, stimulating the patient’s spinal cord tissue. If, for example, the at least one processor 16 does not detect the presence of at least one blood pressure correlated biomarker within the at least one electrical signal detected by the detector 12, the at least one processor 16 may determine that the patient’s blood pressure does not need to be increased. In some cases, the at least one processor 16 determines that the patient’s blood pressure may need to be decreased. In such cases, the at least one processor 16 may control the stimulator device 14 to not stimulate the patient’s spinal cord tissue or to cease stimulating the patient’s spinal cord tissue. In at least one embodiment, not stimulating the patient’s spinal cord tissue comprises controlling the simulator device 14 to stop applying the at least one electrical stimulus signal to the patient’s spinal cord tissue (or to not apply the at least one electrical stimulus signal to the patient’s spinal cord tissue).
[0122] The at least one processor 16 may control for how long the at least one electrical stimulus signal is applied by the stimulator device 14. Additionally, or alternatively, the at least one processor 16 may control one or more parameters of the at least one electrical stimulus signal that is applied by the stimulator device 14 such as an intensity (or amplitude), a frequency at which the at least one electrical stimulus signal is applied by the stimulator device 14, a number of pulses, a directionality of the pulses (such as unidirectional or bi-directional pulses, for example), a pulse duration, an inter-pulse frequency, a pulse width, or combinations of two or more thereof, etc. Such control parameters may vary through out a day as the patient’s blood pressure may need to be regulated to different levels to ensure optimal or at least improved bodily function during daily activities, such as sleeping, exercising, or changing emotional states. For example, anger can lead to increases in mean arterial pressure of up to about 22 mmHg and an increase in heart rate. Conversely, sleep can decrease mean arterial pressure by up to about 13 mmHg.
[0123] The at least one processor 16 may be, or include, one or more microprocessors, one or more analog circuits, one or more configurable logic blocks, one or more application-specific integrated circuits (ASICs), or one or more field programmable gate arrays (FPGAs), for example.
[0124] The blood pressure regulation system 10 also includes a program-codes storage device 17 storing program codes (e.g., software instructions) that, when executed by the at least one processor 16, causes the at least one processor 16 to implement one or more functions of the blood pressure regulation system 10 as those described herein, for example.
[0125] The program-codes storage device 17 may include one or more of the same or different computer-readable and / or computer-writable data-storage media, which in various embodiments may include one or more of a read-only memory (ROM), a random access memory (RAM), a hard disc drive (HDD), a solid-state drive (SSD), and other computer-readable and / or computer-writable data-storage media.
[0126] The at least one processor 16 and program-codes store 17 are an example only, and alternative embodiments may differ. For example, alternative embodiments may include more, fewer, or different components. Also, in alternative embodiments, components described herein may be combined or separated into separate components. Alternative embodiments may include one or more alternatives to components as described herein.
[0127] The detector 12, the stimulator device 14, or both the detector 12 and the stimulator device 14 may be wirelessly communicatively coupled to the at least one processor 16. By coupling at least the detector 12 or the stimulator device 14 wirelessly to the at least one processor 16, the detector 12 and the stimulator device 14 may be positioned relative to the patient independent of one another. For example, the detector 12 may be positioned proximal to the patient’s head while the stimulator device 14 may be positioned proximal to the patient’s back independent of the detector 12. The detector 12, the stimulator device 14, or both the detector 12 and the stimulator device 14 may be wirelessly communicatively coupled to the at least one processor 16 using any suitable wireless communication protocol (such as, Bluetooth or near field communication, for example).
[0128] FIGS. 2Ato 2D schematically illustrate further example embodiments of the blood pressure regulation system 10.
[0129] In the illustrated example embodiment of FIG. 2A, the blood pressure regulation system 10A comprises a first module 18A comprising the detector 12 and the at least one processor 16, and a second module 19A comprising the stimulator device 14. In the illustrated embodiment, the detector 12 is physically connected to the at least one processor 16 while the stimulator device 14 is wirelessly connected to the at least oneprocessor 16. The word module refers to hardware / electronic components herein unless described otherwise.
[0130] In the illustrated example embodiment of FIG. 2B, the blood pressure regulation system 10B comprises a first module 18B comprising the detector 12, and a second module 19B comprising the stimulator device 14 and the at least one processor 16. In the illustrated embodiment, the detector 12 is wirelessly connected to the at least one processor 16 while the stimulator device 14 is physically connected to the at least one processor 16.
[0131] In the illustrated example embodiment of FIG. 2C, the blood pressure regulation system 10C comprises the first module 18B comprising the detector 12, and a second module 19B comprising the stimulator device 14. The at least one processor 16 may be physically separate from the first module 18B and the second module 19A. In the illustrated embodiment, both the detector 12 and the stimulator device 14 are wirelessly connected to the at least one processor 16. In some such embodiments, the at least one processor 16 may be a computing device such as a smartphone or tablet, for example. The detector 12 and the stimulator device 14 may wirelessly communicate with this computing device. Alternatively, in at least one embodiment, the detector 12, the simulator device or both the detector 12 and the stimulator device 14 are physically connected to the computing device.
[0132] In the illustrated example embodiment of FIG. 2D, the blood pressure regulation system 10D comprises the first module 18A comprising the detector 12 and a first processor 16A, and the second module 19B comprising the stimulator device 14 and a second processor 16B. In the illustrated embodiment, the first processor 16A is wirelessly communicatively coupled to the second processor 16B. The detector 12 may be physically connected to the first processor 16A. The stimulator device 14 may be physically connected to the second processor 16B.
[0133] At least one battery may provide electrical power to one or more components of the blood pressure regulation system 10. In at least one embodiment at least a first battery provides electrical power to the detector 12 and at least a second battery provides electrical power to the stimulator device 14. A battery of the blood pressure regulation system 10 may be maintained at a required charge using any suitable charging method (such as a radiofrequency charging method, for example). Alternatively, a battery of the blood pressure regulation system 10 may be replaceable.
[0134] The blood pressure regulation system 10 may be, or include, a brain-spine interface (BSI). The blood pressure regulation system 10 may be operable to titrate bloodpressure correlated neural activity on spinal cord tissue thereby providing situational updates to the patient’s blood pressure and bridging a neural communication gap caused by a spinal cord injury.
[0135] Features and / or aspects of the blood pressure regulation system 10 will now be described in further detail.
[0136] Short-term stabilization of blood pressure may be accomplished by the patient’s arterial baroreflex. This reflex modulates circulation in the body in response to changes in behaviour or activity, including eating, exercise, changes in temperature, and / or emotions. In response to high blood pressure, the baroreflex will inhibit cardiac, renal, and / or vasomotor sympathetic efferents. It can accomplish this through a variety of sensory afferents that inform the brain on the status of the cardiovascular system. The sensory afferents are implemented through arterial baroreceptors (i.e., mechanosensitive neurons with nerve endings on the carotid sinus and aortic arch) which sense changes in vascular stretch of the corresponding arteries. According to the Hagen-Poiseuille equation, the diameter of a blood vessel has a quartic relationship with blood pressure. Contrastingly, blood flow has a linear effect on blood pressure. Thus, the stretch-sensing capability of arterial baroreceptors initiates the patient’s sympathetic circuitry responsible for short-term blood pressure stabilization. The axons from these mechanosensitive neurons are housed in the glossopharyngeal and vagus nerves, conveying baroreceptive afferent inputs to the brainstem. As a result, these afferent signals are not commonly disrupted by spinal cord injury.
[0137] Upon an electrical stimulus signal being applied to the patient’s spinal cord tissue by the stimulator device 14, the patient’s baroreflex may be caused to adjust the patient’s blood pressure accordingly in response to the applied electrical stimulus. As described herein, the baroreflex may also at least partially provide feedback to the brain on a current state of the patient’s blood pressure. In response to the current state, the brain may, for example, signal for further stimulation of the patient’s spinal cord tissue (such as to increase blood pressure, for example) or may signal for stimulation of the patient’s spinal cord tissue to stop (such as, for example, to hold blood pressure at a current value or to decrease blood pressure). In at least one embodiment, the patient’s baroreflex provides sufficient feedback such that the blood pressure regulation system 10 need not comprise another feedback mechanism (such as feedback from the stimulator device 14 being input into the at least one processor 16, for example).
[0138] Visceral afferent signals from mechanoreceptive neurons synapse directly with the Nucleus of the Solitary Tract in the brainstem. The Nucleus of the Solitary Tract may,for example, serve as a centre in the brain where multiple cardiopulmonary afferents from the body project, including, for example, baroreceptors, volume receptors and peripheral chemoreceptors. The Nucleus of the Solitary Tract is also responsive to pain signals. Afferent inputs to the Nucleus of the Solitary Tract have pulse-synchronous discharges, but neurons within the Nucleus of the Solitary Tract are not pulse-mediated, such that the Nucleus of the Solitary Tract may integrate multiple inputs before relaying the signal to other areas of the brainstem for further processing. Moreover, neurons within the Nucleus of the Solitary Tract are modulated by neurons from other cardiovascular-related processing centres of the brain, such as the hypothalamus, amygdala and the ventrolateral medulla. For example, the hypothalamus defense area (i.e., the region of the brain where the “flight-or-fight” response is believed to originate) significantly inhibits the Nucleus of the Solitary Tract when activated.
[0139] The outputs from the Nucleus of the Solitary Tract that are related to the baroreflex project to both the caudal ventrolateral medulla and the rostral ventrolateral medulla. These regions of the brainstem are functionally different, but can also be distinguished between the presence of y-aminobutyric acid (GABA) neurons in the caudal ventrolateral medulla, and the glutamatergic neurons in the rostral ventrolateral medulla. Both these regions of the medulla play a role in the central baroreceptor mechanism. The Nucleus of the Solitary Tract excites neurons in the caudal ventrolateral medulla, which then inhibit excitatory neurons in the rostral ventrolateral medulla when blood pressure elevates. Therefore, the connection and proximity between both medullary regions play a large role in the sympathetic circuity that underlies blood pressure stabilization. Excitation of y-aminobutyric acidergic caudal ventrolateral medulla neurons, triggers a reduction in blood pressure and sympathetic nerve activity, while their inhibition has an opposite effect. Additionally, caudal ventrolateral medulla neurons have pulse-mediated activation at high blood pressure such that they may play an important role in the baroreflex neural circuitry. Neurons that project from the Nucleus of the Solitary Tract monosynaptically to both the caudal ventrolateral medulla and the rostral ventrolateral medulla innervate caudal ventrolateral medulla neurons more quickly. Thus, the caudal ventrolateral medulla may have an important role in mediating the activity of the rostral ventrolateral medulla.
[0140] The rostral ventrolateral medulla may be a critical region for the sympathetic control of blood pressure as it is the final brainstem region that processes and integrates commanding signals before relaying the commanding signals to the spinal cord. As a result, neurons within the rostral ventrolateral medulla may be the primary drivers for blood pressure regulation, cardiac output and sympathetic outflow. Excitatory activity of neuronswithin the rostral ventrolateral medulla may have a notable impact on increasing blood pressure. For example, electrical stimulation of the rostral ventrolateral medulla may cause increases in mean arterial pressure of up to 50 mmHg, depending on stimulation intensity.
[0141] A population of cholinergic neurons within the rostral ventrolateral medulla, known as C1 neurons, are associated within the sympathetic control of blood pressure due to their ability to synthesize epinephrine. Some C1 neurons traverse the spinal cord forming efferent monosynaptic connections that innervate sympathetic preganglionic neurons. These sympathetic preganglionic neurons project to post-ganglionic neurons which directly innervate blood vessels thereby modulating blood pressure.
[0142] Bulbospinal C1 neurons are a subset of the neurons found within the rostral ventrolateral medulla that traverse the spinal cord and monosynaptically innervate sympathetic preganglionic neurons. As such, bulbospinal C1 neurons may be a primary driver of the patient’s baroreflex. These neurons have been characterized based on their electrophysiological activity and by the proteins that they express. They are tonically active at normal blood pressure levels and are proportionally inhibited as blood pressure increases. Their firing rate also increases as blood pressure drops. Bulbospinal C1 neurons contain many of the enzymes necessary for the synthesis of catecholamines within the sympathetic nervous system. This pathway begins with tyrosine, one of the common amino acids used in protein synthesis. Tyrosine is converted to L-Dopa via an enzyme known as tyrosine hydroxylase. L-Dopa is converted to dopamine through an enzyme known as Dopa decarboxylase (or Aromatic-L-amino acid decarboxylase (AADC)). Dopamine is then converted to norepinenphrine with the dopamine-p-hydroxylase enzyme. Lastly, phenylethanolamine-N-methltransferase is an enzyme that converts norepinephrine to epinephrine.
[0143] C1 neurons synthesize epinephrine and thus contain phenylethanolamine-N-methltransferase. As such, bulbospinal C1 neurons have been confirmed to contain phenylethanolamine-N-methltransferase using immunohistochemistry experiments. While many neuronal populations within the rostral ventrolateral medulla express tyrosine hydroxylase, such as for example noradrenergic A1 neurons, bulbospinal C1 neurons also express tyrosine hydroxylase. Moreover, retrograde fluorescent tracing experiments from the spinal cord have demonstrated that bulbospinal C1 neurons also contain dopamine-P-hydroxylase. Thus, the bulbospinal C1 neurons contain the necessary enzymes for the synthesis of both norepinephrine and epinephrine. Given that bulbospinal C1 neurons descend the spinal cord, high-level spinal cord injury disrupts C1 neuronal control of bloodpressure below the level of injury. However, the majority of phenylethanolamine-N-methltransferase+ neurons within the rostral ventrolateral medulla survive losing their spinal projections after a severe spinal cord injury (e.g. about 80% of C1 neurons may survive losing their spinal projections following a spinal cord injury). As such, detection of neural activity from bulbospinal C1 neurons by, for example, the detector 12 may provide a brain computer interface (BCI) for controlling or regulating the patient’s blood pressure following a spinal cord injury. As described above, the detector 12 may be positioned proximal to at least one bulbospinal C1 neuron. In at least one embodiment, the at least one electrode 13 of the detector 12 is positioned proximal to at least one bulbospinal C1 neuron.
[0144] As described above, the detector 12 may comprise at least one electrode 13. In at least one embodiment, the detector 12 comprises a plurality of electrodes 13.
[0145] In at least one embodiment, an electrode 13 comprises a thin-diameter electrode that penetrates a surface of the cortex of the patient’s brain to reach deeper cortical structures (i.e., an invasive electrode). In at least one embodiment, the thin-diameter electrode has a diameter in a range from about 4 pm to about 80 pm. The thin-diameter electrode may be, or include, a needle-like electrode. The needle-like electrode may be implantable within or proximal to the rostral ventrolateral medulla. In at least one embodiment, deep brain stimulation electrodes are used to record neural activity and the electrode 13 comprises a deep brain stimulation electrode. Reaching deeper cortical structures permits the electrode 13 to measure neural activity in close proximity to the electrode 13. In at least one embodiment, the electrode 13 measures neural activity within about 1.5 mm of the electrode 13. In at least one embodiment, the electrode 13 measures neural activity within about 1 mm of the electrode 13 (such as when the electrode 13 is configured to measure local field potentials (LFPs), for example). In at least one embodiment, the electrode 13 measures neural activity within about 300 pm (such as when the electrode 13 is configured to use a spike-based method (e.g. a method based on neural firing rates, etc.), for example). In at least one embodiment, plural electrodes 13 are spatially arranged in an organized structure such as a microelectrode array or cortical array, for example. In at least one embodiment, the cortical array is a high-density cortical array with a density of over 500 electrodes per mm2. Such invasive electrodes 13 may be used to detect neural activity signals that occur at different frequencies.
[0146] One type of neural activity signal that may be detected are low-frequency signals (i.e., having a frequency less than about 500Hz) which may be referred to as local field potentials (LFPs). The local field potentials represent transmembrane currents (orextracellular currents) that are detected in close proximity to each electrode 13 (e.g. within an area of about 100-200 pm around the electrode 13). Several factors may affect these extracellular currents. One such factor may be concentration gradients from ions entering and exiting neurons during action potentials. For example, a strong excitatory current may trigger the release of glutamate into a synapse, triggering the postsynaptic neuron to have an influx of sodium and calcium ions, which in turn affects the local field potential signal. Local field potential signals may be analyzed (or processed), by at least one of the processors described in the various embodiments herein, by considering different frequency spectrum bands associated with neural activity, such as the alpha, beta, and / or gamma frequency bands, to detect, for example, biomarkers correlated to blood pressure. Spectral analyses may be used to determine features for at least one frequency band, or across different frequency bands, such as calculating the area under a periodogram to obtain band-power. In at least one embodiment, a biomarker correlated to blood pressure includes a spectral signature between about 80 to about 200Hz. In at least one embodiment, a biomarker correlated to blood pressure includes a spectral signature between about 120 to about 200Hz. Presence of a biomarker correlated to blood pressure may be detected by performing a frequency analysis of a detected local field potential signal and detecting / scanning for a spectral signature corresponding to the biomarker.
[0147] Additionally, or alternatively, another type of neural activity signal that may be detected is a higher frequency signal (i.e., having a frequency higher than about 500 Hz) which represents action potentials of proximal neurons. For example, an electrode 13 may be operable to detect action potentials from neurons in an area within about 50 pm around the electrode 13. By capturing such action potentials, it is possible to explore how many action potentials fire in a given period of time to quantify neural activity. This process may be referred to as multiunit activity (MUA), which aims to determine how the spiking rate of neurons varies in response to different stimuli. In at least one embodiment, action potentials are further grouped based on shared similarities in their waveforms, such as peak-to-peak amplitudes, wavelet transform coefficients, and / or zero-crossing rates in a process known as spike-sorting. Such process may identify single units. A given single unit may be attributed to an individual neuron close to the electrode. In at least one embodiment, a biomarker correlated to blood pressure is detected using a thresholding analysis. In at least one embodiment, a biomarker correlated to blood pressure is detected if a number of action potentials within a set amount of time (such as one second, for example) is higher than a threshold amount (such as about 4 times the standard deviation or about 5 times the standard deviation higher than a measured baseline value for theaction potential signal, for example). In at least one embodiment, the baseline value for the action potential signal is determined by recording for a duration of time (such as a value of time in a range from about 1 minute to about 5 minutes, for example) the action potential signal without any experimental conditions being applied and calculating (or determining) the biomarker associated with this recorded duration of the action potential signal.
[0148] In at least one embodiment, the biomarker correlated to blood pressure is determined (or detected) at least partially by isolating one or more neurons based on template matching. In at least one embodiment, isolating one or more neurons based on template matching comprises determining a threshold above which a signal captured by an electrode would be considered an action potential. The action potentials may then be processed to consider different features of the waveforms of the action potentials (e.g. a minimum value, a maximum value, a peak-to-peak distance, a width of an action potential, a zero-crossing rate, etc.). The features for each action potential may be combined into a dataset and may be statistically grouped together. The features may, for example, be statistically grouped together using an algorithm such as principal component analysis. The features may be statistically grouped together to attribute a given number of action potentials to a statistically distinct neuron (or group of neurons). A given number of action potentials may be attributed to a statistically distinct neuron (or group of neurons), for example, by using a clustering algorithm such as K-means.
[0149] In at least one embodiment, the at least one electrode 13 comprises an electrocorticography (ECoG) electrode, which may be placed on the surface of the medulla (i.e., does not penetrate the cortical surface). The electrocorticography electrode may detect neural activity similar to the local field potentials. In at least one embodiment, electrocorticography is used to detect neural activity with a spatial resolution of about 5 mm². In at least one embodiment, an electrocorticography signal is analyzed using a method similar to a method used to analyze a local field potential signal.
[0150] In at least one embodiment, the at least one electrode 13 comprises an electroencephalography (EEG) electrode, which may be placed non-invasively on the scalp. In at least one embodiment, electroencephalography is used to detect neural activity with a spatial resolution of about 10 cm². In at least one embodiment, an electroencephalography signal is analyzed using a method similar to a method used to analyze a local field potential signal.
[0151] Stimulating the rostral ventrolateral medulla may lead to an increase in bandpower between about 5 to about 6 Hz. As such, electrocorticography electrodes may beplaced at the dorsal surface of the medulla for detecting local field potential signals from the rostral ventrolateral medulla. Moreover, electroencephalography electrodes may be placed at a region of the scalp closest to the dorsal medulla for correlation (such as spearman correlation or spectral coherence, for example) with signals from multielectrode arrays.
[0152] In at least one embodiment, the at least one electrode 13 comprises an endovascular electroencephalography (EE) electrode, which may be used to capture (or detect) neural signals without the need for invasive surgeries to access the brain. With endovascular electroencephalography, one or more EE electrodes may be placed within blood vessels (such as veins and / or arteries, for example) in close proximity to a cortical structure of interest. In at least one embodiment, EE electrodes may be placed on (or be part of) cardiovascular stents (which may be known as a stentrode, which may be available from Synchron). One or more of these stentrodes may, for example, be delivered to the neural vasculature through a vein or artery. The neural signals acquired from these EE electrodes may be comparable to those obtained using electrocorticography electrodes. In at least one embodiment, the EE signal(s) may have a usable bandwidth up to about 250 Hz. In at least one embodiment, the EE signal(s) may have a spatial resolution of about 3 mm. In some cases, an endovascular electrode is capable of recording neural activity from a blood vessel in the brain where the blood vessel has a diameter of less than about 100 pm.
[0153] Additionally, or alternatively, blood pressure correlated neural activity may be captured (or detected) indirectly. The indirect detection may be non-invasive. For example, the detector 12 may be, or include, a functional near-infrared spectroscopy (fNIRS) detector. At near-infrared (NIR) wavelengths (i.e., about 650 to about 950 nm), the scalp, the skull and the brain are relatively transparent thereby enabling NIR light to reach depths of up to about 35mm. The fNIRS signal may measure light reflected by cortical tissues. At these wavelengths, the most dominant attenuation of light is from hemoglobin located in small blood vessels. As neurons fire more rapidly, there is an increased metabolic demand for oxygen, leading to an increase in local oxygenated blood and glucose (i.e., the hemodynamic response). Subsequently, this also leads to a local decrease in deoxygenated blood shortly thereafter. These changes in local blood flow in response to neural activity are mediated by astrocytes near these neurons in an interaction known as neurovascular coupling. Blood oxygenation changes affect the local concentrations of oxygenated and deoxygenated hemoglobin which can then be measured using functional near-infrared spectroscopy. At wavelengths below about 800 nm there is high absorptionof NIR light by deoxygenated hemoglobin. Conversely, at wavelengths above about 800 nm there is high absorption of NIR light by oxygenated hemoglobin. As a result, these changes in local concentrations of hemoglobin can be used to indirectly capture (or detect) neural activity noninvasively.
[0154] Additionally, or alternatively, rostral ventrolateral medulla activity may be captured (or detected) by measuring oxygenation of the vertebral artery or one of its branches (such as the posterior inferior cerebellar artery, for example). The detector 12 may be, or include, a detector that is operable to directly, or indirectly, measure oxygenation of the vertebral artery or one of its branches.
[0155] In at least one embodiment, the detector 12 uses at least two of the detection modalities described herein to detect the at least one electrical signal.
[0156] Referring now to FIG. 3, program codes stored in the program-codes storage device 17 may include blocks of program codes for performing a method 20 that, when executed by the at least one processor 16, cause the at least one processor 16 to control application of the at least one electrical stimulus signal by the stimulator device 14 in accordance with one example of the teachings herein. It should be noted that while the at least one processor 16 is referred to in describing the method 20, in other embodiments, at least one of processor 16A and 16B may execute software to perform method 20.
[0157] Block 22 includes program codes that, when executed by the at least one processor 16, cause the at least one processor 16 to determine whether at least one blood pressure correlated biomarker (such as any blood pressure correlated biomarker described elsewhere herein, for example) has been captured (or detected) by the detector 12. Determining whether the at least one blood pressure correlated biomarker has been captured (or detected) by the detector 12 may include causing the at least one processor 16 to analyze the signal(s) captured (or detected) by the detector 12 for the at least one blood pressure correlated biomarker. Causing the at least one processor 16 to anaylze the signal(s) captured (or detected) by the detector 12 may, for example, include detecting a threshold number of neural firings or a specific spectral signature of one or more local field potentials. If at least one blood pressure correlated biomarker is detected, the method 20 may continue to block 24. Otherwise, the method 20 may remain at the block 22.
[0158] Block 24 includes program codes that, when executed by the at least one processor 16, cause the at least one processor 16 to determine a stimulus to be applied by the stimulator device 14 based on the detected at least one blood pressure correlated biomarker in accordance with one of the embodiments described earlier. For example, the presence of the at least one blood pressure correlated biomarker may indicate that anincrease in blood pressure is desired. One or more parameters of the stimulus (such as amplitude, duration, and / or frequency, etc.) may be set based on one or more parameters of the at least one blood pressure correlated biomarker. In at least one embodiment, the stimulus is like, or includes, any stimulus described in Phillips et al., “The implantable system that restores hemodynamic stability after spinal cord injury”, May 15, 2024, doi: https: / / doi.org / 10.1101 / 2024.05.10.24306826, which is incorporated by reference herein for all purposes. In at least one embodiment, the stimulus is determined by the at least one processor 16 in a feedforward manner. For example, the at least one processor 16 may input the at least one blood pressure correlated biomarker into a model to determine the stimulus.
[0159] Block 26 includes program codes that, when executed by the at least one processor 16, cause the at least one processor 16 to control the stimulator device 14 to apply the determined stimulus to the spinal cord tissue of the patient for a certain time duration. Upon the stimulus being applied, the blocks 20 may return to the block 22. In at least one embodiment, the blood pressure regulation system 10 is caused to update (i.e. a “tick duration” of the blood pressure regulation system 10) in a range from about every 1 second to about every 3 seconds.
[0160] As described above, the blood pressure regulation system 10 may not include a feedback controller. Rather, the blood pressure regulation system 10 may determine a stimulus to be applied in a feedforward manner based on the captured (or detected) neural activity. Upon the patient’s body being stimulated, the patient’s body may self-regulate the change in blood pressure using, for example, the baroreflex. However, electronic feedback may additionally be used. For example, in at least one embodiment, the blood pressure regulation system 10 comprises at least one feedback controller. For example, the feedback controller may be a Proportional-Integral-Derivative (PID) controller that may be implemented by the at least one processor 16. In such embodiments, a measure of the patient’s blood pressure may be provided as a feedback input into the feedback controller. As another example, detected neural activity from bulbospinal C1 neurons may be added as a compensator feedback input to the PID controller. The feedback controller may at least partially modify the stimulus that is applied to the patient by the stimulator device 14 to regulate blood pressure based on these one or more feedback inputs.
[0161] In at least one embodiment, the feedback controller receives at least one signal captured by a sensor configured to directly measure blood pressure. For example, the feedback controller may receive a two-second sliding window of the at least one signal. The feedback controller may determine an average blood pressure during the window.The feedback controller may also receive as input a user defined blood pressure target (such as 100 mmHg, for example). The feedback controller may determine a difference between the measured blood pressure and the user defined blood pressure target. The feedback controller may at least partially determine an amplitude of an epidural electrical stimulation based on the determined difference between the measured blood pressure and the user defined blood pressure target. Neural signals detected as described herein may be added as an additional input to the feedback controller.EXPERIMENT 1
[0162] In this section various experimental results are presented showing the link between certain neural activity and blood pressure. While changes in blood pressure in these examples is effected by chemical means, the effect provides support of the link between blood pressure changes and certain neural activity. Consequently, the control of neural activity provided by the various embodiments described herein will have a corresponding effect on blood pressure changes.EXPERIMENTAL SETUP
[0163] The experimental results were captured using anesthetized female Sprague Dawley adult rats weighing between about 200-250g. A blood pressure catheter was implanted in the descending aorta of the rat. A craniotomy was performed above the medulla. A multi electrode array was implanted in the rostral ventrolateral medulla (RVLM) of the rat, and the activity of neurons was recorded in response to blood pressure perturbations. These perturbations include sodium nitroprusside (SNP) injections at 20 pg / kg, phenylephrine (PE) injections at 10 mg / kg, lower body negative pressure (LBNP), hypoxia (up to -10% relative O₂ to atmospheric conditions), and hypercapnia (up to +10% relative CO2 to atmospheric conditions).EXPERIMENTAL RESULTS
[0164] FIG. 4 depicts a reduction in neural firing rates in response to increases in BP. Following the injection of phenylephrine, the firing rates of neurons close to the electrodes have a significantly reduced firing rate.
[0165] Conversely, FIG. 5 demonstrates that low blood pressure regimes have an increased firing rate. Here, following the injection of sodium nitroprusside, the firing rates of neurons close to the electrodes have a significantly increased firing rate. In FIG. 5, each binned firing rate during the baseline considers the first 150s prior to the reduction in BP caused by the vasodilator application. The first 60s following the initial drop in BP are considered as the low BP regimes. In both experimental groups, the data is concatenated to create the boxplots.
[0166] Both of the changes in neural firing rates shown in FIGS. 4 and 5 are typical of bulbospinal C1 neurons [53, 1],
[0167] FIG. 6 shows a decrease in neural firing rates following cyclical reductions in blood pressure while FIG. 7 shows a correlation between neural firing rates and systolic blood pressure. The data from FIGS. 6 and 7 show significant correlations between the neural firing rate and both systolic blood pressure (SBP) and mean arterial pressure (MAP). For example, FIG. 7 shows a significant correlation between systolic blood pressure and mean neural firing rates, when binning the data over a period of 8 seconds.
[0168] Referring now to FIG. 8, shown therein are results from an experiment in which inhaled oxygen was reduced, and the firing rate of neurons started to increase notably at -8% relative 02. As blood pressure decreased as well, the firing rate increased further.
[0169] The following data shows LFP analyses as an alternate for firing rate-based analyses. LFP analyses may present better clinical translation, as less invasive neural recording modalities such as the stentrode or ECoG electrodes function on LFPs rather than MUA [170, 73], Datasets were analyzed using both LFP and firing rate-based signal processing methods for determining a novel biomarker that may be correlated with blood pressure. The analysis involved determining the power spectral density (PSD) of different neural signal processing bands (e.g., alpha, beta, delta, gamma, etc.) as described in Table 1. In each band, additionally derived features were determined such as the bandpower, band-power², the ratio between bands and the mean PSD of each band. Each of these features were then correlated to both blood pressure and multiunit firing rates. Band Frequencies (Hz)Delta 1-1Theta 4-8Alpha 8-12Beta 12-30Gamma 30-80High-Gamma- 1 80-125High- Gamma- 2 125-200Spectrum 1-200Table 1: LFP processing bands
[0170] This analysis showed that the band-power in the high-gamma range (e.g., about 80 Hz to about 125 Hz) tended to have strong correlations with blood pressure and multiunit firing rates, which is shown in FIGS. 9-11.
[0171] Referring now to FIG. 9, it was observed that following the injection of a vasopressor (phenylephrine), the mean arterial pressure (MAP) increased sharply (top panel of FIG. 9). This is reflected in the LFP trace (in the middle panel of FIG. 9 in red(and labeled R1)). However, after a relative increase to 170 mmHg, the signal decays considerably. This suggests that there is a relative set-point within the High-Gamma band between 80-200 Hz. In the bottom left panel, there is a strong and significant correlation between the binned High-Gamma power-band and MAP. In the bottom middle panel, the power decreases significantly compared to baseline data using a Mann-Whitney-Wilcoxon test. The correlation between both the high-gamma LFP band-power and MUA firing rates is shown in the bottom right panel of FIG. 9.
[0172] FIG. 10 shows that the correlation between the multi-unit firing rates and the high-gamma LFP band-power in FIG. 9 is quite strong. Here, a linear regression was performed between both parameters, yielding a strong R2 and a highly significant relationship.
[0173] Referring now to Fig. 11, the LFP analysis was repeated with the chemoreflex experimental data, yielding similar results during hypoxia (see FIG. 10 for MUA firing rates on this data set). Here the data captured neural activity during hypoxic conditions. Firing rates increased notably at -8% and -10% (as seen in fourth panel), with a small increase seen in LFP band-power in the high-gamma band. There is a modest negative relationship between high-gamma band-power (bottom left panel) and small correlation with MUA firing rates (bottom right panel). Significant increases in the band-power occurred at higher hypoxic conditions, as seen in the bottom middle panel. These results indicate that the high-gamma band-power may be a novel biomarker for blood pressure when recording within the RVLM.
[0174] The same multiunit pipeline was used on a recent set of experiments and characteristic behaviour of bulbospinal C1 neurons in response to decreases in blood pressure [30, 53] was found. This is seen in the multiunit firing rates in FIG. 12, where, following injections of sodium nitroprusside, the firing rate of neurons captured by an electrode increases proportionally to the decrease in blood pressure. Similar results can be seen with the local field potential band power in a high-gamma band shown in FIG. 13 in which the analyses was done on the same data, as presented in FIG. 12, but using local field potential analysis approaches. In the two top panels of FIG. 13, the mean arterial pressure and the LFP (high-gamma between 125-200Hz) are shown and time synced; Periods labeled R3 (having a US-shape) are notable decreases in blood pressure, which also align to notable increases in band-power in this range. A modest correlation between LFP band-power and MAP is observable in the bottom left panel of FIG. 13, with a significantly increased band-power when compared to baseline periods, as seen in the bottom-centre panel of FIG. 13. Lastly, a small correlation can be seen with the high-gamma LFP band-power and the multi-unit firing rates (as seen in FIG. 12).
[0175] As described previously, a less invasive approach to access the RVLM without the need for a craniotomy may involve the use of endovascular electrodes, such as the Stentrode from Synchron [170, 73, 64], This technology has previously been used in rodent models by creating ultraflexible endovascular electrodes at micrometer scales
[0072] , However, to date, these recordings have primarily been in motor control regions of the brain [72, 73],EXPERIMENT 2Animal Model
[0176] Experiments were conducted on adult female Sprague-Dawley rats (200-300 g body weight, 10-16 weeks of age; Charles River Laboratories, Canada). This strain was selected because the location of the RVLM is well characterized in the literature [19, 35], Female rats were used exclusively, consistent with previous work from this laboratory
[0017] , as observationally female rats exhibited higher post-operative survival following SCI. Housing, surgeries, and euthanasia were performed in compliance with the Canadian Council on Animal Care (CCAC) guidelines. All experimental procedures were approved by the University of Calgary’s Health Sciences Animal Care Committee under ethics protocol numbers AC21 -0230 and AC24-0147.
[0177] Unless otherwise specified, all experiments used a group size of five rats, consistent with prior work from our laboratory [17, 58] and with previous literature exploring preclinical BSIs
[0078] ,Rat Anaesthesia protocol
[0178] Surgical procedures were performed under general anesthesia. Initial induction was achieved using 5% isoflurane. As the experiments were terminal, urethane (1.5 g / kg, intraperitoneally) was administered to provide long-lasting anesthesia (6-10 hours) with minimal cardiovascular and respiratory depression
[0017] , Isoflurane was then lowered to 1 % and maintained throughout all surgical procedures. Rodents were placed on a heating pad to maintain body temperature. The depth of anesthesia was closely monitored during all experimental protocols by assessing withdrawal reflexes and respiratory rate.Hemodynamic monitoring
[0179] BP was recorded on a beat-by-beat basis using a wired catheter (Mikro-Tip SPR-320 or SPR-215, depending on availability; Millar, United States). The catheter was connected to both an analog- to-digital converter (PowerLab, ADInstruments, New Zealand) and a digital signal processor (RZ2 BioAmp Processor, Tucker-Davis Technologies, United States) by splitting the analog signal. Methods for synchronizing dataacross both computers are described below. Calibration of the BP catheter was performed using a pressure control unit (PCU-2000, Millar, United States). BP data was acquired at 1,000 Hz within LabChart (ADInstruments, New Zealand) and stored locally for analysis.
[0180] After anesthesia was induced, the hair on the neck was shaved, and the surgical site was cleaned with alcohol and iodine. The right carotid artery was exposed and isolated using blunt dissection. The rostral portion of the carotid artery was permanently occluded, while the caudal portion was temporarily occluded using silk sutures. A small incision was made in the arterial wall using a bent-tip needle and the catheter was inserted into the lumen of the carotid artery, advanced 5-6 mm, and secured with two silk sutures. In all experiments, the BP catheter was implanted in conjunction with the RVLM MEA to enable simultaneous hemodynamic and neural recordings.RVLM microelectrode array implantation
[0181] RVLM access for MEA implantation was achieved via a craniotomy. Briefly, a midline incision was performed from the frontal bone to the atlanto-occipital junction to expose the calvarium. The calvarium was thinned using a dental drill and irrigated with saline at room temperature over an area of 3 mm by 4 mm. Bone was removed until only the lamina vitrea remained
[0027] ,
[0182] The rat’s head was secured in a stereotaxic frame, and the RVLM was identified using anatomical landmarks based on previous studies
[0035] and stereotaxic coordinates [17, 66], Coordinates used for implantation were -12.2 mm anteroposterior from the bregma, 2 mm lateral, and 7.8 mm ventral from the surface of the cerebellum
[0017] , A custom-manufactured MEA (4-channel ZIF- Clip Microwire Array, Tucker-Davis Technologies, United States) was implanted at the specified coordinates, as outlined in FIGS. 14A-14B. The microwire lengths were staggered by 0.5 mm for concurrent placement at different dorsoventral depths within the RVLM to target C1 neurons
[0035] , The MEA was connected to an amplifier (PZ5 NeuroDigitizer, Tucker-Davis Technologies, United States), routed through a real-time processor (RZ2 BioAmp Processor) and data streamer (RS4), and acquired and stored on a WSH workstation running Synapse (Tucker-Davis Technologies, United States). Neural data was recorded at a sampling rate of 24,414.0625 Hz.
[0183] FIG. 14A shows a schematic overview of the experimental setup. A MEA was implanted dorsally in the RVLM for neural recording. A BP catheter was implanted in the descending aorta, enabling time-aligned measurement of neural activity within the RVLM and arterial BP. FIG. 14B shows a photograph of the custom 4-channel ZIF-Clip Microwire Array (Tucker-Davis Technologies, United States). Recording electrodes were polyimide-insulated tungsten microwires (50 m diameter) cut at 60° tip angles. The array consistedof four evenly spaced recording electrodes with 250 / zm inter-electrode spacing and staggered electrode lengths ranging from 8.5 to 10 mm (0.5 mm increments). The arrays were additionally fabricated with single-ended ground and reference wires, each with 3-inch lengths (7.6 cm) and 0.003-inch diameters (76 / zm). Newly manufactured electrodes were coated with polyethylene glycol to prevent wire deformation during shipping; this coating was removed with saline during surgical insertion.Experimental Design
[0184] While C1 neurons have been extensively studied [1, 3, 6, 16, 23, 25, 30, 35], their characterization using modern neural signal processing techniques remains underexplored. Here, C1 neuron activity was characterized in response to a variety of autonomic perturbations in anesthetized rats. These include pharmacological agents, gas exposure protocols, and physiological perturbations.
[0185] Perturbations were not applied in a fixed order across animals, but were not explicitly randomized. Instead, the order of experiments was determined by experimental constraints and animal vital sign stability. For example, hypotensive perturbation trials were not performed if resting MAP was already low. As all experiments were conducted during acute, terminal preparations, each animal underwent only a subset of the perturbation protocols outlined below.Baroflex perturbations
[0186] In order to quantify how the activity of C1 neurons changed in response to BP perturbations, baroreflex activation and inhibition experiments were performed. These included intravenous administration of sodium nitroprusside (SNP) and phenylephrine (PE), as well as the use of a custom-built lower body negative-pressure (LBNP) chamber
[0017] , Pharmacological injections were administered as a bolus through an intravenous line in the femoral vein. Prior to any perturbation, a 30-minute baseline period was recorded to establish parameters for neural signal processing (e.g., thresholds for multiunit spike detection
[0080] ).
[0187] Sodium nitroprusside: Baroreflex activation was achieved using SNP at a dose of 10-20 jzzg / kg [17, 35], which lowered BP. SNP injections were repeated a minimum of three times per experiment, with 5-10-minute intervals between successive injections.
[0188] Phenylephrine: Baroreflex inhibition was achieved using PE at a dose of 10 / zg / kg
[0035] , which increased BP substantially. Elevated BP following the BP injection was prolonged, and was only performed once or twice per experiment.
[0189] Lower body negative pressure: A custom-built LBNP chamber was used to reduce BP with precise control, as described in previous work by our laboratory
[0017] , which is incrorporated herein in its entirety for all purposes. Briefly, rats were placed prone in a sealed rectangular chamber, from their feet roughly up to their sternum. Negative pressure was induced using a vacuum pump attached to a servo-regulator (ITV2000Q, SMC Corporation, United States), which enabled precise control of pressure inside the chamber. The analog output of the chamber pressure was fed to the computer running LabChart for time-synchronized recordings with the BP catheter, sampled at 1,000 Hz. The LBNP chamber leveraged a custom-built program that incorporated a PI D controller to ensure that a constant negative pressure was applied. Following a calibration of the PID controller, cyclical LBNP was applied at -5 mmHg for 60 seconds followed by 0 mmHg for 60 seconds for at least 10 cycles. In some experiments, LBNP was applied for 60 seconds at randomly selected pressures (e.g., 0, -5, or -10 mmHg). The experimental setup for these experiments is illustrated in FIGS. 15A-15B.
[0190] FIG. 15A shows a schematic overview of the experimental setup for the LBNP experiments in a rodent model. The lower body of the rodent was placed in a custom-built vacuum chamber attached to a vacuum pump and servo- regulator. Negative pressure within the chamber was monitored using a manometer. FIG. 15B shows a schematic overview of the experimental setup for the chemoreflex experiments. Carbon dioxide, oxygen, and nitrogen gasses were connected to a gas mixer and delivered at specified concentrations and fixed flow rates to intubated rats.Chemoreflex perturbations
[0191] In order to assess the extent to which the activity of C1 neurons is modulated by the central and / or peripheral chemoreflex pathways, experiments modulating gas inhalation were performed. Rats were intubated following an intubation protocol
[0115] that was adapted from mouse models for use in rats. Briefly, rats were placed supine on a stereotaxic frame, and the ventral part of the neck was shaved and disinfected with 70% isopropyl alcohol. An incision was made in the neck below the incisor, and skin was carefully retracted with forceps until the ventral wall of the trachea was visible. The tongue was retracted, and a cannula was inserted into the trachea. The catheter was advanced until it was positioned between the vocal cords and the carina. The cannula was then attached to the ventilator with an adaptor.
[0192] The ventilator system (SAR-1000 Small Animal Ventilator, GSM-33-Channel Gas Mixer, Cl / I / E inc., United States) enabled fine-tuned mixing of anesthetic gas (isoflurane),oxygen, nitrogen, and carbon dioxide. Prior to all experiments, a 30-minute baseline was recorded at atmospheric conditions (21% O2, 79% N2). The inhaled gas concentrations (O2, N2, CO2) were streamed to the computer running LabChart for time-synchronized recordings with the BP catheter. The experimental setup for these experiments is illustrated in FIG. 15B.
[0193] Hypercapnia: The levels of hypercapnia exposure were achieved by modifying the percentage of inhaled CO2 relative to atmospheric conditions. Increases in CO2 were matched by equal decreases in N2. First, the relative concentration of CO2 was increased stepwise from 0% to 10% with intervals of 2%. Each modification to the inhaled gas was held for 2 minutes, with a ramp up to 10%, then a ramp back down to 0% with the same intervals and durations. Second, a jump from 0% to 10% and back to 0% of relative CO2 inhalation was performed, with each step lasting 5 minutes. These two interventions were run in separate trials.
[0194] Hypoxia: The levels of hypoxia exposure followed a similar protocol to the above hypercapnia protocol. However, in this case, the relative concentration of inhaled O2 was decreased and replaced with equal increases in N2. First, the relative concentration of O2 was decreased stepwise from 21% to 11% with intervals of 2%, each lasting 2 minutes. Then, the relative concentration of O2 was increased in a stepwise ramp back to 21% with the same intervals and durations. Second, a jump from 21% to 11% and back to 21% of inhaled concentrations of inhaled O2 was performed, with each step lasting 5 minutes. Both these interventions were performed in separate trials.Other pertubations
[0195] To test whether C1 neurons could be modulated in response to non-reflex pathways, direct activation of the sympathetic nervous system was performed. Injections of glucagon-like peptide-1 (GLP-1) at a dose of 1-20 / zg / kg were injected using the intravenous injection protocol described in the baroreflex perturbation section. GLP-1 injections have previously been shown to increase blood corticosterone and increase BP in anesthetized rats [116, 117]. GLP-1 in the brain is associated with activation of the hypothalamic-pituitary-adrenal (HPA) axis, which is related to the stress response
[0116] ,Signal Processing and data analysis
[0196] Signal processing, analysis, and interpretation of the data was done using scripts written in Python 3.12.2. The scripts were written using standard scientific libraries [US-123], the Python Tucker-Davis Technologies application programming interface (Tucker-Davis Technologies, United States), and additional packages, including Param, Panel, and HoloViz, for building interactive dashboards during experiments.
[0197] All analyses were performed within defined epochs to contrast baseline and perturbation conditions (SNP, PE, GLP-1, LBNP chamber pressure, pC>2, PCO2). When there were multiple perturbations (e.g., repeated SNP injections), each perturbation was paired with a baseline period immediately preceding it. During cyclical LBNP sessions, baselines were defined as periods when the chamber pressure was set to 0 mmHg. Chemoreflex experiments had unique epochs for each gas concentration (i.e., 0, 2, 4, 6, 8, and 10% PCO2). The same epoch definitions were applied across BP, neural, and perturbation data for standardized analysis pipelines across all experimental protocols.Blood Pressure
[0198] The analog signal from the BP catheter was split and recorded simultaneously in both computers running LabChart and Synapse. In addition to BP recorded at 1,000 Hz, all auxiliary signals for experimental protocols (e.g., LBNP chamber pressure, gas concentrations) and experimental comments were recorded in LabChart. Synapse recorded BP at 1,015 Hz in parallel with all the neural data recorded from the MEAs implanted in the RVLM.
[0199] Offline, both datasets were resampled to 1,000 Hz, zero-padded to equal length, and cross- correlated to estimate the temporal lag using the BP traces. The fixed lag was computed from the peak of the cross-correlation, and LabChart recordings were shifted by this fixed lag to align with the Synapse dataset. Experimental recordings were split for each protocol (e.g., LBNP, SNP, PE, etc.) and rarely exceeded one hour; therefore, clock drift between computers was considered negligible. LabChart BP was calibrated prior to each experiment and was used as the reference signal for all subsequent analyses.Beat Detection algorithm
[0200] Unbiased detection of systolic and diastolic peaks was obtained from the BP trace in four steps. First, the signal was low-pass filtered using a zero-phase second-order Butterworth filter
[0119] at 20 Hz. Second, candidate systolic peaks were identified using a peak-finding algorithm
[0119] , and the process was repeated on the sign-inverted trace to identify candidate diastolic peaks. Third, a histogram of peak prominences was computed, and a data-driven cutoff (the first non-decreasing point) was selected separately for systolic and diastolic peaks. Finally, peak detection was re-run using these prominence thresholds, and indices were mapped back to the low-pass filtered BP trace. Pilot testingconfirmed that these results were robust to filter cutoffs between 15-20 Hz, and a 20 Hz cutoff was used throughout. This process can be seen in FIGS. 16A-16C.
[0201] FIG. 16A shows a sample BP trace before 1602 and after 1604 applying a low-pass filter. FIG. 16B shows a histogram of systolic peak prominences from the filtered trace. The dashed vertical line indicates the data-driven prominence cut-off to distinguish systolic and diastolic BP from noise or artifacts. FIG. 16C shows identified systolic peaks (red) and diastolic troughs (blue) on the low-passed BP trace used for subsequent analysis.Calculation of derived BP Metrics
[0202] Systolic and diastolic peak series were grouped at the nearest time point to calculate MAP using the Gauer formula
[0124] :
[0203] Pulse pressure was calculated according to its clinical definition
[0125] :PP -SBP -DiiP(2)
[0204] Heart rate was derived from the interval between successive systolic peaks. The A-metrics were computed by subtracting the baseline mean from the corresponding experimental values. This was performed using the first few minutes prior to experimental perturbations, to better account for baseline drift that was observed across experiments, particularly following administration of pharmaceutical agents that modulate BP (PE, SNP). For example:A AP MAP - MAEb-^ inc(3)
[0205] Analogous calculations yielded ASBP, ADBP, APP, and AHR. All derived BP metrics were then resampled to 1,000 Hz, the same as the raw trace. These metrics were later averaged into time-bins of 2-seconds for multimodal analysis with MUA firing rates and LFP band powers.Multiunit AnalysisPreprocessing
[0206] Neural signals recorded from the MEAs implanted in the RVLM were sampled at 24,414.0625 Hz. Each of the four channels was processed independently. First, signals were band-pass filtered between 500 and 7500 Hz using a second-order Butterworth filter using second-order sections
[0119] , Forward (causal) filters were applied, which introducedlag, to match the real-time constraints required for the development of a brain-spine interface. Second, the filtered signal was z-score normalized (i.e., mean of zero and standard deviation of one). If brief high-amplitude noise exceeding 7.5 standard deviations was present in the signal, the values were clipped and replaced with the channel median.Threshold detection
[0207] The MUA was quantified using threshold crossings with a threshold defined a multiple of the median absolute deviation
[0080] :cr - median
[0208] The default threshold was set to 5 a, but was configurable and manually confirmed for each channel for each experiment. The multiple varied between 46- and 5.5< J across all experiments, due to differences in impedance values from the implanted MEAs. A sensitivity analysis demonstrated that results were robust across this range of thresholds. Both positive and negative crossings were considered as spikes within the MUA. To avoid double counting, a minimum separation of successive threshold crossings was enforced at 10 samples (roughly 0.4 ms). Artifacts were rejected if their absolute amplitude exceeded 1,5x the threshold. Crossing indices and timestamps were stored per channel and per epoch.
[0209] Firing Rate estimation: Threshold crossings were converted into firing rate estimates by binning the data into fixed-width 2-second bins (chosen to align with LFP analysis), which produced a MUA firing-rate per channel and per epoch. When analyses with BP were performed, BP was averaged within the same time-bins as the MUA firing rates.Local Field Potentials
[0210] Preprocessing: Neural signals recorded from the MEAs implanted in the RVLM were sampled at 24,414.0625 Hz. Each of the four channels was processed independently. Signals were filtered using a single cascade filter that combined second-order sections for a second-order Butterworth filter (0.01-200 Hz) and notches to suppress line noise (60 and 180 Hz, as only odd harmonics were present in the data)
[0119] , Forward (causal) filters were applied to mirror real-time constraints of filtering data for online applications such as the BSI.
[0211] Calculating band power: Power spectral density (PSD) was estimated per bin using Welch’s method [84, 119], with the window length equal to the bin duration. Bandlimited power was calculated by aggregating PSD within canonical LFP bands shown in Table 2.Table 2: Local field potential frequency bands used within all analysesBand Frequency (Hz)Very-low frequency (VLF) 0.01 - 1Delta 1 - 4Theta 4 - 8Alpha 8 - 12Beta 12 - 30Gamma 30 - 80High-gamma 80 - 200High-gamma1 80 - 125High-gamma2 125 - 200Spectrum 1 - 200
[0212] The Band power was computed as the mean PSD (in V2 / Hz) within each band, computed independently for each electrode and band for each time bin. LFP band powers were grouped into fixed-width bins of 2 seconds, matching the implementation for MUA. Frequencies between 0.01-1 Hz were emphasized for the herein LFP analysis, as they contained Mayer waves in rats (0.4-0.6 Hz) that are known to be related to sympathetic activation
[0126] , This bin duration was chosen as it satisfied the minimal Nyquist criterion to capture Mayer waves in rats. When multimodal analyses with BP were performed, BP was averaged within the same time-bins as the LFP band powers.Statistical Analysis
[0213] All statistical analyses were performed in Python using the SciPy library
[0119] , except for power calculations, which used the statsmodels library
[0127] , Results for individual animals computed a Spearman’s p between BP and LFP band power, while differences between baseline and experimental conditions were tested using the Wilcoxon signed-rank test, and relationships between MUA firing rates and LFP band powers were quantified using the coefficient of determination (R2) from linear regression.
[0214] False discovery rate
[0128] was used to correct for multiple comparisons. For chemoreflex analyses and other multi-condition comparisons, correction was applied across all comparisons within an experiment (i.e., across all frequency bands and conditions per animal). If a result was nonsignificant (p > 0.05), it was replaced with “ns” to simplify figures for single-animal results.
[0215] Z-score normalization was applied for grouped analysis of LFP band powers. The normalization was fit on baseline epochs and applied to subsequent experimental epochsto account for between- animal differences in signal magnitude due to electrode placement and electrode impedance. The mean of each animal’s z-score normalized values was used for statistical comparison, yielding a single estimate per animal per condition. As the baseline values were near-zero by definition following the z-score normalization, the resulting differences were not guaranteed to be normally distributed; statistical comparisons were assessed using the Wilcoxon signed-rank test. Effect sizes were calculated using Cohen’s dzfor paired data
[0129] , and statistical power was estimated post hoc using the observed effect size and the sample size (n=5, p < 0.05).Results
[0216] As a first step toward building a novel BSI to restore hemodynamic stability following SCI, the activity of RVLM neurons (including C1) was characterized to quantify how reliably their activity was modulated by changes in BP. These neurons have previously been associated with conveying baroreceptor efferent signals from the brainstem to the spinal cord [3, 16], Both LFP band power and MUA firing rates were analyzed from signals obtained by implanting BCIs in the RVLM. While LFPs provide less specificity than MUA due to integrating the activity from wider populations of neurons
[0059] , they are considered a more stable signal for clinical applications. Hence, in anticipation of better clinical translatability, a primary focus was to confirm LFP biomarkers and identify any new LFP biomarkers. The same data analysis pipeline was used for all experimental cohorts to ensure standardized characterization results and quantification of neural biomarkers.Baroreflex pertubations
[0217] Toquantify the activity of barosensitive neurons within the RVLM, hypotensive and hypertensive challenges were induced using SNP, PE, and LBNP. Changes to MAP and to neural signals were quantified using 2-second bins for both LFP band powers and MUA firing rates.
[0218] Historically, hypotension via SNP injections were used to identify C1 (or other barosensitive) medullary neurons [1, 30], This was repeated with both the MUA and LFP analyses (see FIGS. 17A-17F). A strong activation was observed in both high-gamma (80-200 Hz) band power and MUA firing rates following injections of SNP (see FIGS. 17A, 17C, 17C). A modest negative correlation was observed between MAP and high-gamma band power, which indicates increased activation from hypotension. A modest positive correlation between high-gamma band power and MUA firing rates was observed, which is supported by the known relationship between gamma LFP band powers and neural firing rates
[0077] ,
[0219] In further detail now, FIG. 17A shows the MAP time series for a single animal with multiple SNP injections. This time series is temporally aligned with the time series in the panels in FIGS. 17B and 17C. Experimental epochs for SNP are highlighted in blue. FIG.17B shows a time series of high-gamma band power binned into 2-second intervals. FIG.17C shows a time series of MUA firing rate sampled into 2-second bins. FIG. 17D shows a correlation between high-gamma band power (2-second bins) and MAP from the same animal’s time series data. A Spearman’s p is reported in the panel title. FIG. 17E shows a bar plot showing mean (± 95% confidence interval) of high-gamma band power (2-second) bins between baseline and experimental periods from the same animal’s time series data. Statistical comparison was performed using Mann-Whitney U test. FIG. 17F shows a regression model fit between high-gamma band power and MUA firing rate. The coefficient of determination and significance are reported in the panel title.
[0220] LBNP was also used for more precise and graded control over hypotensive challenges
[0017] , In FIGS. 18A-18F, 60-second epochs of LBNP were applied to a single rat, cycling randomly between 0, -7, and -14 mmHg. The response to hypotension was consistent with the SNP results; both high-gamma LFP band power and MUA firing rates increased during hypotension. Moreover, the magnitude of these increases was more pronounced at -14 mmHg compared to -7 mmHg (Fig 18E). ‘
[0221] FIG. 18A shows a MAP time series fora single animal with cyclical LBNP applied. This time series is temporally aligned with the time series in the panels in FIGS. 18B-18C. Experimental epochs for LBNP are highlighted in shades of blue corresponding to the magnitude of LBNP applied. FIG. 18B shows a time series of high-gamma band power binned into 2-second intervals. FIG. 18C shows a time series of MUA firing rate sampled into 2-second bins. FIG. 18D shows a correlation between high-gamma band power (2-second bins) and MAP from the same animal’s time series data. A Spearman’s p is reported in the panel title. FIG. 18E shows a bar plot showing mean (± 95% confidence interval) of high-gamma band power (2-second) bins between baseline and experimental periods from the same animal’s time series data. Statistical comparisons were performed using Mann-Whitney U test with false discovery rate correction
[0128] , FIG. 18F shows a regression model fit between high-gamma band power and MUA firing rate. The coefficient of determination and significance are reported in the panel title.
[0222] From the hypotensive results with both LBNP and SNP, a pronounced activation in MUA firing rates and high-gamma band powers was observed during hypotensive episodes, which suggested an inverse relationship with BP. To test whether this relationship further extended to hypertension, PE injections were performed to rapidly inducehypertension, consistent with RVLM and C1 characterization
[0035] , In FIGS. 19A-19F, a single extended PE epoch inhibited both signals. This effect was more pronounced (Spearman’s p =-0.737, FIG. 19Dd) and consistent with previous C1 neuron literature [30, 35],
[0223] FIG. 19A shows a MAP time series for a single animal with a single PE injection. This time series is temporally aligned with the time series in the panels in FIGS. 19B-19C. Experimental epochs for PE are highlighted in red. FIG. 19B shows a time series of high-gamma band power binned into 2-second intervals. FIG. 19C shows a time series of MUA firing rate sampled into 2-second bins. FIG. 19D shows a correlation between high-gamma band power (2-second bins) and MAP from the same animal’s time series data. A Spearman’s p is reported in the panel title. FIG. 19E shows a bar plot showing mean (± 95% confidence interval) of high-gamma band power (2-second) bins between baseline and experimental periods from the same animal’s time series data. Statistical comparison was performed using Mann-Whitney U test. FIG. 19F shows a regression model fit between high-gamma band power and MUA firing rate. The coefficient of determination and significance are reported in the panel title.
[0224] FIGS. 17A-19F demonstrate that there is an inverse relationship in both firing rates and high- gamma band powers between MAP of 40-175 mmHg. These experiments were repeated across 5 rats per experimental cohort to test whether this relationship generalized across animals.
[0225] Changes in high-gamma LFP band powers were significant following increases and decreases to BP across animals, as seen in FIGS. 20A-20C. This inverse relationship was consistent with known barosensitive activity of presympathetic neurons in the RVLM, including C1 neurons
[0035] , As a whole, the high-gamma band (80-200 Hz) had moderate to strong effect sizes: Cohen’s dzvalues were 1.23 for SNP, 2.40 for LBNP, and -1.12 for PE. These corresponded to achieved powers of 0.55, 0.97 and 0.48, respectively. The PID-controlled LBNP provided a sustained hypotensive challenge, rather than the transient effects of bolus injections, which may explain its larger effect size and power.
[0226] FIG. 20A shows a comparison of LFP band powers from specified bands following SNP injections. Each bar represents the group mean of a z-scored band power and error bars representing 95% confidence intervals. Z-score normalization was fit on baseline epochs per animal and applied to experimental epochs in orderto normalize across animals. Individual points correspond to mean values from each animal. Statistical comparisons were performed using Wilcoxon signed-rank test, with false discovery rate correction applied
[0128] ,
[0227] FIG. 20B shows a comparison of LFP band powers from specified bands during cyclical LBNP between 0 and -7 mmHg (1-minute on / off cycles). Each bar represents the group mean of a z-scored band power and error bars representing 95% confidence intervals. Z-score normalization was fit on baseline epochs per animal and applied to experimental epochs in order to normalize across animals. Individual points correspond to mean values from each animal. Statistical comparisons were performed using Wilcoxon signed-rank test, with false discovery rate correction applied
[0128] ,
[0228] FIG. 20C shows a comparison of LFP band powers from specified bands following PE injections. Each bar represents the group mean of a z-scored band power and error bars representing 95% confidence intervals. Z-score normalization was fit on baseline epochs per animal and applied to experimental epochs in order to normalize across animals. Individual points correspond to mean values from each animal. Statistical comparisons were performed using Wilcoxon signed-rank test, with false discovery rate correction applied
[0128] ,
[0229] Further subdividing the high-gamma band into high-gamma1 (80-125 Hz) and high-gamma2 (125-200 Hz) was considered for developing features for the BSI control algorithm. These were significantly modulated via SNP, LBNP, and PE. The achieved power and effect size are summarized in Table 3. In general, the effect size and statistical power were slightly lowerthan considering the entire high-gamma band, but remained sufficient for building a control algorithm. Following these findings, high-gamma band power was further evaluated across additional reflexes that implicate C1 neurons.Table 3: Effect size and achieved statistical power from baroreflex experiments (n=5)Chemoreflex pertubations
[0230] The Chemoreflex pathways were activated using controlled hypercapnic and hypoxic environments. These perturbations were used to test whether the modulation of high-gamma LFP band powers and MUA firing rates extended beyond the baroreflex.
[0231] In FIGS. 21A-21G, stronger hypoxic conditions increased both MUA firing rates and high-gamma band power. The compensatory responses to graded hypoxia are blunted in rats anesthetized with isoflurane
[0130] , which may explain the strong relationship between MAP and high-gamma band power (FIG. 21 E). Thus, it is possible that both the baroreflex and chemoreflex contributed to the RVLM signals acquired during this perturbation.
[0232] FIG. 21A shows a MAP time series for a single animal when hypoxic conditions were applied. This time series is temporally aligned with the time series in the panels in FIGS. 21B-21D. FIG. 21 B shows inhaled O2 concentrations during experimental epochs. Each gas concentration was maintained for approximately two minutes. The baseline epoch was taken at the end of the recording. FIG. 21 C shows a time series of high-gamma band power binned into 2-second intervals. FIG. 21 D shows a time series of MUA firing rate sampled into 2-second bins. FIG. 21 E shows a correlation between high-gamma band power (2-second bins) and MAP from the same animal’s time series data. A Spearman’s p is reported in the panel title. FIG. 21 F shows a bar plot showing mean (± 95% confidence interval) of high-gamma band power (2-second) bins between baseline and experimental periods from the same animal’s time series data. Statistical comparisons were performed using Mann-Whitney U test with false discovery rate correction
[0128] , FIG. 21G shows a regression model fit between high-gamma band power and MUA firing rate. The coefficient of determination and significance are reported in the panel title.
[0233] Across animals (FIGS. 22A-22G), there was a consistent trend of increased band powers in the high-gamma2 (125-200 Hz) band. While this increase was non-significant after correcting for multiple comparisons, it suggests some reactivity to hypoxia. In order to further characterize this reactivity, hypoxic jump protocols were applied by abruptly shifting inspired O2 from 21% to 11%.
[0234] FIG. 22 shows a comparison of LFP band powers from specified bands during hypoxia ramp experiments. Each bar represents the group mean of a z-scored band power and error bars representing 95% confidence intervals. Z-score normalization was fit on baseline epochs per animal and applied to experimental epochs in order to normalize across animals. Individual points correspond to mean values from each animal. Statistical comparisons were performed using Wilcoxon signed-rank test, with false discovery rate correction applied
[0128] ,
[0235] Similar to the graded experiments, sudden changes from 21% to 11% O2 caused increases in high-gamma band power, as shown in FIGS. 22C, 22F and 23. Similarly, MAP decreased with the sudden onset of hypoxia, although the relationshipbetween high-gamma band power and MAP was weaker than in the graded experiments (FIGS. 21 E, 23E), suggesting a peripheral chemoreflex activation of RVLM neurons. Grouped results demonstrate significant increases in high-gamma band powers, as well as a significant decrease in VLF and beta band powers (FIG. 24). The effect sizes and achieved power are summarized in Table 4, demonstrating that the strong modulation of high-gamma bands during hypoxia was sufficiently powered across animals.Table 4: Effect size and achieved statistical power from hypoxia jump experiments (n=5)
[0236] FIG. 23A shows a MAP time series for a single animal when hypoxic conditions were applied. This time series is temporally aligned with the time series in the panels in FIGS. 23B-D. FIG. 23B shows inhaled O2 concentrations during experimental epochs. Each gas concentration was maintained for approximately five minutes. FIG. 23A shows a time series of high-gamma band power binned into 2-second intervals. FIG. 23D shows a time series of MUA firing rate sampled into 2-second bins. FIG. 23E shows a correlation between high-gamma band power (2-second bins) and MAP from the same animal’s time series data. A Spearman’s p is reported in the panel title. FIG. 23F shows a bar plot showing mean (± 95% confidence interval) of high-gamma band power (2-second) bins between baseline and experimental periods from the same animal’s time series data. Statistical comparisons were performed using Mann-Whitney U test with false discovery rate correction
[0128] , FIG. 23G shows a regression model fit between high-gamma band power and MUA firing rate. The coefficient of determination and significance are reported in the panel title.
[0237] FIG. 24 shows a comparison of LFP band powers from specified bands during hypoxia jump experiments. Each bar represents the group mean of a z-scored band power and error bars representing 95% confidence intervals. Z-score normalization was fit on baseline epochs per animal and applied to experimental epochs in order to normalize across animals. Individual points correspond to mean values from each animal. Statistical comparisons were performed using Wilcoxon signed-rank test, with false discovery rate correction applied
[0128] ,
[0238] Responses to hypercapnic conditions within the RVLM mirrored the ramp and jump protocols from the hypoxia experiments. In FIGS. 25A-25G, as the inhaled CO2 increased, both high-gamma band power and MUA firing rates showed significant increases, while MAP remained largely unaffected by the hypercapnic perturbation. The correlation between high-gamma band power and MUA firing rates remained significant and consistent with prior baroreflex and hypoxia experimental results.
[0239] FIG. 25A shows a MAP time series for a single animal when hypercapnic conditions were applied. This time series is temporally aligned with the time series in the panels in FIGS. 25B-25D. FIG. 25B shows inhaled CO2 concentrations during experimental epochs. Each gas concentration was maintained for approximately two minutes. The baseline epoch was taken at the beginning of the recording. FIG. 25C shows a time series of high-gamma band power binned into 2-second intervals. FIG. 25D shows a time series of MUA firing rate sampled into 2-second bins. FIG. 25E shows a correlation between high-gamma band power (2-second bins) and MAP from the same animal’s time series data. A Spearman’s p is reported in the panel title. FIG. 25F shows a bar plot showing mean (± 95% confidence interval) of high-gamma band power (2-second) bins between baseline and experimental periods from the same animal’s time series data. Statistical comparisons were performed using Mann-Whitney U test with false discovery rate correction
[0128] , FIG. 25G shows a regression model fit between high-gamma band power and MUA firing rate. The coefficient of determination and significance are reported in the panel title.
[0240] The increased band power was dose-dependent and reproducible across animals, as shown in FIG. 26. Similar to the hypoxia perturbation experiments, the high-gamma2 band was most strongly modulated, with corresponding effect sizes and achieved statistical power summarized in Table 5. However, this increase in band power also extended to the full high-gamma band (80-200 Hz) and to the high-gamma1 sub-band (80-125 Hz) at higher levels of CO2. This subdivision of the high-gamma band was incorporated into the BSI control algorithm to serve as additional input features for model training and evaluation.Table 5: Effect size and achieved statistical power from hypercapnia ramp experiments (n=5)
[0241] FIG. 26 shows a comparison of LFP band powers from specified bands during hypercapnia ramp experiments. Each bar represents the group mean of a z-scored band power and error bars representing 95% confidence intervals. Z-score normalization was fit on baseline epochs per animal and applied to experimental epochs in order to normalize across animals. Individual points correspond to mean values from each animal. Statistical comparisons were performed using Wilcoxon signed-rank test, with false discovery rate correction applied
[0128] ,
[0242] The hypercapnia jump experiments revealed strong increases in both high-gamma band power and MUA firing rates following the sudden increase in inhaled CO2 (FIGS. 27C, 27F). While there was a small negative correlation between BP and high-gamma power (Spearman’s p = -0.180, FIG.27E), it was smaller than those observed during baroreflex experiments (p = -0.377 for SNP and p = -0.737 for PE, FIGS. 17D and 19D), and consistent with previously observed effects of hypercapnia on BP in nonanesthetized rats
[0131] ,
[0243] FIG. 27A shows a MAP time series for a single animal when hypercapnic conditions were applied. This time series is temporally aligned with the time series in the panels in FIGS. 27B-27D. FIG. 27B shows inhaled CO2 concentrations during experimental epochs. Each gas concentration was maintained for approximately five minutes. FIG. 27C shows a time series of high- gamma band power binned into 2-second intervals. FIG. 27D shows a time series of MUA firing rate sampled into 2-second bins. FIG.27E shows a correlation between high-gamma band power (2-second bins) and MAP from the same animal’s time series data. A Spearman’s p is reported in the panel title. FIG. 27F shows a bar plot showing mean (± 95% confidence interval) of high-gamma band power (2-second) bins between baseline and experimental periods from the same animal’s time series data. Statistical comparisons were performed using Mann-Whitney U test with falsediscovery rate correction
[0128] , FIG. 27G shows a regression model fit between high-gamma band power and MUA firing rate. The coefficient of determination and significance are reported in the panel title.
[0244] Across animals, hypercapnia caused inhibition within most bands, except for high-gamma bands (FIG. 28). FIG. 28 shows a comparison of LFP band powers from specified bands during hypercapnia jump experiments. Each bar represents the group mean of a z-scored band power and error bars representing 95% confidence intervals. Z-score normalization was fit on baseline epochs per animal and applied to experimental epochs in order to normalize across animals. Individual points correspond to mean values from each animal. Statistical comparisons were performed using Wilcoxon signed-rank test, with false discovery rate correction applied
[0128] ,
[0245] Similar to the hypoxia results, high-gamma2 (125-200 Hz) was more strongly modulated, and was consistent with the results from the hypercapnia ramp protocol. The effect sizes and achieved power are summarized in Table 6.Table 6: Effect size and achieved statistical power from hypoxia jump experiments (n=5)
[0246] Together, the baroreflex and chemoreflex perturbations demonstrate that high-gamma band powerswere reliably modulated. Notably, these band powers increased during hypotensive, hypoxic, and hypercapnic challenges, and decreased during hypertensive perturbations. Moreover, in all experiments, there was a consistent small-to-moderate positive coefficient of determination (R2from linear regression) between high-gamma band power and MUA firing rates. In addition to reflex-driven perturbations that implicate the RVLM, direct sympathetic activation was evaluated to determine whether it modulated these two signals.Other Pertubations
[0247] Intravenous GLP-1 injections were used to activate sympathetic pathways. A modest increase to MAP was observed following each injection. However, these did not appear to correspond to any changes in activity within the processed RVLM signals, as seen in FIGS. 29A-29F. Unlike previous experiments, the relationship between high-gamma band power and MUA firing rates was no longer discernable following GLP-1 injections (FIG. 29F). These injections also had inconsistent effects when repeated across multiple animals in FIGS. 30A-30F, leading to no consistent or significant modulation detectable in either LFP band powers or MUA firing rates.
[0248] The findings identified candidate neural biomarkers (notably high-gamma LFP band powers), which inform on the present hemodynamic state. These were then used to develop and test a BSI whose aim was to stabilize BP following SCI.
[0249] FIG. 29A shows a MAP time series for a single animal with multiple GLP-1 injections. This time series is temporally aligned with the time series in the panels in FIGS.29B-29C. Experimental epochs for GLP-1 are highlighted in gold. FIG. 29B shows a time series of high-gamma band power binned into 2-second intervals. FIG. 29C shows a time series of MUA firing rate sampled into 2-second bins. FIG. 29D shows a correlation between high-gamma band power (2-second bins) and MAP from the same animal’s time series data. A Spearman’s p is reported in the panel title. FIG. 29E shows a bar plot showing mean (± 95% confidence interval) of high-gamma band power (2-second) bins between baseline and experimental periods from the same animal’s time series data. Statistical comparison was performed using Mann-Whitney U test. The y-axis was linearly scaled for visual clarity. FIG. 29A shows a regression model fit between high-gamma band power and MUA firing rate. The coefficient of determination and significance are reported in the panel title.BSI Neural Decoder Development
[0250] Results from the above baroreflex experiments revealed a modest negative monotonic relationship via Spearman’s p between high-gamma LFP band power and BP (FIGS. 17D, 19D, and 18D). This relationship was leveraged for the development of a novel BSI that decoded LFP signals from the RVLM to control epidural stimulation commands.
[0251] Several criteria were established to evaluate the feasibility of a given decoder model. First, models needed to achieve high decoding accuracy of RVLM neural activity. Second, online latency is preferable to remain low to allow for rapid feed-forward decisionmaking. Third, calibration with BP is preferable to be kept to a minimum, or ideallyeliminated, to improve the translational potential of the BSI. Finally, interpretable or tuneable models were preferred over black-box models for clinical translatability.Data Standardization
[0252] Neural signals recorded within the RVLM from the previous aim were used to build a dataset to train machine learning models to control EES. The PE, SNP, and LBNP experimental data were used fortraining. These datasets contained normotensive periods paired with either hypotensive or hypertensive epochs with time-aligned RVLM recordings. In order to enrich the data, additional experiments were conducted with a brief baseline followed by SNP injections, and a PE injection using the same experimental methods described in the previous section on the Second Experimental results. As the electrode impedance values altered the magnitude of the recording amplitudes, LFP band powers often resulted in different magnitudes between animals (e.g., between 10-14and 10-12V2 / Hz). To address this variability, data was standardized across each experimental dataset in order to have a sufficiently large dataset for model development and evaluation.
[0253] Initially, the use of commonly used machine learning scalers, such as z-score (mean of 0, standard deviation of 1) or min-max (scaling data such that the minimum value in the dataset is 0 and maximum one is 1) were considered. However, given the nature of the experimental data, the mean or centre value was rarely representative of baseline conditions; experimental epochs represented hypotensive or hypertensive data, which skewed the distribution of LFP band powers above or below baseline, making z-scoring the entire dataset nonviable. Similarly, min-max scaling compressed features into inconsistent ranges depending on the severity of perturbations. For instance, hypertensive conditions (causing low high-gamma LFP band power) had baseline periods near 1 and experimental conditions near zero. Conversely, hypotensive datasets exhibited the opposite effect.
[0254] The goal was to consolidate all experimental data into a single dataset to train a machine learning model. Fitting z-scores on baseline periods and transforming experimental epochs did not solve large variance between animals caused by electrode placement and impedance. To address this, in at least one embodiment, a bespoke setpoint scaler was developed and implemented. This consisted of anchoring the mean of the first / / baseline data points (i.e., / / baseline) to a setpoint (e.g., 0.5) and then scaling the rest of the dataset analogous to min-max scalers relative to that set-point, as seen in Equations 5-8. This approach ensured consistent standardization across animals and experimental conditions that enabled building a dataset suitable for model development.
[0255] The set-point scaler was developed to adhere to a standard Python machine learning framework (Scikit-learn)
[0142] , This approach separated fitting the scaler and transforming the data in two distinct steps. Scaling parameters were computed from baseline data during the fitting step and then applied to the remaining experimental data during the transformation step.
[0256] Given a dataset X e RT*d, with T time points and d features, the baseline mean, baseline minimum, and baseline maximum are defined as such:
[0257] Here, r baseline denotes the number of initial time points used to estimate the baseline mean. A default value of 20 time points was used to provide a stable estimate of baseline neural activity while remaining short enough for calibration use in online, realtime experiments.
[0258] A scale factor and offset were fit on the input baseline:
[0259] Data was transformed using the fitted scale and offset, according to the following vectorized equation:A — X O S + <i,o.(o) where:X e RT*dis the dataset with T time points and d features;s, o e Rdare the scale and offset vectors;O denotes element-wise multiplication with broadcasting across rows; and X' e RT*dis the transformed datasetUnless specified otherwise, a set-point value of 0.5 and r baseline of 20 were used for all analyses.Feature Generation
[0260] The dataset used for model development and evaluation included LFP band powers (2-second windows) from each band outlined in Table 2, as well as additional features derived from these bands. Feature generation was confined to neural data captured from a single electrode because the characterization experimental results revealed that strong barosensitivity was often confined to a single electrode. This may reflect the MEA design, which used four electrodes of different lengths (FIG. 14B), as well as variability in the RVLM size and location between Sprague-Dawley rats
[0019] , The candidate electrode from each dataset was selected based on the strongest Spearman’s p between BP and high-gamma LFP band power.
[0261] Once the best electrode was isolated, ratios between VLF and high-gamma bands and other LFP bands (e.g., VLF: Delta, Theta: High-gamma1) were computed and concatenated with the original band powers to expand the feature set. The high-gamma bands were selected based on the results obtained in the characterization experiments, while VLF was included due to its incorporation of Mayer waves
[0126] ,
[0262] In at least one embodiment, the derivative of this expanded feature set was then numerically estimated for each feature by taking the difference between successive time points, providing an estimate for the rate of change for each feature. Time-lagged features were concatenated into the dataset as additional regressors. The previous five bins were generated by shifting the dataset temporally and concatenating after each shift. This removed five data points from analysis, corresponding to a 10-second reduction of the normotensive baseline period. Finally, second-degree interaction terms from a polynomial expansion of all non-ratio features were included as additional regressors. These feature expansion steps resulted in a dataset with 7,320 features suitable for training both simple and complex machine learning models.
[0263] In at least one embodiment, the ratios between bands, derivatives, time-lagged features, and second-degree interaction features may be computed as a way to expand the feature space for machine learning models. The rations were generated to create a dataset containing both nonlinear and linear features, which granted exploratory models greater degrees of freedom during training.Exploratory models
[0264] Following the creation of the dataset, several exploratory models were considered to provide initial insights into how well LFP-derived features capture BP variation. First, linear models, such as generalized least squares (GLS) and weighted least squares (WLS) were evaluated; these approaches provided interpretable baselines, but showed highautocorrelation between features, which limited the effectiveness of these models. WLS yielded the most promising results when weights were assigned to BP values that were further from the dataset mean (i.e., hypotensive or hypertensive data). This weighting improved prediction accuracy at low and high BP regimes, but did not generalize across all cross-validation results.
[0265] Given these results, nonlinear approaches were evaluated next. A ^-variational autoencoder (?-VAE) was used to reduce autocorrelated features into a more interpretable latent variable space. Integrated gradients were applied to the latent dimensions to identify which input features were most strongly associated with changes in BP. These identified variables were reintroduced into the linear models, which performed well in k-fold cross-validation, but poorly in leave-one-dataset-out (LODO) cross-validation. Although the regression accuracy was insufficient for translational use, these models appeared to reliably determine whether estimated BP was above or below baseline levels.
[0266] Based on these insights, the problem was then reframed as a classification task. Since EES can increase BP [17, 56], the goal of stimulating during hypotension may be solved using logistic regression. This approach aligned with the criteria for model selection, as logistic regression is lightweight and interpretable and so in at least one embodiment, the logistic regression may be implemented as the BSI neural decoding algorithm.Least Squares Model
[0267] The Linear least squares models were first tested to serve as a baseline for subsequent models. Regression models from the statsmodels Python library
[0127] were used to predict BP from LFP-derived features. The assumptions required for ordinary least squares were not met, as many of the derived features were multicollinear (e.g., time-lagged band powers). Thus, GLS was used as a baseline model for comparison with subsequent neural decoding models. Given that a goal of the BSI was to isolate hypotension to turn on EES, and secondarily to avoid stimulation during hypertension, data further from the mean was considered of greater importance. To address this, a custom weight function was developed and applied to the training dataset to emphasize hypotensive and hypertensive data fora WLS model, shown in Equation 9:where:W is the weight vector for the WLS implementation;yt is the BP metric at an individual time point (e.g., MAP, ASBP);y~ is the mean BP metric across the training dataset;oyis the standard deviation of the BP metric across the training dataset; and a is a scaling constant set to 0.85, following a grid search.
[0268] In at least one embodiment predictions from the GLS and WLS models were optionally post-processed with a causal exponentially weighted moving average (EWMA, a = 0.1) to stabilize high-frequency fluctuations for operational use, effectively serving as a temporal prior on the model output. The value for the scaling constant, a, was obtained following a coarse grid search between 0.5 and 2.0 on a test dataset with a 0.05 step size. The post-processing EWMA constant, a, was obtained following a grid search between 0.01 and 0.2 with a step size of 0.01, and served as post-hoc smoothing of rapid changes in BP as a result of the cross-validation paradigms.j3-variational autoencoder
[0269] The least squares models still demonstrated a high degree of multicollinearity. As LFP signals are derived from nonlinear processes
[0082] , it was determined to apply nonlinear feature selection. In at least one embodiment nonlinear feature selection may be applied by using an autoencoder because its architecture nonlinearly reduces highdimensional input data into a low-dimensional latent feature space
[0143] , Specifically, a ft-VAE was used, which refines autoencoder architectures by modelling each latent variable as a distribution and encouraging disentanglement to generate separate and interpretable features
[0144] ,
[0270] The data collected in the baroreflex exploratory experiments was a wide dataset (nfeatures > nobservations). In at least one embodiment, for better performance of the / ?-VAE, mutual information regression may be applied separately on SBP, DBP, and MAP to reduce the number of features, thereby lowering the risk of overfitting. In testing, the mutual information regression retained the top 1,500 features (down from 7,320) for each BP metric, and the features were merged, removing any duplicate features. This resulted in a reduced feature set of 1,848 features, which was used as input to the?-VAE. This model was built in PyTorch
[0145] , with its architecture shown in FIG. 31.
[0271] FIG. 31 shows the encoder mapped features through three fully connected layers (512, 256, 64 units), each with Leaky ReLU activation functions. These layers parameterized the latent space with and o2across 15 latent dimensions. A supervised prediction of BP metrics (MAP, SBP, DBP) was output from the model, reparameterized from the latent space. The canonical decoder and reconstruction output are shown in grey(last 4 columns to the right) for completeness, but were not used in model training or evaluation.
[0272] The output (y) of the latent Gaussian parametrization was reparameterized to enable this regression with BP signals:V - n + t r' (10) where:p is a vector containing the means of the latent variables;is a vector containing the standard deviations of the latent variables; and 8 is a vector containing random noise sampled from a standard normal distribution to add stochasticity during training.
[0273] The loss function used to optimize the supervised?-VAE used both the prediction loss and the Kullback-Leibler Divergence (DKL) loss
[0144] , Mean squared error (MSE) was used as the prediction loss between the model outputs and the BP signals. The DKL loss estimated the distance between the posterior distributions of the latent variables and the prior distribution (Gaussian, p (z) = N (0, 1)). The ft term is a hyperparameter constant that scales the contribution of the DKLloss to encourage disentanglement of the latent space by encouraging less overlap between latent variable distributions:where:n is the batch size;i is the index of a sample within a given batch;d is the dimension of the latent space;I is the index of a latent variable within the latent space;yt is the true BP values for sample i;y~ t is the predicted BP values for sample i;pa. is the mean of the latent variable I for sample i; andfr~< is the variance of the latent variable I for sample i.
[0274] In at least one embodiment, the?-VAE may be optimized using the Adam optimizer with a learning rate of 0.001 and trained for up to 150 epochs with a batch size (n) of 256 and 15 latent dimensions (d). To prevent overfitting, early stopping was implemented(patience of 10 epochs), applied based on the validation loss. The dataset was split 80 / 20 into training and validation sets, ft was set to 2 for the model optimization.Interpretation of / 7-VAE findings
[0275] Following training, the latent space was explored using integrated gradients
[0146] implemented with the Captum Python library
[0147] , Integrated gradients were used to identify how each of the latent variables contributed to the / ?-VAE’s predictions of SBP, DBP, and MAP. This was done by comparing predictions derived from each latent variable (£) against a baseline zero-vector ((?). The gradients along the path defined by the difference between I and C? were used to generate an attribution score that quantified the extent to which each latent variable influenced the predictions from the 7-VAE. The five latent variables that were most strongly associated with SBP, DBP, and MAP were isolated independently.
[0276] Integrated gradients were then rerun to identify which input features most strongly influenced each latent variable. The 25 input features most associated with each of the latent variables from the previous step were then concatenated to create a salient feature set, with any duplicate features removed. WLS was subsequently rerun with this salient feature set. Inputs from the high-gamma1, high-gamma2, and VLF bands were most prominent in the salient feature set. Specifically, time-lagged band powers were more numerous than the current band powers. While polynomial and derivative features were included in the salient feature set, they were excluded from online model development to better align with the low-latency goal of the decoding algorithm. Thus, in at least one embodiment, a feature set that may be used for online, real-time deployment of the BSI decoder may be implemented to only the current and time-lagged band powers from the high-gamma1, high- gamma2, and VLF bands. This feature set contained 18 features: the current band power, and the t-i, t-2,..., t-s time-lagged band powers for the high-gamma1, high-gamma2, and VLF bands.
[0277] Logistic Regression: The final model used for the BSI decoder was logistic regression with L1 regularization, implemented using the statsmodels Python library
[0127] , This model classified whether BP was below baseline according to the criteria for OH [4] (SBP reduction > 20 mmHg), thereby triggering EES to restore BP to baseline levels. A grid search was conducted over different decision thresholds (T e {0.45, 0.5, 0.55, 0.6}) and different L1 regularization strengths (cr e {0, 0.1, 0.2,..., 1}), with the strongest model performance obtained using t = 0.55 and a = 0.4.
[0278] Model performance was evaluated using both k-fold cross-validation and leave-one-dataset-out (LODO) cross-validation. Root mean squared error (RMSE) was used for regression models. Receiver operating characteristic (ROC) curves, precision-recall curves, and confusion matrices were used for classification models. Training datasets were class-balanced during LODO evaluation by ensuring that the non-hypotensive (i.e., hypertensive and normotensive) class contained the same number of samples as the hypotensive class.
[0279] For testing, the final model used as the BSI decoder was a logistic regression model trained on 18 features: the current and time-lagged band powers from the high-gammal, high-gamma2, and VLF bands. The final logistic regression model was then adapted for real-time use.Online BSI Implementation
[0280] In order to support the BSI experiments, a dashboard was built for real-time analytics, data visualization, and modification of experimental parameters. The dashboard was built using the HoloViews family of Python libraries (Panel, Param, HoloViews)
[0148] , It allowed parameter adjustments (e.g., electrode selection, updating the set-point scaler baseline) and provided real- time visualization of LFP band power and stimulation probability from the online decoder. The dashboard and further explanation of the functionality can be seen in the example embodiment shown in FIG. 32.
[0281] In FIG. 32, experiment configuration controls were used for electrode selection and visualization of 2-second binned LFP band power. Baseline controls were used to manage the set-point scaler, either by initializing or updating the baseline, or by resetting it based on the previous 35 bins (70 seconds). TDT controls were connected to the Synapse API for real-time neural data acquisition and allowed for in-dashboard adjustment of stimulation parameters (frequency and amplitude). Test data controls supported offline decoder development using existing datasets. The top right plot showed predicted stimulation probability (blue) and delivered stimulation amplitude (black; stimulation was only delivered if the stimulation toggle was enabled). The bottom right plot showed LFP band power of the user-selected frequency band.Adjustments to decoding algorithms
[0282] The analysis pipeline for LFPs was streamlined from the pipeline developed for RVLM characterization into a lightweight implementation for real-time use. This approach removed batch analysis pipelines, multiprocessing, and epoch-based analysis, reducing support to a single frequency band and single electrode. In at least one embodiment, theonline LFP pipeline may incorporate a 2nd-order Butterworth band-pass cascade filter
[0119] for a given band, including notch filters to remove line noise. The second-order section coefficients were saved for the high-gamma1, high-gamma2, and VLF bands when the dashboard was first initiated, reducing the computational overhead needed for filtering incoming neural data. Whenever possible, functions were vectorized for efficient computations, ensuring LFP features could be rapidly extracted.
[0283] The set-point scaler implementation was modified for online application. As shown in FIG. 33, it was possible to manually trigger the initial set-point calibration and subsequent recalibrations. Pilot experiments revealed that the number of bins used in offline model building produced inconsistent results. This was adjusted to 35 two-second bins, corresponding to a baseline of 70 seconds.Online Performance
[0284] In at least one embodiment, the BSI may have an online constraint to decide whether to stimulate within two seconds, which corresponded to the 2-second bins used in computing LFP band powers. Thus, the decoder, stimulation probability model, and delivery of the stimulation (see EES section below) were operated to comply with this constraint. The online pipeline completed LFP feature extraction and decision-making in about 0.2-0.25 seconds. Streaming neural data from the Synapse API (Tucker-Davis Technologies, United States) added a fixed latency of 0.5-0.7 seconds. The total latency was acceptable for the feed-forward BSI application, remaining within the 2-second constraint. Logging of stimulation metrics (amplitude, probability) and real-time data visualization were handled asynchronously using lightweight buffers and callbacks, ensuring that there was no impact on decoding performance or speed in this example embodiment.
[0285] Stimulation commands were only relayed to Synapse via the API if the stimulation toggle was enabled in the dashboard, shown in FIG. 33. In at least one embodiment, the average of three bins (six seconds) was used to trigger stimulation, which avoided stimulation in response to noise spikes. The decision threshold from the logistic regression was lowered to 0.475, improving performance in pilot experiments. When stimulation was delivered, the amplitude of stimulation was 500 / / A; in at least one embodiment, acquisition of neural data was paused for two seconds to avoid a positive feedback loop caused by stimulation artifacts in the raw neural traces.
[0286] The same animal model was used as in the previous Experimental section (adult female Sprague-Dawley rats: 200-300 g body weight, 10-16 weeks of age; Charles RiverLaboratories, Canada). Housing, anesthesia, hemodynamic monitoring, and RVLM microelectrode array implantation procedures were identical to those described in the previous Experimental section. Similarly, all experiments were conducted as acute and terminal preparations. All experimental procedures were approved by the University of Calgary’s Health Sciences Animal Care Committee under ethics protocol number AC24-0147.
[0287] Spinal cord Injury model: Rats were given a complete T2 spinal transection, as high-thoracic SCI leads to unstable BP [4, 50], This procedure was performed as the final surgical step, following BP catheterization and RVLM MEA implantation. Animals were surgically prepared by shaving the region of interest (T2 / T3 vertebrae), administering saline and buprenorphine for pain management (0.05 mg / kg), and cleaning the area with alcohol and chlorhexidine. A skin incision was made at the injury site using a scalpel blade to expose the muscles. The muscles were separated along the midline and removed from the vertebrae. A small part of the spinal cord was uncovered with a laminectomy of the T2-T3 vertebrae above the injury level. After opening the dura mater, a complete transection was performed with micro-scissors, and the ventral dura was left intact. Any bleeding was controlled with gel foam. The injury site was flushed with saline and the muscle, fascia, and skin were sutured with Vicryl to ensure that there was minimal bleeding for the remainder of the experimental protocol.
[0288] Epidural electrical stimulation (EES) was applied to the hemodynamic hotspot, which is a set of segments of the caudal thoracic spinal cord in rats (T11-T13) that cause a robust pressor effect when stimulated
[0017] , Surgical access to this portion of the spinal cord also utilized laminectomies, similar to the SCI protocol. The region between T9 and T13 vertebrae was shaved and disinfected. Buprenorphine (0.05 mg / kg) was injected for pain management. An incision was made to expose the T11-T13 segments of the spinal cord. Teflon-coated stainless-steel wires, connected to a head stage connector (A19923-001 Omnetics Connector Corporation, United States) were inserted rostrally and passed between the spinal cord and vertebrae to a caudal opening. A small portion of the wire insulation was removed to expose stimulation sites, which were positioned over the T12-T13 spinal cord segments. The head stage was connected to a custom-built adaptor that enabled a ground and reference signal to be passed to the stimulation wires. EES signals were controlled from Synapse, with amplitudes varying based on the BSI-control algorithm (described above). The stimulation was biphasic and set to a frequency of 50 Hz, similar to previous work from this laboratory [17, 58], Stimulation commands were passed through a neurostimulator (SI-4 Subject Interface, Tucker-Davis Technologies, United States),connected through a passive signal splitter (SBox-16, Tucker-Davis Technologies, United States), and finally to the adaptor.
[0289] The BSI experiments followed methods similar to those described in the baroreflex experiments discussed in the previous section. All experiments were performed following the SCI and were repeated once with the BSI active and once with the BSI inactive, which served as a baseline for comparison. BP, neural data, and stimulation outcomes were recorded in LabChart and TDT Synapse throughout each perturbation, as described in the in previous experimental section. Perturbations were not applied in a fixed order and were constrained by physiological stability of the animals. When it was possible, experiment order was randomized across animals.
[0290] A schematic overview of the experimental setup used in the BSI-experiments is shown in FIG. 33. Neural activity was recorded in the RVLM using the same BCIs described in the previous experimental section and streamed to a computer running TDT Synapse, with a custom online decoderand stimulation controller implemented in Python. A complete spinal cord transection was performed at the T2 spinal level, and an EES array was implanted at the T11 spinal level (the hemodynamic hotspot in rodents). Arterial BP was continuously measured using a catheter placed in the descending aorta. RVLM neural activity was used to modulate EES in response to hypotensive and hypertensive perturbations. This schematic is not to scale.
[0291] During these experiments the animals were given 3 doses of Sodium nitroprusside (SNP) (10-20 jtzg / kg), with at least 8 minutes between injections. The animals were also given 3 doses of Phenylephrine (PE) (10 / zg / kg), with at least 20 minutes between injections.
[0292] To better model an episode of OH
[0050] , the Lower body negative pressure (LBNP) chamber was turned on and maintained at a constant pressure (-6 mmHg) for 30 minutes, producing a sustained hypotensive challenge. A group size of 10 animals was used for this experiment, as each animal underwent a single episode with the BSI active and a single episode with the BSI inactive. The order of experiments was randomized between animals.BSI analysis metrics
[0293] Analysis of the BSI experiments used the BP metrics developed in the previous experimental section (AMAP, ASBP, ADBP, APP, AHR), with additional metrics calculated to evaluate BP stability following perturbations. To account for differing baselines between animals, the percentage change was computed between the baseline BP metrics (60seconds prior to each experimental epoch) and those recorded during experimental epochs:Time was normalized for each experimental epoch to ensure that all experimental protocols were temporally aligned. For SNP trials, epochs were aligned such that the nadir occurred within the same 2-second bin (about 14-16 seconds), with a total duration of 80 seconds. For PE trials, alignment was set to the peak within the same 2-second bin (about 14-16 seconds), with a total duration of 200 seconds. For the 30-minute LBNP experiments, only the LBNPON period was considered as the experimental epoch (1800 seconds).
[0294] The area under the curve (AUC) was computed for each of the BP metrics using the trapezoid method, as implemented by the SciPy library
[0119] ,
[0295] The target error was computed for all BP metrics, which was defined as the distance between a given BP metric and a target threshold that was set to the baseline mean:Tar^L-l error, t - X,where:target error, is the target error for a given BP metric for animal i;t is the threshold for the target error, that was set to the mean of the baseline for that given BP metric (Xbaseiine); andXi is the value of the BP metric for animal i.
[0296] All statistical analyses were performed using the same methods described in the previous experimental section. However, paired f-tests were applied for all statistical comparisons, except for target error, which was evaluated using the Wilcoxon signed-rank test. When appropriate, false discovery rate
[0128] was used to account for multiple comparisons.Results
[0297] In at least one embodiment, machine learning models may be used to leverage the high-gamma band power neural biomarker as a decoding target. The success criteria for the decoder were high-accuracy, low latency (aligned with 2-second LFP bins), and interpretability, which would better support clinical translation.
[0298] Once all data was standardized, least-squares models were used to serve as baselines for comparison with future models. Ten-fold cross validation was first performedby concatenating all the datasets after expanding LFP features. This yielded promising results, seen in FIGS. 34A-34C, particularly when using a WLS model that emphasized extreme BP values during hypo- and hypertension. When paired with EWMA, the decoder achieved an R2above 0.5 between neural features and MAP.
[0299] FIG. 34A shows predictions from the GLS and WLS models and their EWMA variants for a sample fold. FIG. 34B shows the RMSE performance of each model, where bars represent the mean RMSE of all folds, the error bars are 95% confidence intervals, and points indicate individual fold results. FIG. 34C shows the Coefficient of determination (R2) for each model, where bars represent the mean R2of all folds, the error bars are 95% confidence intervals, and points indicate individual fold results.
[0300] As shown in FIG. 34A, these models tracked overall BP changes, but EWMA reduced short-term fluctuations in exchange for improved trend tracking. However, this performance did not generalize to LODO cross validation (FIGS. 35A-35C), which better reflected the challenge of applying the decoder to new animals with a BSI. The mean RMSE of the WLS EWMA model was 16.3 mmHg, which is comparable to the reported error of non-invasive BP measurements (-12.2 mmHg, relative to sphygmomanometer readings)
[0149] ,
[0301] The least squares models did not generalize to LODO cross validation, as seen in FIGS. 35A-35C. FIG. 35A shows predictions from the GLS and WLS models and their EWMA variants on a dataset containing an SNP injection immediately followed by a PE injection. FIG. 35B shows RMSE performance of each model, where bars represent the mean RMSE of all folds, the error bars are 95% confidence intervals, and points indicate individual fold results. FIG. 35C shows Coefficient of determination (R2) for each model, where bars represent the mean R2of all folds, the error bars are 95% confidence intervals, and points indicate individual fold results. Compared to the k-fold performance, RMSE increased substantially, and the coefficient of determination approached zero across most models and datasets. Thus, the expanded feature set developed for these models appeared to overfit to individual datasets, and did not generalize well across different experimental conditions.
[0302] To prune the feature set for better performance, in at least one embodiment a nonlinear approach may be used, as neural signals are derived from nonlinear processes
[0082] , A?-VAE neural network was trained to learn a latent representation of the data while simultaneously predicting SBP, DBP, and MAP. Integrated gradient analysis
[0146] was then used to identify which latent variables were most associated with each BP metric. Across the three BP metrics, several latent variables (e.g., to, f2, and fe) had strong contributions,suggesting a shared encoding of physiologically relevant information in these dimensions. The 25 input features from the RVLM dataset with the highest attribution to each of the six unique latent dimensions were isolated. After removing duplicates, 27 features remained, forming a salient feature set. Many of these features were composite features, derived from polynomial expansion (e.g., (d / dt)*HighGamma1t-i * (d / dt)*HighGamma1to).
[0303] WLS / GLS models trained on the salient feature set also did not generalize in LODO cross-validation, predicting MAP close to the mean of the training and test data (not shown). Instead, this analysis was used to refine feature selection rather than serve as a candidate decoding model. The frequency with which each lagged time point and band appeared in this salient feature set is shown in Tables 7 and 8.Table 7: Relative occurrence of time-lagged features in salient datasetTime lag Time lag (s) Percentage of salient features (%)to ["2, 0) 1.6t-1 ["4, -2) 7.2t-2 [-6, -4) 0t-3 [-8, -6) 24.1t-4 [-10, -8) 47.8t-5 [-12, -10) 19.3Table 8: Relative occurrence of LFP bands in salient datasetBand Percentage of salient features (%)VLF 30.8Delta 6.0Theta 12.6HighGamma 26.8HighGammal 11.4HighGamma2 4.0Spectrum 8.4
[0304] In Table 7, the vast majority of the salient features demonstrated a time-lagged dependence with BR Specifically, a 6-12 second delay was observed between changes in BP and the LFP band powers in this salient dataset. While previous literature demonstrates much more rapid barosensitive reactivity of C1 neurons [30, 35], it is possible that using lower-frequency LFP features may account for some of this increased delay in reactivity. Although the use of forward (causal) filtering on the neural data may also have contributed to a small apparent lag. In Table 8, nearly half of the salient features were from the high-gamma band or its subdivisions. Nearly a third of the features contained the VLF band,suggesting that sympathetic information encoded in Mayer waves
[0126] is important for estimating SBP, DBP, and MAP. In at least one embodiment, this information may be used to create a logistic regression model to predict when to stimulate based on time-lagged neural activity from the VLF, high-gamma1, and high-gamma2 bands.
[0305] A logistic regression model was developed to serve as the BSI decoder. Its goal was to determine when to stimulate (i.e., BSION) using neural signals from the RVLM, while minimizing the number of false positives. In FIGS. 36A-36E, the LODO cross validation results and model parameters are detailed. FIG.36A shows true and predicted categories for the logistic regression on a test dataset. Samples were considered as BSION (i.e., the decoder should predict stimulation) if their SBP was at least 20 mmHg below the baseline SBP from the first minute of data. FIG. 36B a forest plot of the odds ratios for the BSI decoder model parameters, sorted in descending order with error bars representing 95% confidence intervals. FIG.36C shows a confusion matrix for the test dataset from FIG. 36A. ROC curves from all the LODO cross validation datasets. A baseline dotted grey line is plotted, denoting chance prediction. FIG. 36E shows accuracy across experiments for the LODO cross validation, separated by the experimental condition.
[0306] Overall, the model performed adequately across all datasets (mean AUG = 0.71). However, these ROC curves did not include any PE experiments, as they only contained a single class, BSIOFF. In those experiments, the model performs well at not stimulating at high blood pressures, consistent with the results observed in the previous experimental section, and in previous literature where barosensitive RVLM neurons are silenced at elevated BP [23, 30, 35], The decoder model accuracy was generally higher for the PE experiments, as seen in FIGS. 36A-36E. The model parameters revealed that the high-gamma2 band power at previous time points was a strong determinant for turning on stimulation, whereas elevated VLF band power more strongly contributed to the decision not to stimulate. This pattern is consistent with increased VLF activity reflecting enhanced sympathetic activation via Mayer waves
[0126] , suggesting that stimulation was unnecessary when sympathetic output was already elevated. This high-gamma band power finding is consistent with the results from the baroreflex perturbation experiments in the previous experimental section. The full decoder model parameters are summarized in Table 9.Table 9: Logistic regression coefficients and odd ratios for the final BSI decoder Standard Odds Parameter Coefficient 95% Cl error ratio
[0307] The high-gamma 1 band power parameters tended toward the model deciding that stimulation was not needed, as shown in Table 9. In the group analyses from the baroreflex perturbation experiments (FIGS. 20A-20C), the effects of the high-gamma1 followed the same trend as the high-gamma2 band, but were slightly muted in comparison. After optimization, the logistic regression yielded coefficients closer to zero, with two dropped coefficients in the high-gamma1 range (lagged times t-i and t-4) and one whose confidence interval that extended above and below an odds ratio of 1 (to). This suggests that in at least one embodiment the LFP features associated with the high-gamma2 range (125-200 Hz) may be used as they are more strongly associated with BP metrics and baroreflex activation.BSI Experiments
[0308] In at least one embodiment, the BSI decoder may be used for real-time feedforward control of EES, e.g., in anesthetized rats, with complete SCI. This was evaluated to validate the efficacy of this system during baroreflex perturbations.
[0309] The BSI attenuated the severity of acute hypotension following injections of SNP, as shown in FIGS. 37A-37F and Table 10. Notably, the BSI activated stimulation during the first ten seconds of most trials, which contributed to a blunted MAP nadir (at roughly 15 seconds in FIGS. 37A-37B). While the BSI was unable to completely counteract the hypotension caused by SNP injections, it sustained elevated BP throughout the experimental epoch. Translationally, EES has been able to maintain stable BP duringorthostatic challenges
[0056] , which suggests that this effect may result from competing mechanisms between the vasoconstriction induced by EES [17, 56] and the vasodilation triggered by SNP boluses. While the changes in HR and pulse pressure were underpowered, they trended toward non-significance, suggesting that the BP increases observed from EES likely involved vascular mechanisms to a greater extent than cardiac mechanisms.
[0310] FIG. 37A shows a time-aligned AMAP traces for BSION and BSIOFF trials normalized to the onset of hypotension caused by SNP injections. The bolded traces are the mean across all animals, and the thinner lines represent the mean of three SNP injection trials per animal. Below a time-aligned raster plot shows the stimulation events (2-second bins) for all 15 trials. FIG. 27B shows the percentage change in MAP between BSION and BSIOFF trials at the nadir (about 15 seconds in FIG. 37A) following SNP injections. FIG. 37C shows the mean change in heart rate between BSION and BSIOFF trials. FIG. 37D shows the percentage change in MAP, SBP, and DBP between BSION and BSIOFF trials. Statistical comparisons were performed with false discovery rate correction applied
[0128] , FIG. 37E shows the AUC of the AMAP trace in panel a between the BSION and BSIOFF trials. FIG. 37F shows the pulse pressure percentage change between the BSION and BSIOFF trials.Table 10: Effect size and achieved statistical power for BSI-mediated responses following SNP injections (n=5)Comparison Effect size (Cohen’s dz) PowerChange at nadir (%) 2.39 0.97Change in HR (%)-0.04 0.05AUC of SNP traces (mmHg) 1.18 0.52Change in pulse pressure -0.38 0.10(%)Change in AMAP (%) 2.03 0.92Change in ADBP (%) 2.03 0.92Change in ASBP (%) 1.83 0.82
[0311] The acute hypertension BSI experiments served as a positive control for the BSI decoder to minimize stimulation during elevated BP. After correcting for multiple comparisons, no significant differences were observed between BSION and BSIOFF trials following injections of PE, shown in FIGS. 38A-38F. While some trials exhibited periodic stimulation events, the BSI decoder rarely activated stimulation during hypertension. In FIG. 38D, there was a 10-15% increase in BP metrics, trending toward significance, which may have resulted from variability in baseline BP across experiments. The AUG of the PEtraces (FIG. 38E) was non-significant but slightly lower during PE trials, suggesting a greater influence of baseline BP values on these results.
[0312] FIG. 38A shows time-aligned AMAP traces for BSION and BSIOFF trials normalized to the onset of hypertension caused by PE injections. The bolded traces are the mean across all animals, and the thinner lines represent the mean of three PE injection trials per animal. Below a time-aligned raster plot shows the stimulation events (2-second bins) for all BSION trials. FIG. 38B shows the percentage change in MAP between BSION and BSIOFF trials at the peak (about 15 seconds in FIG. 38A. following PE injections. FIG. 38C shows the mean change in heart rate between BSION and BSIOFF trials. FIG. 38D shows the percentage change in MAP, SBP, and DBP between BSION and BSIOFF trials. Statistical comparisons were performed with false discovery rate correction applied
[0128] , FIG. 38E shows the AUC of the AMAP trace in panel a between the BSION and BSIOFF trials. FIG. 38F shows the pulse pressure percentage change between the BSION and BSIOFF trials.
[0313] The prolonged 30-minute LBNP sessions were designed to simulate a typical OH episode faced by individuals with SCI. The results shown in FIGS. 39A-39F, demonstrate that the BSI successfully stabilized BP for extended periods of time. FIG. 39A shows time-aligned AMAP traces for BSION and BSIOFF trials normalized to the onset of hypotension caused by LBNP. The bolded traces are the mean across all animals, and the thinner lines represent a single 30-minute LBNP trial for each experiment. Below a time-aligned raster plot shows the stimulation events (2-second bins) for each BSION trial. FIG. 39B shows the target error in mmHg between BSION and BSIOFF trials. A Wilcoxon signed-rank test was used for statistical comparison. FIG. 39C shows the mean change in heart rate between BSION and BSIOFF trials. FIG. 39D shows the percentage change in MAP, SBP, and DBP between BSION and BSIOFF trials. Statistical comparisons were performed with false discovery rate correction applied
[0128] , FIG. 39E shows the AUC of the AMAP trace in FIG. 39A between the BSION and BSIOFF trials. FIG. 39F shows the pulse pressure percentage change between the BSION and BSIOFF trials.
[0314] The BSI increased MAP within approximately 5% of baseline, with several trials stabilizing at or above baseline. Across all BSION trials, there were rapid increases in MAP immediately following the onset of LBNP, indicating quick decoder reactivity. This was followed by a slight dip and then a gradual rise in MAP, which remained higher than the BSIOFF trials. These fluctuations generally coincided with temporary deactivation of stimulation after several pulses before resuming it, suggesting that it attempted to maintain BP near a set point, which is consistent with previous hypotheses of baroreflex regulation [9], In FIG. 39B, the target error indicated that, on average, the BSI restored MAP to within 2 mmHg of baseline prior to the induction of LBNP. However, there was a singleexperiment where the target error worsened compared to the BSIOFF trial. Although the mean MAP, SBP, and DBP increased during this experiment, there was a pronounced period of hypotension that occurred during low stimulation activity. Similar to the SNP experiments, vascular mechanisms appeared to contribute more strongly to BP increases rather than cardiac mechanisms (FIGS. 390, 39F), however these findings remained underpowered (Table 12).Table 11: Effect size and achieved statistical power for BSI-mediated responses following PE injections (n=5)Comparison Effect size (Cohen’s dz) PowerChange at peak (%) -0.31 0.08Change in HR (%) -0.06 0.05AUG of PE traces (mmHg) -0.87 0.32Change in pulse pressure (%) -0.18 0.06Change in AMAP (%) 1.24 0.56Change in ADBP (%) 1.35 0.62Change in ASBP (%) 1.14 0.49Table 12: Effect size and achieved statistical power for BSI-mediated responses during LBNP experiments (n=10)Comparison Effect size (Cohen’s dz) PowerTarget error (mmHg) 0.98 0.79Change in HR (%)0.27 0.12AUG of LBNP traces (mmHg) 1.07 0.85Change in pulse pressure (%) 0.52 0.32Change in AMAP (%) 1.27 0.94Change in ADBP (%) 1.02 0.82Change in ASBP (%) 1.17 0.91
[0315] In summary, the BSI successfully activated EES during hypotension, and minimized it during hypertension. Moreover, it demonstrated that these biomarkers in the RVLM are stable following acute SCI, establishing a bridge between neural pathways and autonomic regulation that may enable future neuroprosthetic development.Discussion
[0316] The embodiments discussed in accordance with the teachings herein employed direct and invasive implantation of MEAs within the RVLM to drive control of EES. This was a deliberate decision made to capture the integrated output of multiple supraspinal autonomic pathways that converge in the RVLM to encode sympathetic drive for BP regulation [3, 16], For instance, this includes the pathways explored from thecharacterization experiments (chemoreflex, HPA, and baroreflex) described previously, as well as pathways related to postural control, arousal, hemorrhagic response, and psychological stress [3, 8, 13, 16, 29, 37],Decoder design and online control strategy
[0317] The neural decoder leveraged the LFP findings from the baroreflex characterization experiments (FIGS. 20A-20C). In at least one embodiment, the LFP signal may be chosen over MUA firing rates because it aggregates the activity across a wider population of neurons
[0081] , which may make the reliance on electrode placement close to barosensitive RVLM neurons less stringent. Moreover, LFPs may be preferred as a proposed translational step toward less invasive access to the RVLM, such as endovascular electrodes that have primarily leveraged LFP signals for BCI control [71-73], The results from the chemoreflex and GLP-1 experiments were not used for training the neural decoder to minimize the possible confounding effects of multiple neural pathways beyond the very well characterized baroreflex pathway [1, 3],
[0318] Having constrained the training data for the decoder to baroreflex perturbations, additional features were generated to better address the nonlinear processes that comprise the LFP signal [59, 82], This feature expansion was motivated by both experimental results from the previous chapter and physiological priors. For instance, high-gamma band power in the RVLM was significantly modulated by changes in BP (FIGS. 20A-20C). Features that incorporated the VLF band were also included, based on the relevance of Mayer waves (0.4-0.6 Hz) in rats to the baroreflex
[0126] , These features may improve decoder accuracy, despite no significance observed in the results from the baroreflex characterization experiments.
[0319] In at least one embodiment which uses the VLF band, two-second windows may be used for calculating LFP band powers. In rats, the upper end of the Mayer wave frequency range minimally satisfies the Nyquist criterion with this chosen window duration. In humans, Mayerwaves occur around 0.1 Hz
[0126] , Translating the use of VLF for RVLM-based control in humans would introduce major latency to the system and represents a major translational barrier to overcome with this approach.
[0320] Nonlinear feature pruning leveraged a custom / ?-VAE, which was trained to encode a latent representation predictive of SBP, DBP, and MAP. Integrated gradient analysis revealed that the majority of latent dimensions with the strongest attribution to these BP metrics were shared. This finding may arise, in part, from the chosen lossfunction (Equation 11) that assigned equal importance to SBP, DBP, and MAP, despite MAP being a deterministic function of systolic and diastolic BP.
[0321] The initial decoding models sought to learn a relationship between neural activity in the RVLM and EES amplitude, as outlined in FIGS. 34A-34C and 35A-35C. The linear regression models performed adequately with k-fold cross validation, but did not generalize to LODO cross validation. This latter result suggested that the online performance would be insufficient to titrate EES amplitude for BP stabilization after an SCI transection, as hypothesized. This motivated the decision to reformulate the control paradigm as a classification problem rather than continuing with regression models.BSI performance
[0322] The final online decoding algorithm for the BSI included a two-class logistic regression model with features outlined in FIGS. 36A-36E. The decoding accuracy of the BSI discussed herein, as estimated using the area under the ROC curve (FIG. 36D), was 71%. This modest AUC may reflect the cyclical nature of the LBNP protocol that was used as training data for the model’s development; some epochs labelled as baseline likely contained residual neural activation within the RVLM, despite no negative pressure being applied. This overlap between baseline and hypotensive neural states may have reduced separability between the two classes of the logistic regression model, limiting performance. Moreover, the VLF and high-gamma features with the most predictive power were all lagged by three-to-five time points (6-12 seconds), as shown in FIG. 36B, which may have further contributed to this reduced separability in the LBNP datasets. This final model contained solely present and lagged-time features from VLF and high-gamma bands (FIG. 36B and Table 9). The VLF band exhibited significant and negative odds ratios on the logistic regression model, indicating that this band inhibited stimulation decisions. The high-gamma finding is consistent with the results obtained in the experiments described previously.
[0323] The BSI performance was validated against both hypotensive and hypertensive perturbations (FIGS. 37A-37F, 38A-38F and 39A-39F). These perturbations were chosen deliberately to evaluate decoder accuracy and decision speed, and to model prolonged OH episodes [4, 50], While the BSI was unable to completely mitigate the vasodilation from SNP, shown in FIG. 37A, it significantly blunted the drop in MAP. This was most evident at the nadir, which occurred approximately 15 seconds following injection. In contrast, a design objective was to minimize stimulation during hypertension, such as an episode of AD [4, 58], Across all PE trials (FIG. 38A-38F), the BSI rarely stimulated, whichwas consistent with the strongest relative performance observed when the decoder was trained in FIG. 36E. Furthermore, this behaviour aligns with observed 01 neuron firing rate reductions at elevated BP [6, 30, 35], Lastly, the BSI stabilized BP during 30-minute LBNP trials (FIGS. 39A-39F), demonstrating a sustained performance over extended durations. Together, these findings support the feasibility of using RVLM activity in real-time for physiologically relevant control as a strategy for mitigating hypotension after SCI.
[0324] As part of the model building, a comprehensive feature engineering step was performed. This intentionally supplemented the high-gamma band power finding from the previous chapter (FIGS. 17A-20C). The choice of using polynomial expansion, lagged-time features, derivatives, and ratios between bands created a risk of redundant and autocorrelated features. This was mitigated by considering mutual information to prune features, and nonlinear feature selection through the combination of the?-VAE and integrated gradient analysis. Nevertheless, integrating prior physiological knowledge (prioritizing VLF and high-gamma band powers) to the feature expansion at earlier stages may have reduced the computational burden and simplified the analysis. This step may have improved the model performance over simply using the significant results obtained from the RVLM characterization experiments by identifying the importance of lagged-time features and the VLF band.
[0325] It will be appreciated by those skilled in the art that changes could be made to the various aspects of the subject application described above without departing from the inventive concept thereof. It is to be understood, therefore, that this subject application is not limited to the particular aspects disclosed, but it is intended to cover modifications as defined by the appended claims.ReferencesThe following list of references are all incorporated by reference in their entirety.[1] William W. Blessing. The lower brainstem and bodily homeostasis. Oxford University Press, New York, 1997.[3] P. G. Guyenet, “The sympathetic control of blood pressure,” Nature Reviews Neuroscience, vol. 7, no. 5, pp. 335-346, May 2006.[4] A. A. Phillips and A. V. Krassioukov, “Contemporary Cardiovascular Concerns after Spinal Cord Injury: Mechanisms, Maladaptations, and Management,” Journal of Neurotrauma, vol.32, no. 24, pp. 1927-1942, Dec. 2015.[6] A. M. Schreihoferand P. G. Guyenet, “The Baroreflex And Beyond: Control Of Sympathetic Vasomotor Tone By Gabaergic Neurons In The Ventrolateral Medulla,” Clinical and Experimental Pharmacology and Physiology, vol. 29, no. 5-6, pp. 514-521, May 2002.[8] G. M. P. R. Souza, H. Thakkalapally, F. E. Berry, U. M. Atongazi, D. S. Stornetta, and S. B. Abbott, “Control of Blood Pressure Variability Across Behavioral States by Brainstem Adrenergic Neurons,” May 2025.
[0013] D. Kline, T. King, J. Austgen, C. Heesch, and E. Hasser, “Sensory afferent and hypoxia-mediated activation of nucleus tractus solitarius neurons that project to the rostral ventrolateral medulla,” Neuroscience, vol. 167, no. 2, pp. 510-527, May 2010.
[0016] P. G. Guyenet, R. L. Stornetta, G. Bochorishvili, S. D. DePuy, P. G. R. Burke, and S. B. G. Abbott, “C1 neurons: The body’s EMTs,” American Journal of Physiology-Regulatory, Integrative and Comparative Physiology, vol. 305, no. 3, pp. R187-R204, Aug. 2013.
[0017] . W. Squair, M. Gautier, L. Mahe, J. E. Soriano, A. Rowaid, A. Bichat, N. Cho, M. A. Anderson, N. D. James, J. Gandar, A. V. Incognito, G. Schiavone, Z. K. Sarafis, A. Laskaratos, K. Bartholdi, R. Demesmaeker, S. Komi, C. Moerman, B. Vaseghi, B. Scott, R. Rosen- treter, C. Kathe, J. Ravier, L. McCracken, X. Kang, N. Vachicouras, F. Fallegger, I. Jelescu, Y. Cheng, Q. Li, R. Buschman, N. Buse, T. Denison, S. Dukelow, R. Charbonneau, I. Rigby, S. K. Boyd, P. J. Millar, E. M. Moraud, M. Capogrosso, F. B. Wagner, Q. Barraud, E. Bezard, S. P. Lacour, J.Bloch, G. Courtine, and A. A. Phillips, “Neuroprosthetic baroreflex controls haemodynamics after spinal cord injury,” Nature, vol. 590, no. 7845, pp. 308-314, Feb. 2021.
[0019] J. P. Card, J. C. Sved, B. Craig, M. Raizada, J. Vazquez, and A. F. Sved, “Efferent projections of rat rostroventrolateral medulla C1 catecholamine neurons: Implications for the central control of cardiovascular regulation,” The Journal of Comparative Neurology, vol. 499, no. 5, pp.840-859, Dec. 2006.
[0023] J. Manuel, N. Farber, D. A. Gerlach, K. Heusser, J. Jordan, J. Tank, and F. Beissner, “Deciphering the neural signature of human cardiovascular regulation,” eLife, vol. 9, p. e55316, Jul. 2020.
[0025] A. F. Sved, S. Ito, C. J. Madden, S. D. Stocker, and Y. Yajima, “Excitatory Inputs to the RVLM in the Context of the Baroreceptor Reflex,” Annals of the New York Academy of Sciences, vol. 940, no. 1, pp. 247-258, Jan. 2006.
[0027] E. V. Golanov, D. A. Ruggiero, and D. J. Reis, “A brainstem area mediating cerebrovascular and EEG responses to hypoxic excitation of rostral ventrolateral medulla in rat,” The Journal of Physiology, vol. 529, no. 2, pp. 413-429, Dec. 2000.
[0029] S. B. G. Abbott, S. D. DePuy, T. Nguyen, M. B. Coates, R. L. Stornetta, and P. G. Guyenet, “Selective Optogenetic Activation of Rostral Ventrolateral Medullary Catecholaminergic Neurons Produces Cardiorespiratory Stimulation in Conscious Mice,” Journal of Neuroscience, vol. 33, no. 7, pp. 3164-3177, Feb. 2013.
[0030] D L Brown and P G Guyenet. Electrophysiological study of cardiovascular neurons in the rostral ventrolateral medulla in rats. Circulation Research, 56(3):359-369, March 1985.
[0035] A. J. M. Verberne, R. L. Stornetta, and P. G. Guyenet, “Properties of C1 and other ventrolateral medullary neurones with hypothalamic projections in the rat,” The Journal of Physiology, vol. 517, no. 2, pp. 477-494, Jun. 1999.
[0037] G. M. Souza, R. L. Stornetta, D. S. Stornetta, P. G. Guyenet, and S. B. Abbott, “Adrenergic C1 neurons monitor arterial blood pressure and determine the sympathetic response to hemorrhage,” Cell Reports, vol. 38, no. 10, p. 110480, Mar. 2022.
[0050] M. Hubli, C. M. Gee, and A. V. Krassioukov, “Refined Assessment of Blood Pressure Instability After Spinal Cord Injury,” American Journal of Hypertension, vol. 28, no. 2, pp. 173— 181, Feb. 2015.
[0053] Anthony J. M. Verberne, Ruth L. Stornetta, and Patrice G. Guyenet. Properties of C1 and other ventrolateral medullary neurones with hypothalamic projections in the rat. The Journal of Physiology, 517(2):477-494, June 1999.
[0056] A. A. Phillips et al., An implantable system to restore hemodynamic stability after spinal cord injury,” Nature Medicine, vol. 31, no. 9, pp. 2946- 2957, Sep. 2025.
[0058] J. E. Soriano, R. Hudelle, L. Mahe, M. Gautier, A. Y. Y. Teo, M. A. Skinnider, A. Laskaratos, S. Ceto, C. Kathe, T. Hutson, R. Charbonneau, F. Girgis, S. Casha, J. Rimok, M. Tso, K. Larkin-Kaiser, N. Hankov, A. Gandhi, S. Amir, X. Kang, Y. Vyza, E. Martin-Moraud, S. Lacour, R. Demesmaeker, L. Asboth, Q. Barraud, M. A. Anderson, J. Bloch, J. W. Squair, A. A. Phillips, and G. Courtine, “A neuronal architecture underlying autonomic dysreflexia,” Nature, Sep.2025.
[0059] G. Buzsaki, C. A. Anastassiou, and C. Koch, “The origin of extracellular fields and currents - EEG, ECoG, LFP and spikes,” Nature Reviews Neuroscience, vol. 13, no.6, pp.407-420, Jun.2012.
[0064] Thomas J Oxley, Nicholas L Opie, Sam E John, Gil S Rind, Stephen M Ronayne, Tracey L Wheeler, Jack W Judy, Alan J McDonald, Anthony Dornom, Timothy J H Lovell, Christopher Steward, David J Garrett, Bradford A Moffat, Elaine H Lui, NawafYassi, Bruce C V Campbell, Yan T Wong, Kate E Fox, Ewan S Nurse, Iwan E Bennett, Sebastien H Bauquier, Kishan A Liyanage, Nicole R Van Der Nagel, Piero Perucca, Arman Ahnood, Katherine P Gill, Bernard Yan, Leonid Churilov, Christopher R French, Patricia M Desmond, Malcolm K Horne, Lynette Kiers, Steven Prawer, Stephen M Davis, Anthony N Burkitt, Peter J Mitchell, David B Grayden, Clive N May, and Terence J O’Brien. Minimally invasive endovascular stent-electrode array for high-fidelity, chronic recordings of cortical neural activity. Nature Biotechnology, 34(3):320-327, March 2016.
[0066] G. Paxinos and C. Watson, The Rat Brain in Stereotaxic Coordinates, 6th ed. Amsterdam: Elsevier / Academic Press, 2007.
[0072] Anqi Zhang, Emiri T. Mandeville, Lijun Xu, Creed M. Stary, Eng H. Lo, and Charles M. Lieber. Ultraflexible endovascular probes for brain recording through micrometer-scale vasculature. Science, 381 (6655):306— 312, July 2023.
[0073] Peter Mitchell, Sarah C. M. Lee, Peter E. Yoo, Andrew Morokoff, Rahul P. Sharma, Daryl L. Williams, Christopher Maclsaac, Mark E. Howard, Lou Irving, Ivan Vrljic, Cameron Williams, Steven Bush, Anna H. Balabanski, Katharine J. Drummond, Patricia Desmond, Douglas Weber, Timothy Denison, Susan Mathers, Terence J. O’Brien, J. Mocco, David B. Grayden, David S. Liebeskind, Nicholas L. Opie, Thomas J. Oxley, and Bruce C. V. Campbell. Assessment of Safety of a Fully Implanted Endovascular Brain-Computer Interface for Severe Paralysis in 4 Patients: The Stentrode With Thought-Controlled Digital Switch (SWITCH) Study. JAMA Neurology, January 2023.
[0077] J. R. Manning, J. Jacobs, I. Fried, and M. J. Kahana, “Broadband Shifts in Local Field Po- tential Power Spectra Are Correlated with Single-Neuron Spiking in Humans,” The Journal of Neuroscience, vol. 29, no. 43, pp. 13613-13620, Oct. 2009.
[0078] M. Bonizzato, G. Pidpruzhnykova, J. DiGiovanna, P. Shkorbatova, N. Pavlova, S. Micera, and G. Courtine, “Brain-controlled modulation of spinal circuits improves recovery from spinal cord injury,” Nature Communications, vol. 9, no. 1, p. 3015, Aug. 2018.
[0080] R. Q. Quiroga, Z. Nadasdy, and Y. Ben-Shaul, “Unsupervised Spike Detection and Sorting with Wavelets and Superparamagnetic Clustering,” Neural Computation, vol. 16, no. 8, pp.1661-1687, Aug. 2004.
[0081] A. Schilling, R. Gerum, C. Boehm, J. Rasheed, C. Metzner, A. Maier, C. Reindl, H. Hamer, and P. Krauss, “Deep learning based decoding of single local field potential events,” NeuroImage, vol. 297, p. 120696, Aug. 2024.
[0082] G. T. Einevoll, C. Kayser, N. K. Logothetis, and S. Panzeri, “Modelling and analysis of local field potentials for studying the function of cortical circuits,” Nature Reviews Neuroscience, vol.14, no. 11, pp. 770-785, Nov. 2013.
[0084] P. Welch, “The use of fast Fourier transform for the estimation of power spectra: A method based on time averaging over short, modified periodograms,” IEEE Transactions on Audio and Electroacoustics, vol. 15, no. 2, pp. 70-73, Jun. 1967.
[0115] N. Limjunyawong, J. Mock, and W. Mitzner, “Instillation and Fixation Methods Useful in Mouse Lung Cancer Research,” Journal of Visualized Experiments, no. 102, p. 52964, Aug.2015.
[0116] M. Gil-Lozano, D. Perez-Tilve, M. Alvarez-Crespo, A. Mart'is, A. M. Fernandez, P. A. F. Catalina, L. C. Gonzalez-Matias, and F. Mallo, “GLP-1(7-36)-amide and Exendin-4 Stimulate the HPA Axis in Rodents and Humans,” Endocrinology, vol. 151, no. 6, pp. 2629-2640, Jun.2010.
[0117] E. Bojanowska and B. Stempniak, “Effects of centrally or systemically injected glucagon-like peptide-1 (7-36) amide on release of neurohypophysial hormones and blood pressure in the rat,” Regulatory Peptides, vol. 91, no. 1-3, pp. 75-81, Jul. 2000.
[0118] C. R. Harris, K. J. Millman, S. J. Van DerWalt, R. Gommers, P. Virtanen, D. Cournapeau, E. Wieser, J. Taylor, S. Berg, N. J. Smith, R. Kern, M. Picus, S. Hoyer, M. H. Van Kerkwijk, M. Brett, A. Haldane, J. F. Del R'IO, M. Wiebe, P. Peterson, P. Gerard-Marchant, K. Sheppard, T. Reddy, W. Weckesser, H. Abbasi, C. Gohlke, and T. E. Oliphant, “Array programming with NumPy,” Nature, vol. 585, no. 7825, pp. 357-362, Sep. 2020.
[0119] P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, S. J. van der Walt, M. Brett, J. Wil- son, K. J. Millman, N. Mayorov, A. R. J. Nelson, E. Jones, R. Kern, E. Larson, C. J. Carey, I. Polat, Y. Feng, E. W. Moore, J. VanderPlas, D. Laxalde, J. Perktold, R. Cimrman, I. Henriksen, E. A. Quintero, C. R. Harris, A. M. Archibald, A. H. Ribeiro, F. Pedregosa, and P. van Mulbregt, “SciPy 1.0: Fundamental algorithms for scientific computing in Python,” Nature Methods, vol.17, no. 3, pp. 261-272, Mar. 2020.
[0120] W. McKinney, “Data Structures for Statistical Computing in Python,” in Proceedings of the 9th Python in Science Conference, vol. 445, Austin, Texas, USA, 2010, pp. 56-61.
[0121] J. D. Hunter, “Matplotlib: A 2D Graphics Environment,” Computing in Science & Engineering, vol. 9, no. 3, pp. 90-95, 2007.
[0122] M. Waskom, “Seaborn: Statistical data visualization,” Journal of Open Source Software, vol. 6, no. 60, p. 3021, Apr. 2021.
[0123] F. Charlier, M. Weber, D. Izak, E. Harkin, M. Magnus, J. Lalli, L. Fesnais, M. Chan, N. Markov, O. Amsalem, S. Proost, A. Krasoulis, and S. Repplinger, “Statannotations,” Zenodo, 2022.
[0124] T. G. Papaioannou, A. D. Protogerou, D. Vrachatis, G. Konstantonis, E. Aissopou, A. Argyris,. Nasothimiou, E. J. Gialafos, M. Karamanou, D. Tousoulis, and P. P. Sfikakis, “Meanarterial pressure values calculated using seven different methods and their associations with target organ deterioration in a single-center study of 1878 individuals,” Hypertension Research, vol. 39, no. 9, pp. 640-647, Sep. 2016.
[0125] J. E. Hall and M. E. Hall, “Guyton and Hall textbook of medical physiology,” in Guyton and Hall Textbook of Medical Physiology, fifteenth edition, ed. Philadelphia, PA: Elsevier, 2025 - 2026.
[0126] C. Julien, “The enigma of Mayer waves: Facts and models,” Cardiovascular Research, vol. 70, no. 1, pp. 12-21, Apr. 2006.
[0127] S. Seabold and J. Perktold, “Statsmodels: Econometric and statistical modeling with python,” in Proceedings of the 9th Python in Science Conference, S. J. Van Der Walt and J. Millman, Eds., vol. 445, Austin, Texas, USA, 2010, pp. 92-96.
[0128] Y. Benjamini and Y. Hochberg, “Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing,” Journal of the Royal Statistical Society Series B: Statistical Methodology, vol. 57, no. 1, pp. 289-300, Jan. 1995.
[0129] D. Lakens, “Calculating and reporting effect sizes to facilitate cumulative science: A practical primer for t-tests and ANOVAs,” Frontiers in Psychology, vol. 4, 2013.
[0131] I. C. Wenker, C. Abe, K. E. Viar, D. S. Stornetta, R. L. Stornetta, and P. G. Guyenet, “Blood Pressure Regulation by the Rostral Ventrolateral Medulla in Conscious Rats: Ef- fects of Hypoxia, Hypercapnia, Baroreceptor Denervation, and Anesthesia,” The Journal of Neuroscience, vol. 37, no. 17, pp. 4565-4583, Apr. 2017.
[0142] F. Pedregosa, F. Pedregosa, G. Varoquaux, G. Varoquaux, N. Org, A. Gramfort, A. Gramfort, V. Michel, V. Michel, L. Fr, B. Thirion, B. Thirion, O. Grisel, O. Grisel, M. Blondel, P. Pret- tenhofer, P. Prettenhofer, R. Weiss, V. Dubourg, V. Dubourg, J. Vanderplas, A. Passos, A. Tp, and D. Cournapeau, “Scikit-learn: Machine Learning in Python,” MACHINE LEARNING IN PYTHON, 2011.
[0143] G. E. Hinton and R. R. Salakhutdinov, “Reducing the Dimensionality of Data with Neural Networks,” Science, vol. 313, no. 5786, pp. 504-507, Jul. 2006.
[0144] I. Higgins, L. Matthey, A. Pal, C. Burgess, X. Glorot, M. Botvinick, S. Mohamed, and A. Lerchner, “beta-vae: Learning basic visual concepts with a constrained variational framework,” in International conference on learning representations, 2017.
[0145] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “PyTorch: An Imperative Style, High-Performance Deep Learning Library,” Dec. 2019.
[0146] M. Sundararajan, A. Taly, and Q. Yan, “Axiomatic Attribution for Deep Networks,” Jun.2017.
[0147] . Kokhlikyan, V. Miglani, M. Martin, E. Wang, B. Alsallakh, J. Reynolds, A. Melnikov, N. Kliushkina, C. Araya, S. Yan, and O. Reblitz-Richardson, “Captum: A unified and generic model interpretability library for PyTorch,” CoRR, vol. abs / 2009.07896, 2020.
[0148] P. Rudiger et al. “Holoviz-holoviews: Version 1.21.0,” Zenodo, Jun. 2025.
[0149] H. M. Cheng, D. Lang, C. Tufanaru, and A. Pearson, “Measurement accuracy of noninvasively obtained central blood pressure by applanation tonometry: A systematic review and meta-analysis,” International Journal of Cardiology, vol. 167, no. 5, pp. 1867-1876, Sep.2013.
[0170] Sam E. John, Nicholas L. Opie, Yan T. Wong, Gil S. Rind, Stephen M. Ronayne, Giulia Gerboni, Sebastien H. Bauquier, Terence J. O’Brien, Clive N. May, David B. Grayden, and Thomas J. Oxley. Signal quality of simultaneously recorded endovascular, subdural and epidural signals are comparable. Scientific Reports, 8(1 ):8427, May 2018.
[0171] Henri Lorach, Andrea Galvez, Valeria Spagnolo, Felix Martel, Serpil Karakas, Nadine Intering, Molywan Vat, Olivier Faivre, Cathal Harte, Salif Komi, et al. Walking naturally after spinal cord injury using a brain-spine interface. Nature, pages 1-8, 2023.
Claims
CLAIMS1. A blood pressure regulation system comprising:a detector operable to detect at least one electrical signal from brain tissue;a stimulator device operable to apply at least one electrical stimulus signal to spinal cord tissue to regulate blood pressure; andat least one processor configured to control application by the stimulator device of the at least one electrical stimulus signal to the spinal cord tissue based at least partially on a presence of at least one blood pressure correlated biomarker in the detected at least one electrical signal.
2. The system of claim 1, wherein the at least one processor is further configured to scan the at least one electrical signal for the presence of the at least one blood pressure correlated biomarker.
3. The system of claim 1 or 2, wherein the at least one processor is further configured to, in response to detecting the at least one blood pressure correlated biomarker in the detected at least one electrical signal, determine at least one parameter of the at least one electrical stimulus signal based at least partially on the at least one blood pressure correlated biomarker.
4. The system of any one of claims 1 to 3, wherein the detector is operable to detect the at least one electrical signal from medullary neurons.
5. The system of any one of claims 1 to 4, wherein the detector is operable to detect the at least one electrical signal from ventrolateral medulla neurons.
6. The system of any one of claims 1 to 5, wherein the detector is operable to detect the at least one electrical signal from rostral ventrolateral medulla neurons.
7. The system of any one of claims 1 to 6, wherein the detector is operable to detect the at least one electrical signal from bulbospinal C1 neurons.
8. The system of any one of claims 1 to 7, wherein the at least one blood pressure correlated biomarker comprises a spectral signature of a local field potential, wherein the spectral signature was a current band power or a time-lagged power.
9. The system of claim 8, wherein the spectral signature is between about 80±10% Hz to about 200±10% Hz.
10. The system of claim 8 or 9, wherein the spectral signature is between about 120±10% Hz to about 200±10% Hz.
11. The system of claim 8, wherein the spectral signature is between about 80±10% Hz to about 120±10% Hz.
12. The system of claim 8, wherein the spectral signature is between about 0.01 ±10% Hz to about 1 ±10% Hz.
13. The system of any one of claims 1 to 12, wherein the at least one blood pressure correlated biomarker comprises a number of neural firings above a threshold value.
14. The system of claim 13, wherein the threshold value is four standard deviations greater than a baseline number of neural firings.
15. The system of claim 13, wherein the threshold value is five standard deviations greater than a baseline number of neural firings.
16. The system of any one of claims 13 to 15, wherein the detected at least one electrical signal is from a local field potential (LFP).
17. The system of any one of claims 1 to 12, wherein the at least one blood pressure correlated biomarker is determined at least partially by isolating neurons based on template matching.
18. The system of any one of claims 1 to 17, wherein the detector is invasive.
19. The system of claim 18, wherein the detector comprises at least one electrode operable to detect the at least one electrical signal.
20. The system of claim 19, wherein the at least one electrode of the detector is implantable within the brain tissue.
21. The system of claim 20, wherein the at least one electrode of the detector is implantable within the rostral ventrolateral medulla.
22. The system of claim 19, wherein the at least one electrode of the detector is an electrocorticography electrode.
23. The system of claim 19, wherein the at least one electrode of the detector is an endovascular electroencephalography electrode.
24. The system of any one of claims 1 to 17, wherein the detector is non-invasive.
25. The system of claim 24, wherein the detector comprises at least one electrode operable to detect the at least one electrical signal.
26. The system of claim 25, wherein the at least one electrode of the detector is an electroencephalography electrode.
27. The system of any one of claims 1 to 26, wherein the stimulator device is invasive.
28. The system of any one of claims 1 to 26, wherein the stimulator device is non-invasive.
29. The system of any one of claims 1 to 28, wherein the stimulator device comprises at least one electrode operable to apply the at least one electrical stimulus signal to the spinal cord tissue.
30. The system of any one of claims 1 to 29, wherein the stimulator device is operable to apply the at least one electrical stimulus signal to spinal cord tissue between a T10 spinal cord level and a T12 spinal cord level.
31. The system of any one of claims 1 to 30, wherein the detector communicates wirelessly with the at least one processor.
32. The system of any one of claims 1 to 31, wherein the stimulator device communicates wirelessly with the at least one processor.
33. The system of any one of claims 1 to 32, further comprising a feedback controller.
34. The system of claim 33, wherein the feedback controller is configured to receive as input a measure of blood pressure.
35. The system of claim 33 or 34, wherein the feedback controller is configured to receive as input the at least one electrical signal.
36. The system of any one of claims 33 to 35, wherein the feedback controller is a Proportional-Integral-Derivative (PID) controller.
37. The system of any one of claims 33 to 35, wherein the feedback controller is configured to use a weighted least squares linear model to control application by the stimulator device of the at least one electrical stimulus signal to the spinal cord tissue.
38. The system of any one of claims 33 to 35, wherein the feedback controller is configured to implement a logistic regression model to control application by the stimulator device of the at least one electrical stimulus signal to the spinal cord tissue.
39. The system of claim 38, wherein the logistic regression model was trained on features including current and time-lagged band powers from high-gamma1, high-gamma2 and VFL spectral bands.
40. The system of any one of claims 38 to 39, wherein the logistic regression model is paired with a weighted least squares (WLS) model or an exponentially weighted moving average (EWMA) model.
41. The system of any one of claims 1 to 40, wherein an average over time of the detected at least one electrical signal was used to trigger stimulation to avoid stimulation in response to noise spikes.
42. A method of regulating blood pressure, the method comprising:detecting presence of at least one blood pressure correlated biomarker in at least one electrical signal from brain tissue; andbased on the detected presence of the at least one blood pressure correlated biomarker, applying at least one electrical stimulus signal to spinal cord tissue to regulate the blood pressure.
43. The method of claim 42, wherein the detecting is performed using a detector and at least one processor, and the at least one electrical stimulus signal is applied using a stimulator device controlled by the at least one processor.
44. The method of claim 42 or claim 43, wherein the at least one electrical signal is detected from rostral ventrolateral medulla neurons.
45. The method of claim 44, wherein the at least one electrical signal is detected from bulbospinal C1 neurons.
46. The method of any one of claims 42 to 45, wherein the at least one electrical stimulus signal is applied between a T10 spinal cord level and a T12 spinal cord level.
47. The method of any one of claims 42 to 46, wherein the method is performed using the system defined by any one of claims 1 to 41.
48. Use of the system of any one of claims 1 to 41 as a blood pressure regulator.
49. A non-transitory computer readable medium storing program code instructions that, when executed by at least one processor, configure the at least one processor to perform a method defined according to any one of claims 42 to 47 for regulating blood pressure.
50. A blood pressure regulation system comprising:a detector operable to detect at least one blood-pressure correlated signal from brain tissue;a stimulator device operable to apply at least one electrical stimulus signal to spinal cord tissue to regulate blood pressure; andat least one processor configured to control application by the stimulator device of the at least one electrical stimulus signal to the spinal cord tissue based at least partially on a presence of at least one blood pressure correlated biomarker in the detected at least one blood-pressure correlated signal.
