Methods and devices for the non-invasive assessment of cerebro-spinal dynamics and diagnosis of brain pathophysiologies

Abstract

Systems and methods for assessment of fluid dynamics in a human subject involve: one or more physiological sensors configured to non-invasively record one or more physiological signals from the human subject; one or more microprocessors in communication with the one or more physiological sensors and configured to extract one or more features from the one or more physiological signals and to implement a machine learning algorithm configured to estimate one or more aspects of fluid dynamics from one or both of the one or more physiological signals and the one or more features; and an output device in communication with the microprocessors and configured to output the one or more aspects of fluid dynamics. The fluid dynamics may be cerebrospinal fluid dynamics. The fluid dynamics may be glymphatic fluid dynamics.

Description

PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT METHODS AND DEVICES FOR THE NON-INVASIVE ASSESSMENT OF CEREBRO-SPINAL DYNAMICS AND DIAGNOSIS OF BRAIN PATHOPHYSIOLOGIES STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

[0001] This invention was made with government support under Grant No. W81XWH-21-09-0016 awarded by the Department of Defense, Army Medical Research and Development Command. The government has certain rights in the invention. BACKGROUND

[0002] Hydrocephalus is a common neurological disorder that is associated with abnormal dynamics and accumulation of cerebrospinal fluid (CSF) in the brain ventricles. Hydrocephalus affects one in every 1,000 live births. Treating the condition costs an estimated $50,000 per patient per year and costs the U.S. health care system about $1 billion per year. The excess fluid observed in hydrocephalus increases intraventricular pressure and causes the ventricles to widen, resulting in a harmful increase in pressure on the brain's tissues. Hydrocephalus may be present at or shortly after birth, or may result over time from damage, injury, or inflammation of the brain. Overt neurological symptoms of hydrocephalus appear much delayed and can vary significantly from person to person, mostly depending on age. Hydrocephalus that is untreated causes severe disability and even death and in aged people with neurological decline.

[0003] Hydrocephalus is most often diagnosed after the first symptoms have appeared. The diagnosis is usually done with clinical neurological tests and by using brain imaging techniques. Among neurological tests, a spinal tap (lumbar puncture) allows a clinician to estimate CSF pressure and analyze the fluid by inserting a needle in the lower back and removing and testing some of the fluid. IntracranialPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT pressure monitoring uses a small pressure monitor inserted into the brain or ventricles to measure the pressure and detect the amount of swelling that may have occurred in the brain. Fundoscopic examination uses a special device to view the optic nerve at the back of the eye. It can show evidence of swelling that suggests elevated intracranial pressure, which can be a result of hydrocephalus.

[0004] Among imaging techniques, ultrasound is often the first method used to diagnose infants because it is relatively simple and low risk. When used during routine prenatal exams, ultrasound may also detect hydrocephalus in unborn babies. Computed tomography (CT) can be used to determine whether the ventricles are enlarged or if there is an obstruction. Magnetic resonance imaging (MRI) can determine if the ventricles are enlarged, assess the CSF flow, and provide information about the brain tissue surrounding the ventricles. MRI is often the initial and preferred test used to diagnose adults; it is able accurately assess both anatomical and functional features of the brain by measuring the velocity and bulk flow of CSF high resolution.

[0005] All of the imaging techniques are cumbersome, expensive, and typically only applied after the first overt neurological symptoms have appeared. In particular, an MRI is non-portable due to the requirement of large magnets and data acquisition systems, thus not readily available in patients living in rural areas and in developing countries. As such, no noninvasive technology is available that may be used to diagnose patients suffering from hydrocephalus or from other abnormalities in CSF accumulation and dynamics in a continuous, non-invasive, and comfortable way. BRIEF DESCRIPTION OF DRAWINGS

[0006] Specific embodiments of the disclosed technology will now be described in detail with reference to the accompanying figures. Like elements in the various figures are denoted by like reference numerals for consistency.

[0007] FIG. 1 shows the placement of a chest patch, a PPG sensor, and a piezoelectric belt, in accordance with one or more embodiments.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0008] FIG. 2 shows a sensor detecting shunt occlusion, in accordance with one or more embodiments.

[0009] FIG. 3 shows a flowchart for methods, in accordance with one or more embodiments.

[0010] FIG. 4 shows a velocity of cerebrospinal fluid, in accordance with one or more embodiments.

[0011] FIG. 5 shows bulk flow of cerebrospinal fluid, in accordance with one or more embodiments.

[0012] FIG. 6 shows a cerebral aqueduct area, in accordance with one or more embodiments.

[0013] FIG.7 shows a blood flow of the middle cerebral artery, in accordance with one or more embodiments.

[0014] FIG.8A shows a graph structure where the nodes of the graph correspond to electrodes, in accordance with one or more embodiments.

[0015] FIG. 8B shows the positions of transcranial probes and headgear, in accordance with one or more embodiments.

[0016] FIG. 9 shows a framework for integration of different modalities, in accordance with one or more embodiments.

[0017] FIG. 10 shows correlations between a transcranial doppler, CSF flow from an MRI, and signals from a sleep dataset, in accordance with one or more embodiments.

[0018] FIG. 11 shows a schematic of a proposed REG, in accordance with one or more embodiments.

[0019] FIG. 12 shows a neural network, in accordance with one or more embodiments.

[0020] FIG. 13 shows a computer system, in accordance with one or more embodiments.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0021] FIG. 14 shows the number of splines required to fit a TCD curve, in accordance with one or more embodiments.

[0022] FIG. 15 shows an average duration of the TCD cycles from a right MCA plotted against the width of a CSF velocity curve, in accordance with one or more embodiments.

[0023] FIG. 16 shows a diastolic deceleration of a TCD from a left MCA plotted against a width of a CSF velocity curve, in accordance with one or more embodiments.

[0024] FIG.17 shows Sample 7 tesla phase contrast MRI images, in accordance with one or more embodiments.

[0025] FIG. 18 shows a process that fits an ellipse to a region of interest, in accordance with one or more embodiments.

[0026] FIG. 19 shows a pulse signal with a clear systolic and diastolic peak, in accordance with one or more embodiments.

[0027] FIG. 20 shows pulsatile signal obtained from human finger and frequency domain analysis, in accordance with one or more embodiments.

[0028] FIG.21 shows the change in fluorescence activity, in accordance with one or more embodiments.

[0029] FIG. 22 shows a training error as a function of epochs, in accordance with one or more embodiments.

[0030] FIG. 23 shows results of the validation dataset, in accordance with one or more embodiments.

[0031] FIG. 24A shows an E-field envelope following a ǻI frequency offset, in accordance with one or more embodiments.

[0032] FIG. 24B shows another E-field envelope following another ǻI^ IUHTXHQF\^ offset, in accordance with one or more embodiments.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0033] FIG. 25 shows a patch antenna consisting of two dielectric layers and three metal layers, in accordance with one or more embodiments.

[0034] FIG.26 shows an antenna driven by a signal propagation chain, consisting of a signal generator, a power amplifier, a coupler, and an isolator, in accordance with one or more embodiments.

[0035] FIG.27 shows a measured antenna input reflection coefficient, in accordance with one or more embodiments.

[0036] FIG. 28 shows circuit simulations, in accordance with one or more embodiments.

[0037] FIG. 29 shows a 200K ohm shunt resistor, in accordance with one or more embodiments.

[0038] FIG.30 shows a weighted graph structure obtained from data, in accordance with one or more embodiments. SUMMARY

[0039] This summary is provided to introduce a selection of concepts that are further described below in the detailed description. This summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to be used as an aid in limiting the scope of the claimed subject matter.

[0040] In one aspect, embodiments disclosed herein relate to a system for assessment of cerebrospinal fluid dynamics in a human subject, the system including: one or more physiological sensors configured to non-invasively record one or more physiological signals from the human subject; one or more microprocessors in communication with the one or more physiological sensors; an output device in communication with the one or more microprocessors. The one or more microprocessors may be configured to extract one or more features from the one or more physiological signals, and implement a machine learning algorithm, wherein the machine learning algorithm is configured to estimate one or more aspects of cerebrospinal fluid dynamics from one or both of the one or more physiologicalPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT signals and the one or more features. The output device may be configured to output the one or more aspects of cerebrospinal fluid dynamics.

[0041] The one or more microprocessors of the system for assessment of cerebrospinal fluid dynamics may be configured to use the one or more aspects of cerebrospinal fluid dynamics to perform a diagnosis of a neurological condition, and the output device is configured to display the diagnosis. The diagnosis may be selected from the group consisting of hydrocephalus, traumatic brain injury, Parkinson’s, Alzheimer’s, epilepsy, an impaired shunt used in a treatment of hydrocephalus, and combinations thereof.

[0042] The one or more physiological sensors of the system for assessment of cerebrospinal fluid dynamics may be configured to continuously record the one or more physiological signals. The one or more physiological sensors may be configured to intermittently record the one or more physiological signals.

[0043] The output device of the system for assessment of cerebrospinal fluid dynamics may be in wireless communication with the one or more microprocessors. The output device may be a display device configured to display the one or more aspects of cerebrospinal fluid dynamics.

[0044] The one or more physiological sensors of the system for assessment of cerebrospinal fluid dynamics and the one or more microprocessors of the system for assessment of cerebrospinal fluid dynamics may be combined in a wearable device. The wearable device may include or be a headgear. The wearable device may include or be a chest patch. The wearable device may be in wireless communication with the output device.

[0045] The one or more physiological signals of the system for assessment of cerebrospinal fluid dynamics may be selected from the group consisting of cardiac sensors, photoplethysmogram sensors, piezoelectric sensors, transcranial probes, transcranial doppler ultrasound sensors, electrocardiogram sensors, cardiac function sensors, cardiac cycle sensors, blood volume sensors, blood gas sensors, electroencephalography sensors, photoplethysmogram sensors, near infraredPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT spectroscopygram sensors, rheoencephalography sensors, orbital sonography sensors, photoplethysmogram sensors, respiratory sensors, and combinations thereof.

[0046] The one or more features of the system for assessment of cerebrospinal fluid dynamics may be selected from the group consisting of a heart rate, a mean respiratory rate, a heart rate variability, a second derivative of photoplethysmogram signals, a photoplethysmogram intensity ratio, a respiration rate, a mean of the respiration rate, a standard deviation of the respiration rate, an area under a respiration curve, a maximum inhalation slope, a maximum inhalation slope, a blood volume, a partial pressure of oxygen, a mean R wave peak-to-peak distance, a heart rate variability, a root mean square of successive differences between normal heartbeats, a percentage of successive RR intervals that differ by more than 50 ms, a low frequency (LF) power, a high frequency (HF) power, a ratio between low frequency to high frequency power, a normalized low frequency power, a normalized high frequency power, a root mean square of successive differences (RMSSD), a percentage of adjacent NN intervals that vary from each other by more than 50 ms (PNN50), a very low frequency (VLF) power, an LF / HF ratio, a normalized LF power, a normalized HF power, a pulse transit time, a pulse intensity ratio, a photoplethysmogram heart rate variability, a percent dicrotic, a respiration rate, a respiration standard deviation, an inhale slope, an exhale slope, an area on a curve, a permutation entropy, an SVD entropy, a spectral entropy, an approximate entropy, a sample entropy, a Hjorth mobility, a Hjorth complexity, a Katz fractal dimension, a detrended fluctuation, a Higuchi fractal dimension, a Petrosian fractal dimension, a Hurst exponent, an arterial tone, a mutual information, a maximum transcranial doppler (TCD) velocity, a minimum TCD velocity, a mean TCD velocity, a pulsatility index, a resistance index, a systolic acceleration, a duration of cycle, a number of splines, a length of cycle, a diastolic deceleration, a number of maxima and minima, a delta wave power, K-complex waveforms, sleep spindles, saw-tooth waveforms, a degree of centrality, and combinations thereof. The one or more features may be calculated from a Sylvian aqueduct. The one or more featuresPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT may be calculated from sleep stages. The one or more features may include one or more graph features selected from the group consisting of a centrality, a page rank, a clustering coefficient, a global efficiency, a number of partitions, a shortest path length, an in degree, an out degree, and combinations thereof.

[0047] The machine learning algorithm of the system for assessment of cerebrospinal fluid dynamics may include a neural network trained to learn a relationship between the one or more physiological signals, the one or more features, and the one or more aspects of cerebrospinal fluid dynamics.

[0048] In another aspect, embodiments disclosed herein relate to a method for assessment of cerebrospinal fluid dynamics in a human subject including non- invasively recording, using one or more physiological sensors, one or more physiological signals from a human subject; extracting, using one or more microprocessors in communication with the one or more physiological sensors, one or more features from the one or more physiological signals; estimating, with a machine learning algorithm, one or more aspects of cerebrospinal fluid dynamics from one or both of the one or more physiological signals and the one or more features; and outputting, using an output device in communication with the one or more microprocessors, the one or more aspects of cerebrospinal fluid dynamics.

[0049] The method for assessment of cerebrospinal fluid dynamics may include: diagnosing, using the one or more microprocessors and the one or more aspects of cerebrospinal fluid dynamics, a neurological condition, and displaying, using the output device, a diagnosis. The method may include selecting the diagnosis from the group consisting of hydrocephalus, traumatic brain injury, Parkinson’s, Alzheimer’s, epilepsy, an impaired shunt used in a treatment of hydrocephalus, and combinations thereof.

[0050] The method for assessment of cerebrospinal fluid dynamics may include continuously recording the one or more physiological signals. The method may include intermittently recording the one or more physiological signals.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0051] The method for assessment of cerebrospinal fluid dynamics may include displaying the one or more aspects of cerebrospinal fluid dynamics on the output device, where the output device is a display device. The output device may be in wireless communication with the one or more microprocessors.

[0052] The method for assessment of cerebrospinal fluid dynamics may include combining the one or more physiological sensors and the one or more microprocessors may be combined in a wearable device. The wearable device may include or be a headgear. The wearable device may include or be a chest patch. The wearable device may be in wireless communication with the output device.

[0053] The method for assessment of cerebrospinal fluid dynamics may include selecting the one or more physiological signals from the group consisting of cardiac sensors, photoplethysmogram sensors, piezoelectric sensors, transcranial probes, transcranial doppler ultrasound sensors, electrocardiogram sensors, cardiac function sensors, cardiac cycle sensors, blood volume sensors, blood gas sensors, electroencephalography sensors, photoplethysmogram sensors, near infrared spectroscopygram sensors, rheoencephalography sensors, orbital sonography sensors, photoplethysmogram sensors, respiratory sensors, and combinations thereof.

[0054] The method for assessment of cerebrospinal fluid dynamics may include selecting the one or more features from the group consisting of a heart rate, a mean respiratory rate, a heart rate variability, a second derivative of photoplethysmogram signals, a photoplethysmogram intensity ratio, a respiration rate, a mean of the respiration rate, a standard deviation of the respiration rate, an area under a respiration curve, a maximum inhalation slope, a maximum inhalation slope, a blood volume, a partial pressure of oxygen, a mean R wave peak-to-peak distance, a heart rate variability, a root mean square of successive differences between normal heartbeats, a percentage of successive RR intervals that differ by more than 50 ms, a low frequency (LF) power, a high frequency (HF) power, a ratio between low frequency to high frequency power, a normalized low frequency power, aPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT normalized high frequency power, a root mean square of successive differences (RMSSD), a percentage of adjacent NN intervals that vary from each other by more than 50 ms (PNN50), a very low frequency (VLF) power, an LF / HF ratio, a normalized LF power, a normalized HF power, a pulse transit time, a pulse intensity ratio, a photoplethysmogram heart rate variability, a percent dicrotic, a respiration rate, a respiration standard deviation, an inhale slope, an exhale slope, an area on a curve, a permutation entropy, an SVD entropy, a spectral entropy, an approximate entropy, a sample entropy, a Hjorth mobility, a Hjorth complexity, a Katz fractal dimension, a detrended fluctuation, a Higuchi fractal dimension, a Petrosian fractal dimension, a Hurst exponent, an arterial tone, a mutual information, a maximum transcranial doppler (TCD) velocity, a minimum TCD velocity, a mean TCD velocity, a pulsatility index, a resistance index, a systolic acceleration, a duration of cycle, a number of splines, a length of cycle, a diastolic deceleration, a number of maxima and minima, a delta wave power, K-complex waveforms, sleep spindles, saw-tooth waveforms, a degree of centrality, and combinations thereof. The one or more features may be calculated from a Sylvian aqueduct. The one or more features may be calculated from sleep stages. The one or more features may include one or more graph features selected from the group consisting of a centrality, a page rank, a clustering coefficient, a global efficiency, a number of partitions, a shortest path length, an in degree, an out degree, and combinations thereof.

[0055] The machine learning algorithm of the method for assessment of cerebrospinal fluid dynamics may include a neural network trained to learn a relationship between the one or more physiological signals, the one or more features, and the one or more aspects of cerebrospinal fluid dynamics.

[0056] In yet another aspect, embodiments disclosed herein relate to a method for assessment of glymphatic fluid dynamics in a human subject including non- invasively recording, using one or more physiological sensors, one or more physiological signals from a human subject; extracting, using one or more microprocessors in communication with the one or more physiological sensors, onePATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT or more features from the one or more physiological signals; estimating, with a machine learning algorithm, one or more aspects of glymphatic fluid dynamics from one or both of the one or more physiological signals and the one or more features; and outputting, using an output device in communication with the one or more microprocessors, the one or more aspects of glymphatic fluid dynamics.

[0057] The method for assessment of glymphatic fluid dynamics may include: diagnosing, using the one or more microprocessors and the one or more aspects of glymphatic fluid dynamics, a neurological condition, and displaying, using the output device, a diagnosis. The method may include selecting the diagnosis from the group consisting of hydrocephalus, traumatic brain injury, Parkinson’s, Alzheimer’s, epilepsy, an impaired shunt used in a treatment of hydrocephalus, and combinations thereof.

[0058] The method for assessment of glymphatic fluid dynamics may include continuously recording the one or more physiological signals. The method may include intermittently recording the one or more physiological signals.

[0059] The method for assessment of glymphatic fluid dynamics may include displaying the one or more aspects of glymphatic fluid dynamics on the output device, where the output device is a display device. The output device may be in wireless communication with the one or more microprocessors.

[0060] The method for assessment of glymphatic fluid dynamics may include combining the one or more physiological sensors and the one or more microprocessors may be combined in a wearable device. The wearable device may include or be a headgear. The wearable device may include or be a chest patch. The wearable device may be in wireless communication with the output device.

[0061] The method for assessment of glymphatic fluid dynamics may include selecting the one or more physiological signals from the group consisting of cardiac sensors, photoplethysmogram sensors, piezoelectric sensors, transcranial probes, transcranial doppler ultrasound sensors, electrocardiogram sensors, cardiac function sensors, cardiac cycle sensors, blood volume sensors, blood gas sensors,PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT electroencephalography sensors, photoplethysmogram sensors, near infrared spectroscopygram sensors, rheoencephalography sensors, orbital sonography sensors, photoplethysmogram sensors, respiratory sensors, and combinations thereof.

[0062] The method for assessment of glymphatic fluid dynamics may include selecting the one or more features from the group consisting of a heart rate, a mean respiratory rate, a heart rate variability, a second derivative of photoplethysmogram signals, a photoplethysmogram intensity ratio, a respiration rate, a mean of the respiration rate, a standard deviation of the respiration rate, an area under a respiration curve, a maximum inhalation slope, a maximum inhalation slope, a blood volume, a partial pressure of oxygen, a mean R wave peak-to-peak distance, a heart rate variability, a root mean square of successive differences between normal heartbeats, a percentage of successive RR intervals that differ by more than 50 ms, a low frequency (LF) power, a high frequency (HF) power, a ratio between low frequency to high frequency power, a normalized low frequency power, a normalized high frequency power, a root mean square of successive differences (RMSSD), a percentage of adjacent NN intervals that vary from each other by more than 50 ms (PNN50), a very low frequency (VLF) power, an LF / HF ratio, a normalized LF power, a normalized HF power, a pulse transit time, a pulse intensity ratio, a photoplethysmogram heart rate variability, a percent dicrotic, a respiration rate, a respiration standard deviation, an inhale slope, an exhale slope, an area on a curve, a permutation entropy, an SVD entropy, a spectral entropy, an approximate entropy, a sample entropy, a Hjorth mobility, a Hjorth complexity, a Katz fractal dimension, a detrended fluctuation, a Higuchi fractal dimension, a Petrosian fractal dimension, a Hurst exponent, an arterial tone, a mutual information, a maximum transcranial doppler (TCD) velocity, a minimum TCD velocity, a mean TCD velocity, a pulsatility index, a resistance index, a systolic acceleration, a duration of cycle, a number of splines, a length of cycle, a diastolic deceleration, a number of maxima and minima, a delta wave power, K-complex waveforms, sleep spindles, saw-tooth waveforms, a degree of centrality, and combinations thereof. The one orPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT more features may be calculated from a Sylvian aqueduct. The one or more features may be calculated from sleep stages. The one or more features may include one or more graph features selected from the group consisting of a centrality, a page rank, a clustering coefficient, a global efficiency, a number of partitions, a shortest path length, an in degree, an out degree, and combinations thereof.

[0063] The machine learning algorithm of the method for assessment of glymphatic fluid dynamics may include a neural network trained to learn a relationship between the one or more physiological signals, the one or more features, and the one or more aspects of cerebrospinal fluid dynamics.

[0064] In still another aspect, embodiments disclosed herein relate to a system for assessment of glymphatic fluid dynamics in a human subject, the system including: one or more physiological sensors configured to non-invasively record one or more physiological signals from the human subject; one or more microprocessors in communication with the one or more physiological sensors; an output device in communication with the one or more microprocessors. The one or more microprocessors may be configured to extract one or more features from the one or more physiological signals, and implement a machine learning algorithm, wherein the machine learning algorithm is configured to estimate one or more aspects of glymphatic fluid dynamics from one or both of the one or more physiological signals and the one or more features. The output device may be configured to output the one or more aspects of glymphatic fluid dynamics.

[0065] The one or more microprocessors of the system for assessment of glymphatic fluid dynamics may be configured to use the one or more aspects of glymphatic fluid dynamics to perform a diagnosis of a neurological condition, and the output device is configured to display the diagnosis. The diagnosis may be selected from the group consisting of hydrocephalus, traumatic brain injury, Parkinson’s, Alzheimer’s, epilepsy, an impaired shunt used in a treatment of hydrocephalus, and combinations thereof.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0066] The one or more physiological sensors of the system for assessment of glymphatic fluid dynamics may be configured to continuously record the one or more physiological signals. The one or more physiological sensors may be configured to intermittently record the one or more physiological signals.

[0067] The output device of the system for assessment of glymphatic fluid dynamics may be in wireless communication with the one or more microprocessors. The output device may be a display device configured to display the one or more aspects of glymphatic fluid dynamics.

[0068] The one or more physiological sensors of the system for assessment of glymphatic fluid dynamics and the one or more microprocessors of the system for assessment of glymphatic fluid dynamics may be combined in a wearable device. The wearable device may include or be a headgear. The wearable device may include or be a chest patch. The wearable device may include or be a piezoelectric belt. The wearable device may be in wireless communication with the output device.

[0069] The one or more physiological signals of the system for assessment of glymphatic fluid dynamics may be selected from the group consisting of cardiac sensors, photoplethysmogram sensors, piezoelectric sensors, transcranial probes, transcranial doppler ultrasound sensors, electrocardiogram sensors, cardiac function sensors, cardiac cycle sensors, blood volume sensors, blood gas sensors, electroencephalography sensors, photoplethysmogram sensors, near infrared spectroscopygram sensors, rheoencephalography sensors, orbital sonography sensors, photoplethysmogram sensors, respiratory sensors, and combinations thereof.

[0070] The one or more features of the system for assessment of glymphatic fluid dynamics may be selected from the group consisting of a heart rate, a mean respiratory rate, a heart rate variability, a second derivative of photoplethysmogram signals, a photoplethysmogram intensity ratio, a respiration rate, a mean of the respiration rate, a standard deviation of the respiration rate, an area under a respiration curve, a maximum inhalation slope, a maximum inhalation slope, a bloodPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT volume, a partial pressure of oxygen, a mean R wave peak-to-peak distance, a heart rate variability, a root mean square of successive differences between normal heartbeats, a percentage of successive RR intervals that differ by more than 50 ms, a low frequency (LF) power, a high frequency (HF) power, a ratio between low frequency to high frequency power, a normalized low frequency power, a normalized high frequency power, a root mean square of successive differences (RMSSD), a percentage of adjacent NN intervals that vary from each other by more than 50 ms (PNN50), a very low frequency (VLF) power, an LF / HF ratio, a normalized LF power, a normalized HF power, a pulse transit time, a pulse intensity ratio, a photoplethysmogram heart rate variability, a percent dicrotic, a respiration rate, a respiration standard deviation, an inhale slope, an exhale slope, an area on a curve, a permutation entropy, an SVD entropy, a spectral entropy, an approximate entropy, a sample entropy, a Hjorth mobility, a Hjorth complexity, a Katz fractal dimension, a detrended fluctuation, a Higuchi fractal dimension, a Petrosian fractal dimension, a Hurst exponent, an arterial tone, a mutual information, a maximum transcranial doppler (TCD) velocity, a minimum TCD velocity, a mean TCD velocity, a pulsatility index, a resistance index, a systolic acceleration, a duration of cycle, a number of splines, a length of cycle, a diastolic deceleration, a number of maxima and minima, a delta wave power, K-complex waveforms, sleep spindles, saw-tooth waveforms, a degree of centrality, and combinations thereof. The one or more features may be calculated from a Sylvian aqueduct. The one or more features may be calculated from sleep stages. The one or more features may include one or more graph features selected from the group consisting of a centrality, a page rank, a clustering coefficient, a global efficiency, a number of partitions, a shortest path length, an in degree, an out degree, and combinations thereof.

[0071] The machine learning algorithm of the system for assessment of glymphatic fluid dynamics may include a neural network trained to learn a relationship between the one or more physiological signals, the one or more features, and the one or more aspects of cerebrospinal fluid dynamics.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0072] Other aspects and advantages of the claimed subject matter will be apparent from the following description and the appended claims. DETAILED DESCRIPTION

[0073] In the following detailed description of embodiments of the disclosure, numerous specific details are set forth in order to provide a more thorough understanding of the disclosure. However, it will be apparent to one of ordinary skill in the art that the disclosure may be practiced without these specific details. In other instances, well-known features have not been described in detail to avoid unnecessarily complicating the description.

[0074] Throughout the application, ordinal numbers (e.g., first, second, third, etc.) may be used as an adjective for an element (i.e., any noun in the application). The use of ordinal numbers is not to imply or create any particular ordering of the elements nor to limit any element to being only a single element unless expressly disclosed, such as using the terms “before,” “after,” “single,” and other such terminology. Rather, the use of ordinal numbers is to distinguish between the elements. By way of an example, a first element is distinct from a second element, and the first element may encompass more than one element and succeed (or precede) the second element in an ordering of elements.

[0075] In the following description of FIGs. 1-30, any component described with regard to a figure, in various embodiments disclosed herein, may be equivalent to one or more like-named components described with regard to any other figure. For brevity, descriptions of these components will not be repeated with regard to each figure. Thus, each and every embodiment of the components of each figure is incorporated by reference and assumed to be optionally present within every other figure having one or more like-named components. Additionally, in accordance with various embodiments disclosed herein, any description of the components of a figure is to be interpreted as an optional embodiment which may be implemented in addition to, in conjunction with, or in place of the embodiments described with regard to a corresponding like-named component in any other figure.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0076] It is to be understood that the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise. Thus, for example, reference to “chest patch” includes reference to one or more of such chest patches.

[0077] Terms such as “approximately,” “substantially,” etc., mean that the recited characteristic, parameter, or value need not be achieved exactly, but that deviations or variations, including for example, tolerances, measurement error, measurement accuracy limitations and other factors known to those of skill in the art, may occur in amounts that do not preclude the effect the characteristic was intended to provide.

[0078] It is to be understood that one or more of the steps shown in the flowcharts may be omitted, repeated, and / or performed in a different order than the order shown. Accordingly, the scope disclosed herein should not be considered limited to the specific arrangement of steps shown in the flowcharts.

[0079] Although multiple dependent claims are not introduced, it would be apparent to one of ordinary skill that the subject matter of the dependent claims of one or more embodiments may be combined with other dependent claims.

[0080] A method and system are presented that consist of wearable technologies and computer algorithms that allow for convenient, economical, comfortable, and non-invasive monitoring of cerebrospinal (CSF) dynamics and diagnosing pathophysiological changes. The computer algorithms may include conventional signal processing methods as well as modern machine learning methods. The system may include a chest patch, physiological sensors, signal processing algorithms to extract information from signals received by the physiological sensors, and machine learning algorithms for predicting the behavior of CSF. The method and system may be used to diagnose pathophysiologies including hydrocephalus, traumatic brain injury, Parkinson’s, Alzheimer’s, epilepsy, or an impaired shunt used in a treatment of hydrocephalus.

[0081] The physiological signals may be selected from the group consisting of cardiac sensors, photoplethysmogram sensors, piezoelectric sensors, transcranialPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT probes, transcranial doppler ultrasound sensors, electrocardiogram sensors, cardiac function sensors, cardiac cycle sensors, blood volume sensors, blood gas sensors, electroencephalography sensors, near infrared (NIR) sensors, spectroscopygram sensors, rheocephalography sensors, orbital sonography sensors, respiratory sensors, and combinations thereof.

[0082] As shown in FIG. 1, a chest patch (102) may be used to adhere to a patient’s skin and receive and process input from one or more wearable physiological sensors. The physiological sensors may be placed on the patient’s body for the continuous or intermittent monitoring of one or more physiological parameters, including cardiac cycle, blood volume, pO2, and respiration. A photoplethysmogram (PPG) may be measured, in particular, with the PPG sensor (104). Signal processing algorithms are used to extract information from the physiological sensors that are recording continuously or intermittently. For the signal processing algorithms, a wide array of techniques from time series analysis, frequency domain analysis, information theory, and graph theory are used to discover a novel subset of biomarkers that help estimate CSF flow non-invasively. A machine learning prediction algorithms estimate the dynamics of a patient’s CSF, including, but not limited to, the CSF flow, velocity, and time to peak / trough.

[0083] As used herein, measures of CSF flow and its changes over time are considered different aspects of CSF dynamics, including, but not limited to, its volume and changes in volume (accumulation / depletion), its bulk movement and changes in bulk movement, its flow and changes in flow, its velocity and changes in velocity, its time and speed to peak, its time and speed to trough, the area and changes in area through which the CSF movement occurs, and the combination of any two or more of these aspects. Typical average CSF flow rates are between 0.1- 0.2 cm / s when not in the range of suspected pathologies. Glymphatic flow dynamics may represent CSF flow dynamics in the perivascular space.

[0084] MRI may be used to directly non-invasively measure different aspects of the CSF dynamics in different areas of the brain. Certain neurological disordersPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT are associated with abnormal dynamics of CSF. For example, hydrocephalus is demonstrated by abnormal accumulation of CSF in brain ventricles.

[0085] Data related to CSF and glymphatic flow may be obtained from one or more physiological sensors and one or more microprocessors combined in a wearable device. The wearable device may include headgear, a chest patch, or other devices attachable to the human body. These devices may be in wireless communication with an output device that can display the data and information obtained from the data, such as features and diagnoses.

[0086] As an example of recording data related to CSF dynamics, three physiological signals may be recorded simultaneously during a sleep study: cardiac cycle by electrocardiogram (ECG), blood volume and pO2 by PPG, and respiration by piezoelectric belt (106). The ECG is a non-invasive tool for assessing the health of the heart, and is used as a diagnostic tool in hospital and clinical settings. It is obtained by placing skin electrodes on different locations of the torso in order to measure global electrical cardiac signals produced by the heart. In this example, only one channel is used to record an ECG signal. The ECG allows the recording of heart rate, mean R wave peak-to-peak distance (“RR interval”), heart rate variability, the root mean square of successive differences between normal heartbeats, percentage of successive RR intervals that differ by more than 50ms, low frequency (LF) and high frequency (HF) power, the ratio of low frequency to high frequency power, and the normalized low and high frequency power. Properties and statistics that may be calculated from a signal will be referred to as mathematical features, or simply “features” throughout this document.

[0087] The PPG sensor (104) measures by a optical technique that detects fluctuations in blood volume. Either red or infra-red light is illuminated on the skin of a subject and the reflected light is then measured using a photodetector. This reflected light is proportional to changes in blood volume via the Beer-Lambert law. The use of this technology has increased in popularity due to its simplicity, non- invasiveness, and comfort when placed on a subject. PPG signals allow for thePATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT calculation of the pulse intensity ratio, heart rate variability, percent of dicrotic waves, and the length of the PPG waveform. The total number of features for the PPG feature set may be from 0 to 1000, for example, 80.

[0088] The respiration signal from the piezoelectric belt (106) detects the breathing pattern of the subject over time. Respiration is known to modulate and affect the heart rate, which in turns affects the partial pressure of oxygen (pO2), and both respiration and cardiac cycle are known to affect CSF flow. Thus, from the respiration waveform the mean and standard deviation of the respiration rate are calculated, along with the maximum inhalation slope, maximum exhalation slope, and the area under the respiration curve. The total number respiration features obtained from the respiration waveform may be from 0 to 1000, for example, 85.

[0089] The most common and effective treatment of patients diagnosed with hydrocephalus involves the surgical insertion of a drainage system called a shunt. This consists of a long, flexible tube with a valve that keeps CSF from the brain flowing in the right direction and at the proper rate. The shunt is placed in the ventricles, the tube is tunneled under the patient’s skin to another part of the body, usually the abdomen, where it can be easily absorbed. While effective, this treatment suffers from complications. Shunt systems often stop draining the CSF or poorly regulate the drainage because of mechanical problems, infections, or blockage. There is currently no effective way to determine whether an implanted shunt continues to be fully operational, and often shunt occlusions are detected only when patients manifest overt neurological symptoms and are hospitalized. The treatment, in this case, is usually another surgery to replace the defective shunt.

[0090] FIG.2 therefore shows an embodiment of the system that may diagnose the occlusion of shunts implanted to treat hydrocephalus before any overt neurological symptoms are displayed. The wearable chest patch (202) may be equipped with wireless communication; it may send signals to an external device (208) that has ability to perform signal processing / machine learning and output estimates and / or diagnosis.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0091] To begin the processing of signals received by the systems described above, the ECG, PPG, and respiration signals may be first low-pass filtered with a 6th order Butterworth filter with a cutoff frequency to eliminate powerline, motion, and electromyographic noise, and then sampled at a sampling frequency. The cutoff frequency may be from 40-60 Hz, for example, 50 Hz. The sampling frequency may be from 80 to 250 Hz, for example, 200 Hz. Normal filtering may be done causally (in the forward direction in time), thus creating a time delay and a change in the phase of the signal. To compensate for these changes, the signal may be subsequently filtered backwards in time (anti-causally). The combined effect of symmetric causal and anti-causal filters is a zero-phase filter, which is beneficial for preserving information in time without introducing distortions. Signals may be resolved at frequencies up to a resolution frequency resultant from the Nyquist theorem. The resolution frequency may be from 80-120 Hz, for example, 100 Hz.

[0092] Following the filtering with the Butterworth filter, the data are segmented into 30 second non overlapping blocks aligned with labels. (For instance, during sleep, a somnologist may label each block of data.) Each overlapping block may have duration from 0 to 60 seconds, for example, 30 seconds. Segmented blocks associated with the same sleep stage are grouped together for feature extraction. This allowed analysis of how the performance of a regressor changes with different sleep stages. From the segmented and grouped blocks of data, feature extraction is performed. Some of the extracted features for each data modality (e.g., ECG, PPG, respiration) are listed in Table 1.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT Table 1

[0093] The ECG, PPG and respiration are all time varying waveforms and are first analyzed using waveform features calculated in the time domain. For the ECG, heart rate (HR) of a subject is measured by counting the number of R waves that appear in a set amount of time. R waves are a feature of the ECG signals that denote the peak depolarization of the heart ventricle. Also measured is the time between R waves to obtain and measure of the heart rate variability (HRV), an important metric shown to be useful in tracking cognitive decline. It is widely believed that pulsations of the heart and vascular system are the primary cause of movement of the CSF even at the arteriolar level. HRV can also be thought of as a surrogate for general vascular control which is extremely important for studying CSF flow. Due to the importance of HRV, features related to HRV are sought that might be useful for CSF flow estimation. These features are calculated per every 30-second epoch; only features that are capable of estimation in that span of time are used.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0094] When analyzing signals, the root mean square of successive differences between normal heartbeats (RMSSD) may be calculated. This feature is calculated by first finding all RR intervals, then squaring them, then summing, and finally taking the square root. Next these intervals are squared, then summed, and finally square rooted. This procedure is detailed in the following equation

[0095] where N is the number of heartbeats and RR is the calculated RR interval between successive heartbeats. The HRV, the frequency spectrum of the PPG, the second derivative of the PPG signal, and lastly the PPG intensity ratio (PIR) may be investigated in the context of CSF flow. The HRV can be calculated from the PPG by tracking the peak to peak distance over time, as in the case of the ECG. The PIR is the ratio of PPG peak intensity to PPG valley intensity of one cardiac cycle. This gives an indirect measure of arterial diameter change. It is known that these metrics correlate to CSF flow throughout the brain to varying degrees when measured by fMRI, with the PIR having the greatest correlation value across the regions. The percent of dicrotic notches per epoch may also be calculated. Dicrotic notch is a small downward deflection in the arterial pulse that separates the systolic and diastolic phase in the PPG; it exists as a consequence of the closing of the aortic valve. The appearance of this notch in the PPG signal changes with arterial stiffness and aging, both of which have a relationship with CSF flow. Dicrotic notches were detected using simple peak detection between the systolic peak and the end of the beat. The number of detected notches is then divided by the total number of heartbeats in the epoch. Lastly, the average length of each PPG heartbeat is calculated with the following equationPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0097] where a and b are the endpoints of the PPG heartbeat cycle and is the derivative of the PPG waveform. To calculate the derivative, a cubic spline is first fit to the PPG waveform and then used to calculate a derivative. The length is used as a measure of complexity in the PPG signal. The idea behind using this metric is from the visual observation that the morphology of the PPG varies between patients and over time.

[0098] It has been shown that respiration plays a role in the influx of CSF into the periarterial space. Methods for measuring respiration rate are grouped by the way the respiratory rate is measured. However, a problem is that during most MRI acquisitions to measure CSF flow directly, the acquisition process is not gated to the respiratory rate. This means that the effect of respiration on CSF flow can be averaged out, and it was shown that when a respiratory gated MRI was used, the net CSF flow quantity through the aqueduct was affected by breathing in healthy subjects. Thus, the present method may include analyzing the respiration signal more extensively for CSF flow estimation. The present method may include calculating the maximum exhalation and inhalation slope. This metric gives a measure of how deeply a subject inhales or exhales during respiration. The present method may include calculating the mean area under the respiration curve to give a metric proportional to the total volume of air inhaled. The present inventors believe that these metrics are important descriptors of breathing and thus are relevant features.

[0099] After analyzing signals in the time domain, the power spectrum of the ECG signal may be analyzed. Low frequency (LF) band (which may be, for instance, 0.04–0.15 Hz) and high frequency (HF) band (which may be, for instance, 0.15–0.4 Hz) of the ECG may be used for studying HR oscillations. The LF and HF bands are associated with heart beat rhythms affected by respiration. In addition, the ratio of the power in these bands may estimate the ratio between sympathetic nervous system and parasympathetic nervous system activity under controlled conditions.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0100] In addition to linear features and power spectral analysis, nonlinear features in the ECG, PPG, and respiration time series may be examined. A metric to do this analysis is the entropy. Entropy is a measure of how many distinct states a system can take, with the understanding that the more states a system can take, the more complexity it contains. There are several different entropies calculated in the literature that use different mathematical derivations to approximate a version of entropy. Examples include permutation entropy, singular value decomposition (SVD) entropy, spectral entropy, approximate entropy, and sample entropy. Entropy is a measure of how many distinct states a system can take, with the understanding that the more states a system can take, the more complexity it contains. In practical terms, it is usually not possible to directly measure entropy as one may not observe every possible state a system takes. As a result, there are several different entropies calculated in the literature that use different mathematical derivations to approximate a version of entropy. In this work, permutation entropy, singular value decomposition (SVD) entropy, spectral entropy, approximate entropy, and sample entropy are used.

[0101] Fractal dimension is a metric used to capture the underlying dimensionality of a self-similar object. This concept may be extended to the analysis of time series data by treating these waveforms as a geometric object. The fractal dimension of a time series may lie between 1 and 1.5. The fractal dimension of biosignals may be capable of distinguishing different brain states characterized by an increase in waveform complexity. Known computational implementations to calculate the fractal dimension include the Petrosian, Katz, and Higuchi fractal dimension algorithms. One or more of these algorithms may be used. Fractal dimension is a metric used to capture the underlying dimension of a self-similar object. This concept has been extended to the analysis of time series data by treating these waveforms as a geometric object. Since waveforms are somewhere between simple lines and a plane, the fractal dimension of any time series should lie between 1 and 1.5. It has been shown in the literature that the fractal dimension calculation, when applied to biosignals, is capable of distinguishing different brain statesPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT characterized by an increase in waveform complexity, such as during a seizure. There are three common computational implementations to calculate the fractal dimension: the Petrosian, Katz, and Higuchi fractal dimension algorithms.

[0102] Hjorth parameters are time domain features that may be used for analyzing EEG time series signals. Hjorth mobility and complexity may be calculated on the present signals. The Hjorth mobility is the square root of the ratio of the variance of the first derivative of the signal to the variance of the signal itself. This metric can be interpreted as the variance of the frequency power spectrum of the signal. The Hjorth complexity is dimensionless and derived as the ratio between the mobility of the first derivative of the signal and the mobility of the signal itself. The Hjorth mobility is the square root of the ratio of the variance of the first derivative of the EEG to the variance of the EEG itself. This metric can be interpreted as the variance of the frequency power spectrum of the EEG. The Hjorth complexity is dimensionless and derived as the ratio between the mobility of the first derivative of the EEG and the mobility of the EEG itself.

[0103] The Hurst exponent is a measure of the autocorrelation of a time series, specifically it quantifies the relative tendency of a time series either to regress strongly to the mean or to cluster in a direction. The Hurst exponent takes value between 0 and 1. A value of 0.5 means the data looks like a random walk. A value less than 0.5 means the time series is persistent to trending in one direction. If the time series takes a low value, then one would expect at the next time instance the signal will take a high value. If the Hurst exponent is greater than 0.5 then a persistence of the trend is expected. The Hurst exponent is a measure of the autocorrelation of a time series, specifically it quantifies the relative tendency of a time series either to regress strongly to the mean or to cluster in a direction.

[0104] The detrended fluctuation parameter is similar to the Hurst exponent as it also measures long term dependence of a varying time series. The main difference is that the calculation does not assume stationarity of the signal and is thus more general and can capture more complex behavior.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0105] When used together, the features from Table 1 may achieve the objective to accurately and robustly predict CSF dynamics. Combining signals from different sources to boost their predictive power is a goal of multimodal machine learning. The act of combining data from different sensors is known as “data fusion.” Thus, in addition to the individual feature sets obtained from each physiological sensors as described above, all extracted features from all modalities may be concatenated into one feature set, and then used as input to various machine learning models.

[0106] A process of training a machine learning algorithm is illustrated in FIG. 3. The machine learning algorithm is trained to estimate and CSF flow and velocity from all the non-invasive modalities. Reconstructed CSF flow curves and CSF velocity curves may be analyzed for abnormalities and used for diagnoses. Box (301) represents three modalities (ECG, PPG, and respiration), i.e., a subset of the maximum possible number of features that might be used in estimation (i.e., prediction). In Box (302), these data sets are preprocessed to removed outliers and noise. In Box (304), the concatenated feature set is created. Box (306) populates the edges (304) and nodes (302) of a machine learning model with the appropriate values and begins the iterative process of minimizing the error between predicted (^^) and real (^) CSF flow, shown in Box (308). The ground truth CSF flow data comes from MRI data, and is shown in Box (310).

[0107] The schematic in FIG. 3 may be framed as a supervised regression problem since CSF flow is a continuous variable, rather than a ordinal or categorical variable. The input are features from the modalities and the label is CSF flow measured from the ground truth of MRI. The goal is to learn a model predicts CSF flow by minimizing the error between the estimated CSF flow (^^) and the true CSF flow (^).

[0108] Another application of the methods above is to the effective clearance of metabolites and waste products out of the brain; this is necessary to maintain proper brain physiology and function. Such clearance is driven by the glymphatic flow, which consists in the bulk movement of cerebrospinal fluid (CSF) from thePATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT subarachnoid space, along peri-arterial spaces, into the interstitial space surrounding neurons and glial cells in the neuroparenchyma, and from there to the peri-venous space and back into the blood stream. Thus, physiological changes in brain activity and vascular tone will affect the clearance of metabolic waste. For example, cerebral vasodilation as it occurs during robust synchronous neuronal activation during wake activity, is expected to decrease perivascular space and therefore decrease glymphatic flow, while phasic vascular pulsatility during sleep will promote it. Also, pathophysiological changes to either the structure and / or function of the vasculature and / or of the neuroparenchyma are expected to decrease the glymphatic flow and thus clearance of metabolites. Many types of injuries to the brain, such as traumatic or hypoxic injuries, are known to cause variably persistent impairment in the physiology of the brain parenchyma and the connected vasculature, which are expected to result in impaired CSF and glymphatic flows, thus in the accumulation of waste metabolites. This accumulation in turn will compound the pathophysiology of the injury and results in further brain dysfunction, cognitive, behavioral, and sleep impairment.

[0109] To best diagnose and treat the underlying medical problem, it is thus desirable to first assess the dysfunction in CSF and glymphatic flows. However, this is not easily achieved. A method to directly and non-invasively measure CSF and glymphatic flows is based on high resolution magnetic resonance imaging.

[0110] MRI allows for the fine assessment of both anatomical and functional features, and thus measures of velocity and bulk flow of CSF and glymphatic flow. While this technique is non-invasive, it is expensive, time consuming, and non- portable because it requires large magnets and acquisition systems. Exemplary MRI measurements are shown in FIG.4, FIG.5, and FIG.6 discussed further below.

[0111] However, tone and function of the arterial vasculature of the brain may be assessed non-invasively by transcranial doppler ultrasound which measures changes in the velocity of blood flow in cerebral arteries such as the middle cerebralPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT artery. Exemplary transcranial doppler ultrasound measurements are shown in FIG. 7 discussed further below.

[0112] This non-invasive technique is well established and relatively portable, depending on the size of the ultrasound probes. Similarly, scalp electroencephalography (EEG) is a well-established non-invasive electrophysiological method to measure brain activity and thus the function of the brain parenchyma. This technique is commonly used for the assessment of brain physiology and pathophysiology, such as seizure or sleep disorders, and relies on electrodes placed on the scalp of the patient and prolonged monitoring by multiple electrodes mapping brain activity. Both techniques are reliably used for clinical diagnosis and to generate real-real time signals that can be relatively easily acquired, digitized and processed.

[0113] As described above, signal processing and machine learning are commonly used to extract meaningful information from physiological time series. First, the amount of useful information detectable in the signals is maximized by filtering to eliminate frequencies outside of the useful range and removing noise and its. Filtering is performed in the forward and reverse time direction to ensure phase distortions are not introduced. From the filtered signals, mathematical features are calculated in the time and frequency domain. Linear features include mean and standard deviation, while non-linear features include more complex statistical information of the time series such as entropy and mutual information.

[0114] A set of time series signals may be further converted into a graphical representation to leverage graph theory for signal information extraction. For example, a set of signals can be used to form a graph where each node is an electrode and edges are drawn between the nodes based on a given metric to determine how connected two nodes are. One example would be to use a simple calculation of correlation to determine how similar two signals are and use this calculation to define the weight of an edge between two nodes. Once the time series data are reformulated into a graphical representation, graph metrics, like centralities, may bePATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT leveraged to capture useful information from the original signals. Centralities are a method to calculate how important a node is in a graph, where the notion of importance is different depending on the specific centrality. The degree centrality would just be the sum of weighted edges into a node, where a higher degree centrality indicates that the node is well connected to the other nodes in the graph, and lower degree centrality indicates a node that is separated from the rest of the graph.

[0115] Once features have been extracted from the cleaned data, features may be systematically correlated from one physiological time series to another. This allows us to find relationships between different physiological time series, even if once believed to be independent of each other, and to robustly predict one physiological time series, from one or more of the others. This process requires the careful segmentation of the data into a training, validation, and test set. The training set is fed into a machine learning algorithm that attempts to learn the relationship between different physiological time series. The robustness of this learned relationship may be checked by examining the results of the validation and test set. If a machine learning model performs well on the test set then one may be confident that a true predictive model has been found. One physiological time series may then be replaced with its prediction based on the machine learning-based analysis of other correlated physiological time series. This process, is highly nontrivial especially when replacing the need for an expensive and non-portable MRI with an easily portable headgear with EEG and ultrasound capabilities.

[0116] The physiological relationship between glymphatic / CSF flows, neuronal activity and vascular tone suggests the functional state of CSF and glymphatic flow could be estimated non-invasively from the direct measurements of vascular tone and brain function, thus avoiding a need for their direct measurement by MRI. The present method may estimate the CSF / glymphatic flow from the independent measurements of non-invasive brain activity by scalp EEGPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT and / or arterial tone by transcranial doppler ultrasound, as well as portable and convenient device to deploy and use this technology.

[0117] The one or more features may include one or more graph features. The graph features may be selected from the group consisting of a centrality, a page rank, a clustering coefficient, a global efficiency, a number of partitions, a shortest path length, an in degree, an out degree, and combinations thereof. The graph features may correspond to sensors. FIG.8A shows a graph structure where the nodes of the graph correspond to electrodes.

[0118] In a basic form an application of the methods above to monitoring glymphatic flow and clearing waste products out of the brain includes headgear, mathematic features, and mathematical predictors.

[0119] The headgear may be a headband or helmet that is customized to prop on the patient’s scalp / hair and place one or more EEG electrodes and / or ultrasound probes in the required position to measure brain activity and cerebral blood flow, respectively (FIG. 8B). In panel (a) of FIG. 8B, general schematics are shown of possible positions of transcranial doppler (TCD) probes (800) for the detection of cerebral blood flow through the middle cerebral artery, and of EEG electrodes (802) for the detection of brain activity, for the estimate of CST / glymphatic flow. In panel (B), one of the embodiments is shown of the headgear (804) placing the EEG electrodes and the TCD probes for the estimate of CSF / glymphatic flow.

[0120] The mathematical features may be used to extract information and correlate the time series acquired by MRI, EEG and TCD. These are: MRI: flow and velocity of CSF / glymphatic (FIGs. 4 and 5, Table 2); EEG (Table 3); TCD (Table 4).

[0121] The mathematical predictors may be trained to estimate the CSF / glymphatic flows from the TCD and EEG data (anything from regressions to trained machine learning models). Table 2PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCTTable 3Table 4PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0122] Results of application of the methods above to this application are detailed in Example 1, below.

[0123] The list of modalities used in the methods in this disclosure is presented in Table 5. FIG. 9 depicts the framework for integration of different modalities leading to an estimate of the CSF flow. Table 5

[0124] FIG. 9 shows the new multimodal machine learning applied to physiological signals for CSF flow prediction. The model starts multiple independent data streams from multiple noninvasive modalities. Each modality mayPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT be processed for relevant features. Each modality may have a different number, N, of relevant features and each feature can have a different data type. For example, the TCD has image-based features of vasculature and time series features of hemodynamics. These features may be grouped into M distinct groups. Note that group features may be from different modalities. A choice of M=1 would imply a concatenation of all features. This grouping step may be seen as a simple form of feature fusion as well as an engineering necessity. Each modality may not have the same sampling rate or data type, meaning that concatenation is not always possible and synchronization can become an issue. Grouping similar time scale features may mitigate this issue. Next, the groups are fed into their own regressors for CSF flow prediction, giving M regressors. These regressors have not been specified as the optimal choice may only be determined from implementation and comparison. It is unlikely that non neural network based regressors like support vector regression will outperform bidirectional recurrent neural networks and transformers if an adequate amount of training data can be acquired. The grouping and regression stage are the feature fusion portion of the framework. All predictions are then concatenated into a prediction vector and are fed into a decision fusion block for final prediction. The optimal decision fusion algorithm may be an extended Kalman filter since this allows a nonlinear combination of the individual predictions and is designed for tracking a continuous valued signal over time. The final prediction is then subtracted with the ground truth value to obtain the error of the model. This model is then fed back into the regressors and decision fusion block for learning and refinement of model parameters. Note that one may design a framework where the error also feeds into the grouping step in order to optimize the grouping of features. However, this would likely force a user to use the same regressor for each group in order for proper optimization, thus hand specifying the groups may allow for greater flexibility of the regressors.

[0125] One may also feed the raw data from noninvasive modalities into a representation learning algorithm, like a transformer network, to learn new features from the raw data. This allows for one to discover new features that exist in eachPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT modality and across the modalities. Using an open BCI platform, ECG, EEG, and PPG may be measured simultaneously in various postures. Glymphatic flow has been shown to be significantly affected by posture, therefore analyzing the features of the three modalities in different postures can allow us to glean some knowledge about which features will be useful in further analysis. All features shown in Table 1 were extracted and used simple feature fusion to predict the posture of the patient. The EEG spectral features were most informative and gave perfect accuracy with a linear support vector machine.

[0126] Various of the measurements described herein may be related to others of the various measurements described herein. FIG. 10 shows exemplary correlations between a transcranial doppler, CSF flow from an MRI, and signals from a sleep dataset discussed further below.

[0127] In another application, the methods above are applied to non-invasive monitoring of brain activity with high spatial and temporal resolution; this is important for several areas in neuroscience such as brain disease diagnosis, brain function analysis, and the emerging brain-machine interface. During decades, several methods have been developed towards this goal, like fMRI, PET, MEG, and fNIRS. Among these methods, fNIRS stands out with multiple advantages, e.g., relatively high spatial temporal resolution, lower cost, portability and harmless non- ionizing radiation.

[0128] fNIRS utilizes measurements obtained from source-detection pairs on the surface to recover the information of underlying hemodynamics caused by brain activities. Based on the type of source light, fNIRS can be divided into continuous wave (CW), frequency domain (FD), and time domain (TD), among which TD fNIRS allows accurate quantification of hemodynamics and potential better spatial resolution due to the richest information in time-of-flight measurements TD fNIRS uses a short pulse of laser light and measures the arrival times of photons emerging from the tissue. Using an array of sources and detectors, this technology can collect different projections of light transport through different locations and depths atPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT wavelengths that correspond to oxygenated and deoxygenated hemoglobin. Using a forward model that mimics the physics of the light transport, one may obtain an accurate reconstruction of absorption changes in the two wavelengths, and thus an accurate hemoglobin dynamic indicating brain activities.

[0129] The fNIRS method usually uses light within the “optical window” of biological tissue which is around 700 - 1300 nm. In this range, there is relatively weak absorbance of photons by the primary chromophores in biological tissue (water, lipids, and hemoglobin), which allows photons to penetrate as deep as several centimeters. The reconstruction of hemodynamics relies on the varying extinction coefficients of the two types of hemoglobin, which requires at least two wavelength measurements.

[0130] Based on the requirements, a two-wavelength TD-fNIRS system with multiple source-detection pairs was designed. In the system paradigm, the SuperK FIANIUM (FIR-20) laser was used as the source. The fiber coupled output is split equally using a beam splitter into two light paths. Then two spectral filters (SF) were used to get two specific wavelengths at 740 nm and 850 nm. The wavelengths are chosen to achieve relatively small system error based on previous literatures. The two-wavelength light is separately directed by two sets of Galvo Mirrors (GM) and projected onto the specific position of the common end of the customized fiber bundle to turn on the corresponding source fiber iteratively. With high-speed Galvo mirrors, one may achieve approximately 50 Hz frame rate for two wavelength measurements. The measurements are then collected by detection fibers and then projected onto the SPAD (single photon avalanche diode) array (PF32, from Photon Force). And thus, the SPAD array can obtain time-of-flight information of separate pixels corresponding to specific detection fibers at different positions at the same time. The results of applying this system are detailed in Example 1, below.

[0131] Yet another application of the above methods was to designing a rheoencephalography (REG) sensor on a breadboard and performed circuit simulations to verify the working principle. The breadboard-version of the REGPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT sensor was characterized and then its efficiency was improved by measuring impedance at different frequencies. REG is one of the most notable electrophysiological techniques for the evaluation of CBF. An increase (or decrease) in CBF usually leads to a decrease (or increase) in CSF. Thus, REG serves as an indirect marker of the CSF flow. The conductivity of the brain changes during blood flow circulation, and REG measures such conductivity changes through two scalp electrodes. To measure the impedance, a current is sourced by one of the electrodes, and the voltage difference between the two electrodes is recorded. Then resistance can be calculated by the following equation: R = V / I. The baseline brain impedance varies from person to person but generally falls between 200 - ^^^^ȍ^^+RZHYHU^^WKH^ impedance change caused by the CBF flow falls in the range of 0.1 - ^^^^^ȍ^^ZKLFK^ is more than 2,000 times smaller than the baseline impedance, causing a significant challenge to be directly measured. A very high precision high speed ADC would be suitable for this application. All the sensors are integrated with open source hardware from OpenBCI, which does not have a very high speed and precision ADC. For this reason, to alleviate these issues an AFE was built to relax the ADC requirements. The AFE is built in such a way that it amplifies the low frequency changes caused by CBF while removing the baseline voltage to prevent amplifier saturation. A schematic of a REG is shown in FIG.11. The current source generates a 1mA current which is delivered to the electrodes attached to the skin of the patient. Due to the current passing through the resistance and the CBF related impedance oscillations, there may be a voltage induced across V+ and V-. This voltage is sensed by a differential amplifier which is then high pass filtered to remove any DC offset that is present between the two electrodes. This voltage is then passed through an envelope detector and a peak detector. The envelope detector tracks the amplitude changes of the voltage changes caused by CBF whereas the peak detector detects the baseline. The envelope detector output contains information about the baseline and oscillating resistance whereas the peak detector carries information only about the baseline impedance. By subtracting the output voltages from the two blocks, the resulting signal carries information only related to the oscillating resistance. ThePATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT resulting voltage is amplified and sent to the OpenBCI hardware for processing. In the post processing phase, the read voltage is divided by the overall gain of the entire AFE chain to be able to quantify the input voltage oscillations. Then, the resulting voltage is divided by the current from the current source in order to be able to quantify the impedance oscillations due to CBF.

[0132] Another application of the above methods is to restore glymphatic flow; the technique applies EMvelope stimulation, a minimally invasive DBS setup using temporally interfering GHz EM waves. At GHz frequencies, antenna arrays may be created at the scale of a few centimeters or less that may be endocranially implanted to enable longitudinal stimulation and circumvent signal attenuation due to the scalp and skull. Furthermore, owing to the small wavelength of GHz EM waves, both amplitudes and phases of the EM waves may be optimized to achieve high intensity and focal stimulation at targeted regions within the safety limit for exposure to EM waves. A simulation framework has been developed for investigating the propagation of GHz EM waves generated by line current antenna elements and the corresponding heat generated in the brain tissue and proposed two optimization flows to identify antenna current amplitudes and phases for either maximal intensity or maximal focality transmission of the interfering electric fields with EM waves safety constraint. A representative result of this study is that with two endocranially implanted arrays of size 4.2 cm × 4.7 cm each, one can achieve an intensity of 12 V / m with a focality of 3.6cm at a target deep in the brain tissue. In this proof-of- principle study cited below, the idea of EMvelope stimulation is shown to merit further investigation as it can be a minimally invasive way of stimulating deep brain targets and offers benefits not shared by prior methodologies of electrical or magnetic stimulation.

[0133] A common factor across all modalities described above is the optimization of stimulation dosimetry parameters, like amplitude, phase, and the location of one or more transducers, which controls the efficacy and precision of stimulation. Many methods tackle the optimization of continuous parameters likePATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT amplitude and phase effectively. However, the transducer location, a discrete parameter, is either selected in an ad-hoc manner or by assuming that the transducers lie on a specific grid and implicitly optimizing for locations by optimizing for the amplitude of stimulation. However, this approach cannot exploit the full range of possibility of location optimization to achieve focal and intense stimulation. Another approach used is to utilize methods like branch and bound (BnB) to find locations that satisfy a specific criterion, like number of electrodes, etc. However, for BnB to be solved in a reasonable time, the optimization criterion must be convex. Additionally, even for commercial non-convex MINLP software like BARON, the dimensions of the problem they can be practically handled is small.

[0134] A general framework may be developed for optimizing transducer amplitude, phase, and locations. No constraints are placed on the convexity / non- convexity or size of the problem (in this case the size is restricted by machine resources only). This process is divided into the following steps:

[0135] 1) First, known antenna impulse response values are used to approximate the underlying Maxwell’s equations using neural networks.

[0136] 2) Second, continuous optimization algorithms, like stochastic gradient descent, are used to optimize the transducer location on a continuous grid.

[0137] A key insight is that, once an accurate model is obtained that can provide the antenna impulse response at any antenna location, continuous optimization algorithms, like stochastic gradient descent, may be used to optimize the transducer location on a continuous grid. Even though the framework of EMvelope stimulation is used to show the efficacy of this method, the general principles remain the same and the framework is also extendable to other forms of brain stimulation.

[0138] As the term is used above and throughout this disclosure, signal processing (also known as “time series analysis,” when applied exclusively to signals recorded in time) may involve analyzing, modifying, and synthesizing signals. Signal processing may be performed in the original domain of data collection / generation. That is, if a signal is measured as a function of time, thePATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT processing of that signal may also be done by modifying the values of the signal in time. However, it has proved useful and illuminating to transform signals into other domains. For example, the Fourier transform moves a signal into a domain where each coefficient represents the strength of a sinusoidal signal of a given frequency within the original signal. Fourier transforms (and other signal transforms) are often created by the point-by-point multiplication of the original signal with another, canonical signal, and then summing the result to produce a coefficient. For certain classes of signals (such as sinusoids) the resulting set of coefficients has many useful properties, such as maintaining the total energy of the original signal. Operations upon the coefficients in a transformed domain will modify the original signal upon back-transformation. Modifying a signal in a transformed domain may be referred to as “filtering.” In the case of the Fourier transform, manipulating the coefficients in the transformed domain may result in a smoother or rougher signal upon back- transformation. Furthermore, certain classes of signal transformations may be accomplished by very fast algorithms. Some of these algorithms, such as the fast Fourier transform (FFT) and the wavelet transform, may be performed with a number of operations proportional to the number of data points in the original signal.

[0139] Other signal processing operations do not require transforming a signal to another domain. For instance, signals may be clipped so that they do not exceed or go below a certain value. Another operation is automatic gain control, which either increases or decreases the amplitude of a signal based on a root mean square value within a specified window. Stretching / squeezing and interpolating / extrapolating signals are common applications. Compressing signals is also a common task in signal processing. These examples do not limit the extent of signal processing; a person of ordinary skill in the art would understand the full breadth and depth of methods used in the field.

[0140] Machine learning may be thought of as a modern extension of classical signal processing techniques. Conversely, classical signal processing techniques may be defined as simple applications of machine learning. Indeed, the demarcationPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT line between the two disciplines is fuzzy, with the former associated more with fast algorithms for signal manipulation, while the latter associated with artificial intelligence applications. Some methods, however, clearly sit on the side of machine learning. For instance, FIG.12 shows a neural network, a common ML architecture for prediction / inference. At a high level, a neural network (1200) may be graphically depicted as comprising nodes (1202), shown here as circles, and edges (1204), shown here as directed lines connecting the circles. The nodes (1202) may be grouped to form layers, such as the four layers (1208, 1210, 1212, 1214) of nodes (1202) shown in FIG. 12. The nodes (1202) are grouped into columns for visualization of their organization. However, the grouping need not be as shown in FIG.12. The edges (1204) connect the nodes (1202). Edges (1204) may connect, or not connect, to any node(s) (1202) regardless of which layer (1205) the node(s) (1202) is in. That is, the nodes (1202) may be fully or sparsely connected. A neural network (1200) will have at least two layers, with the first layer (1208) considered as the “input layer” and the last layer (1214) as the “output layer.” Any intermediate layer, such as layers (1210) and (1212) is usually described as a “hidden layer”. A neural network (1200) may have zero or more hidden layers, e.g., hidden layers (1210) and (1212). However, a neural network (1200) with at least one hidden layer (1210, 1212) may be described as a “deep” neural network forming the basis of a “deep learning method.” In general, a neural network (1200) may have more than one node (1202) in the output layer (1214). In this case the neural network (1200) may be referred to as a “multi-target” or “multi-output” network.

[0141] Nodes (1202) and edges (1204) carry additional associations. Namely, every edge is associated with a numerical value. The numerical value of an edge, or even the edge (1204) itself, is often referred to as a “weight” or a “parameter”. While training a neural network (1200), numerical values are assigned to each edge (1204). Additionally, every node (1202) is associated with a numerical variable and an activation function. Activation functions are not limited to any functional class, but traditionally follow the form:PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0143] where ^ is an index that spans the set of “incoming” nodes (1202) and edges (1204) and ^ is a user-defined function. Incoming nodes (1202) are those that, when viewed as a graph (as in FIG.8A), have directed arrows that point to the node (1202) where the numerical value is computed. Functional forms of ^ may includethe linear function ^(^)= ^, sigmoid function ^(^)=^ష^, and rectified linear unit function ^(^)= max(0, ^), however, many additional functions are commonly employed in the art. Each node (1202) in a neural network (1200) may have a different associated activation function. Often, as a shorthand, activation functions are described by the function ^ by which it is composed. That is, an activation function composed of a linear function ^ may simply be referred to as a linear activation function without undue ambiguity.

[0144] Relating Eq.1 to signal processing, if the activation functions are linear, this reduces a layer (1214) of a neural network (1200) to the point-by-point multiplication of a signal (represented by node values) with another discrete function (represented by the edge values) followed by summation. This process is a fundamental part of many signal processing applications, such as, e.g., transforming the signal to the Fourier domain; in this case the latter function would be a sinusoid and the resulting coefficient would represent the strength of a certain frequency in the original signal. In this way, one may view the classical signal processing transformations as trivial versions of single-layer neural networks.

[0145] When the neural network (1200) receives an input, the input is propagated through the network according to the activation functions and incoming node (1202) values and edge (1204) values to compute a value for each node (1202). That is, the numerical value for each node (1202) may change for each received input. Occasionally, nodes (1202) are assigned fixed numerical values, such as the value of 1, that are not affected by the input or altered according to edge (1204) values and activation functions. Fixed nodes (1202) are often referred to as “biases” or “bias nodes” (1206), and are depicted in FIG.12 with a dashed circle.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0146] In some implementations, the neural network (1200) may contain specialized layers (1205), such as a normalization layer, or additional connection procedures, like concatenation. One skilled in the art will appreciate that these alterations do not exceed the scope of this disclosure.

[0147] As noted, the training procedure for the neural network (1200) comprises assigning values to the edges (1204). To begin training, the edges (1204) are assigned initial values. These values may be assigned randomly, assigned according to a prescribed distribution, assigned manually, or by some other assignment mechanism. Once edge (1204) values have been initialized, the neural network (1200) may act as a function, such that it may receive inputs and produce an output. As such, at least one input is propagated through the neural network (1200) to produce an output. Recall that a given data set will be composed of inputs and associated target(s), where the target(s) represent the “ground truth”, or the otherwise desired output. The neural network (1200) output is compared to the associated input data target(s). The comparison of the neural network (1200) output to the target(s) is typically performed by a so-called “loss function”; although other names for this comparison function such as “error function” and “cost function” are commonly employed. Many types of loss functions are available, such as the mean- squared-error function. However, the general characteristic of a loss function is that it provides a numerical evaluation of the similarity between the neural network (1200) output and the associated target(s). The loss function may also be constructed to impose additional constraints on the values assumed by the edges (1204), for example, by adding a penalty term, which may be physics-based, or a regularization term. Generally, the goal of a training procedure is to alter the edge (1204) values to promote similarity between the neural network (1200) output and associated target(s) over the data set. Thus, the loss function is used to guide changes made to the edge (1204) values, typically through a process called “backpropagation.”

[0148] The loss function will usually not be reduced to zero during training. And, once trained, it is not necessary or required that the neural network (1200)PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT exactly reproduce the output elements in the training data set when operating upon the corresponding input elements. Indeed, a neural network (1200) that exactly reproduces the output for its corresponding input may be perceived to be “fitting the noise.” In other words, it is often the case that there is noise in the training data, and a neural network (1200) that is able to reproduce every detail in the output is reproducing noise rather than true signal. The price to pay for using such a “perfect” neural network (1200) is that it will be limited to fitting only the training data and not able to generalize to produce a realistic output for a new and different input that has never been seen by it before. An analog of this problem occurs when fitting a polynomial to data points. The higher the degree of the polynomial, the closer the resulting curve will be to fitting all the points (a high enough polynomial is guaranteed to fit all the points). However, higher degree polynomials will tend to diverge quickly away from the fit data point values – hence, a high degree polynomial will not exhibit generalizability.

[0149] FIG. 13 further depicts a block diagram of a computer system (1302) used to provide computational functionalities associated with described algorithms, methods, functions, processes, flows, and procedures as described in this disclosure, according to one or more embodiments. The illustrated computer (1302) is intended to encompass any computing device such as a server, desktop computer, laptop / notebook computer, wireless data port, smart phone, personal data assistant (PDA), tablet computing device, one or more processors within these devices, or any other suitable processing device, including both physical or virtual instances (or both) of the computing device. Additionally, the computer (1302) may include an input device, such as a keypad, keyboard, touch screen, or other device that can accept user information, and an output device that conveys information associated with the operation of the computer (1302), including digital data, visual, or audio information (or a combination of information), or a GUI.

[0150] The computer (1302) can serve in a role as a client, network component, a server, a database or other persistency, or any other component (or a combinationPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT of roles) of a computer system for performing the subject matter described in the instant disclosure. The illustrated computer (1302) is communicably coupled with a network (1330). In some implementations, one or more components of the computer (1302) may be configured to operate within environments, including cloud- computing-based, local, global, or other environment (or a combination of environments).

[0151] At a high level, the computer (1302) is an electronic computing device operable to receive, transmit, process, store, or manage data and information associated with the described subject matter. According to some implementations, the computer (1302) may also include or be communicably coupled with an application server, e-mail server, web server, caching server, streaming data server, business intelligence (BI) server, or other server (or a combination of servers).

[0152] The computer (1302) can receive requests over network (1330) from a client application (for example, executing on another computer (1302)) and responding to the received requests by processing the said requests in an appropriate software application. In addition, requests may also be sent to the computer (1302) from internal users (for example, from a command console or by other appropriate access method), external or third-parties, other automated applications, as well as any other appropriate entities, individuals, systems, or computers.

[0153] Each of the components of the computer (1302) can communicate using a system bus (1303). In some implementations, any or all of the components of the computer (1302), both hardware or software (or a combination of hardware and software), may interface with each other or the interface (1304) (or a combination of both) over the system bus (1303) using an application programming interface (API) (1312) or a service layer (1313) (or a combination of the API (1312) and service layer (1313)). The API (1312) may include specifications for routines, data structures, and object classes. The API (1312) may be either computer-language independent or dependent and refer to a complete interface, a single function, or even a set of APIs. The service layer (1313) provides software services to thePATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT computer (1302) or other components (whether or not illustrated) that are communicably coupled to the computer (1302). The functionality of the computer (1302) may be accessible for all service consumers using this service layer. Software services, such as those provided by the service layer (1313), provide reusable, defined business functionalities through a defined interface. For example, the interface may be software written in JAVA, C++, or other suitable language providing data in extensible markup language (XML) format or another suitable format. While illustrated as an integrated component of the computer (1302), alternative implementations may illustrate the API (1312) or the service layer (1313) as stand-alone components in relation to other components of the computer (1302) or other components (whether or not illustrated) that are communicably coupled to the computer (1302). Moreover, any or all parts of the API (1312) or the service layer (1313) may be implemented as child or sub-modules of another software module, enterprise application, or hardware module without departing from the scope of this disclosure.

[0154] The computer (1302) includes an interface (1304). Although illustrated as a single interface (1304) in FIG. 13, two or more interfaces (1304) may be used according to particular needs, desires, or particular implementations of the computer (1302). The interface (1304) is used by the computer (1302) for communicating with other systems in a distributed environment that are connected to the network (1330). Generally, the interface (1304) includes logic encoded in software or hardware (or a combination of software and hardware) and operable to communicate with the network (1330). More specifically, the interface (1304) may include software supporting one or more communication protocols associated with communications such that the network (1330) or interface’s hardware is operable to communicate physical signals within and outside of the illustrated computer (1302).

[0155] The computer (1302) includes at least one computer processor (1305). Although illustrated as a single computer processor (1305) in FIG. 13, two or more processors may be used according to particular needs, desires, or particularPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT implementations of the computer (1302). Generally, the computer processor (1305) executes instructions and manipulates data to perform the operations of the computer (1302) and any algorithms, methods, functions, processes, flows, and procedures as described in the instant disclosure.

[0156] The computer (1302) also includes a memory (1306) that holds data for the computer (1302) or other components (or a combination of both) that can be connected to the network (1330). For example, memory (1306) can be a database storing data consistent with this disclosure. Although illustrated as a single memory (1306) in FIG.13, two or more memories may be used according to particular needs, desires, or particular implementations of the computer (1302) and the described functionality. While memory (1306) is illustrated as an integral component of the computer (1302), in alternative implementations, memory (1306) can be external to the computer (1302).

[0157] The application (1307) is an algorithmic software engine providing functionality according to particular needs, desires, or particular implementations of the computer (1302), particularly with respect to functionality described in this disclosure. For example, application (1307) can serve as one or more components, modules, applications, etc. Further, although illustrated as a single application (1307), the application (1307) may be implemented as multiple applications (1307) on the computer (1302). In addition, although illustrated as integral to the computer (1302), in alternative implementations, the application (1307) can be external to the computer (1302).

[0158] There may be any number of computers (1302) associated with, or external to, a computer system containing computer (1302), wherein each computer (1302) communicates over network (1330). Further, the term “client,” “user,” and other appropriate terminology may be used interchangeably as appropriate without departing from the scope of this disclosure. Moreover, this disclosure contemplates that many users may use one computer (1302), or that one user may use multiple computers (1302).PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT EXAMPLES EXAMPLE 1

[0159] FIGs. 4, 5, and 6 illustrate measurements of CSF flow based on high resolution magnetic resonance imaging. FIG. 4 shows the velocity (400) of CSF in the Sylvian aqueduct as measured by 7 tesla phase contrast MRI in 20 different human subjects. Twenty frames (402) were acquired during the cardiac cycle. Note the large subject-to-subject variability in CSF velocity (400). FIG. 5 shows bulk flow (500) of CSF through the Sylvian aqueduct as measured by 7 tesla phase contrast MRI in 20 different human subjects. Twenty frames (402) were acquired during the cardiac cycle. Note the large subject-to-subject variability in bulk flow (500) of CSF. FIG.6 shows Calculated cerebral aqueduct area from 7T PC-MRI for all patients. The range of values falls within the values seen in literature. An interesting observation is the significant fluctuation in area for some patients.

[0160] All data used in the analysis and estimation of CSF flow are from healthy volunteers. Data used in mild traumatic brain injury (mTBI) detection and mTBI’s impact on CSF flow are from the online data set FITBIR. FIGs. 4, 5 and 6 depict 7 Tesla MRI data collected to illustrate the CSF flow. FIG.4 measures velocity (400) of CSF and area of aqueduct from 7 Tesla MRI images leading to the measurement of the flow (500) in FIG. 5 defined as the velocity of the CSF times the area (600). Each curve represents a subject across a time period. Using a Phase-Contrast technique the bidirectional velocity of CSF was measured along the major perivascular spaces. This aspect of the study was completed with normal volunteers.

[0161] FIG. 7 shows the velocity of blood flow in cerebral arteries (700) as measured during cardiac cycle by transcranial doppler ultrasound in a human subject. Multiple phases can be recognized: systolic acceleration, diastolic deceleration, and the complex oscillating phase in between them that is dependent on the patient’s vascular tone.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT EXAMPLE 2

[0162] For the application to monitoring glymphatic flow and clearing waste products out of the brain, various mathematical predictors have been developed based on various models of linear regression. In FIG. 14, the number of splines required to fit the TCD curve is plotted against the width of the CSF velocity curve (1400). Splines are used to fit piecewise polynomials to non analytical functions. The splines are fit to the TCD curve in order to observe how many splines are needed to accurately fit the curve. This gives a proxy of complexity of the waveform. The number of splines (1402) captures how many peaks and troughs the TCD has and how complex the waveform is. The Pearson correlation coefficient is displayed in the top left and indicates a strong linear relationship between the variables. The number attached to each data point is the patient number (1404).

[0163] In FIG.15, the average duration of the TCD cycles (1502) from the right MCA is plotted against the width of the CSF velocity curve (1000). The average duration is essentially the length of each heartbeat, though lagged since this is measured in the brain’s middle cerebral artery. The number attached to each data point is the patient number (1004).

[0164] In FIG. 16, the diastolic deceleration (1602) of the TCD from the left MCA is plotted against the width of the CSF velocity curve (1000). The diastolic deceleration captures the speed of blood flow leaving the brain. Again, the number attached to each data point is the patient number (1004).

[0165] In one experiment, the input to the predictor model included the features in 5 different time series of data (physiological, TCD, EEG, Sleep Stages, and Sleep Stages graph; Table 6). The output was the actual waveform associated with each row (velocity, area flow). The number of patients for each column differs based on how many patients have had both their MRI and the data acquired in the column name. Subjects were randomly allocated to train the predictor (50% of subjects) or test its performance (50% of subjects); ridge regression was also performed. This is a specific type of linear regression that adds a penalty on the value of the weights.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT The Pearson correlation coefficient was calculated between each test reconstruction and the actual CSF curve. The mean was then determined over the test data samples. The predictor was observed to perform very well except, as expected, for area, as the area curves are the most erratic shaped waveforms. Table 6

[0166] The impact of different stages and positions of sleep on CSF flow was investigated. A significant correlation between the TCD and MRI was identified. Specifically, the number of splines in the TCD waveforms was calculated, which essentially measures how complex the TCD waveform is, and how correlated it to the width of the CSF velocity curve which is the duration of time that CSF velocity is positive. Table 7 presents an overview of all data collected by Methodist and Baylor. Labels for the Sleep data refer to sleep stages and body position throughout sleep. Table 7PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0167] FIG. 17 presents Sample 7 tesla phase contrast MRI images. The left image shows an anatomical image of the brain used for locating the cerebral aqueduct. The phase image on the right is used for calculating velocity and flow of CSF as the phase is proportional to velocity. FIG. 18 locates the aqueduct by using a custom algorithm that fits an ellipse to the region of interest. The area and bounding box are calculated for each frame, allowing us to also see how the area of the aqueduct changes over the cardiac cycle. Preliminary analysis of the data that shows strong correlations between the transcranial doppler, CSF flow from the MRI, and signals from the sleep dataset (FIG. 10). Due to these strong correlations, it is reasonable to assume one could predict the outcome of one modality from the other two given there is enough data. FIG. 4 is a calculated velocity curve from 7T PC- MRI for all patients. Differences between patients are seen, but a general trend is observable. The velocity peaks in the middle which is when the cardiac cycle is at its lowest activity, further confirming the finding that CSF movement and blood flow are inversely proportional. FIG. 5 is a calculated flow curve from 7T PC-MRI for all patients. Flow is calculated by multiplying the velocity and area curves together for each patient. FIG. 6 is the calculated cerebral aqueduct area from 7T PC-MRI for all patients. The range of values falls within the values seen in literature. An interesting observation is the significant fluctuation in area for some patients.

[0168] Table 8 shows an overview of the number of correlations between the CSF features and respective datasets. BP indicates body position. N denotes the number of patients that have both the MRI data and the data in the respective column. Threshold indicates the minimum Pearson correlation needed to count as a significant correlation. Naturally the number of significant features decreases as thePATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT threshold is increased. Interestingly, EEG measured during sleep may be seen to be extremely correlated with CSF movement, revealing many new relationships that were previously unknown. Table 8

[0169] Preliminary results of an experiment with the two-wavelength TD- fNIRS system with multiple source-detection pairs, detailed above, are presented here. The experiment was conducted on a human finger with a single source-detector pair of 1 cm separation to test the sensitivity of the device. A single path of 740nm of ~5.35 mW was used. Setting SPAD array integration time as 25ms one may achieve a 40 Hz frame rate measurement, which is important to illustrate the full pulsatile signal. From the original signal in FIG.19, an obvious pulse signal may be observed along with clear systolic and diastolic peaks. The original signal may then be processed with a band-pass filter to reduce the high frequency as well as background noise. Fast Fourier transform followed to analyze the signal in the frequency domain. The spectral peak is around 1.07Hz which is very close to the ground truth heart rate around 67-70 / min measured by commercial pulse oximetry. The preliminary in vivo test shows the temporal resolution of the current setup isPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT able to reveal the pulsatile signal. FIG. 20 shows pulsatile signal obtained from human finger and frequency domain analysis.

[0170] Another novel technique is proposed to modulate the brain activity in a non-invasive / minimally invasive manner using temporally interfering electromagnetic waves, which is termed as EMvelope stimulation. A simple cnidarian polyp, Hydra vulgaris, was studied, as no ICUC protocol is needed for experimenting with Hydra. Since EMvelope stimulation relies on the low-pass filtering properties of the neurons, an intermediate step was to study the RC circuit response of Hydra cells.

[0171] The goal is to study how a primitive nervous system, like that of Hydra, responds to electrical stimulation and how it is correlated with behavior. Moreover, the correlation between electrical stimulation and hydra neural activity was investigated. For this purpose, Hydra is a lucrative animal model as one may study whole body behavior as well as single cell resolution activity. For example, amongst two neurons in different parts of the body, one may observe which neuron fires first.

[0172] In this context, the first challenge to overcome is that the cell membrane time constant of Hydra is unknown due to challenges in traditional patch clamp methods. Cells across phylogeny are considered filters to input signals. As a proxy, the lowpass filter (RC circuit) model is used to approximate the membrane time constant of Hydra. FIG. 21 VKRZV^WKH^FKDQJH^LQ^IOXRUHVFHQFH^DFWLYLW\^ǻ)^)^^ZKHUH^ F is intensity of fluorescence signal. Notice that even though generally Hydra responds faster to lower frequencies, it responds to 100 Hz faster than 10 Hz, showing a possible frequency preference.. In FIG. 21, notice that even though generally Hydra responds faster to lower frequencies, it responds to 100 Hz faster than 10 Hz, indicating a possible sweet spot frequency.

[0173] Regarding the general framework for transducer amplitude, phase, and locations detailed above, a U-net was used to model the underlying finite-element- modeling method. Antennas were place at 38 locations inside the brain tissue and used to train the U-net. FIG. 22 shows that the training is successful as, by the endPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT of training at epoch 3000, the U-net model is able to effectively learn impulse response of the antennas. It may be seen that at epoch 3000, the U-net’s training output matches the ground truth impulse response generated in the physics-based COMSOL Multiphysics software. COMSOL numerically solves Maxwell's equations by finite-element-method and hence, for this reason serves as the ground- truth. FIG.22 shows Training Error as a function of epochs. The training error goes down as the epochs increase, showing that the model is able to emulate the training data effectively. Note that as the epochs increase (at epoch 3000), the training output from the model matches the ground truth from the physics-based software COMSOL.

[0174] After training the model, the model was tested on unseen data to determine if it may generalize well to examples it has not seen before. FIG.23 shows the result of the validation dataset. For creating the validation dataset, 20 random antenna locations were selected inside the brain tissue and their corresponding ground truth impulse responses were generated in COMSOL Multiphysics by solving Maxwell’s equations. Note that the validation output matches the ground truth impulse response for all the test labels / locations. This shows that the model even generalizes to unseen data. This means, impulse responses may be created for essentially any location inside the brain tissue. FIG.23 shows Validation Error as a function of epochs. The validation error also goes down as the epochs increase, showing that the model generalizes well and creates impulse response for essentially any location inside the brain tissue.

[0175] Another experimental goal is to measure the E-field envelope distribution in a brain-tissue-simulating phantom and compare the measurement results against theoretical prediction. Two antennas are mounted on a 19-cm glass petri dish, which is filled with a head-simulating liquid (HBBL600-10000V6). The liquid has a similar electrical property (i.e., the permittivity and conductivity) as the human brain in the frequency range of 600 MHz to 10 GHz. Two antennas are driven E\^ WZR^ EHQFKWRS^ VLJQDO^ JHQHUDWRUV^ DW^ IUHTXHQFLHV^ I^^ DQG^ I^^ ^^ ǻI^^ UHVSHFWLYHO\^^PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT Meanwhile, three electrical field probes are mounted on a motorized 3-D translational stage sequentially to measure the E-field distribution in the liquid with D^VSDWLDO^UHVROXWLRQ^RI^^^^P^^7KH^WKUHH^(-field probes consist of two time domain- sensor (TDS) probes E1TDSx and E1TDSz, which measure the relative time- domain waveforms in the x, y, and z directions, and a specific absorption rate (SAR) probe EX3DVx, which measures the absolute E-field intensity. As shown in the FIG. 24A and FIG. 24B, the E-ILHOG^ HQYHORSH^ IROORZV^ WKH^ ǻI frequency offset, aligning well with theoretical prediction. FIGs.24A and 24 B show measured time- domain E-field waveforms when the first antenna radiates at f1 = 1.5 GHz and the VHFRQG^DQWHQQD^UDGLDWHV^DW^I^^^^ǻI, where, for FIG. 24, ǻI^ ^^^N+], and, for FIG> 24B, ǻI^ ^^^^^+]^

[0176] The off-the-shelf antennas currently used in the phantom experiment are designed for free-space radiation, which has poor efficiency and radiation performance when delivering EM waves into the brain. The ideal antenna design should occupy a small size, have a matched impedance to the brain tissue, and allow for safe operation. In a first attempt, a patch antenna topology was chosen due to its low profile and ease of fabrication for rapid prototyping. The patch antenna consists of two dielectric layers and three metal layers, as shown in FIG. 25. An aperture- coupled feeding structure is designed to achieve good impedance matching. The feedline is on metal layer 3. The EM wave is coupled to the patch on metal layer 1 through the slot on the ground plane (metal layer 2). Antenna miniaturization is achieved by choosing a substrate with a high dielectric constant (Rogers RO3010, İU^ ^^^^. During the antenna fabrication, the copper on the two substrates was first etched away. Then copper tapes are cut into the desired geometries and then attached the tapes to the substrates. Finally, epoxy to glue is used to glue the two substrates together. The fabricated antenna was measured using a vector network analyzer N5225B, capable of achieving a good matching around 1.8 GHz. These custom- designed patch antennas are used to replace the off-the-shelf antennas to continue the phantom experiment.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0177] The signal generators used in the phantom experiment are not strong enough to deliver the output power necessary for animal studies. Thus, another experimental setup has been created for animal studies. In addition to higher output power, this setup consists of three antennas (instead of two in the phantom experiment) to improve the focality. Each antenna is driven by a signal propagation chain, consisting of a signal generator, a power amplifier (PA), a coupler, and an isolator, as shown in FIG. 26. The coupled port of the coupler is connected to an oscilloscope, which displays an attenuated replica of the antenna waveform in real time. The three signal generators used for the animal studies need to be synchronized and output three 120°-phase-shifted signals. FIG.26 shows a schematic of the signal propagation chain of each antenna.

[0178] The three dipole antennas that will be used for animal studies are designed with an operating frequency of 2 GHz when radiating into the brain. The operating frequency may be tuned by changing the length of the wires and the gap between the wires. FIG.27 shows the measured antenna input reflection coefficient S11.

[0179] The design of the breadboard-version of the REG sensor has been completed. Simulations in the circuit simulator software LTSpice have been performed to validate and debug the design. For the Current Source designed in the 2nd quarter, the current was designed to be 0.1 mA. After further investigation, it was concluded that a higher current is required in practice in order to increase the sensitivity of the REG sensor. Thus, the Current Source output was increased to 1 mA. The impedance change caused by the CBF is in the range of 0.1 to 0.25 ohm, whereas the baseline impedance is in the range of 200 ohm. The unique advantage of the architecture is that it can remove the baseline impedance at its output and only detect the minute impedance change caused by the CBF.

[0180] To verify the proposed REG sensor, circuit simulations were performed in LTSpice by modeling the bio-impedance using the resistor R7 in FIG. 28. The resistance value is configured to have a baseline impedance (Rbase) of 200 ohm andPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT varying impedance (Rchange) with a peak-to-peak amplitude of 0.4 ohm and a sinusoidal frequency of 2 Hz. The output voltage of the REG sensor is plotted in FIG. 28. The theoretical output voltage change can be found using the following equation: Vout = 0.4 ohm (Rchange) * 1 mA (current source) * 650 (overall gain) * cos(2*pi*2Hz*t) After 9 seconds of settling time, the simulated REG output reaches the theoretical value as was expected and it does not contain any artifacts derived from the baseline impedance. FIG. 28 shows REG sensor circuit simulations in LTSpice. The orange curve is the input bio-impedance (Rbase + Rchange) between the two electrodes. The cyan curve is the REG sensor output, which follows the minute impedance changes after 9 seconds of settling time. The amplitude of the REG sensor output aligns with the theoretical prediction.

[0181] To measure the breadboard-version REG sensor, ideally, a custom testing setup was created. In this setup, the impedance at the input of the board is manually changed from 200 to 199.98 ohm. The way to cause this minute change is to introduce another 200K ohm shunt resistor, as shown in FIG. 29. The switch in FIG. 29 was opened and closed manually to change the input impedance. The breadboard-version REG sensor produces a voltage waveform following the impedance change, which is then read by the Cyton + Daisy OpenBCI board. The data is eventually transferred to a computer. The measurement data is saved and processed to plot the input impedance using the following equation: Z = 3.3(Power Supply) * (Received_Data - 512) / 1024 / (660(Gain)*1mA(Current))+200(Rbase).

[0182] As can be seen in FIG. 28, the impedance change aligns with expectations, showing that both the breadboard-version REG sensor and the OpenBCI analog-to-digital interface work well.

[0183] To restore Glymphatic flow, the idea of minimally invasive DBS using temporally interfering gigahertz (GHz) electromagnetic (EM) waves has been explored. The idea of EMvelope stimulation was investigated, a minimally invasive DBS setup using temporally interfering gigahertz (GHz) electromagnetic (EM) waves. At GHz frequencies, one may create antenna arrays at the scale of a fewPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT centimeters or less that can be endocranially implanted to enable longitudinal stimulation and circumvent signal attenuation due to the scalp and skull. Furthermore, owing to the small wavelength of GHz EM waves, one may optimize both amplitudes and phases of the EM waves to achieve high intensity and focal stimulation at targeted regions within the safety limit for exposure to EM waves. A simulation framework was developed by investigating the propagation of GHz EM waves generated by line current antenna elements and the corresponding heat generated in the brain tissue; two optimization flows were proposed to identify antenna current amplitudes and phases for either maximal intensity or maximal focality transmission of the interfering electric fields with an EM waves safety constraint. A representative result of the study is that with two endocranially implanted arrays of size 4.2 cm × 4.7 cm each, one may achieve an intensity of 12 V / m with a focality of 3.6cm at a target deep in the brain tissue. As this proof-of- principle study cites, EMvelope stimulation was shown to merit further investigation as it may be a minimally invasive way of stimulating deep brain targets and offers benefits not shared by prior methodologies of electrical or magnetic stimulation. EXAMPLE 3

[0184] As another example, a test cohort was formed that contained 22 healthy control subjects that underwent a sleep study and 7T PC-MRI acquisition for elucidation of ground truth CSF flow through the Sylvian aqueduct. Three machine learning regression models were trained per input feature set. 7 input feature sets were derived from the sleep study data. These included features derived from the ECG, PPG, and respiration. Lastly, all feature sets were concatenated together to generate one fusion feature set. From these feature sets the machine learning models learn a function to estimate the full CSF flow curve as derived from the 7T PC-MRI dataset. The models were trained in a leave one out (LOO) fashion, meaning that the models train on all subjects but one and then are tested on the excluded subject. This process was redone for every subject in the development cohort. A Pearson product-PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT moment correlation coefficient was calculated between the estimated CSF flow curve and the ground truth CSF flow curve as a measure of prediction quality.

[0185] A quantification of the mean and median correlation coefficients is shown in Tables 9 and 10. Table 9 shows results on development cohort flow. Table 10 shows results on an independent validation cohort flow. In general, LASSO and Ridge regression methods outperformed the PLS model. It is noteworthy that all input feature sets yield mean and median correlation coefficients above 0.9 for the LASSO and ridge regression models. The best modality for CSF estimation therefore depended upon which regressor was used. Fusion of all features resulted in the highest correlation coefficient for the ridge model. Table 9PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT Table 10EXAMPLE 4

[0186] Cerebrospinal fluid (CSF) is an ultrafiltrate of plasma predominantly secreted by the choroid plexus and contained within the ventricles of the brain and the subarachnoid spaces within cranium and spine. Adult CSF volume is estimated to be approximately 150 ml, with a distribution of 5 / 6th within the subarachnoid spaces and 1 / 6th within the ventricles. CSF secretion varies between individuals, usually ranging from 400 to 600 ml per day, which contributes to complete CSF renewal four to five times per day. The flow of CSF provides nourishment and waste removal to the whole central nervous system. Thus, pathophysiological changes in the movement and accumulation of CSF within the ventricles are expected to affect the function of the central nervous system parenchima, and they will precede overt neurological symptoms.

[0187] Under normal conditions the CSF is in constant motion due to various physiological processes. The choroid plexus adds CSF volume, while the arachnoid villi reabsorb CSF into the bloodstream, though generating a hydrostaticPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT pressure that pushes CSF through the brain and spinal cord. In addition, the motion of CSF is significantly affected by the cardiac and respiratory cycles. With each heartbeat the arterial pressure varies leading to pulsatile motion of CSF. Similarly, changes in intrathoracic pressure caused by breathing affect the movement of CSF because the negative thoracic pressure during inspiration increases venous return from the brain to the heart. In addition, changes in intrathoracic pressure are transmitted through the spinal canal which is filled with CSF and connects the intracranial subarachnoid space with the spinal subarachnoid space. Thus, changes in any of these physiological functions may result in abnormal accumulation and / or movement of CSF, and subsequent neurological deficits.

[0188] Magnetic Resonance Imaging (MRI) allows for the non-invasive assessment of both anatomical and functional features, and allows one to measure the dynamics of CSF (FIGs. 4, 5, and 6). However, this technique it is expensive, time consuming, and non-portable because it requires large magnets and acquisition systems. Thus, novel technologies to non-invasively measure and monitor CSF movement will aid in the early diagnosis of neurological disorders and facilitate their early detection and treatment.

[0189] Embodiments disclosed herein relate to a system of wearable technologies and machine learning algorithms that allow the convenient, economical, comfortable and non-invasive monitoring of CSF dynamics and the early diagnosing of their pathophysiological changes. One or more embodiments disclose:

[0190] 1) A chest patch that adheres to the patient’s skin and receives and processes input from one or more wearable physiological sensors (FIG.1).

[0191] 2) Physiological sensors placed on the patient’s body for the continuous or intermittent monitoring of one or more physiological parameters, including cardiac cycle, blood volume, pO2, and respiration (FIG. 1).PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0192] 3) Mathematical features used to extract information from- and correlate the time series acquired by- one or more physiological sensors recording continuously or intermittently the physiological parameters (FIGs.1 and 2).

[0193] 4) Mathematical predictors (from simple regressions to machine learning models) that estimate the dynamics of a patient’s CSF including, but not limited to, the CSF velocity, flow, and time to peak / trough (FIGs.4 and 5).

[0194] Because existing techniques for estimating CSF flow from non-invasive modalities did not exist, a wide array of features from time series analysis, frequency domain analysis, information theory, and graph theory were used to discover a novel subset of biomarkers that help estimate CSF flow non-invasively.

[0195] For the initial studies, three physiological signals were recorded simultaneously during a sleep study: cardiac cycle by electrocardiogram (ECG), blood volume and pO2 by photoplethysmogram (PPG), and respiration by piezoelectric belt. The ECG, PPG, and respiration signals were first low-pass filtered with a 6th order Butterworth filter with a cutoff frequency of 50 Hz, to eliminate powerline, motion, and electromyographic noise. Normal filtering is done only in the forward direction in time, which causes a time delay and a change in the phase of the signal. To compensate for this, the signal was filtered backwards in time. This is called a zero-phase filter and is beneficial here to preserve all information in time without introducing distortions.

[0196] Next, the data were segmented into 30 second non overlapping blocks aligned with the sleep scoring labels. All segmented blocks associated with the same sleep stage are grouped together for feature extraction. This allowed us to analyze how the performance of the regressor changes with different sleep stages. From the segmented and grouped blocks of data feature extraction was performed (Table 1).

[0197] The ECG is a non-invasive tool for assessing the health of the heart. The ECG is a diagnostic signal obtained traditionally in a hospital or clinical setting. It is obtained by placing skin electrodes on different locations of the torso in order to measure global electrical signals produced by the heart. In this work only onePATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT channel of the ECG signal was used. For the ECG the following were calculated: heart rate, mean R wave peak to peak distance (RR interval), heart rate variability, root mean square of successive differences between normal heartbeats, percentage of successive RR intervals that differ by more than 50ms, low frequency and high frequency power, the ratio of low frequency (LF) to high frequency (HF) power, and the normalized low and high frequency power. These features are calculated for every epoch and then are aggregated and averaged over the five sleep stages. This strategy is implemented for all four physiological modalities. The total number of features for the ECG feature set was 110.

[0198] PPG is a simple optical technique that detects fluctuations in blood volume. Either red or infra-red light is illuminated on the skin of a subject and the reflected light is then measured using a photodetector. This reflected light is proportional to changes in blood volume via the Beer-Lambert law. The use of this technology has increased in popularity due to its simplicity, non-invasiveness, and comfort when placed on a subject. For the PPG, the following were calculated: pulse intensity ratio, heart rate variability, percent of dicrotic waves, and the length of the PPG waveform. The total number of features for a PPG feature set in this example was 80.

[0199] The respiration signal detects the breathing pattern of the subject over time. Respiration in known to modulate and affect the heart rate, which in turns affects the pO2, and both respiration and cardiac cycle are known to affect CSF flow. Thus, from the respiration waveform the mean and standard deviation of the respiration rate, the maximum inhalation slope, maximum exhalation slope, and the area under the respiration curve were calculated. The total number of features for the respiration feature set for this example was 85.

[0200] It is unlikely that one feature or one modality is sufficient to accurately and robustly predict CSF dynamics, but, when used together, they could better achieve the objective. Combining signals from different sources to boost their predictive power is the main goal of the field of multimodal machine learning. ThePATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT act of combining data from different sensors is known as data fusion. Thus, in addition to the individual feature sets created obtained from each physiological sensors as described, the extracted features from all modalities were concatenated into one set and used as input to various machine learning models. The goal was to learn to estimate CSF flow and velocity from those non-invasive modalities. The problem may be classified as a supervised regression problem since CSF flow is a continuous quantity as opposed to a distinct class or category (FIG.3). The input are features from the modalities and the label is CSF flow measured from the ground truth of MRI. The goal is to then learn a model that minimizes the error between the estimated CSF flow and the true CSF flow.

[0201] The development cohort contained 22 healthy control subjects that underwent a sleep study and 7T PC-MRI acquisition for elucidation of ground truth CSF flow through the Sylvian aqueduct (FIGs.4, 5, and 6). Three machine learning regression models were trained per input feature set. There are 7 input feature sets derived from the sleep study data. This includes features derived from the ECG, PPG, and respiration. Lastly, all the feature sets were concatenated together to generate one fusion feature set. From these feature sets the machine learning models learn a function to estimate the full CSF flow curve as derived from the 7T PC-MRI dataset. The models are trained in a leave one out (LOO) fashion meaning that the models train on all subjects but one and then are tested on the excluded subject. This process is redone for every subject in the development cohort. The Pearson product- moment correlation coefficient is calculated between the estimated CSF flow curve and the ground truth CSF flow curve.

[0202] The movement of CSF, including flow, velocity and time to peak and trough, has been estimated in the cerebral aqueduct of healthy human control subjects using a set of non-invasive modalities rather than MRI. First, a novel semi- automatic pipeline was developed for calculating cerebral aqueduct area and associated velocity and flow from 7T PC-MRI images. The ground truth CSF flow curves showed a large variance in both the morphology and amplitude of the curvesPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT even within a cohort of healthy control subjects. The cerebral aqueduct area was shown to vary throughout the cardiac cycle and does so to different extents in different subjects.

[0203] Next, the ability to estimate CSF flow was demonstrated in both a development cohort of subjects and an entirely new independent validation cohort of subjects to highlight the robustness and generalizability of the machine learning method. This was achieved with extensive and novel feature extraction pipelines across several modalities. High estimation accuracy was observed, as assessed by the correlation coefficient and mean squared error, across the modalities. This indicates that it is possible to estimate CSF flow with different non-invasive modalities. In addition, the machine learning models used do not contain millions of parameters as in neural network methods, implying that the method can be easily translated into portable hardware for CSF flow estimation. The performance of the model was shown to not change drastically when confined to features calculated within a single sleep stage. This implies that CSF dynamics may be estimated in both an awake and sleep state, thus increasing the applicability of this machine learning method to allow the tracking of CSF flow as a biomarker for the development of neurological disorders over time.

[0204] In summary, the methods and system described above resulted in a means for estimating the movement of CSF, including flow, velocity, and time to peak and trough, in the cerebral aqueduct of healthy human control subjects using a set of non-invasive modalities (rather than MRI). First, a novel semi-automatic pipeline was developed for calculating cerebral aqueduct area and associated velocity and flow from 7T PC-MRI images. Ground truth CSF flow curves showed a large variance in both the morphology and amplitude of the curves even within a cohort of healthy control subjects. The cerebral aqueduct area was shown to vary throughout the cardiac cycle and does so to different extents in different subjects. It was also shown that the performance of the model does not change drastically when confining the calculated features to be within a single sleep stage. This implies thatPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT one can estimate CSF dynamics in both an awake and sleep state, thus increasing the applicability of the machine learning method and allowing the tracking of CSF flow as a biomarker for the development of neurological disorders over time.

[0205] The ability to estimate CSF flow in both a development cohort of subjects and an entirely new independent validation cohort of subjects demonstrated the robustness and generalizability of the machine learning method. This was achieved with feature extraction pipelines across several modalities. High estimation accuracy was observed, as assessed by a correlation coefficient and a mean squared error, across the non-invasive modalities. This indicates that it is possible to estimate CSF flow with different non-invasive modalities than ECG, PPG, and respiration, such as EEG, sleep stage time series, and sleep stage graphical models. In addition, the machine learning models contained a sufficiently low number of parameters that the method can be implemented in portable hardware for CSF flow estimation. Each model contained one hyperparameter and then the total number of learned parameters is equal to the number of features plus one for the regression intercept. This yields 276 parameters for the fusion of ECG, PPG respiration in addition to one hyperparameter. EXAMPLE 5

[0206] A benefit of using simple machine learning models to analyze data, is the ease of analyzing the learned weights to identify the contribution of the different calculated features in the model’s eventual estimation of CSF flow. In an application, weights learned in the ridge regression model, when trained on the development cohort, may be used for feature importance analysis.

[0207] First, the top features used for each of the 7 input feature sets may be identified. However, a top feature must be defined. The ridge regression model may be trained in a LOO (leave one out) fashion, leading to 22 models per input feature set since the development cohort has 22 patients. For each of the 22 models, the absolute value of the learned ridge coefficients (which has a size of the number of features by the length of the CSF estimate which is 20) may be taken and summedPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT over the second dimension. This produces a vector with a length equal to the number of input features. The values within the vector give an estimate of the importance of each feature because features that are not useful will have a small magnitude in the learned ridge weights, while important features will have larger weights. This is referred to as the feature importance vector. Taking the absolute value of the weights may indicate that both positive and negative correlations between the input features and the learned ridge weights are important. Next, the ten features with the highest values are selected, and this is done for all 22 models. Next, the ten features that appear the most across the 22 models are picked, and referred to as the top features. The top ten features for each input feature set are shown in Table 11. The features are ranked in order by the number of times they appeared in the top ten features across the 22 different models from the LOO train test scheme. Each feature first states the sleep stage where the data originated from for feature calculation, and then follows with the feature name. Table 11PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0208] For the ECG and PPG feature set, it is observed that all top features are time domain specific features. For the ECG, it may be seen that RMSSD, heart rate, mean RR, and HRV are the most impactful features. Heart rate has already been shown to be correlated to CSF flow in the literature, but HRV has not. These results give credence to the idea that HRV’s relationship to vascular control make it a useful feature for estimating CSF flow. It may also be seen that HRV is also impactful in the PPG feature set, taking up five of the top ten spots. It should be noted that the top feature for the PPG was the length, and it appears four times within the top ten features. The length of a waveform increased when there was more curvature, thus this feature may be a surrogate for detecting the dicrotic notch. One may also see that pulse intensity ratio is one of the top features.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0209] For the respiration signals, it was observed that the top features are Hjorth complexity, respiration rate, and respiration rate standard deviation. Hjorth complexity can be interpreted as the similarity of the shape of a signal to a pure sine wave. The respiration signal does look like a simple sine wave, especially if the breathing pattern is consistent and regular. Therefore, this feature may imply that the regularity in breathing is correlated to CSF flow. The standard deviation of the respiration rate may also be capturing a similar aspect of respiration since a decreased standard deviation would imply a more consistent breathing pattern. Respiration rate has already been studied in the literature in the context of CSF flow, but with mixed results on its importance. These may show that respiration rate may be correlated with CSF flow. The top ten features for the EEG are all ratios of power of frequency bands – particularly, the two frequency bands of delta and high gamma. Additionally, all features are from the N3 sleep stage, further supporting a correlation between delta power and CSF flow. The top eight features for the STS feature set are all counts of sleep stages or body positions. The length of time asleep may be important for CSF flow estimation rather than the percentage of time spent within any specific sleep stage. The number of stage transitions and Hjorth complexity are also in the top ten features, indicating that the smoothness of the hypnogram and body position waveforms are important for CSF flow estimation.

[0210] The top features for the SSG dataset are page rank, number of self-loops, betweenness centrality, shortest path length, and eigen centrality. The page rank was important specifically for stage N1 indicating that this sleep stage and its position within the sleep stage graph structure is important for CSF flow estimation since the page rank gives a measure of importance to a node. It may also be seen that the eigen centrality into sleep stage N3 is important, which is highly related to page rank mathematically. The number of self-loops measures the ability for a subject to sustain sleep within a sleep stage. This feature may be seen for the wake state, REM state, and N3 state. This may be linked to the observation that sleep disruption and deprivation is linked to dysfunctions in CSF flow. Betweenness centrality shows up as top feature for all sleep stages except N3. This metric identifies nodes located onPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT the most traveled paths by measuring the number of shortest pathways in the network that passes through a given node. Average shortest path length also appears as a top feature, which again is a measure of the efficiency of information or mass transport on a network. The betweenness centrality and average shortest path length being important features hints that the ability to easily transition between sleep stages is important for CSF flow estimation.

[0211] For the fusion feature set, five of the top features are from EEG, three from STS, and two from respiration. A majority of the features in the fusion feature set are from the EEG, indicating that the top ten features are not all from the EEG feature set. This demonstrates that the machine learning model is fusing information across modalities to form a CSF flow estimate. These results validate the importance of some physiological features already described in the literature and also demonstrate some new features that may be important in the context of CSF flow that have not been explored prior. The fusion feature set allows for investigation into the interplay between multiple modalities for CSF estimation. A summary of the key findings is shown in Table 12. For the six feature sets that makeup the fusion feature set the number of features that each feature set contains is shown. This number is divided by the total number of features across the feature sets and multiplied by 100 to give the feature composition percentage. A naive regressor would use each feature set proportionally to its feature composition percentage. To investigate this, the feature contribution percentage is calculated which quantifies the percentage of magnitude in the ridge regression model weights for each feature set. It may be observed that the feature composition percentage and the feature contribution percentage differ, indicating that the regressor selectively favors some feature sets possibly due to their strong relationship with CSF flow. This effect is quantified by dividing the contribution and composition percentages. It may be observed that the STS feature set has the most disproportionate contribution the model estimation. Table 12PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0212] First, looking at the number of features contributed by each feature set in the fusion feature set, it may be seen that EEG has the most features by far and thus takes on a disproportionate percentage of the total features. The number of features per modality is divided by the total number of features to give the feature composition percentage. Learned ridge regression weights may be obtained for each of the 22 models. From this, one may take the absolute value and sum over the second dimension to give the feature importance vector of length equal to 8840 (the sum of the number of features column in Table 12). The weights associated to one feature set may be summed and divided by the total sum of the feature importance vector to produce a notion of the contribution of the feature set to the overall ridge regression estimation. This contribution may then then averaged over the 22 models and the results displayed in the column entitled feature contribution percentage. It may be noticed that the feature composition percentage is not the same as the feature contribution percentage, indicating that the model is not using all features in a proportionate fashion but instead uses features across modalities in a non-obvious manner. The feature contribution percentage may be divided by the feature composition percentage to give a ratio that indicates how important a specific feature set was for CSF flow estimation. A value of 1 indicates that the model uses the feature set in a proportional manner, a value above 1 indicates that the feature set is used more than expected, and a value below 1 indicates that the feature set is usedPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT less than expected. It may be seen that the feature set with the highest ratio is the STS feature set, meaning that this feature set is used 7x more than what would be proportional to the number of features it makes up in the fusion feature set. We also see a high ratio value for the respiration feature set. The top ten features for the fusion feature set in Table 11 came from the EEG, STS, and respiration feature sets, further highlighting the importance in respiration features and sleep stage features. EEG has a ratio below one, indicating that many of the calculated features are not useful for CSF flow estimation. The SSG dataset has the lowest ratio, possibly indicating that analyzing sleep stages with a time series analysis approach is superior to a graphical approach. The SSG dataset also did not contain any analysis on body positions which may have limited its expressive power.

[0213] The last feature analysis performed was to analyze the contribution of features calculated in the five sleep stages for the ECG, PPG, respiration, and EEG feature sets for CSF flow estimation. To do this, feature importance vectors were calculated for each of the four feature sets shown in Table 13 and the summed values associated with one sleep stage and divide by the sum of the whole vector. These values were averaged across the 22 models to give the values shown in Table 13. The highest percentage values per row is put in boldface to call attention to highest contributing sleep stage for CSF flow estimation. The ECG and EEG features calculated during the N3 sleep stage are the most impactful for model estimation, while the wake state is the most impactful for the PPG and respiration signal. Table 13PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT

[0214] The sleep stages with the highest contribution differ between the modalities and are only the wake or N3 state. The highest contribution of the features for EEG come from the N3 sleep stage, further confirming that deep sleep and the EEG are important in the context of CSF flow. The ECG features calculated during N3 also show the highest contribution like EEG, but with a greater percentage than for the EEG. PPG and respiration have the highest feature contributions during the wake state. For the respiration, it may be seen that the feature contribution decreases along the row, indicating that respiration features are more important in lighter stages of sleep. This also fits with the top ten features shown in Table 10, where most top features came from the wake, N1, and N2 stages.

[0215] FIG.30 shows a weighted graph structure obtained from data illustrating an example sleep stage sequence, determination of a transition matrix for subject 17 from the sleep stage sequence, and determination of a weighted directed graph of sleep stages for subject 17 from the transition matrix.

[0216] Although only a few example embodiments have been described in detail above, those skilled in the art will readily appreciate that many modifications are possible in the example embodiments without materially departing from this invention. Accordingly, all such modifications are intended to be included within the scope of this disclosure as defined in the following claims.

Claims

PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT CLAIMS What is claimed is:

1. A system for assessment of cerebrospinal fluid dynamics in a human subject, the system comprising: one or more physiological sensors configured to non-invasively record one or more physiological signals from the human subject; one or more microprocessors in communication with the one or more physiological sensors and configured to: extract one or more features from the one or more physiological signals; and implement a machine learning algorithm, wherein the machine learning algorithm is configured to estimate one or more aspects of cerebrospinal fluid dynamics from one or both of the one or more physiological signals and the one or more features; and an output device in communication with the one or more microprocessors and configured to output the one or more aspects of cerebrospinal fluid dynamics.

2. The system of claim 1, wherein: the one or more microprocessors are configured to use the one or more aspects of cerebrospinal fluid dynamics to perform a diagnosis of a neurological condition, and the output device is configured to display the diagnosis.

3. The system of claim 2, wherein the diagnosis is selected from the group consisting of hydrocephalus, traumatic brain injury, Parkinson’s, Alzheimer’s, epilepsy, an impaired shunt used in a treatment of hydrocephalus, and combinations thereof.

4. The system of claim 1, wherein the one or more physiological sensors configured to continuously record the one or more physiological signals.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT 5. The system of claim 1, wherein the one or more physiological sensors configured to intermittently record the one or more physiological signals.

6. The system of claim 1, wherein the output device is in wireless communication with the one or more microprocessors.

7. The system of claim 1, wherein: the one or more physiological sensors and the one or more microprocessors are combined in a wearable device.

8. The system of claim 7, wherein the wearable device comprises a headgear.

9. The system of claim 7, wherein the wearable device comprises a chest patch.

10. The system of claim 7, wherein the wearable device comprises a piezoelectric belt.

11. The system of claim 7, wherein the wearable device is in wireless communication with the output device.

12. The system of claim 1, wherein the one or more physiological signals are selected from the group consisting of cardiac sensors, photoplethysmogram sensors, piezoelectric sensors, transcranial probes, transcranial doppler ultrasound sensors, electrocardiogram sensors, cardiac function sensors, cardiac cycle sensors, blood volume sensors, blood gas sensors, electroencephalography sensors, near infrared spectroscopygram sensors, rheoencephalography sensors, orbital sonography sensors, respiratory sensors, and combinations thereof.

13. The system of claim 1, wherein the one or more features are selected from the group consisting of a heart rate, a mean respiratory rate, a heart rate variability, a second derivative of photoplethysmogram signals, a photoplethysmogram intensity ratio, a respiration rate, a mean of the respiration rate, a standard deviation of the respiration rate, an area under a respiration curve, a maximum inhalation slope, a maximum inhalation slope, a blood volume, a partial pressure of oxygen, a mean R wave peak- to-peak distance, a heart rate variability, a root mean square of successivePATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT differences between normal heartbeats, a percentage of successive RR intervals that differ by more than 50 ms, a low frequency (LF) power, a high frequency (HF) power, a ratio between low frequency to high frequency power, a normalized low frequency power, a normalized high frequency power, a root mean square of successive differences (RMSSD), a percentage of adjacent NN intervals that vary from each other by more than 50 ms (PNN50), a very low frequency (VLF) power, an LF / HF ratio, a normalized LF power, a normalized HF power, a pulse transit time, a pulse intensity ratio, a photoplethysmogram heart rate variability, a percent dicrotic, a respiration rate, a respiration standard deviation, an inhale slope, an exhale slope, an area on a curve, a permutation entropy, an SVD entropy, a spectral entropy, an approximate entropy, a sample entropy, a Hjorth mobility, a Hjorth complexity, a Katz fractal dimension, a detrended fluctuation, a Higuchi fractal dimension, a Petrosian fractal dimension, a Hurst exponent, an arterial tone, a mutual information, a maximum transcranial doppler (TCD) velocity, a minimum TCD velocity, a mean TCD velocity, a pulsatility index, a resistance index, a systolic acceleration, a duration of cycle, a number of splines, a length of cycle, a diastolic deceleration, a number of maxima and minima, a delta wave power, K-complex waveforms, sleep spindles, saw-tooth waveforms, a degree of centrality, and combinations thereof.

14. The system of claim 1, wherein the one or more features are calculated from a Sylvian aqueduct.

15. The system of claim 1, wherein the one or more features are calculated from sleep stages.

16. The system of claim 1, wherein the one or more features comprise one or more graph features selected from the group consisting of a centrality, a page rank, a clustering coefficient, a global efficiency, a number of partitions, a shortest path length, an in degree, an out degree, and combinations thereof.PATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT 17. The system of claim 1, wherein the machine learning algorithm comprises a neural network trained to learn a relationship between the one or more physiological signals, the one or more features, and the one or more aspects of cerebrospinal fluid dynamics.

18. A method for assessment of cerebrospinal fluid dynamics in a human subject comprising: non-invasively recording, using one or more physiological sensors, one or more physiological signals from a human subject; extracting, using one or more microprocessors in communication with the one or more physiological sensors, one or more features from the one or more physiological signals; estimating, with a machine learning algorithm, one or more aspects of cerebrospinal fluid dynamics from one or both of the one or more physiological signals and the one or more features; and outputting, using an output device in communication with the one or more microprocessors, the one or more aspects of cerebrospinal fluid dynamics.

19. A method for assessment of glymphatic fluid dynamics in a human subject comprising: non-invasively recording, using one or more physiological sensors, one or more physiological signals from a human subject; extracting, one or more microprocessors in communication with the one or more physiological sensors, one or more features from the one or more physiological signals; implementing a machine learning algorithm, using the one or more microprocessors, wherein the machine learning algorithm is configured estimate one or more aspects of glymphatic fluid dynamics from one or both of the one or more physiological signals and the one or more features; andPATENT APPLICATION ATTORNEY DOCKET NO.17500-229WO1 CLIENT REF. NO.2023-035-PCT outputting, using an output device in communication with the one or more microprocessors, the one or more aspects of glymphatic fluid dynamics.

Images