System and method for determination of blood pressure using an implantable device

Abstract

There is provided a system and method for determination of blood pressure using an implantable device. The method including: receiving electrical signals from one or more electrodes, the electrical signals including pulse artifacts generated from deformation of the respective electrode; digitizing the electrical signals received from the one or more electrodes into temporal segments; predicting systolic and diastolic blood pressure values from the digitized signals using a trained machine learning model, the machine learning model trained using a training dataset including digitized signals and associated measured blood pressure values; and outputting the predicted systolic and diastolic blood pressure values.

Description

SYSTEM AND METHOD FOR DETERMINATION OF BLOOD PRESSURE USING AN IMPLANTABLE DEVICETECHNICAL FIELD

[0001] The following relates, generally, to implantable medical devices; and more particularly, to a system and method for determination of blood pressure using an implantable device.BACKGROUND

[0002] Blood pressure measurement is an essential index for predicting cardiovascular disease and for performing health evaluation. There is a direct relationship between blood pressure and cardiovascular disease, which suggests that blood pressure data can provide useful information about the heart and circulatory function. However, there are currently very limited options available for long-term continuous monitoring of arterial blood pressure. For example, wearable photoplethysmogram (PPG) sensors can continuously monitor blood pressure, but are not able to perform such measurements over a long period of time.

[0003] The auscultatory method is the most common clinical approach for measuring blood pressure and involves the use of a sphygmomanometer and stethoscope. As the pressure of the inflated rubber arm cuff is gradually reduced, Korotkoff sounds are used to identify the systolic (SBP) and diastolic (DBP) blood pressure values. This is limited to a single measurement that is performed in a clinical setting and can be prone to errors. There can be a tendency to underestimate the SBP, while it has been shown that discrepancies occur when compared to intra-arterially measured values. This approach becomes difficult to use for subjects with organ damage or elders due to the auscultatory gap phenomenon where the Korotkoff sounds become inaudible between SBP and DBP. The type of sphygmomanometer can also affect the accuracy of measurements. Mercury-based sphygmomanometers provide the least amount of error but have technical problems, defectiveness, and observer bias that can impact accuracy. Modern aneroid devices use a mechanical system for measurement and are less accurate compared to mercury. Furthermore, in-clinic methods of measuring blood pressure can be affected by ‘white coat syndrome’, where a patient may get a high blood pressure reading in a clinician's office but a normal reading at home.

[0004] Oscillometry is another blood pressure measuring technique, which analyzes pressure oscillations caused by deflating a sphygmomanometer cuff. The measured pressuresuperimposes the cuff pressure and oscillations of blood flow through the artery. The signal is high-pass filtered and used to predict SBP and DBP by fitting the data to empirically-derived coefficients or by applying learning algorithms. Oscillometry has a general propensity for measurement errors in certain patient populations with cardiovascular issues, such as those with arteriosclerosis.

[0005] Photoplethysmographic (PPG) can be used to measure blood pressure when used in conjunction with a secondary device. For example, a finger cuff can be used to measure a continuous PPG signal, which consists of alternating-current (AC) and direct-current (DC) components that represent the arterial pressure pulse and the average blood volume, respectively. Generally, such approach requires pressure to be applied by the finger cuff or other device (i.e., to provide counter pressure). This may require that the device be continuously adjusted in a closed-loop fashion such that the blood volume (DC component of PPG) remains constant. The time-varying pressure applied by the cuff therefore corresponds to the arterial blood pressure; and can be referred to as a vascular unloading technique.

[0006] In other cases, the AC component of the PPG can be used to measure pulse transit time (PTT), which is defined by the time it takes for a pulse pressure to propagate through a given length of an arterial path. The PPT is inversely proportional to the arterial blood pressure, or conversely, the conduction velocity of the pressure pulse is proportional to blood pressure. It is suggested that the pulse arrival time (PAT) may be used as a proxy for estimating the SBP and DBP, where the time delay between the R-wave of the electrocardiogram (ECG) and a pressure pulse (i.e., PPG) can be measured.

[0007] Tonometry is another approach to predict blood pressure by partially compressing or splinting an artery against a bone and placing an external pressure sensor (tonometer) on the skin surface. The contact stress between the skin and the tonometer is equal to the instantaneous intraluminal pressure. The position of the tonometer can affect the accuracy of the measurement, where the best accuracy is obtained when the pressure sensor is directly aligned with the blood vessel. Tonometry has been shown to provide better accuracy for SBP estimation than for estimating mean arterial pressure (MAP) and DBP, and tends to underestimate pressure values under hypertensive conditions.SUMMARY

[0008] In an aspect, there is provided a system for determination of blood pressure using an implantable device, the implantable device comprising one or more electrodes positioned on or around a vessel, the system comprising one or more processors and a data memory, the one or more processors in communication with the data memory to receive instructions to execute: a signal processing module to receive electrical signals from the one or more electrodes, the electrical signals comprising artifacts generated from mechanical deformation of the electrodes, the signal processing module digitizes the signals received from the one or more electrodes into temporal segments; an analysis module to predict systolic and diastolic blood pressure values from the digitized signals using a trained machine learning model, the machine learning model trained using a training dataset comprising digitized signals and associated measured blood pressure values; and an output module to output the predicted systolic and diastolic blood pressure values.

[0009] In a particular case of the system, the machine learning model comprises a convolutional neural network, and wherein digitizing the electrical signals received from the one or more electrodes further comprises transforming each temporal segment into three-channel images for input into the machine learning model.

[0010] In another case of the system, digitizing the electrical signals comprises performing a Hilbert-Huang Transform to decompose the received signals into intrinsic mode functions.

[0011] In yet another case of the system, the three-channel image comprises a raw digitized values in a first channel, a first derivative of the digitized values in a second channel, and second derivative of the digitized values in the third channel.

[0012] In yet another case of the system, digitizing the electrical signals further comprises normalizing the first derivative and the second derivative for each dimension.

[0013] In yet another case of the system, the machine learning model comprises a neural network sequential model, and wherein digitizing the electrical signals comprises determining a set of feature vectors from the received signals.

[0014] In yet another case of the system, the feature vectors comprise one or more of a measure of standard deviation of a time series, a measure of variability of the time series, and a measure of asymmetry and characteristics of distribution of variations in the time series.

[0015] In another aspect, there is provided a method for determination of blood pressure using an implantable device, the implantable device comprising one or more electrodes positioned on or around a vessel, the method comprising: receiving electrical signals from the one or more electrodes, the electrical signals comprising artifacts generated from mechanical deformation of the electrodes; digitizing the electrical signals received from the one or more electrodes into temporal segments; predicting systolic and diastolic blood pressure values from the digitized signals using a trained machine learning model, the machine learning model trained using a training dataset comprising digitized signals and associated measured blood pressure values; and outputting the predicted systolic and diastolic blood pressure values.

[0016] In a particular case of the method, the machine learning model comprises a convolutional neural network, and wherein digitizing the electrical signals received from the one or more electrodes further comprises transforming each temporal segment into three-channel images for input into the machine learning model.

[0017] In another case of the method, digitizing the electrical signals comprises performing a Hilbert-Huang Transform to decompose the received signals into intrinsic mode functions.

[0018] In yet another case of the method, the three-channel image comprises a raw digitized values in a first channel, a first derivative of the digitized values in a second channel, and second derivative of the digitized values in the third channel.

[0019] In yet another case of the method, the method further comprising normalizing the first derivative and the second derivative for each dimension.

[0020] In yet another case of the method, the machine learning model comprises a neural network sequential model, and wherein digitizing the electrical signals comprises determining a set of feature vectors from the received signals.

[0021] In yet another case of the method, the feature vectors comprise one or more of a measure of standard deviation of a time series, a measure of variability of the time series, and a measure of asymmetry and characteristics of distribution of variations in the time series.

[0022] In yet another case of the method, the method further comprising performing electrical nerve stimulation by the one or more electrodes.

[0023] In yet another case of the method, the electrical nerve stimulation is performed to down- regulate nerve activity or to up-regulate nerve activity based on the predicted systolic and diastolic blood pressure values.

[0024] In another aspect, there is provided an implantable device for determination of blood pressure, the implantable device comprising: one or more electrodes positionable on or around a vessel, the electrodes generating an electrical signal due to mechanical deformations of the electrodes from periodic pulsations of the vessel; electronic circuitry to relay the signals captured by the implantable device to an external computing device, the external computing device digitizes the signals into temporal segments and predicts systolic and diastolic blood pressure values from the digitized signals using a trained machine learning model.

[0025] In a particular case of the implantable device, the one or more electrodes have the form of a cuff, spiral, or helix.

[0026] In another case of the implantable device, the device further comprising a flexible enclosure to support the one or more electrodes in position on or around the vessel.

[0027] In yet another case of the implantable device, the one or more electrodes comprises two or more electrodes separated by a predetermined distance.

[0028] These and other aspects are contemplated and described herein. It will be appreciated that the foregoing summary sets out representative aspects of the system, method, and device to assist skilled readers in understanding the following detailed description.DESCRIPTION OF THE DRAWINGS

[0029] A greater understanding of the embodiments will be had with reference to the Figures, in which:

[0030] FIG. 1 A illustrates an example form factor of an implantable device, in accordance with an embodiment, with two annular-shaped transducers;

[0031] FIG. 1 B illustrates another example form factor of the implantable device with four annularshaped transducers;

[0032] FIG. 1C illustrates another example form factor of the implantable device with two adjacent helical-shaped transducers with self-coiling properties;

[0033] FIG. 1 D illustrates another example form factor of the implantable device with two separated helical-shaped transducers with external enclosure;

[0034] FIG. 1 E illustrates another example form factor of the implantable device with a wrapped self-sizing enclosure with a plurality of transducers embedded on the inner surface;

[0035] FIG. 2A illustrates an example of the implantable device coupled to a device interface that includes electronic circuitry to relay the signals captured by the implantable device that includes a cuff-type transducer;

[0036] FIG. 2B illustrates another example of the implantable device coupled to the device interface where the implantable device includes a non-cuff-type transducer;

[0037] FIG. 2C illustrates another example of the implantable device coupled to the device interface by a direct connection via having the device interface embedded within, or directly in physical contact, the cuff-type transducer;

[0038] FIG. 2D illustrates another example of the implantable device coupled to the device interface by having the device interface embedded within an intra-vascular type transducer.

[0039] FIG. 3 illustrates a diagram of a system for determination of blood pressure using the implantable device, in accordance with an embodiment;

[0040] FIG. 4 illustrates results of example experiments showing in vitro testing of electrical signals generated by electrodes for a book-type cuff electrode (top) and a helical-type cuff electrode (bottom);

[0041] FIG. 5A is a chart showing sample data from in vivo example experiments that shows a raw electrovasculogram (EVG) signal;

[0042] FIG. 5B is a chart showing sample data from the in vivo example experiments that shows a filtered EVG signal;

[0043] FIG. 5C is a chart showing sample data from the in vivo example experiments that shows intra-arterial blood pressure;

[0044] FIG. 5D is a chart showing sample data from the in vivo example experiments that shows an electrocardiogram (ECG);

[0045] FIG. 6A is a chart showing recorded signals from the in vivo example experiments, illustrating measurement of time delay between the R-wave of the ECG and the EVG signal;

[0046] FIG. 6B is a chart showing recorded signals from the in vivo example experiments, illustrating an overlay of normalized EVG and intra-arterial blood pressure signals;

[0047] FIG. 7 is a chart showing correlation between pulse arrival times (PAT) obtained using a blood pressure signal (ordinate) and an EVG signal (abscissa) for example experiments;

[0048] FIG. 8 is a chart showing a relationship between mean arterial blood pressure and pulse arrival times obtained from an EVG signal for example experiments;

[0049] FIG. 9 is a chart showing a relationship between mean arterial blood pressure and pulse arrival times obtained from an intra-arterial pressure signal for the example experiments;

[0050] FIG. 10 is a flowchart for a method for determination of blood pressure using the implantable device, according to an embodiment;

[0051] FIG. 11 is a flowchart for one case of the method of FIG. 10 using a convoluted neural network (CNN) model;

[0052] FIG. 12 is a chart showing an example of a 227x227x3 image used as input to the CNN model of FIG. 11;

[0053] FIG. 13A and 13B illustrate charts, each from different example experiments, showing segments with a highest cross-correlation;

[0054] FIG. 14 illustrates two graphs visualizing the model predictions compared to true diastolic and systolic blood pressure values;

[0055] FIGS. 15A and 15B illustrate charts showing a sample of individual artifact signals recorded from in vitro example experiments, where the procedure was performed in de-ionized water for FIG. 15A and 0.9% saline for FIG. 15B;

[0056] FIGS. 16A and 16B illustrate charts showing a sample of artifact signals recorded from post-mortem example experiments, where the procedure was performed in de-ionized water for FIG. 16A and 0.9% saline for FIG. 16B;

[0057] FIG. 17 is a graph showing a comparison of signal amplitudes generated by changes in intraluminal pressure for the example experiments;

[0058] FIG. 18 is a diagram showing an experimental setup of further in vivo example experiments;

[0059] FIG. 19 is a flowchart showing an example approach for data preprocessing to classification for the further example experiments; and

[0060] FIG. 20 are graphs illustrating the effects of increasing the size of a training data set on the accuracy of the CNN model for the further example experiments.DETAILED DESCRIPTION

[0061] For simplicity and clarity of illustration, where considered appropriate, reference numerals may be repeated among the Figures to indicate corresponding or analogous elements. In addition, numerous specific details are set forth in order to provide a thorough understanding of the embodiments described herein. However, it will be understood by those of ordinary skill in the art that the embodiments described herein may be practised without these specific details. In other instances, well-known methods, procedures and components have not been described in detail so as not to obscure the embodiments described herein. Also, the description is not to be considered as limiting the scope of the embodiments described herein.

[0062] Various terms used throughout the present description may be read and understood as follows, unless the context indicates otherwise: “or” as used throughout is inclusive, as though written “and / or”; singular articles and pronouns as used throughout include their plural forms, and vice versa; similarly, gendered pronouns include their counterpart pronouns so that pronouns should not be understood as limiting anything described herein to use, implementation, performance, etc. by a single gender. Further definitions for terms may be set out herein; these may apply to prior and subsequent instances of those terms, as will be understood from a reading of the present description.

[0063] Any module, unit, component, server, computer, terminal or device exemplified herein that executes instructions may include or otherwise have access to computer readable media such as storage media, computer storage media, or data storage devices (removable and / or nonremovable) such as, for example, magnetic disks, optical disks, or tape. Computer storage media may include volatile and non-volatile, removable and non-removable media implemented in anymethod or technology for storage of information, such as computer readable instructions, data structures, program modules, or other data. Examples of computer storage media include RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by an application, module, or both. Any such computer storage media may be part of the device or accessible or connectable thereto. Further, unless the context clearly indicates otherwise, any processor or controller set out herein may be implemented as a singular processor or as a plurality of processors. The plurality of processors may be arrayed or distributed, and any processing function referred to herein may be carried out by one or by a plurality of processors, even though a single processor may be exemplified. Any method, application or module herein described may be implemented using computer readable / executable instructions that may be stored or otherwise held by such computer readable media and executed by the one or more processors.

[0064] Cuff electrodes have been extensively used over multiple decades as a device capable of chronically recording and electrically evoking peripheral nerve activity; for example, they are used extensively in vagus nerve stimulation (VNS) therapy for treatment in epilepsy, depression, pain management, cardiovascular disease, and more. While a cuff electrode can be used to collect information to continuously estimate blood pressure (BP), when used as a recording device, cuff electrodes are very susceptible to various types of large signal artifacts. These can be attributed to external electrical sources or movement of the electrode. When implanted around an arterial vessel, each pressure wave transmits a rapidly changing force through the arterial wall, resulting in mechanical changes in the electrode and the electrode-tissue interface. This signal is analogous to a pulse artifact commonly seen with respect to electroencephalograms (EEG). While this signal is undesirable in EEG analysis, the present inventors have determined that such pulse artifacts are in fact useful to predict BP.

[0065] Embodiments of the present disclosure provide an implantable, flexible pulse sensor (FPS) that is wrapped around an arterial blood vessel. In some cases, a wheatstone bridge circuit can be used to sense changes in pressure. The mechanism of the bridge circuit relies on pulsatile deformation of the flexible bridge circuit wrapped around an artery. Deformation of the bridge circuit causes a change in its resistance value which, in turn, generates a change in voltage output that mimics the change in arterial blood pressure. In this way, the implantable device of the present embodiments utilizes changes in the electric potential measured between electrodecontacts (which can be referred to as a transducer) that are embedded on the inner surface of a flexible cuff and is placed in contact to the surface of the arterial vessel. Changes in electrical potential are caused by movement of the electrode contacts relative to the blood vessel surface and / or flexion or movement of the insulating material (cuff or enclosure) relative to the wire.

[0066] Embodiments of the present disclosure generate an electrical signal corresponding to mechanical perturbations of arterial vessels (e.g., pressure pulses), which can then be recorded. This electrical signal can be referred to as an electrovasculogram (EVG). The EVG signal can be processed and analyzed, either alone or in combination with an ECG signal, electrical impedance, electromyogram, or the like, to predict one or more of the systolic (SBP), diastolic (DBP), and mean (MEP) arterial pressures; among other possible derivative measures of cardiovascular function. Higher order analyses of the recorded EVG may be performed to obtain additional information regarding autonomic function; for example, variability of heart rate or blood pressure and changes in organ function.

[0067] Various approaches can be used to predict blood pressure using machine learning to analyze physiological signals. For example, using support vector machines (SVM) from photoplethysmography (PPG) and ECG signals. In another example, using a nonlinear autoregressive model with exogenous input (NARX) to predict blood pressure using ECG signals. These approaches generally require complex signal processing and feature extraction and are very limited in their ability to provide continuous blood pressure estimates. Other deep learning approaches have been used to predict blood pressure using a convolutional neural network (CNN) with Hilbert-Huang transforms (HHT) to obtain spectral data of PPG as input to predict hypertensive and normotensive states. While this approach can achieve high accuracy for predicting blood pressure states, there are significant technical challenges associated with obtaining continuous blood pressure estimates. For example, implantable PPG sensors injected subcutaneously are generally subject to long-term performance problems, such as encapsulation and migration of the device.

[0068] An implantable device 50 is provided, in accordance with the present embodiments, to interface with an arterial blood vessel 10 used to measure blood pressure values. The implantable device 50 is sized to accommodate the location, size, and pulsatile motion of a targeted vessel 10. FIGS. 1A to 1 E illustrate example embodiments of the present invention showing various potential form factors of the implantable device 50 located on an outer surface of an arterial vessel 10. The implantable device includes one or more transducers 52, each comprised of an electrodecontact, to measure an electrical signal, as described herein. The transducers 52 can have any suitable shape or form, for example, cuff, spiral, or helical as illustrated in the embodiments of FIGS. 1A to 1 E. Mechanical deformations of the transducer 52 will generate an electrical signal that is associated with periodic pulsations of said vessel 10. It is understood that the mechanical deformations that generate the pulse artifacts can be due to movement of the electrodes relative to each other, changes in the contact pressure at an interface of the tissue and a respective electrode, and / or changes in the contact area at the interface of the tissue and the respective electrode.

[0069] The one or more transducers 52 may consist of materials with various suitable electrical conductivity properties; for example, high conductivity (gold, platinum, or platinum-iridium), intermediate conductivity (conductive hydrogel), or low conductivity (rubber). In other cases, the one or more transducers 52 may include piezoelectric materials that, when connected to an appropriate circuit, can generate EVG signals corresponding to mechanical movements caused by arterial pulsations.

[0070] The electrode contact transducers 52 can be made from rigid material or flexible material, as suitable to generate a readable artifact of the electrical signal, as described herein. As also described herein, in addition to receiving the electrical signal containing the artifacts, the electrode contact transducers 52 can also be used to deliver electrical stimulation to nervous tissue.

[0071] In addition to the one or more transducers 52, the implantable device 50 can also include a flexible enclosure 54, which can support the electrode contact transducers 52 in position and serve a number of purposes. The enclosure 54 can consist of electrically insulating or conducting material (e.g., silicone elastomer or thin metals) or other flexible materials that can accommodate the pulsatile movements of arterial vessels 10. The enclosure 54 can consist of multiple contiguous segments that have different electrically conductive materials, depending on the embodiment of the electrode.

[0072] In some cases, EVGs can be simultaneously measured by two or more transducers 52, which are each separated by a suitable distance. The example embodiments illustrated in FIGS. 1A, 1 B, and 1 D, show arrangements with multiple transducers 52. Arrangements with multiple transducers 52 can be used to provide directional information regarding the pressure pulse, used to determine a velocity of the arterial pressure pulse, used to determine pulse arrival time when combined with ECG. With respect to determining the velocity of the arterial pressure pulse, theinformation can be used to determine a pulse transit time (PTT), which is a time delay between two EVG signals recorded from two different electrode recording sites. With respect to determining pulse arrival time, the information can be used to determine PAT, which is a time delay between R-wave and EVG the signal. Further, arrangements with multiple transducers 52 could provide functional redundancy if one or more transducers were to become inoperable. Further, arrangements with multiple transducers 52 can provide an approach for recording ECG signals. Further, arrangements with multiple transducers 52 could provide an approach for detecting movement of a patient, such as using detected artifactual electromyography (EMG) signals. Further, arrangements with multiple transducers 52 could provide an approach for selectively reducing or eliminating electrical noise (e.g., signal artifact) that may be generated by electrical sources; such sources can include muscles or external electronic devices.

[0073] As shown in the example embodiments of FIGS. 1 A to 1 E, the transducers 52 can occupy discrete locations along the arterial vessel 10, or in other cases, can interface with extended regions along the arterial vessel 10. The transducers 52 can be in close physical contact with the vessel 10 or, in other cases, there may be a small gap between the outer wall of the vessel 10 and the transducer 52. The transducers 52 can have similar mechanical flexibility with that of the enclosure 54, or, in other cases, can have slightly different mechanical flexibility than that of the enclosure 54. In some cases, the enclosure 54 can be constructed as a hollow cylinder with an inner diameter that is equal to, or slightly larger, than the blood vessel (for example, similar to a cylindrical cuff). In other cases, the enclosure 54 can have self-sizing properties (for example, similar to a spiral or helical cuff) that enable the inner diameter of the cuff to adjust in response to pressure pulses flowing through the blood vessel. In some cases, such as the embodiment illustrated in FIG. 1 E, the transducer 52 can have self-sizing properties that adjust the diameter of the transducer 52 in response to pressure pulses flowing through the blood vessel. Electrode contacts may be strategically located on the external and internal surfaces of the self-sizing substrate, such that they create an overlapping area when the device is implanted on an artery. In this embodiment, the overlapping area of the electrode contacts (also causing mechanical friction) will change as arterial pressure pulses cause the self-sizing device to expand and then return to its resting state.

[0074] In some cases, there may be multiple implantable devices 50 that are each located at different locations along the same arterial vessel 10 or located on different vessels 10. Using multiple implantable devices 50 allows for the generation of multiple EVG signals with a predictable time difference (e.g., lag or lead time).

[0075] As illustrated in the example embodiments of FIGS. 2A to 2D, the implantable device 50 can be coupled to a device interface 60 that includes electronic circuitry to relay the signals captured by the implantable device 50 to an external computing device or system, such as the system described herein. Such electronic circuitry can include, for example, one or more amplifiers, one or more analog-to-digital (A / D) converters, a wireless communication module, one or more batteries, and the like. In some cases, the one or more batteries may include replaceable batteries or a rechargeable battery. The wireless communication module can be used to communicate, via wireless connection (e.g., radio-frequency, Bluetooth, or optical) with other devices, such as a communication interface 106, as described herein. In some cases, further modules or devices can be located intermediate to the device interface 60 and the communication interface 106, such as external relays 65 that can be placed on the skin surface or located within some proximity of to the skin or body. A connection between the implantable device 50 and the device interface 60 can be implemented in any suitable way; for example:• by a direct wire connection between the device interface 60 and the cuff-type transducer 52, as illustrated in FIG. 2A;• by a direct wire connection between the device interface 60 and a non-cuff-type transducer 52 electrode, as illustrated in FIG. 2B;• by a direct connection via having the device interface 60 embedded within, or directly in physical contact, the cuff-type transducer 52, as illustrated in FIG. 2C; or• by having the device interface 60 embedded within an intra-vascular (e.g., mesh) type transducer 52, as illustrated in FIG. 2D.

[0076] The implantable device 50 can be located at any location that is appropriate for obtaining signals. As an example, the implantable device 50 can be located at an upper extremity for continuously measuring EVG signals, where the transducer 52 is implanted on the radial or brachial arteries. This may allow efficient interfacing (e.g., charging and / or communication) with one or more external relays 65 embedded with a wearable garment placed around the wrist or upper arm. At the upper extremity, it may also allow the use of a smart watch or other type of wearable device that has a built-in ECG monitoring function. In another case, EVG signals may be obtained from locations that provide specific reference to the brain or its substructures (e.g., internal carotid artery) or may be obtained from locations that allow for effective simultaneous recording of EVG and ECG signals (e.g., subclavian artery). Selecting relatively superficial arterialvessels may provide additional advantages; for example: (1) providing minimally-invasive and less complex surgical procedures (both implant and explant), (2) providing more efficient wireless communication between the device interface 60 and the external relay 65, and (3) providing choice of different locations for implanting multiple devices.

[0077] In some cases, the external relay 65 can be used and placed on the skin surface by self- adhesive material or wearable apparatus, at a location sufficiently close to the device interface 60. The external relay 65 can be used to wirelessly charge the implanted battery, amplify signals generated by the device interface 60, collect data in lieu of a further device (e.g., tablet, smartphone, etc.), and program stimulation protocols for neuromodulation applications. The module may include antennae, processors, amplifiers, memory storage, and a power source. In certain cases, the external relay 65 may include a telecommunication component (e.g., 5G transceiver) that can be used to alert medical or hospital staff of critical changes in cardiovascular function of a patient.

[0078] Turning to FIG. 3, a diagram of a system 100 for determination of blood pressure using the implantable device 50, according to an embodiment, is shown. As shown, the system 100 has a number of physical and logical components, including one or more processors 102, data memory 104, a communications interface 106, a network interface 108, and a local bus 114 enabling the processor(s) 102 to communicate with the other components. In some cases, at least some of the one or more processors can be graphical processing units. The processor(s) 102 execute various modules, as described below, and potentially other software, such as an operating system. The communications interface 106 enables an administrator or user to provide input via an input device, for example a keyboard and mouse or a touchpad. The communications interface 106 can also output information to output devices to the user, such as a display and / or speakers. The communications interface 106 can also interact with other devices, such as the device interface 60 and / or other modules. The network interface 108 permits communication with other systems, such as other computing devices and servers remotely located from the system 100, such as for a typical cloud-based access model. The data memory 104 stores the computerexecutable instructions for implementing the functions of the modules, as well as any data used by these modules. Additional stored data can be stored in a database 116.

[0079] In an embodiment, the system 100 includes a number of conceptual modules to be executed on the one or more processors 102, including a signal processing module 120, an analysis module 122, and an output module 124. In further cases, the functions and / or operationsof the modules can be combined or executed on other modules, or executed on remote or cloud based systems.

[0080] The present inventors determined that an electrical artifact is generated by the electrode transducer 52 when a mechanical deformation is experienced. In an example experiment, in a saline bath, a bipolar cuff electrode was fully immersed while the lead wires were connected to a low-noise differential amplifier. The signals were displayed and digitally stored. A non-conducting rod was used to mechanically deform the cuff electrodes by periodically applying light pressure for brief time periods. As illustrated in the charts of FIG. 4, which demonstrate results of example experiments, these deformations generated electrical artifact signals (denoted with an arrow) that exhibited transient changes in electric potentials, with peak-to-peak signal amplitudes in the order of 5 to 10 mV.

[0081] Turning to FIG. 10, a flowchart for a method 200 for determination of blood pressure using an implantable device, according to an embodiment, is shown. With the implantable device 50 implanted around a suitable vessel, at block 202, the signal processing module 120 receives electrical signals from the one or more electrode contact transducers 52 via the device interface 60 and the communications interface 106. The signals received represent electrovasculogram (EVG) with electrical artifacts due to the deflection caused by the expansion and / or contraction of the electrode. In some cases, the EVG signal can be obtained by differentially recording an electric potential difference between two of the electrode contacts 52. In this way, rapid positive deflection corresponds to an arterial pressure pulse passing through the particular electrode contact transducer 52.

[0082] Accordingly, the EVG is a signal that captures the pulse artifact by recording electrical activity from a cuff electrode implanted on or around a pulsating arterial vessel. It should be noted that the terms artifact signal, pulse signal artifact, electrical artifact, BP artifact, or the like, can be used interchangeably. Generally, the pulse artifact refers to the electrical signal generated by the electro-mechanical interaction between the cuff electrode and a pulsating artery. These signals (i.e. , EVG) capture several features that are present in blood pressure signals obtained from, for example, an intra-arterial pressure sensor. This includes maxima and minima that correspond to systolic and diastolic pressures, respectively. Short multi-phasic components which are consistent with the dichrotic notch can also be observed in the EVG.

[0083] At block 204, the signal processing module 120 digitizes the signals received from the one or more electrode contact transducers 52, using any suitable approach, into distinct temporal segments. In some cases, prior to, or after, digitization, the signal processing module 120 can condition the signal processing module 120; for example, by filtering and / or amplification. In an example, the signals can be filtered with a bandpass of 5 Hz to 20 Hz in order to provide the best signal while minimizing high frequency noise. In some cases, the digitized signals can be transformed into RGB (Red-Green-Blue) images for input into a machine learning model, as exemplified herein.

[0084] At block 206, the analysis module 122 provides the digitized signals as input to a machine learning model trained to predict systolic and diastolic blood pressure values from the inputted signals. In a particular case, the machine learning model can comprise a convolutional neural network (CNN) using the Hilbert-Huang Transform (HHT). The CNN architecture can include several layers that perform convolution, pooling, and fully connected operations. The CNN framework can learn, through training, underlying features of the signal artifacts that are associated with blood pressure changes. In other cases, the machine learning model can be any suitable type of model, for example, a neural network sequential model or a random forest regressor model.

[0085] In an example, as illustrated in the flowchart of FIG. 11 , the signal processing module 120 can digitize the received signal by splitting the data into successive temporal segments, for example, split into successive 1.135 second segments. With samples at 1.135 seconds in length, in the time domain, there are approximately three pressure pulse artifacts per sample. Each temporal segment of the EVG signal can be further pre-processed by down-sampling, for example, to 200Hz; as may be required for input to the CNN model.

[0086] HHT can be performed on intrinsic mode functions of the down-sampled signal in order to decompose the non-stationary time signal into multiple signals to output frequency spectral data. If one assumes that the EVG signal can change due to multiple factors (e.g., blood pressure, sympathetic tone, etc.), then the spectral plots at each time interval provides an approximation of the EVG; which can be provided for the first and second derivatives as described herein.

[0087] In some cases, in order to provide a matrix input to the machine learning model, such as in the form of an RGB image, first and second derivatives of the signal can be determined by the signal processing module 120 to generate second and third dimensions of the matrix. The firstand second derivatives correspond to velocity and acceleration of the change in arterial pressure, respectively; which are generally important physical characteristics of a pressure pulse traveling through an artery. In some cases, the matrix values can be normalized per dimension of each derivative by the signal processing module 120, and in some cases, can also be zero-padded. In this way, each signal within the matrix can be normalized with respect to the maximum value within each sample. As illustrated in an example shown in FIG. 12, a 227x227x3 RGB image can be used as input to the machine learning model; whereby the input data consists of spectral values of the raw values (red), first derivative values (blue), and second derivative values (green) processed from the received signal.

[0088] Training the machine learning model can include a split of training data and testing data; for example 70% training data to 30% testing data.

[0089] At block 208, the output module 124 outputs the systolic and diastolic blood pressure values for each temporal segment. In some cases, to a user via the communications interface 106, or in other cases, to the database 116 or other systems via the network interface 108.

[0090] In example experiments, analysis of data received from the implantable device 50 in anesthetized rats showed that accurate prediction of systolic and diastolic arterial blood pressures (Grade A or B), as shown in the results of TABLE 1. The example experiments used criteria for hospital grade blood pressure measuring devices. In these experiments, the percentage of blood pressure values predicted by the CNN-HHT model exhibited estimation errors of < 5mmHg, < lOmmHg, < 15mmHg in at least 50%, 75%, and 90% of samples, respectively (i.e. , grade B). In experiments 1 , 4 and 9, it is noted that the systolic blood pressure values predicted by the CNN- HHT model were assessed as grade C (i.e., estimation errors of < 5mmHg, < lOmmHg, < 15mmHg were achieved in at least 40%, 65%, and 85% of samples, respectively).TABLE 1

[0091] Using the same sampled data from the above example experiments, blood pressure was also independently estimated. A Catch22 (CAnonical Time-series Characteristics) algorithm was applied to each temporal segment to obtain a set of feature vectors (for example, 24 features) to classify different features of the time-series EVG signal, which can then be used in the machine learning model. A subset of such features was selected, which were deemed relevant to the analysis: (1) standard deviation of the time series data, (2) variability of the time series data, and (3) asymmetry and characteristics of the distribution of variations in the time series data. A neural network sequential model was used; which was composed of one input dense layer (64 nodes), 11 hidden dense layers (3 layers with 256 nodes; and the remaining layers with 128 nodes), and one dense output layer (2 nodes to represent systolic and diastolic pressures, respectively). The neural network sequential model was trained on the three selected features, using 100 epochs with batch size set to 12 and validation split to 20%. The neural network sequential model performed predictions on test sets in the example experiments and outputted a mean-square- error (MSE) and estimated blood pressure values. As shown in TABLE 2, these example experiments yielded improved estimates of systolic and diastolic pressure values. In nine animals (90% of experiments), a BHS grade of A was achieved for predicted both systolic and diastolic pressures.TABLE 2

[0092] The implantable device 50 may be surgically implanted to target a specific organ by placing the transducer electrode contacts 52 on a single artery or placed on multiple arterial blood vessels entering an organ. The implantable device 50, at such locations, can be used to measure arterial blood pressure and / or biomarkers of cardiovascular function, in order to provide information regarding the physiological function of said organ or a biological system related to and including said organ. In some cases, the determined information of physiological function can be used to provide closed-loop neuromodulation therapy; where electrical nerve stimulation can be applied using the same electrode contacts of the implantable device 50 (e.g., where such contacts are located around a nerve in conjunction with the vessel) or using additional electrode contacts. Autonomic nerve fibers located in close proximity to the stimulating electrode contacts (e.g., fibers embedded or traveling superficially along an arterial vessel) may be stimulated to achieve different functional outcomes, depending on a clinical indication being addressed. As one example, the implantable device 50 that includes a cuff electrode may be implanted on a renal artery to predict (continuously or periodically) arterial blood pressure via simultaneous measurement of the EVG and ECG signals. One or more of these signals can provide a measure of arterial blood pressure, which may include systolic, diastolic, and mean pressures, that in turn can trigger delivery of electrical pulses to renal nerve fibers. Electrical pulses or signals may be delivered to down-regulate renal nerve activity (e.g., DC current, high frequency stimulation above 1 kHz, or radiofrequency ablation) or to up-regulate renal nerve activity (e.g., trains of biphasic pulses delivered between 1 Hz and 500 Hz).

[0093] In some cases, where the transducer electrode contacts 52 are also used to provide electrical stimulation, such contacts may be configured to elicit nerve action potentials that propagate in a unidirectional manner. Unidirectional activation may include all axons orsubpopulations of axons within a targeted renal nerve, where the electrically stimulated fibers may consist of specific types of axons (e.g., myelinated or unmyelinated fibers). As a consequence, up-regulation of renal nerve activity may be achieved by generating action potentials that are directed in either the efferent (towards the kidney) or afferent (towards the central nervous system) directions. Using this unidirectional stimulation, down-regulation of renal nerve activity may alternatively be achieved by a collision blocking technique, where physiological efferent renal nerve activity is reduced or prevented from reaching the kidney by electrically activating action potentials that travel in the afferent direction; i.e., opposite direction. Similarly, physiological afferent renal nerve activity may be blocked or reduced by electrically eliciting efferently-directed action potentials. The use of electrical nerve stimulation delivered to the renal nerve may be used to modulate one or more physiological functions that are modulated by the kidney, such as blood pressure, diuresis, sodium excretion / reabsorption, or ureteric peristalsis. The location of an implanted electrode on the renal artery may be further advantageous for simultaneous recording of ECG signals that can be picked up by the transducer electrode contacts 52 (i.e., PAT or PTT). It should be understood by a person skilled in the art that closed-loop neuromodulation approaches can similarly be used; for example, instrumenting arteries or arterial vessels supplying internal organs (such as the liver, spleen, heart, pancreas, gall bladder, stomach, large intestine, small intestine, bladder, or lungs) where the autonomic nerve fibers or ganglia are located within or in close proximity to the arterial vessel wall.

[0094] In an example of a closed-loop electrical neuromodulation approach for controlling blood pressure (BP), transducer electrode contacts 52 in the form of a nerve cuff can be implanted around a cervical vagal neurovascular bundle located in the neck region. The nerve electrode can be used to record vagal nerve activity that corresponds to cardiac function and electrical stimulation can be applied to the vagal nerve fibers to modulate cardiac function. Vagal nerve stimulation can be used to modulate cardiac function or treat cardiac related conditions, such as heart failure. The implanted device can be used to deliver electrical stimulation to vagal fibers to treat, for example, hypertension or heart failure.

[0095] As an example, the helical cuff electrode of the present embodiments can be placed at the carotid sinus to measure and predict blood pressure. Subsequently, electrical stimulation can be applied to baroreceptor nerves at the carotid sinus to cause a reflex reduction in blood pressure, and thus, downregulate blood pressure. Or alternatively, used to treat heart failure or other medical conditions related to cardiovascular regulation. In this way, blood pressure can be monitored continuously by measuring and analyzing the EVG signal and then electricalstimulation therapy can be delivered for treating heart failure. For example, if a patient exhibits low blood pressure where a systolic or diastolic value falls below a certain threshold, then electrical stimulation can be titrated down; and if a patient exhibits high blood pressure (e.g., systolic greater than 160 mmHg) then electrical stimulation can be titrated up.

[0096] The present inventors conducted example experiments to verify the substantial advantages of the present invention. Particularly, recording of the electrical artifact signals - the electrovasculogram (EVG) - were tested in preclinical studies involving anesthetized rats; where an electrocardiogram (ECG) (lead II configuration), arterial blood pressure (intra-arterial catheter inserted into the left carotid artery), and EVG (cuff electrode implanted around the right carotid artery) were simultaneously recorded. In this example experiment, the implanted electrode consisted of four platinum wire electrical contacts that were located along the inner surface of a cylindrical silicone cuff. An EVG signal was obtained by differentially recording the electric potential difference between two contacts of the electrode (e.g., contacts one and three). The electrical ground was connected to the fourth electrode contact. After conditioning the signal (in this case, filtering and amplifying the signal), the EVG recording was identified by a rapid positive deflection corresponding to the arterial pressure pulse passing through the cuff electrode.

[0097] Sample EVG data are shown in FIGS. 5A and 5B. The arterial pressure pulse artifact is shown as the raw EVG, which is then filtered (frequency range = 5 Hz to 20 Hz). FIG. 5C shows the corresponding intra-arterial blood pressure signal measured from the contralateral carotid artery, with the ECG shown in FIG. 5D. As shown in FIG. 6A, the EVG signal provides an electrical measure of each pressure pulse that occurs following every ventricular contraction (i.e. , R-wave of the ECG). The time delay between a ventricular contraction and the arrival of the pressure pulse at the cuff electrode (i.e., PAT) can be quantitively expressed by either (1) the peak arrival time (the time delay between R-wave peak and the positive peak of deflection in the EVG) or (2) the trough arrival time (the time delay between the R-wave peak and the minimum value immediately before the positive deflection in the EVG). As shown in FIG. 6B, there is a strong resemblance in the waveform between the EVG signal and the actual carotid arterial pressure signal (labeled as ‘Blood Pressure’), which indicates that the peak of the EVG signals corresponds to the systolic peak of the blood pressure and the trough of the EVG signal corresponds to the diastolic peak of the blood pressure. In this example, the differential voltage between contacts one and three were used to measure the EVG signal.

[0098] In the example experiments, a correlation between the signal received from the implantable device 50 and the actual blood pressure signal was assessed by plotting the trough arrival time in comparison to the blood pressure (diastolic peak) arrival time. As shown in FIG. 7, there was a strong positive correlation between the two variables, indicating that the signal received from the implantable device 50 provides a close approximation of the pressure pulse traveling through the implanted cuff electrode. As depicted in FIG. 8, analysis of the mean arterial pressure compared to the trough arrival time showed that the pressure pulse velocity increased as the mean arterial pressure was increased. The same relationship between the mean arterial pressure and the actual blood pressure (diastolic peak) arrival time is shown in FIG. 9; which provides validation that the received signal can be used as a measure of arterial blood pressure when applied to techniques such as PAT, PTT, or machine learning approaches to approximate systolic, diastolic or mean pressure values.

[0099] Further example experiments illustrated that EVG signals recorded from the carotid artery showed similarities to blood pressure waveforms and were able to capture essential cardiovascular features like the dicrotic notch. In such experiments, animals were anesthetised and following a ventral incision along the neck, the left vagal neurovascular bundle was dissected, and the carotid artery was catheterized to record continuous BP. The right vagal neurovascular bundle was then dissected and the carotid artery was separated from the vagus nerve. A tetrapolar nerve cuff with an inner diameter of 1 ,5mm was implanted solely around the right carotid artery. The cuff has 250pm platinum contacts with 3mm contact spacing, and 1 mm edge spacing. Three needles were inserted intramuscularly to provide recordings for a lead II ECG configuration. A bipolar configuration was used where configuration A used a differential voltage between contacts 2 and 3 and a configuration B used the differential voltage between contacts 1 and 3; and where contact 4 was used as a ground for both recordings. The EVG signals were digitally filtered with a 5-20Hz bandpass filter prior to processing. The left arterial catheter was connected to a pressure transducer that was linked to a bridge amplifier and the ECG signal was conditioned through a separate amplifier. Data segments were downsampled from 10 kHz to 1 kHz to allow for efficient processing and analysis of the data while preserving essential information.

[0100] In these example experiments, to ensure consistent measurements the ECG and EVG signals were normalized. Peak analysis was used to precisely count the heartbeats. A comparison of the number of heartbeats in both signals was carried out to ensure validity of the signals. Peak analysis was also performed on BP recordings to acquire systolic and diastolic pressure measures along each heartbeat. To facilitate efficient analysis and modeling, the EVG data wassegmented into smaller, more manageable samples. Segments of 20 seconds were utilized for validation testing, allowing for a comprehensive evaluation of the signal. For the subsequent implementation of the CNN framework, segments of 1.135 seconds were selected, optimizing the computational efficiency while capturing essential features of the data. Importantly, data segments demonstrating a strong correlation of 0.8 or higher with the BP signal were retained for further analysis.

[0101] In these example experiments, validation was conducted to assess the consistency and correlation between the EVG signal and BP measurements. In a first approach, a comparative analysis of the pulse arrival time (PAT) derived from both the BP waveform and the EVG signal. The PAT is defined as the time delay between the peak of the R-wave in the ECG and arrival of the corresponding pressure pulse at a distal location. In this context, the local minimum point of the BP waveform (diastolic pressure) and the first trough in the EVG signal served as the respective landmarks for calculating PAT. By comparing the PAT values obtained from both signals, the example experiments were able to evaluate the degree of synchronization and consistency between the EVG signal and direct BP measurements. A high level of correlation in the PAT values would suggest a strong association with and utility of the EVG signal as a surrogate for estimating BP changes. Secondly, the cross-correlation between the EVG signal and the BP waveform was investigated to further validate the EVG signal. Cross-correlation is a statistical measure that assesses the similarity and temporal alignment between two signals. A high cross-correlation coefficient indicated a strong correlation and alignment between the two signals, showing that the EVG signal captured relevant BP artifacts. Additionally, by examining the time lag between the signals, insights can be gained into potential delays or discrepancies, providing valuable information about the propagation and transmission of pressure pulses within the vascular system.

[0102] Through the evaluation of PAT and cross-correlation, the example experiments validated that the EVG signal is a reliable and informative indicator of BP artifacts, establishing the consistency and temporal relationship between the EVG signal and direct BP measurements.

[0103] In these example experiments, the EVG samples were processed to be represented as red-green-blue (RGB) images and fed into a CNN model; in this case, AlexNet; which requires a 227x227 image size. To satisfy this requirement, the EVG data was downsampled from 10000Hz to 200Hz, then split the data into 1.135 s (227-^2) samples. This resulted in a 1x227 vector. Next, HHT was performed on the vector to generate a 101x227 spectral matrix and then zero paddedto result in a 227x227 matrix. These steps were also performed on the first and second derivative of these EVG sampled data to obtain the green and blue dimensions, respectively, in the final image. Finally, the values in all three matrices were normalized and then combined for a 227x227x3 RGB image.

[0104] The CNN architecture featured interconnected layers tailored for robust feature extraction and representation. The input layer used preprocessed EVG signal images with dimensions of 227x227 pixels and 3 color channels. Successive convolutional layers with ReLLI activation function learned hierarchical features within the data, followed by max pooling to reduce spatial dimensions and capture dominant patterns. Global average pooling condensed extracted features, while two dense layers captured intricate relationships and nonlinearities. The final dense layer produced two continuous outputs corresponding to systolic and diastolic blood pressure predictions.

[0105] For optimization, the Adam algorithm was employed, dynamically adjusting learning rates during training to enhance convergence efficiency and generalization. Adam, short for adaptive moment estimation, dynamically adjusted learning rates during training. This adaptability facilitated the optimization process, enabling efficient convergence and effective generalization, which is useful for accurate predictions. Root mean squared error (RMSE) served as the loss function, quantifying the average discrepancy between predicted and actual blood pressure values. To ensure the model's robustness and generalization capacity, 70% of the data was used for training (with a 20% validation split), and 30% was used for testing. This division allowed for effective model training, hyperparameter tuning based on validation performance, and unbiased model evaluation on the test set.

[0106] The validation of the EVG signal with respect to BP was conducted to assess the relationship between the PAT derived from the EVG signal and the PAT measured from BP. FIGS. 13A and 13B illustrate ten plots, each representing a segment from a different experiment, showcasing the segments with the highest cross-correlation between the EVG and BP signals. These plots demonstrate visually strong correlations between the two signals, with correlation coefficients ranging from 0.93 to 0.99. To further explore the relationship between the EVG and BP signals, PAT analyses were conducted. The PAT values derived from BP and EVG fell in similar ranges, where the mean PAT for each group differed by 2.7ms.

[0107] Predictions were comprehensively tested against the grading standards set by the British Hypertension Society (BHS) for BP devices to evaluate the system’s performance. The EVG signals recorded during these experiments exhibited a voltage range of 10-200pv, capturing a wide range of signal amplitudes. All models in the example experiments achieved hospital-grade accuracy in predicting diastolic pressure, while 7 out of 10 experiments satisfied the requirements for systolic pressure predictions.

[0108] FIG. 14 illustrates graphs visualizing the CNN predictions compared to true diastolic (top) and systolic (bottom) blood pressure values. This model was trained and tested on data from a first experiment and achieved grade B and C prediction accuracies for diastolic and systolic recordings respectively.

[0109] Furthermore, to assess the generalizability of the models, additional analysis was performed. A CNN was trained on the combined data from the seven experiments that achieved hospital-grade BP predictions for both systolic and diastolic pressures. The training dataset comprised 34,653 samples, representing approximately 655 minutes of data. The systolic BP ranged from 44 to 153 mmHg with a standard deviation of 15 mmHg, while the diastolic BP ranged from 25 to 103 mmHg with a standard deviation of 15 mmHg. To ensure an unbiased representation of the training dataset, the data were randomized while maintaining an equal distribution of BP ranges. The training-testing data split were manipulated to maintain an even distribution of BP ranges during training. Notably, even when models were trained on 45% of available data, they achieved prediction accuracies that met the hospital-grade criteria according to the BHS standards.

[0110] The present inventors conducted further example experiments to verify the substantial advantages of the present invention. In these experiments, EVG signals were simulated by artificially replicating the mechanical interaction between a recording cuff electrode and a pulsating blood vessel. Two different approaches were used, where a bipolar self-coiling helical cuff electrode (inner diameter » 1 mm) was implanted on (1) a custom fabricated silicone elastomeric tube (outer diameter » 1mm) and (2) the abdominal aorta (outer diameter » 1 mm) in post-mortem rodents. In both cases, the tube / aorta was catheterized with polyethylene (PE50) tubing which was then connected in series with a pressure transducer and syringe. Individual boluses were applied by briefly depressing the plunger such that the intraluminal pressure increased from 80 mmHg to 120 mmHg (approximation of systole of the cardiac cycle). The signalwas recorded by connecting the two lead wires to a low-noise preamplifier and digitally storing the data on a computer.

[0111] In vitro experiments were performed (n = 4) by affixing the silicone tube to the bottom surface of a petri dish (90 mm x 15 mm) and filling it with water. This enabled the implanted electrode to be fully immersed within the solution. Artifact signals corresponding to the systolic phase of an arterial pulse were generated by initially setting the starting (diastolic) pressure at 80 mmHg (using an air-filled syringe). This caused the silicone tube to distend such that its outer diameter approximated the inner diameter of the electrode. Manual boluses were then applied such that the intraluminal pressure increased to approximately 120 mmHg. The electrode was immersed in 0.9% saline and in de-ionized water.

[0112] FIGS. 15A and 15B illustrate a sample of artifact signals recorded from in vitro experiments, where a series of 5 boluses were applied to increase the intraluminal pressure from 80 to 120 mmHg. Individual traces of the recorded voltage are plotted for the procedure performed in de-ionized water for FIG. 15A and 0.9% saline for FIG. 15B. The dashed segment within each line represents the period of pressure change, whereas the solid line corresponds to when the target (systolic) pressure of approximately 120 mmHg was maintained.

[0113] As shown in FIGS. 15A and 15B, the artifact signal exhibited a rapid rise during the bolus (dashed segment), which lasted up to 0.3 ms, and was followed by a gradual return to baseline. In this example, the peak amplitude of the artifact signal was notably larger when the procedure was performed in de-ionized water (0.12 ± 0.02 mV) as compared to 0.9% saline (0.08 ± 0.01 mV); and subsequently returned to baseline.

[0114] Artifact signals were also simulated in post-mortem experiments (n = 4), where a selfcoiling helical electrode was implanted on the surgically dissected abdominal aorta in euthanized rats. The implanted electrode was immersed in fluid by filling the surgically exposed area with either deionized water or 0.9% saline. The syringe used to provide boluses was also filled with the same fluid used to fill the abdominal cavity (i.e. , fully immerse the implanted electrode). The sample data, illustrated in FIGS. 16A and 16B, show a similar effect; where the amplitude of the artifact is markedly larger in de-ionized water (0.12 ± 0.01 mV) than in 0.9% saline (0.04 ± 0.01 mV). During the bolus (dashed segment), it is noted that the response to the pressure changes is smaller in amplitude, when compared to the silicone tube, shown in FIGS. 15A and 15B. This isattributed to the difference in the physical and / or anatomical properties of a silicone tube and a post-mortem abdominal aorta.

[0115] FIGS. 16A and 16B show a sample of artifact signals recorded from post-mortem experiments, where a series of 5 boluses were applied to increase the intraluminal pressure from 80 to 120 mmHg. Individual traces of the recorded voltage are plotted for the procedure performed in de-ionized water for FIG. 16A and 0.9% saline for FIG. 16B. The dashed segment within each line represents the period of pressure change, whereas the solid line corresponds to the target (systolic) pressure of approximately 120 mmHg.

[0116] FIG. 17 shows a comparison of the signal amplitudes generated by changes in intraluminal pressure (80 mmHg to 120 mmHg), where the silicone tube exhibited larger signals than the abdominal aorta in post-mortem experiments (n=4 in both groups). In de-ionized water, the signal amplitudes were 0.34 ± 0.15 mV and 0.12 ± 0.02 mV for the silicone tube and abdominal aorta, respectively. Similarly, in 0.9% saline, signal amplitudes were 0.05 ± 0.01 mV and 0.03 ± 0.01 mV for the silicone tube and abdominal aorta, respectively.

[0117] As depicted in FIG. 17, experiments performed using de-ionized water generated markedly larger artifact signals when compared to 0.9% saline. This was observed in both types of experiments, suggesting that the electrical properties of the de-ionized water can provide a mechanism for increasing the amplitude of the recorded signal. There was also a measurable difference in the signal amplitude when comparing the different tubular structures. Distension of the silicone tube yielded larger artifact signals when compared to the abdominal aorta, in both saline and de-ionized water. These results suggest that modifications to the recording electrode could further improve the amplitude of the artifact signal when measured in vivo. The electrode contacts could be coated with material(s) that approximate the electrical properties of de-ionized water (e.g., hydrogel, porous layer) and electro-mechanical properties of electrically insulating material (e.g., silicone elastomer, polymer, etc).

[0118] The present inventors also investigated the use of other types of machine learning models; and in particular, a random forest regressor model. In such example experiments, ten male Sprague-Dawley rats (weight = 450-700 g, Charles River) were anesthetized under isoflurane. Following a ventral incision along the neck, the left vagal neurovascular bundle was blunt dissected, and the carotid artery was catheterized to record continuous BP. The right vagal neurovascular bundle was then dissected and the carotid artery was separated from the vagusnerve. A cylindrical cuff electrode with plantinum contacts (250 m) and 3mm contact spacing and 1mm edge spacing was implanted around the right carotid artery. The EVG signal was differentially recorded between pairs of contacts that were connected to a low-noise pre-amplifier. A differential voltage was determined between two contacts (2 and 3) in configuration A, and differential voltage was determined between two contacts (1 and 3) in configuration B. The most caudal electrode contact (contact 4) served as the electrical ground for both recording configurations. The EVG signal was digitally filtered using a 5-20 Hz bandpass filter.

[0119] FIG. 18 illustrates the experimental setup of the above example experiments. Item (A) is a nerve cuff electrode implanted on the right carotid artery used to measure the electro-vascular- gram (EVG). Item (B) shows catherization of the left carotid artery for continuous blood pressure monitoring. Item (C) shows heparinized saline dispensed every hour. Item (D) shows a lead II ECG recording. Item (E) shows isoflurane adjusted to control the blood pressure. Finally, item (F) is a schematic diagram of the cuff electrode with contacts labelled 1 to 4.

[0120] The catheter inserted into the left carotid artery was connected to a pressure transducer that was linked to a bridge amplifier. The ECG was measured using a lead II configuration, where the signal was conditioned through an amplifier. Arterial BP was manipulated during the example experiment by adjusting the percent concentration of the inhaled isoflurane. This was done in accordance with established art that showed a negative correlation between the anesthetic dose and arterial BP. This allowed the recording of EVG under hypotensive, hypertensive, and baseline conditions.

[0121] Collected EVG signals were visually inspected to identify segments that exhibited artifacts corresponding to arterial pressure. Selected segments were extracted and down-sampled from 10 kHz to 1 kHz to allow for better computational performance while still preserving important information about the signal. The ECG and EVG were normalized to eliminate any potential signal drift. The ECG was used to ensure each heartbeat (R-wave) preceded the BP waves and EVG signals in the extracted data segments. The systolic and diastolic pressure was also extracted to obtain beat-by-beat BP values. EVG data was segmented into 20s windows for validation as it allowed for more efficient analysis (i.e., smaller, more manageable sizes of the data) and for better evaluation of the EVG with the BP.

[0122] The EVG signal was validated with respect to the intra-arterial BP signal using a crosscorrelation analysis, which assessed the similarity and temporal alignment between the twosignals. A high cross-correlation coefficient indicated a strong alignment between the EVG and BP signals, showing that the EVG can capture relevant features of the BP signal.

[0123] The EVG signal was used to predict the systolic and diastolic BP by examining the signal using a feature set known as Catch22. This set of features (summarized in Table I) has shown to achieve best on average performance for a variety of timeseries data. The extracted features were then fed into a random forest regressor. The Catch22 algorithm was applied (using the 24 feature implementation) to extract 24 features. These features (Table I) were then used as input into a random forest regressor implemented using RandomForestRegressor from sklearn in Python. The hyperparameters were tuned using a training set specific to systolic BP. The best parameters from the sweep were then used to train random forest regressor models to predict systolic BP, diastolic BP, and systolic BP and diastolic BP simultaneously. The data was split into 70% training and 30% testing within each rat. FIG. 19 illustrates a flowchart of an example approach used in the examples experiments.

[0124] In another approach using a CNN, the data was further processed by finding the HHT to generate a 101x277 spectral matrix which was then zero-padded on both sides to form the final 227x227 matrix. Then to mimic the 3 channels of an image the first and second derivatives of the EVG were also used following the same steps described above. This resulted in a 227x227x3 matrix which had its values normalized. The CNN was developed in Keras using a Tensorflow backend. The input was a 227x227x3 image, and this input layer was followed by 2 blocks consisting of a convolutional layer and a max-pooling layer. The first block used a 16-filter convolutional layer followed by a 2x2 max-pooling layer, whereas the second block used a 32- filter convolutional layer followed by a 2x2 max-pooling layer. The output of the second block was then sent to a global-average layer followed by 2 fully connected layers (both 64 nodes). The output was then a regression value of the systolic and diastolic BP. The activation function for each layer was ReLU and optimization was done using the Adam algorithm and a root-mean- squared error loss function. The CNN architecture is summarized in Table II. The data was divided into 70% training and 30% testing sets within each rat. The training data in this case was also further split by using 20% of this data as a validation set.

[0125] Performance in both approaches within each animal were evaluated using the grading standards set by the British Hypertension Society (BHS) for BP devices. This standard categorizes estimation into four grades: Grade A requires (60,85,95) % of estimations to be within (5,10,15) mmHg or less, respectively. Grade B requires (50,75,90) %, Grade C requires(40,65,80) % and Grade D is any result worse than Grade C. Mean absolute error was also provided for the best performing algorithm using EVG compared to that of the BP signal. This was to provide an indication of how close the values would be from the observed BP values.

[0126] The top 5 features were tracked from each rat to determine which features contributed the most to the classification. The top 3 most frequently occurring features were then investigated to determine whether fewer features can be used without compromising performance. Such an approach would be especially relevant for instances where the system 100 is embedded and an overall reduction in the number of features and algorithm complexity and power consumption is desired. The top3 features were inputted into a neural network (64 x 128 x 2 nodes) for predicting the systolic BP, diastolic BP, and systolic BP and diastolic BP simultaneously. The input and hidden layers had a ReLU activation function. The hidden layers used an L1 regularized with a regularization factor of 0.01. Optimization was done using Adam with a learning rate of 0.001 and a mean-squared error loss function was used. Training and testing was done on a 70-30 split, with 20% of the training being further split to form a validation set.

[0127] A CNN was trained on the combined data from the example experiments and achieved hospital-grade BP predictions for both systolic and diastolic blood pressure (7 / 10 experiments). This training data was comprised of 34,653 training samples, with the range of systolic BP from 44 to 153 mmHg (standard deviation: 15mmHg) and diastolic BP ranging from 25 to 103 mmHg (standard deviation: 15 mmHg). Testing was performed using subset(s) of the training set to determine the minimum amount of data needed to achieve hospital grade performance.

[0128] Additional analysis was performed to assess the generalizability of the machine learning approaches. Using the catch22 approach, a leave-one-rat-out approach was performed to observe if data from other rats could predict blood pressure on the rat of interest.

[0129] As shown in Figure 3, there was a high degree of alignment between the minimum / maximum peaks of the EVG signal and the diastolic / systolic peaks of the BP signal, respectively. The cross-correlation analysis showcases a high degree of correlation with coefficients ranging between 0.93 to 0.99. It is noted that the EVG signal was able to capture the dichroitic notch that follows the systolic peak typically seen in the BP signal.

[0130] FIGS. 13A AND 13B shows a comparison between samples of normalized EVG and BP signals that were obtained from each animal in the example experiments. The windows with the highest cross-correlation between normalized EVG and BP signals from each experiment areshown. As shown in FIGS. 13A AND 13B, there was a high degree of alignment between the minimum / maximum peaks of the EVG signal and the diastolic / systolic peaks of the BP signal, respectively. The cross-correlation analysis showcases a high degree of correlation with coefficients ranging between 0.93 to 0.99. It is noted that the EVG signal was able to capture the dichroitic notch that follows the systolic peak typically seen in the BP signal.

[0131] Both machine learning approaches were able to predict computationally the systolic and diastolic BP values for each sample of the testing data set. FIG. 14 shows visualizations of the prediction for the CNN (A,B) and random forest regressor (C,D) compared to the true diastolic and systolic blood pressure values. In this example, the CNN model achieved grades B and C and the random forest achieved grades A and A. When compared to the CNN method (A,B), the random forest model achieved BP predictions with markedly smaller errors, as depicted by the narrower bands of sample points (C,D).

[0132] The differential accuracy of the two approaches is presented in Table 3, where BHS-based assessment of SYS and DIA pressure predictions both achieved a grade A in 30% and 100% of animals using CNN and random forest methods, respectively. In three animals, the CNN model exhibited a grade C in the accuracy of predicting SYS pressure. It is noted that the size of the sample data collected from across experiments ranged from 24 - 178 mins, and the BP values (systolic and diastolic) exhibited a difference of at least 16 mmHg between their maximum and minimum values.TABLE 3

[0133] The random forest applied to the BP signal (control) showed the theoretical ceiling of the performance. The prediction means and standard deviations of systolic: 98.7 ± 2.1 , 99.8 ± 0.3, 99.9 ± 0.1 and diastolic: 98.4 ± 2.5, 99.8 ± 0.4, 99.9 ± 0.1 for <5 mmHg, <10 mmHg, <15 mmHg respectively. The random forest approach on the EVG signal was able to perform at a similar level with systolic: 94.9 ± 6.7, 98.7 ± 1.7, 99.6 ± 0.6 and diastolic: 95.8 ± 6.3, 98.9 ± 1.7, 99.6 ± 0.5 for<5 mmHg, <10 mmHg, <15 mmHg respectively. The CNN approach performed much worse than the BP signal approach with systolic: 64.6 ± 20.3, 88.9 ± 9.1, 95.9 ± 4.1 and diastolic: 73.3 ± 17.5, 93.7 ± 6.3, 98.0 ± 2.5 for <5 mmHg, <10 mmHg, <15 mmHg respectively. In Table 4, the mean absolute error can be seen between the use of the original BP signal to the best performing classifier (random forest regressor using Catch22 features). The BP signal provides a tighter spread on predicting BP, but the EVG approach is close in performance.TABLE 4

[0134] The efficiency of the models was examined by performing additional analyses. First, the input into the random forest model was reduced to the top 3 features that were extracted from the catch22 algorithm (shown in Table I). As shown in Table 5 (left columns), the overall performance was comparable to that achieved using all 24 features (i.e., only two experiments failed to achieve a grade A BHS score). Next, the accuracy of the BP predictions was assessed as the size of the training set used in the CNN model was varied. FIG. 20 shows charts illustrating the effects of increasing the size of the training data set on the accuracy of the CNN model. This analysis included 7 / 10 experiments that achieved BHS grades of A for predicting systolic and diastolic pressure. FIG. 20 illustrates that medical-grade accuracy (min grade B in both SYS and DIA) can be achieved by using 45% of the available data as the training set. Lastly, the performance of the leave-one-rat-out approach is also seen in Table 5 (right columns).TABLE 5

[0135] These particular example experiments demonstrated that the systolic and diastolic BP could be predicted with high levels of accuracy. Firstly, the physical nature of the EVG is differentthan what can be seen in peripheral nerve recordings using cuff electrodes to measure baroreceptive activity from vagal nerve fibers. The EVG relies on the electrical potential generated by the movement of the cuff electrode relative to the arterial vessel based on pulse signal artifacts. The observed EVG waveform presented similarities in all aspects of the pressure pulse generated by a BP recording excluding its amplitude. The cross-correlation results ranged from 0.93 - 0.99; which suggests that the EVG can be used as a surrogate signal for BP. In addition, the EVG signal exhibits prominent features typically observed in a BP waveform, including the dicrotic notch. In the PPG domain, this is often a feature of interest due to its usefulness in many analyses; such as showing an association between age and cardiovascular disease. Presence of these prominent features in the EVG signal advantageously provides potential utility for various biomedical analyses related to cardiovascular function.

[0136] The overall performance in BP prediction was excellent from the two approaches with the random forest regressor, generally outperforming the CNN in every prediction. This suggests that the time-series features extracted via the catch22 algorithm is sufficient to describe the important information embedded within the EVG signals for predicting BP. The feature reduction analysis further shows that the BP can be predicted with only a select number of features (which also outperforms the CNN approach) suggesting that the complexity of a CNN is not necessarily needed.

[0137] In the example experiments, the list of 24 input features were as provided in Table 6, below:TABLE 6ACF - autocorrelation function

[0138] The top 3 features that most frequently occurred in decreasing order of occurrence were: • DN_Spread_Std, • PD_PeriodicityWang_th0_01 • CO_trev_ 1_num

[0139] DN_Spread_Std calculates the standard deviation of the timeseries. This describes the variability of data points over time and indicate how spread out the data is from the mean value. This suggest that variability present in the EVG signal can play a crucial role in predicting systolic and diastolic BP values.

[0140] PD_PeriodicityWang_th0_01 returns the first peak of the autocorrelation function after detrending the timeseries using a three-knot cubic regression spline. Removing long-term trends in a timeseries allows the models to focus more on the fluctuations or patterns that are not explained by the trend. This makes identifying and analyzing events that are more irregular more clearly. As this feature is also capturing variation in the EVG signal, it shows that variation or variability play a key role in BP prediction.

[0141] CO_trev_1_num is a feature that calculates the average across the timeseries of the cube of successive timeseries differences. This provides information on the distribution of successiveincreases or decreases, where the value reaches zero when the successive increases match the number of successive decreases. This means that distributions that are more symmetric would present values closer to 0. The values would be positive if increases tend to be larger in magnitude and negative if vice-versa. Thus, this feature may be able to capture overall rises or drops in BP.

[0142] Impressively, all rats using the random forest regressor approach could achieve a BHS standard device grade of A and in the CNN approach only 3 cases of grade C were observed. This is relevant as the BHS considers any device graded B or above to be acceptable for use in a hospital setting.

[0143] Advantageously, the present embodiments provide an implantable device, which includes transducer electrode contacts, placed around an arterial blood vessel, that is able to be used to measure blood pressure values. While common electrical nerve stimulation cuffs will generally filter out signal artifacts caused by deformation of the electrode contacts, the present embodiments advantageously use such artifacts for a highly accurate measurement of blood pressure values. The present embodiments can be provided on a single arterial vessel or on a plurality of arterial vessels, for example, vessels associated with the pancreas, liver, kidney, thyroid, brain, spleen, gallbladder, or the like.

[0144] In contrast to other approaches in the art, the present embodiments advantageously make use of pulse artifacts. The phenomenon of pulse artifacts occurs in EEG and ECG recordings when one or more of the surface electrodes are positioned in close proximity to an underlying arterial vessel. In these situations, the appearance of pulse artifacts is often unpredictable and their characteristics (e.g., waveform, amplitude) will generally vary depending on the location of the electrode and the types of tissue located between the blood vessel and the surface electrode. The present embodiments are able to capture this artifact in a more consistent manner by placing a recording cuff electrode around an arterial blood vessel; such that suitable signal processing approaches can be employed to utilize the artifact to analyze cardiovascular functions. Unlike other approaches in the art, the present embodiments apply machine learning models to generate predicted values of arterial blood pressure, such as the systolic, diastolic, and mean arterial pressures, using measurements from the recording cuff electrode. Any suitable machine learning model can be used, such as random forest models that have been shown to yield better prediction of blood pressure values than, for example, convoluted neural network approaches; however, advantageously, any suitable machine learning model can be used to predict blood pressure from the recorded EVG (pulse artifact) signals.

[0145] The EVG signal of the present embodiments can potentially be used for closed-loop BP modulation systems. Generally, there is a great importance for continuous BP monitoring for mitigating complications associated with BP variability. Being able to continuously and accurately monitor BP can prevent many health issues and tracking individual’s overall cardiovascular health. Catheterization, as used in other approaches, is highly accurate but is very invasive and requires expertise of physicians and doctors to conduct the measurement; and thus, is restricted to hospital settings. In contrast, the present embodiments offer a less intrusive approach for continuously monitoring BP once implanted.

[0146] The EVG signal captured by the present embodiments exhibits many similarities to the intra-arterial BP signal, which includes amplitude peaks corresponding to SYS and DIA pressures and the characteristic dicrotic notch. The accuracy of BP predictions, for example, as achieved by the random forest regressor model, shows that an implanted cuff electrode can be used to continuously measure BP. Thus, the present embodiments can be used to improve BP monitoring and management by offering a particularly advantageous alternative to non-invasive wearables and intra-arterial catheters.

[0147] Although the foregoing has been described with reference to certain specific embodiments, various modifications thereto will be apparent to those skilled in the art without departing from the spirit and scope of the invention as outlined in the appended claims.

Claims

Claims1. A system for determination of blood pressure using an implantable device, the implantable device comprising one or more electrodes positioned on or around a vessel, the system comprising one or more processors and a data memory, the one or more processors in communication with the data memory to receive instructions to execute: a signal processing module to receive electrical signals from the one or more electrodes, the electrical signals comprising pulse artifacts generated from mechanical deformation of the electrodes, the signal processing module digitizes the signals received from the one or more electrodes; an analysis module to predict blood pressure values from the digitized signals using a trained machine learning model, the machine learning model trained using a training dataset comprising digitized electrical signals comprising pulse artifacts and associated measured blood pressure values; and an output module to output the predicted blood pressure values.

2. The system of claim 1, wherein the machine learning model comprises a convolutional neural network, and wherein digitizing the electrical signals received from the one or more electrodes further comprises transforming each temporal segment into a three- channel image for input into the machine learning model.

3. The system of claim 2, wherein digitizing the electrical signals comprises performing a Hilbert-Huang Transform to decompose the received signals into intrinsic mode functions.

4. The system of claim 2, wherein the three-channel image comprises a raw digitized values in a first channel, a first derivative of the digitized values in a second channel, and second derivative of the digitized values in the third channel.

5. The system of claim 4, wherein digitizing the electrical signals further comprises normalizing the first derivative and the second derivative for each dimension.

6. The system of claim 1, wherein the machine learning model comprises a neural network sequential model, and wherein digitizing the electrical signals comprises determining a set of feature vectors from the received electrical signals.

7. The system of claim 6, wherein the feature vectors comprise one or more of a measure of standard deviation of a time series, a measure of variability of the time series, and a measure of asymmetry and characteristics of distribution of variations in the time series.

8. A method for determination of blood pressure using an implantable device, the implantable device comprising one or more electrodes positioned on or around a vessel, the method comprising: receiving electrical signals from the one or more electrodes, the electrical signals comprising pulse artifacts generated from mechanical deformation of the one or more electrodes; digitizing the electrical signals received from the one or more electrodes; predicting blood pressure values from the digitized signals using a trained machine learning model, the machine learning model trained using a training dataset that comprises digitized electrical signals comprising pulse artifacts and associated measured blood pressure values; and outputting the predicted blood pressure values.

9. The method of claim 8, wherein the machine learning model comprises a convolutional neural network, and wherein digitizing the electrical signals received from the one or more electrodes further comprises transforming each temporal segment into a three- channel images for input into the machine learning model.

10. The method of claim 9, wherein digitizing the electrical signals comprises performing a Hilbert-Huang Transform to decompose the received signals into intrinsic mode functions.

11. The method of claim 9, wherein the three-channel image comprises a raw digitized values in a first channel, a first derivative of the digitized values in a second channel, and second derivative of the digitized values in the third channel.

12. The method of claim 11, further comprising normalizing the first derivative and the second derivative for each dimension.

13. The method of claim 8, wherein the machine learning model comprises a neural network sequential model, and wherein digitizing the electrical signals comprises determining a set of feature vectors from the received signals.

14. The method of claim 13, wherein the feature vectors comprise one or more of a measure of standard deviation of a time series, a measure of variability of the time series, and a measure of asymmetry and characteristics of distribution of variations in the time series.

15. The method of claim 8, further comprising performing electrical nerve stimulation by the one or more electrodes.

16. The method of claim 15, wherein the electrical nerve stimulation is performed to down- regulate nerve activity or to up-regulate nerve activity based on the predicted blood pressure values.

17. An implantable device for determination of blood pressure, the implantable device comprising: one or more electrodes positionable on or around a vessel, the electrodes generating an electrical signal representative of pulse artifacts due to mechanical deformations of the electrodes from periodic pulsations of the vessel; electronic circuitry to relay the signals captured by the implantable device to an external computing device, the external computing device digitizes the signals and predicts blood pressure values from the digitized electrical signals using a trained machine learning model.

18. The implantable device of claim 17, wherein the one or more electrodes have the form of a cuff, spiral, or helix.

19. The implantable device of claim 17, further comprising a flexible enclosure to support the one or more electrodes in position on or around the vessel.

20. The implantable device of claim 17, wherein the one or more electrodes comprise two or more electrodes separated by a predetermined distance.

Images