Seizure identification using neural activity information

Abstract

A vagus nerve stimulator (VNS) can be used for the treatment of epilepsy, such as in response to detecting that a seizure is occurring. An aspect of the present subject matter can include or can perform seizure detection, such as facilitating a closed-loop VNS therapy delivery approach. Such an approach can include use of personalized detection or therapy selection (or both), such as a dynamic scheme where one or more of detection criteria or therapy can be varied over time. As an illustration, a system can be configured to provide a machine-learning-based approach for seizure detection or therapy delivery (or both). Neural activity can be sensed electrically and used as an input variable. Other input variables can be used to enhance robustness of seizure detection.

Description

SEIZURE IDENTIFICATION USING NEURAL ACTIVITY INFORMATIONCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application is related to and claims priority to United States Provisional Application No. 63 / 470,252, filed on June 1, 2023, and entitled “Seizure Identification using Neural Activity Sensing such as for Personalized Stimulation Therapy Control,” the entirety of which is incorporated herein by reference; and this application is related to and claims priority to United States Provisional Application No. 63 / 543,384, filed on October 10, 2023, and entitled “Seizure Identification using Neural Activity Sensing such as for Personalized Stimulation Therapy Control,” the entirety of which is incorporated herein by reference.FIELD OF THE DISCLOSURE

[0002] This document pertains generally, but not by way of limitation, to seizure identification and treatment using implantable systems.BACKGROUND

[0003] Neurological disorders encompass a range of conditions that impact the nervous system, including the brain, spinal cord, and peripheral nerves. These conditions may present with a variety of symptoms, such as cognitive deficits, issues with motor skills, sensory disturbances, and abnormalities in the autonomic nervous system. The origins of these disorders are varied, encompassing genetic predispositions, environmental factors, traumatic injuries, and ongoing illnesses.

[0004] Management of neurological disorders depends on the particular disorder and its intensity. Treatment options may involve medications, physical rehabilitation, and surgical interventions. However, these treatments do not always yield positive results for all patients, with some experiencing adverse reactions or only modest improvements in their condition.

[0005] Epilepsy is characterized by disrupted nerve cell activity in the brain, leading to seizures. These seizures can cause unusual behaviors, sensations, and sometimes loss of consciousness. Treatment for epilepsy can include pharmaceuticals, surgical interventions, implanted devices, or dietary modifications. While medications can manage some seizures, alternative treatments such as neurostimulation therapy may be explored if medications are insufficient.SUMMARY

[0006] The systems and methods described herein can include seizure detection and closed-loop therapy delivery, such as using an implantable device. The systems described herein can include or use a vagus nerve stimulator (VNS) system. The VNS system can be used for treatment of epilepsy or other conditions or diseases. In an example, the VNS system can administer a treatment or therapy in response to detecting an adverse event or disease progression. For example, the VNS system can provide treatment in response to detecting that a seizure is imminent or occurring.

[0007] In an example, the systems and methods discussed herein can be configured to perform personalized seizure detection (i.e., using patientspecific detection criteria), personalized therapy selection and delivery, or both. The detection and / or therapy can be updated or adjusted over time to help meet evolving patient needs. In an example, a machine learning-based approach for seizure detection or therapy delivery can be used.

[0008] The present inventors have recognized, among other things, that various challenges exist in detecting a seizure occurrence. For example, the use of an increase in heart rate above a specified (e.g., user-selected or clinician-determined) threshold to indicate that a seizure is imminent or occurring can present a high false positive rate versus other inputs because seizures can be uncorrelated with an increase in heart rate. The present inventors recognized that a high false positive rate may lead to premature depletion of VNS system resources (e.g., reduced battery life) or decreased therapy effectiveness (e.g., due to patient habituation to therapy).

[0009] The present subject matter can help provide a solution to these and other problems, such as by sensing patient neural activity, optionally in combination with other physiological signal information, to detect one or more of likelihood of onset of a seizure, a present or ongoing seizure, termination of a seizure, or combinations thereof.BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS

[0010] To easily identify the discussion of any particular element or act, the most significant digit or digits in a reference number refer to the figure number in which that element is first introduced.

[0011] FIG. 1 illustrates generally an example of a neurostimulation system.

[0012] FIG. 2A illustrates generally an example that includes implantable electrodes.

[0013] FIG. 2B illustrates generally an example of electrodes disposed at a nerve target.

[0014] FIG. 3 illustrates generally an example of a patient status evaluation system.

[0015] FIG. 4 illustrates generally an example of a vagus activity chart.

[0016] FIG. 5A illustrates generally an example representation of first neural activity.

[0017] FIG. 5B illustrates generally an example representation of second neural activity.

[0018] FIG. 6A illustrates generally a first example representation of neural activity associated with respective sensing electrodes.

[0019] FIG. 6B illustrates generally a second example representation of neural activity associated with respective sensing electrodes.

[0020] FIG. 7 illustrates generally a data visualization example including multiple variables.

[0021] FIG. 8 illustrates generally an example of a first method that can include using a machine learning-based model to analyze physiologic status information about a patient.

[0022] FIG. 9 illustrates generally an example of a machine that can comprise one or more of the systems discussed herein, or that can be configured to perform any one or more of the methods discussed herein.DETAILED DESCRIPTION

[0023] An implantable vagal nerve stimulation (VNS) system may perform seizure detection through a multimodal approach in which information from two or more physiologic signals is used together to identify a seizure event or to identify a likelihood that a seizure event is imminent. In some examples, information from multiple signals may be used together as inputs to a machine learning-based model that is configured to identify or predict seizure events. To improve battery longevity, information from one sensor may be used to provide an initial detection, which is then confirmed by one or more other signals from the same or other sensor.

[0024] The system can be configured to perform seizure detection for the purpose of reporting seizure burden, or it can be configured to respond to a detection event by altering a parameter that defines a stimulation therapy provided to the patient. For example, the system may respond to a detection event by increasing stimulation intensity for a programmable or specified period of time. In an example, stimulation parameters can be selected based on an intensity, duration, or other characteristic of a detected seizure.

[0025] In various examples, the VNS system can communicate wirelessly with an external programming device (e.g., a mobile phone, a tablet, a computing device, or the like). The programming device can include a display configured to receive and display data from an implanted device, including individual sensor data and seizure detection annotations. The VNS system may calculate and display seizure burden, display an event log, or the like. The VNS system may include a patient device to allow patients toconfirm events, add comments, or the like (e.g., through an application, or the like).

[0026] In an example, the system can tailor seizure detection to the individual patient in response to feedback (explicit or automated feedback) about false positives and false negatives. For example, the system can include or use default or specified thresholds for event detection (e.g., using different or respective threshold values for respective types of sensor data), and the thresholds can be updated (e.g., maintained, raised or lowered) for the individual patient based on patient feedback or clinician feedback.

[0027] The VNS system can include or use a patient device for the patient to indicate whether seizures were accurately detected. For example, the patient device can include a magnet, and the patient can use a swipe of the magnet to indicate the incidence of a seizure. In an example, the patient device can include a handheld device with a screen, such as a tablet or smartphone. The patient device can be configured to receive information from the patient that confirms or denies detected seizure events, or to indicate the incidence of a seizure that was not detected (i.e., automatically detected based on sensed physiologic status information).

[0028] In response to an incorrect detection (e.g., a false positive or a false negative), the system can be configured to adjust seizure detection parameters (e.g., detection thresholds) to tune detection to the individual patient and improve sensitivity and specificity, for example.

[0029] The VNS system may include various operational modes. For example, if certain input variables are not available (e.g., sensors not connected or not transmitting data), a machine learning model instance can suppress VNS or otherwise perform seizure detection in a fallback or regression mode using the remaining available input variables.

[0030] An illustrative (but non-restrictive) example, as shown in FIG. 1, includes a system 100 for providing neurostimulation to a vagus nerve 102, or vagus nerve stimulation (VNS). In an example, the system 100 can be configured to sense nerve activity or other electrical activity or motion. The example of the system 100 includes an implantable device 112 such as cancomprise a processor circuit 114 and a signal generator 116. The processor circuit 114, or control circuit, can control operation of the signal generator 116 according to various therapy delivery algorithms or therapy signaldefining parameters. The signal generator 116 can be configured to generate neurostimulation signals or pulses according to parameters or instructions from the control circuit. In an example, the signal generator 116 includes independent current sources and controllers to enable independent and simultaneous output of multiple respective therapy signals.

[0031] In an example, the implantable device 112 comprises or is coupled to one or more physiologic status sensors that are configured to sense information about a patient. For example, the system can include a sensor 118. The implantable device 112 can further include a recording unit 120 configured to record information from the sensor 118. The sensor 118 can comprise a portion of the implantable device 112 or can be coupled to a lead that is coupled to the implantable device 112. In other examples, the sensor 118 can be an external sensor that is coupled to, or otherwise configured to receive information from, the patient. In an example, the sensor 118 comprises an accelerometer configured to sense motion information about the patient. In an example, the sensor 118 comprises one or more electrodes configured to sense electrical signals from the patient. In an example, the one or more electrodes can be implanted at or near a vagus nerve of the patient and can be configured to deliver electrical signals to, or receive electrical signals from, the vagus nerve (or other neural target). For example, the one or more electrodes can be configured to sense vagal activity information from the patient.

[0032] In an example, the system 100 includes an external device 122 that can communicate with the implantable device 112. The external device 122 can include a patient device or clinician device that is configured to receive information from, or provide information to, the implantable device 112. In an example, the external device 122 can include a display configured to receive and display data from the implantable device 112, including individual sensor data and seizure detection annotations. The system 100 orthe external device 122 may calculate and display seizure burden, and / or display an event log. In an example, the external device 122 includes an interface that allows patients to confirm events and / or add comments or annotations to detected events.

[0033] For example, the external device 122 can be used to set one or more neurostimulation parameters for a neurostimulation therapy that is provided by the implantable device 112. In an example, the external device 122 can be used to report information to a patient or clinician about one or more therapies provided by the implantable device 112. In an example, the external device 122 includes one or more sensors that are configured to monitor physiologic or behavioral information about the patient.

[0034] The interaction between the external device 122 and the implantable device 112 is facilitated through a bidirectional communication link using a wireless coupling 124 that allows for the continuous exchange of data and commands between the two devices. The communication is established using wireless technology protocols that are specifically designed for medical devices, ensuring secure and reliable data transmission.

[0035] The external device 122 is equipped with various sensors, including a high-resolution camera, a microphone, and an accelerometer, which collect a wide array of physiological and environmental data, as described below. This data includes visual and audio records of the patient’s movements, vocalizations, and surrounding environment, as well as quantitative measurements such as detected motion patterns, heart rate, and / or respiration rate. The collected data is then processed (e.g., at the external device 122, at the implantable device 112, or elsewhere) using advanced algorithms to identify potential seizure events.

[0036] In an example, the processor circuit 114 is programmed with a set of parameters that define thresholds for initiating or adjusting VNS therapy. Upon receiving a therapy-indicating signal from the external device 122 or based on information from the sensor 118, the processor circuit 114 can be configured to analyze the sensor data against the predefined parameters. If the data indicates that a seizure is occurring or imminent, the implantabledevice 112 adjusts the neurostimulation therapy parameters accordingly. This adjustment may involve changing the intensity, frequency, duration, or other characteristic of the electrical impulses delivered to the vagus nerve to provide an appropriate therapeutic response.

[0037] In an example, seizure detection and VNS can include or use one or two vagus nerve sensing electrodes (e.g., “recording cuff’ or helical electrodes), such as located in different longitudinal positions along the cervical vagus region, relative to a stimulation site. Separate stimulating electrodes (e.g., an anode and a cathode) can be positioned to provide VNS. In the example of FIG. 1, the system 100 includes a first electrode 104, a separate second electrode 106, a separate third electrode 108, and a separate nth electrode 110 positioned at or near the vagus nerve 102. The various electrodes can be used in various combinations to provide an epilepsy therapy or to sense activity from the vagus nerve or other nerve. In an example, the multiple electrodes comprise respective portions of a single lead, or multiple leads can be used, with each lead comprising one or more electrode.

[0038] The count and position of electrodes in the example of FIG. 1 is merely illustrative. For example, an implantable device can include circuitry for sensing (e.g., recording) neural activity (e.g., an action potential or compound action potential), along with circuitry for generating VNS. In such an example, a machine learning approach, such as an instance of a machine learning -based model (e.g., such as can be referred to as an artificial intelligence or “AI”-based technique) can be instantiated by the implant circuitry or the processor circuit 114. Such a machine learning-based model can be used for detection of a seizure, or for therapy control in response thereto, or both.

[0039] In an example, the sensing electrodes and related circuitry can be separate from the stimulating electrodes and the sensing electrodes can be monitored by a separate unit (e.g., an external assembly) that can be used in an acute or temporary manner, such as supporting the implant procedure or implantable device configuration. For example, in the case that the sensingand stimulating electrodes are separate, the sensing electrode may be explanted acutely as a portion of a first procedure or soon after the first procedure. In yet another example, there can be three or more electrodes, each configurable as either a stimulating electrode or a sensing electrode at any time. For example, two electrodes closest to a brain of a patient could be assigned as an anode and a cathode, respectively, and another electrode that is located more distally could be assigned as a sensing electrode to detect efferent nerve activation. As another illustration, two electrodes most distal to the brain could be assigned as an anode and a cathode, respectively, and an electrode more or most proximal to the brain could be assigned as a sensing electrode to detect afferent activity.

[0040] FIG. 2A illustrates generally components of an electrode assembly. The components can include a first electrode array 202 and ring electrode 208. The first electrode array 202 can comprise multiple, separately addressable electrodes, such as including a first electrode 204 and a second electrode 206. One or more of the first electrode array 202 and the ring electrode 208 can optionally be coupled to a support structure 210 or scaffold, and the support structure 210 can be provided about a target nerve or tissue. The first electrode array 202 or the ring electrode 208 can comprise examples or embodiments of any one or more of the first electrode 104, the second electrode 106, the third electrode 108, and / or the nth electrode 110 from the example system 100 of FIG. 1.

[0041] FIG. 2B illustrates generally an example of an electrode assembly 212. The electrode assembly 212 includes a first electrode array 218, a second electrode array 220, a third electrode array 222, and a fourth electrode array 224. Each of the arrays can comprise multiple, separately addressable electrodes. Each of the arrays comprise an example or embodiment of the electrodes of the example system 100 of FIG. 1.

[0042] The electrode assembly 212 can be provided about a nerve 214. In an example, the nerve 214 comprises a portion of a vagus nerve, such as in a cervical region of a patient. In an example, the electrode assembly 212 can be secured about the nerve 214 using one or more sutures 216.

[0043] The electrodes of the electrode assembly 212 can be supported by a monolithic, shared scaffold or support, or the electrodes can be individually supported at their respective locations along the nerve 214. Implanting a monolithic structure at, near, or around a nerve has the advantage of generally being easier and faster for the implanting physician, as compared to implantation of individual discrete electrode structures.

[0044] In an example, the first electrode array 218 comprises electrodes that are configured to be arranged or disposed at discrete locations about a circumference of a nerve. The electrodes can optionally be arranged about a twisted or helical path. The electrodes can be selectively energized to stimulate nearby nerve fibers.

[0045] The use of a plurality of electrodes in an array, or in a series of ring structures, or in other arrangements, facilitates the programmability of different spatial arrangements of neural activity sensing, stimulation, or both. In various examples, an array configuration provides redundancy in case of loss of sensing or a reduction of stimulation efficacy of a particular electrode. Such a multi-electrode configuration can be used to provide sensing modalities or stimulation electrode configurations that can vary over time to maintain the effectiveness of VNS therapy, seizure detection, or both.

[0046] FIG. 3 illustrates an example of a portion of a patient status evaluation system 300. In an example, the patient status evaluation system 300 can be configured for detection of a seizure (e.g., likelihood of onset of a seizure, ongoing seizure, termination of a seizure, or combinations thereof), for therapy control in response to the detection of the seizure, or both, according to an embodiment. The patient status evaluation system 300 can comprise a component or portion of the system 100. For example, one or more portions of the patient status evaluation system 300 can be implemented using the processor circuit 114 of the implantable device 112, using the external device 122, or using another remote system or service. In an example, the patient status evaluation system 300 includes or uses one or more machine learning models to process information from or about thepatient, such as from the sensor 118, the recording unit 120, the external device 122, or other source.

[0047] The patient status evaluation system 300 includes a training module 310 and a prediction module 320. In an example, the training module 310 provides training data 302 to a feature determination module 304. The feature determination module 304 can be configured to determine or identify or more features 306 or characteristics from the training data 302. For example, the training data 302 can include physiologic signal information from the sensor 118. The feature determination module 304 can identify the training data features 306, or characteristics, of the sensor signal information. For example, the training data 302 can comprise heart rate (or other cardiac motion) information from a heart rate sensor, and a feature or characteristic of the heart rate information can include a heart rate, a heart rate variability, a maximum heart rate, a heart rate trend, or other characteristic information that can be derived or determined using the heart rate sensor information. In an example, the features 306 include information that is, or that may be, used to help predict or confirm a seizure for a patient. The features 306 can be selected or determined manually or automatically, for example, based on previous seizure events for the patient or for similarly situated patients (e.g., patients having similar age, weight, health condition, or the like).

[0048] In an example, the training data 302 includes patient-specific physiologic information about a patient, or includes information about a population of patients. The training data 302 can include information from or about different patient states (e.g., a baseline state, a seizure state, a stimulation state, etc.). Referring to machine learning-based approaches, a “feature” can refer to a set of multivariate parameters that occur during, or closely in association with, a specific patient state.

[0049] A machine learning algorithm 308 can be configured to use the features 306 to provide a prediction model 318. In an example, the machine learning algorithm 308 provides the prediction model 318 by learning from data (e.g., the features 306) to make predictions or decisions, optionallywithout being explicitly programmed to perform the task. A quality or accuracy of the prediction module 320 can be assessed based on feedback about the prediction model 318 results. For example, a processing result from the prediction model 318 can be processed or evaluated by a validation module 322 to determine whether updates are needed to the machine learning algorithm 308 and / or the prediction model 318. In an example, the validation module 322 can be configured to receive an input from a patient or clinician about a validity or correctness of the prediction model 318 results. If the prediction model 318 correctly predicted the patient status, as validated by the input to the validation module 322, then the prediction model 318 can continue to be used without changes. If, however, the prediction model 318 incorrectly predicted the patient status, then the prediction model 318 can be updated based on the actual patient status and the one or more features corresponding to that patient status.

[0050] In an example, the machine learning algorithm 308 can comprise a supervised or unsupervised machine learning algorithm. Examples of supervised learning algorithms include artificial neural networks, Bayesian networks, instance-based learning, support vector machines, decision trees (e.g., Iterative Dichotomiser 3, C4.5, Classification and Regression Tree (CART), Chi-squared Automatic Interaction Detector (CHAID), and the like), random forests, linear classifiers, quadratic classifiers, k-nearest neighbor, linear regression, and hidden Markov models, among others. Examples of unsupervised learning algorithms include expectationmaximization algorithms, vector quantization, and information bottleneck method. In an example, a multi-class logistical regression model is used.

[0051] In the prediction module 320, the patient data 312 can comprise physiologic or other data related to seizure events for a particular individual patient, data related to vagus nerve activity or stimulation provided to the patient, or other information monitored from or about the patient. The patient data 312 can be provided to a feature determination module 314. The feature determination module 314 can be configured to identify monitoring data features 316 of the patient data 312 for use in the prediction model 318.The monitoring data features 316 determined by the prediction module 320 can be the same or different as the features 306 determined by feature determination module 304 in the training module 310. In some examples, the feature determination module 304 and the feature determination module 314 can comprise instances of the same processor-implemented module. The features 316 can be provided to the prediction model 318.

[0052] In an example, the training module 310 can be configured to operate in an offline manner to train the seizure prediction model 318. The prediction module 320, however, can be configured to operate in an online or ongoing manner as the patient data 312 is acquired and evaluated over time.

[0053] The seizure prediction model 318 can be intermittently or periodically updated via additional training and / or feedback (e.g., via the validation module 322). The feedback can include patient feedback (e.g., including patient responses to questions about whether a seizure occurred, a severity of a seizure event, or the like). The feedback can include sensor information, such as can be automatically received and processed by the validation module 322, and the sensor information can include information about changes in a patient physiologic status. In an example, a patient who experienced a seizure event can provide an explicit response indicating that stimulation was requested or necessary, or a clinician may provide feedback during review of the recorded data of the VNS system. Such responses can be used by the patient status evaluation system 300 as additional training data for updating the seizure prediction model 318.

[0054] In various examples, the system 100 can be used to administer or control vagus nerve stimulation timing and parameters based on the predictions related to seizure events made by the patient status evaluation system 300. In various examples, the system 100 can thus use a trained artificial intelligence (Al) model to determine features of different states of the patient, and selectively respond to such states. In an example, the model establishes (or controls, or both) the timing and parameters of the vagus nerve stimulation. In various examples, the model is dynamically trained andthe determination of the features of the different states is updated accordingly.

[0055] The system 100 can use artificial intelligence (e.g., including machine learning-based approaches, such as described for the patient status evaluation system 300) to learn or determine the feature(s) of different physiologic statuses or “states” for a patient or for a patient population comprising multiple patients. The approach can include or use a model instance (e.g., the prediction model 318) to establish or otherwise control stimulator timing and parameters. As an illustrative example, recording cuff electrodes can read vagus nerve activity (e.g., using electrical signal sensing) and obtain a profile of the neural firings during different states. In an example, the approach can include receiving the vagus activity information together with one or more other signals from physiologic sensors or information reported from a patient for one, multiple, or all of the states.

[0056] A first state, or Statel, can correspond to a patient baseline or nonseizure status or status. A machine learning-based model can be trained using input variables (e.g., the training data 302), optionally including vagus nerve activity information, during a reference time, such as when no seizure or stimulation from the implanted device is occurring. In an example, the information for Statel can be collected at different times during the day to find a general pattern of these variables including any variation associated with time of day or activities of daily life.

[0057] A second state, or State2, can correspond to therapy delivery. During VNS therapy delivery, a machine learning-based model can be trained using input variables to find a pattern of these variables during the VNS therapy. VNS efficacy can be assessed using, for example, variations in therapy parameters such as neurostimulation signal amplitude, frequency, pulse width, pulse shape, or pattern of the stimulation waveform, or combinations thereof. VNS efficacy can be assessed using various on-off time ratios that describe when therapy is provided. For example, therapy can be provided for 1 minute and then withheld for 10 minutes, or provided for 5 minutes and then withheld for an hour, and so on. In an example, VNSefficacy can be assessed using various on-off times relative to a phase of other input variables. For example, the phase for a respiration input variable could be, phase 1: during inhalation, phase 2: during exhalation, or for a cardiac cycle input variable could be: phase 1: during contraction, phase 2: during systole, phase 3: during diastole, phase 4: during relaxation. Physiologic input variables or patient input can be used to assess whether an adverse physiological response is occurring in response to VNS, such as unwanted laryngeal activity or cardiac activity, or other discomfort.

[0058] A third state, or State3, can correspond to a seizure event. The model, or the machine learning algorithm 308, can use the various input variables, such as monitored before, during, and after a seizure, to find a pattern of these input variables related to a seizure occurrence.

[0059] A fourth state, or State4, can correspond to an imminent seizure. Identification of State4 can be particularly helpful for preemptive therapeutic interventions. In this state, the machine learning model uses input variables that indicate a transition from a baseline (or other) state to one where a seizure is highly likely to occur soon. These variables might include sudden changes in the frequency or amplitude of neural activity, or specific patterns detected in the wavelet-transformed data that historically correlate with the onset of a seizure. By identifying these predictive patterns, the model can trigger an alert or initiate preemptive therapy, potentially averting the full onset of a seizure or mitigating its severity. This proactive approach aims to enhance patient safety and improve the overall effectiveness of seizure management strategies.

[0060] One or multiple input variables can be monitored in relation to the states mentioned above. Generally, as shown and described herein, neural activity will be monitored (e.g., sensed electrically), either alone or in combination with other input variables.

[0061] In an example, an input variable can include vagus spike train data in the frequency domain. The spike time series can be transformed into the frequency domain (or vectorized using another approach such as waveletanalysis). Specific patterns in the frequency domain may be associated with one or more of the specified states mentioned above.

[0062] In an example, an input variable can include vagus spike train data in the wavelet domain. That is, spike time series data from neural activity can be transformed into a wavelet domain. Specific patterns in the wavelet domain may be associated with one or more of the specified states mentioned above. Transforming the data into the wavelet domain allows extraction of features that may be indicative of seizures. These features can include changes in the frequency content over time or the presence of specific wavelet coefficients that correlate with seizure activity. The machine learning algorithm 308 can then use these features to classify segments of neural activity and detect potential seizures.

[0063] In an example, an input variable can include cardiac motion information, such as heart rate or heart rate variability. As mentioned above, heart rate alone may not be a robust indicator of seizure, such as due to a high false positive rate. Heart rate metrics, when combined with other indicia, may be more robust for detection of a seizure. For example, a machine learning approach might find that the seizure only happens when the heart rate is increasing and, contemporaneously, a signature of sensed neural activity is occurring (e.g., increased activity on a respective neural activity sensing electrode channel). In an example, electrodes on a lead or a pulse generator can be configured to measure ECG and detect shifts in heart rate and / or heart rate variability. In an example, an accelerometer (e.g., comprising the sensor 118) can be used to sense information about a patient heart rate or heart rate variability.

[0064] In an example, an input variable can include blood pressure-related parameters. Such parameters can include a systolic pressure, a diastolic pressure, mean blood pressure, or contractility, as illustrative examples. Sensing can be performed using an external cuff or using an implantable pressure sensor. Different locations can be used to provide data indicative of heart rate or blood pressure. For example, a sensor could be placed at or near a carotid artery to sense one or more of heart rate or blood pressure.

[0065] In an example, an input variable can include respiration information. Respiration phase, cycle or pattern information, such as combined with other input variables, could be used as another marker to detect seizures. Respiration monitoring could be performed using an external band, an accelerometer, or using an impedance-based approach, as illustrative examples.

[0066] In an example, an input variable can include EEG information. For example, specific electroencephalography (EEG) patterns in different regions of the brain (alpha, beta, gamma band) can be used. EEG signals generally contain energy in different frequency bands, and such bands correspond to different functional characteristics. In adults, illustrative examples of frequency bands and their approximate spectral boundaries are delta (1-3 Hz), theta (4-7 Hz), alpha (8-12 Hz), beta (13-30 Hz), and gamma (30-100 Hz). Information indicative of activity in one or more bands may be used by a machine learning-based approach to predict onset or occurrence of a seizure.

[0067] In an example, an input variable can include time of day information. Time can be a variable that can play a role in seizure occurrence. In an example, a machine learning-based approach might find that there can be a higher likelihood of a seizure occurring during a specific time or duration, and such an indicia can be combined with other input variables, such as to enhance prediction or detection of a seizure.

[0068] Input variables can optionally include or use other physiological parameters that can be related to a seizure, such as but not limited to limb or torso motion, skin conductivity, skin temperature, or odor molecules detected in breath or sweat. In an example, an accelerometer (e.g., a 3D or multiple-axis accelerometer) can be used to detect patient motion, falls, or other spasmodic movement. In an example, an implanted device may communicate wirelessly with an external device that is configured to measure one or more of electrodermal activity (galvanic skin response), skin temperature, activity level or motion, orientation or posture, respiratory rate, and / or tidal volume. In some examples, the external device can include awrist band, a device that adheres or couples to a patient’s chest and / or a chest band.

[0069] To increase detection sensitivity and specificity, the system 100 can perform seizure detection using a multimodal approach in which multiple signals (e.g., comprising information about respective different input variables or physiologic status information) are used in combination. To improve battery longevity, one sensor may be used to provide an initial detection, which is then confirmed by one or more other signals. In some examples, information from multiple signals may be used together.

[0070] In response to an incorrect detection (e.g., a false positive or a false negative), the system 100 can be configured to adjust seizure detection parameters (e.g., detection thresholds) to tune detection to the individual patient and improve sensitivity and specificity.

[0071] Various operational modes are possible. For example, if certain input variables are not available (e.g., sensors not connected or not transmitting data), then the machine learning algorithm 308 can suppress VNS or otherwise perform seizure detection in a fallback or regression mode using one or more remaining available input variables.

[0072] FIG. 4 illustrates generally an example of a vagus activity chart 400. The vagus activity chart 400 can include time domain VENG information about a global or local firing pattern for a vagus nerve.

[0073] The example of FIG. 4 shows a pattern of multiple intermittent neural discharges between about 400 seconds and 600 seconds, followed by a substantially continuous neural discharge from about 600 seconds to 800 seconds. This pattern of intermittent neural discharges followed by a continuous discharge could, in an example, be defined as a feature of interest. For example, if this neural activity pattern leads to a seizure (e.g., as reported by the patient or detected at, e.g., data visualization example 700 seconds in the example of the vagus activity chart 400), then whenever the system 100 detects the same or similar neural activity pattern during future vagal nerve monitoring, a seizure detection event can be logged and the implantable device 112 can automatically respond. For example, theimplantable device 112 can respond by enabling a VNS therapy or by altering a VNS parameter, such as increasing a neurostimulation signal intensity.

[0074] Various characteristics of neural activity firing patterns can be identified as features. For example, the vagus activity chart 400 can include a first firing pattern 402 that can be identified as a first feature, and a second firing pattern 404 that can be identified as a second feature. The first and second features can comprise examples of the features 306 of the training module 310 or the features 316 of the prediction model 318.

[0075] The firing pattern, or firing rate, can corresponding to a particular number of compound action potentials occurring during a certain interval. Other features or signal characteristics can include, for example, interspike intervals (i.e., a time difference between spikes), a change in the interspike intervals (e.g., a burst, or the like), or the like. In various examples, the system 100 uses multiple sensing channels and the firing pattern feature(s) include a synchrony among neuronal activities from separate electrodes of the multiple sensing channels.

[0076] In an example, a feature of interest can include a combination of the first firing pattern 402 and the second firing pattern 404. For example, if the combination of the first firing pattern 402 and the second firing pattern 404 leads to a seizure, then whenever the prediction model 318 detects these features or patterns, a seizure detection event can be logged and the system 100 may respond accordingly (e.g., to enable VNS, alter a VNS parameter, such as increasing VNS intensity, or the like).

[0077] In an example, the system 100 comprises or uses a plurality of electrodes, as described herein, to selectively monitor activity in or stimulate the vagus fibers. Multiple electrodes of the plurality of electrodes can be used to receive or monitor respective vagus activity information. In various examples, the activity information received by each electrode will differ because each corresponding portion of the vagus nerve can carry different information. In other words, if an electrode or lead configuration with multiple electrodes is used, such as the electrode assembly 212 shown in theexample of FIG. 2B, respective sensing electrodes (or pairs of sensing electrodes) can be used to obtain a record similar to that shown in FIG. 4. These time-series signals will differ from each other, even if recorded over the same duration, because different vagus nerve bundles carry different information. The machine learning-based algorithms and models discussed herein can be configured to account for or leverage such differences in sensed vagal activity information. For example, a feature used in model generation or application can include or use timing, electrode location, and / or neural activity information together to improve seizure detection.

[0078] FIG. 5A illustrates generally an example of a graphical representation of first neural activity 500 over time. In the example of FIG.5 A, the horizontal or x-direction corresponds to a time axis. Events, indicated in the neural activity chart as vertical lines or bars, can correspond to a neural activation or firing at a particular time, rate, intensity, or other characteristic.

[0079] The first neural activity 500 can be based on vagus (or other nerve) activity information from one or multiple electrode pairs, or can comprise composite information determined based on neural activity information together with other sensor information. In an example, each spike or bar in the neural activity chart indicates an event corresponding to at least a specified threshold level of neural activity at a particular time. Spacings between adjacent bars can indicate a relative frequency of the events. For example, an interval with relatively fewer bars and / or more spaced apart bars per unit time can indicate a relatively lower neural activity level, and an interval with relatively more bars and / or less space between bars per unit time can indicate a relatively higher neural activity level. In other words, the bars and spacings between bars in the neural activity chart can indicate a firing rate or firing pattern of neural activity. Generally, a firing pattern in a recorded time-series of neural activity has several indices or parameters, such as a firing rate (e.g., a count of compound action potentials for a specified duration), interspike intervals (e.g., a time difference between thespikes), and a change in the interspike intervals (e.g., having a burst-like appearance).

[0080] The example of FIG. 5A includes a first interval 504 and a subsequent second interval 506. During the first interval 504, the first neural activity 500 exhibits relatively less neural activity, or a relatively lower neural firing rate, as indicated by the four bars distributed over the first interval. During the second interval 506, the first neural activity 500 exhibits relatively more neural activity, or a relatively higher neural firing rate, as indicated by the many bars with little spacing in between. In an example, the change in firing pattern, or the firing rate observed during one or more of the intervals, can be a pattern of interest, or a feature, that can be used for seizure detection.

[0081] FIG. 5B illustrates generally an example of a graphical representation of second neural activity 502. The second neural activity 502 includes a third interval 508 and a subsequent fourth interval 510. In the example of FIG. 5B, the third and fourth intervals have the same duration. The neural activity firing rate is the same per interval, for example, because the number of events that satisfy a threshold condition is the same in each of the third interval 508 and the fourth interval 510. However, the firing pattern is different in each interval due to different interspike intervals. Even though the overall per-interval rate of firing is the same, the variability in the timing of the spikes can suggest different neural activity or states. This difference in the firing pattern can comprise a feature that can be used to detect the likelihood of a seizure event.

[0082] The neural activity information represented in FIG. 5A and FIG. 5B can be particularly useful in algorithms that focus on temporal patterns in data, such as time-series analysis or recurrent neural networks. These algorithms can be configured to learn to differentiate between the patterns of neural activity that may not be apparent from rate alone. For instance, a machine learning model can be trained on historical data where interspike interval patterns associated with seizures are labeled. The model can then learn the complex sequences of neural activity that typically occur before aseizure. Once deployed, a system using such a model can continuously monitor the patient's neural activity and apply the learned model to detect these patterns in real-time. Upon detecting a pattern that matches the seizure-associated signature, the system can trigger an alert or initiate a therapy.

[0083] In an example, neural activity sensing can be performed using various different electrode combinations. The resulting respective timeseries signals can be different from each other, even if recorded over the same duration, because different nerve bundles (e.g., vagus nerve bundles or branches) can carry different information. Using information about a local firing pattern (e.g., sensed neural activation that is spatially localized to a particular region), a bundle that carries the most important information about each state (for example seizure, or VNS states) can be identified, or a combined pattern related to a state can be used. For example, increased activity at a first electrode detected concurrently with decreased activity at a second electrode can indicate that a seizure is likely to occur.

[0084] Each local firing pattern can be defined by various indices or parameters, such as a firing rate (e.g., a count of compound action potentials during a specified duration), interspike intervals (e.g., a time difference between the spikes), and a change in the interspike intervals (e.g., having a burst-like appearance). In addition, if multiple sensing channels are used, synchrony among the neuronal activities from separate electrodes can be analyzed.

[0085] FIG. 6A illustrates generally a graphical representation of third neural activity 602 sensed using a first electrode (Electrode 1) and fourth neural activity 604 sensed using a second electrode (Electrode 2). In an example, the third neural activity 602 and fourth neural activity 604 represent activity at respective different portions of the vagus nerve, however, the neural activity can additionally or alternatively represent activity at other nerve(s).

[0086] In the example of FIG. 6A, neural activity sensed by the first electrode includes information about, or that can be correlated with, aseizure, for example at a fifth interval 606 or a sixth interval 608. Neural activity sensed by the second electrode, however, may not include information about the seizure, or may be uncorrelated with seizure events. For example, neural activity detected using the second electrode may comprise information related to other physiologic functions, such as gut activity, which may be unrelated to seizure.

[0087] In an example, the system 100 can be configured to differentiate between diverse patterns of neural activity using parameters such as firing rates, interspike intervals, and an overall pattern of neural discharges. Using signal processing and machine learning techniques, the system 100 can discern the subtle differences between neural signals that indicate a seizure and those that are indicative of normal physiological functions. The signal differentiation can be performed using information about temporal and spatial characteristics of the neural signals, which can be a function of the specific nerve bundles or branches being monitored. In an example, the system can evaluate the synchrony between the neural activities recorded by multiple electrodes. Synchrony, or a lack thereof, can provide additional insights into the state of the nervous system. In some examples, a high degree of synchrony between neural activities sensed at different electrodes during a specific time frame might suggest a coordinated neural response typical of a seizure event, or asynchronous firing patterns could indicate disparate physiological processes occurring independently of each other.

[0088] FIG. 6B illustrates generally a graphical representation of fifth neural activity 610 and sixth neural activity 612. The fifth neural activity 610 can be sensed using a third electrode (Electrode 3) and the sixth neural activity 612 can be sensed using a fourth electrode (Electrode 4). In the example of FIG. 6B, a seizure event can be identified in correlation with particular features of the fifth neural activity 610 and the sixth neural activity 612 together. For example, the seizure event may not be correlated with the firing patterns represented by the fifth neural activity 610 alone or by the sixth neural activity 612 alone. However, the seizure event can be correlated with a pattern that includes a period of increased neural activitysensed by the fourth electrode, such as at a seventh interval 614, that is followed by (e.g., after or within a specified duration) a period of increased neural activity sensed by the third electrode, such as at an eighth interval 616. In another example, the seizure event can be correlated with an increasing firing rate detected by the third electrode concurrently with a decreasing firing rate detected by the fourth electrode.

[0089] FIG. 5A, FIG. 5B, FIG. 6A, and FIG. 6B illustrate generally that the system 100 can be configured to distinguish between different types of neural activity patterns, for example, based on a rate of neural firings, based on a pattern or timing of neural firings, or both, such as can be applied using one or more neural activity sensing electrodes. Neural firing or activity detection and pattern classification helps enable more precise and personalized seizure detection, particularly when processed by a machine learning-based model as discussed herein. That is, machine learning-based algorithms performed by or using the system 100 can be configured to analyze such patient-specific features using machine learning techniques to determine if they match the patient-specific seizure-related features or characteristics previously identified. If a match is found, then the system can automatically proceed with an appropriate programmed response, such as notifying the patient or adjusting neurostimulation therapy parameters to potentially prevent or mitigate severity, duration, or other effects of a seizure.

[0090] FIG. 7 illustrates generally a graphical example of data that can be analyzed using a machine learning-based model. The example of FIG. 7 includes a data visualization example 700 that includes feature data, or physiologic parameter data, in clusters to help in understanding the multidimensional nature of physiological data that can be collected from a patient. In an example, each cluster represents or can be labeled as a different state or condition of the patient, such as baseline neural activity, seizure activity, or post-stimulation neural responses. By training a machine learning model with these (or other) labeled data clusters, the system 100can learn to recognize the subtle nuances that characterize the onset of a seizure, thereby enabling timely and precise intervention.

[0091] In machine learning-based modeling and data analysis approaches, a “feature” can refer to a set of multivariate parameters that occur during, or closely in association with, a specific state. For example,can denote a state “Y” at the time “t.” State Y can be defined as an n- dimensional variable. Variables xi,...xn can represent respective input variables. Where n> 3, visualization of the parameter space becomes challenging. However, for purposes of illustration, assume that three input variables can be identified that reliably differentiate between different states (e.g., baseline, present seizure, VNS delivery). In an example, xi can correspond to a neural (e.g., vagus) firing rate, X2 can correspond to local EEG activity, and r, can correspond to heart rate. When a machine learningbased model is trained by the baseline, seizure, and VNS data, it can determine that the baseline data forms a cluster (see, e.g., the baseline data cluster 706, the rightmost cluster in FIG. 7, with “\” hatching), that seizure data forms a cluster (see, e.g., the seizure data cluster 704, the middle clusterhatching), and the VNS data forms a cluster (see, e.g., the VNS delivery data cluster 702, the left cluster in FIG. 7 with finer “\\” hatching). A trained model instance could then, for example, later identify a y value in or near the middle region (“feature”), and use such identification to declare that a seizure is likely. In response, for example, VNS therapy could be enabled, or parameters altered, as a closed-loop therapy system. Reduction of dimensionality to a specified reduced set of input variables can be performed using a cascaded approach, such as performing Principal Component Analysis (PCA) on a larger group of input variables, and then reducing a count of input variables for machine learning model training after performing PCA or other evaluation.

[0092] As another illustration, initially, a stimulator (e.g., the implantable device 112) can be used as an open-loop stimulation source. In an open-loop mode, the stimulator can be in an always-on therapy mode, or can duty cycle a therapy between on and off states according to a deterministic or fixedduty cycle and not reacting to a seizure detection. In an example, the openloop mode can be used before an algorithm, or artificial intelligence-based (Al-based) model, is trained to identify different states in a patient. In a training mode, an Al-based (e.g., machine learning-based) model can be configured to adapt to differentiate between the baseline, VNS, and seizure features. After training, or optionally while training and model refinement is ongoing, the model can be used to detect a feature that is not similar to the baseline, such as closer to a locus of the seizure cluster, and in response, VNS can be enabled or altered. Such features may evolve over time.

[0093] For example, features can change over time as the nervous system can remodel over time or electrode effectiveness can vary over time such as due to scar tissue formation. In an example, an Al-based approach can be trained over time (e.g., continuously, intermittently, or periodically), or according to other specified triggering criteria, such as to learn the new features when the features are changing. Such an approach contrasts with a static non-AI programmed closed-loop system that may use predefined values or a range of values for each input variable for performing seizure detection.

[0094] As an illustration, assuming that at a time of the implant, a simultaneous high heart rate and high neuronal activity at an electrode are indicative of an occurrence of a seizure. Such occurrence could be detected based on historical clinical data that showed that this combination of features or parameters indicated a seizure was occurring, or could be based on patient confirmation that a seizure occurred via input into the implanted device via a patient input such as a magnet, or user input on an external device (e.g., the external device 122, or an external programmer or application running on a mobile device). A non-AI programmed closed-loop stimulator may monitor the heart rate and the electrode activity, and in response to detecting that both are increasing (e.g., corresponding to a statically defined feature), a system can enable VNS therapy, or update a VNS therapy parameter. In this illustration, assume that after a month, the high-heart rate and high electrode activity are no longer predictively associated with seizure. Instead, asimultaneous high heart rate and high electrode activity at a different electrode (e.g., corresponding to a newly defined feature) are now associated with the seizure. In this example, a non-AI programmed closed-loop stimulator may need to be manually reprogrammed based on the new feature definition, which may not be feasible in an ambulatory setting, because the new feature either cannot be detected by the system or the programming may not be modified.

[0095] By contrast, if an Al-based approach is used, particularly where a model instance is re-trained or personalized, the Al-based approach can be configured to observe and analyze the input variables for new features that can help it to establish or modify seizure detection criteria, or responsive therapy (or both). For example, if on Month 1, high xi and high x (e.g., respective features) are indicative of a seizure, then a machine learningbased approach can detect a seizure and deliver the VNS when both of the input variables are triggered (e.g., “high’'). If in Month 2, high xi and low X2 are indicative of a seizure, then the machine learning-based approach can adapt to deliver the VNS only when the xi is high-valued and X2 is low valued.

[0096] Model training can be performed on an ongoing basis. For example, training can be manually triggered, such as through confirmation from a patient or a clinician that a seizure was recognized and detected (e.g., using the validation module 322), or training can be triggered in an automated manner where the model instance or other portion of the system detects that a seizure was not actually occurring despite a previously identified seizure pattern. In response to detecting that a seizure has not actually occurred (e.g., in verification of a false-positive scenario), a machine learning-based approach can be triggered to re-train or perform model refinement to identify a new or updated seizure detection pattern.

[0097] In an example, VNS therapy or neurostimulation parameters (e.g., pulse width, signal amplitude, etc.) can be set by or using an Al-based approach. For example, an Al-based therapy update approach can include detecting an impact or result of VNS therapy on the detected feature whenthe therapy is applied at different amplitudes. Such operation can represent a self-stimulation-dose titration process that occurs at specified intervals or at the request of a physician or patient. For example, if a machine learningbased model detects that before a seizure, with VNS at amplitude 2 mA and pulse width 250 microseconds brings the vagus neural feature closer to the baseline feature compared to therapy delivered at 1.8 mA and 300 microseconds, then the model can indicate a preference for 2 mA, 250 microsecond settings for ongoing VNS therapy, or can modify such VNS parameters (e.g., subject to patient or clinician verification).

[0098] In an example, active electrodes for VNS set by an Al-based approach, such as a machine learning-based model implemented by the system 100, can change based on the input variables. For example, using electrode numbers 1, 2, 8, and 9 that are merely illustrative, the system 100 can identify that elevated firing rates on either electrode 1 or electrode 8 is associated with seizures. The system 100 can determine that applying VNS using electrode 2, which is closer to electrode 1 than other electrodes, is more effective in mitigating seizures triggered by high firing rates on electrode 1. Similarly, the system 100 can identify that VNS on electrode 9, which is closer to electrode 8, is more effective in countering seizures induced by high firing rates on electrode 8. In a more complex scenario where both electrode 1 and electrode 8 exhibit heightened firing rates, the system 100 can deduce that bipolar VNS between electrode 2 and electrode 9 yields the most effective stimulation to prevent seizure occurrences.Accordingly, the VNS could be delivered at different sites (electrodes) at specific times based on a stimulation protocol that is, at least in part, established using an Al-based approach.

[0099] FIG. 8 illustrates an example first method 800 for using a VNS system, such as the system 100, to provide a VNS therapy to a patient. Although the example first method 800 depicts a particular sequence of operations, the sequence may be altered without departing from the scope of the present disclosure. For example, some of the operations depicted may be performed in parallel or in a different sequence that does not materiallyaffect the function of the first method 800. In other examples, different components of an example device or system that implements the first method 800 may perform functions at substantially the same time or in a specific sequence.

[0100] At operation 802, the first method 800 includes sensing physiologic status information about a patient using one or more sensors coupled to an implantable vagus nerve stimulation (VNS) system. In an example, operation 802 includes receiving physiologic status information from a patient using one or more sensors (e.g., the sensor 118) that are coupled to, or comprise a portion of, an implantable Vagus Nerve Stimulation (VNS) system (e.g., the implantable device 112). In an example, the one or more sensors are configured to monitor various respective physiological parameters that can include, but are not limited to, nerve activity information, heart rate, heart rate variability, blood pressure, and respiration information, among other things. In an example, data collected by these sensors can be used as an input for a seizure detection algorithm or therapy titration algorithm. In an example, the seizure detection algorithm can be a machine learning-based model or algorithm.

[0101] At operation 804, the first method 800 includes training a machine learning-based model to detect a seizure event for the patient using first physiologic status information received from the one or more sensors. In an example, the model can be trained or applied using one or more features derived from the sensed physiologic status information from operation 802.

[0102] Training, at operation 804, can include analyzing historical physiologic status information about the patient, or about a population of patients, to identify patterns and correlations that indicate seizure events or a likelihood of an imminent seizure event. The model can be trained to recognize the unique physiological signatures associated with a patient's seizure activity, which may vary from one individual patient to another. During the training phase, the model can receive patient-reported or clinician-reported seizure event information to enhance efficiency or accuracy of the training process. In an example, other sensor data (orfeatures derived from such sensor data) that is indicative of a seizure event can be used, such as additionally or alternatively to the patient or clinician input, to identify actual seizure events. Operation 804 can thus provide a trained machine learning-based model that provides a personalized seizure detection algorithm tailored to each patient's specific seizure manifestation. In an example, at least a portion of the model can be based on data from one or more other patients, or a population of patients.

[0103] At operation 806, the first method 800 includes applying the trained machine learning-based model to second physiologic status information from the one or more sensors to detect a seizure event or to determine that a seizure event is imminent for the patient. The second physiologic status information can be received subsequently to the first physiologic status information used to train the model at operation 804. In an example, operation 806 includes using the model to process and analyze the second physiologic status information substantially in real-time with acquisition of the data to achieve early detection of seizures or timely prediction of imminent seizure events. In an example, the model applies the learned physiologic status information patterns to the incoming sensor data to determine whether the physiological signals align with the characteristics of a seizure. If the model detects a match or identifies a high probability of an impending seizure, then it triggers a specified response to address the event.

[0104] At operation 808, the first method 800 includes controlling the VNS system to provide a VNS therapy signal to the patient to treat a seizure event. In response to a result of the analysis performed using the machine learning-based model at operation 806, the VNS system can be activated to generate and provide a therapeutic neurostimulation signal to the vagus nerve. The parameters of this VNS therapy signal, such as intensity, frequency, duty cycle, waveform, and duration, among other parameters, can be automatically adjusted to the patient's immediate needs to effectively treat the seizure. Accordingly, timely and patient-specific intervention can be provided, which may help prevent the seizure from occurring, reduce its severity, or shorten its duration.

[0105] In an example, after operation 808, the machine learning-based model can be iteratively refined to further enhance the efficacy of the seizure management system over time. The model can be configured to continuously adapt to changes in the patient's physiologic manifestation of a seizure event or to changes in the patient's response to the VNS therapy delivered at operation 808.

[0106] For example, as the implantable device 112 or other VNS system provides therapy, the sensor 118 coupled to the implantable device 112 can continue to collect data, capturing the patient's physiological state during and after therapy. This data includes immediate responses to the therapy, such as changes in neural activity patterns, heart rate variability, and other relevant physiological parameters that may indicate the effectiveness of the intervention. The machine learning-based model can be configured to use this data to further train and update its performance. This re-training process can include analyzing the patient's response to the therapy (e.g., using direct patient feedback or input, or using physiologic status information automatically sensed about the patient) and determining whether the therapeutic intervention achieved the desired outcome, such as seizure cessation or reduction in severity. If the response deviates from the expected outcome, then the model can learn from this discrepancy and adjust its seizure event predictions and therapy parameter recommendations accordingly.

[0107] This ongoing training loop allows the model to adapt to each patient's unique physiological responses and to refine its seizure detection and prediction capabilities. It also enables the model to optimize therapy parameters for future interventions, ensuring that each patient receives the most effective personalized treatment over time.

[0108] FIG. 9 is a block diagram illustrating a machine in the example form of machine 900, within which a set or sequence of instructions may be executed to cause the machine to perform any one of the methodologies discussed herein, according to an example embodiment. In alternative embodiments, the machine operates as a standalone device or may beconnected (e.g., networked) to other machines. In a networked deployment, the machine may operate in the capacity of either a server or a client machine in server-client network environments, or it may act as a peer machine in peer-to-peer (or distributed) network environments. The machine may be an onboard vehicle system, wearable device, personal computer (PC), a tablet PC, a hybrid tablet, a personal digital assistant (PDA), a mobile telephone, or any machine capable of executing instructions (sequential or otherwise) that specify actions to be taken by that machine. Further, while only a single machine is illustrated, the term “machine” shall also be taken to include any collection of machines that individually or jointly execute a set (or multiple sets) of instructions to perform any one or more of the methodologies discussed herein. Similarly, the term “processor-based system” shall be taken to include any set of one or more machines that are controlled by or operated by a processor (e.g., a computer) to individually or jointly execute instructions to perform any one or more of the methodologies discussed herein. In an example, the system 100 and / or its components can comprise respective instances of the machine 900 or components thereof.

[0109] The example machine 900 includes at least one processor 902 (e.g., a central processing unit (CPU), a graphics processing unit (GPU) or both, processor cores, compute nodes, or the like), a main memory 904, and a static memory 906, which communicate with each other via a link 908 (e.g., bus). The machine 900 may further include a display unit 910 (e.g., comprising the external device 122), an input device 912 (e.g., a keyboard), and a user interface UI navigation device 914 (e.g., a mouse). In one embodiment, the display unit 910, input device 912, and UI navigation device 914 are incorporated into a single device housing such as a touch screen display. The machine 900 may additionally include a storage device 916 (e.g., a drive unit), a signal generation device 918 (e.g., a speaker), a network interface device 920, and one or more sensors 930, such as an inertial measurement unit (IMU), a global positioning system (GPS) sensor, compass, accelerometer, or other sensors, such as can comprise examples of the sensor 118 of the system 100. The machine 900 may include an outputcontroller 928, such as a serial (e.g., universal serial bus (USB), parallel, or other wired or wireless (e.g., infrared (IR), near field communication (NFC), or the like) connection to communicate or control one or more peripheral devices (e.g., a monitor, a printer, card reader, a computing device, or the like).

[0110] The storage device 916 includes a non-transitory machine-readable medium 922 on which is stored one or more sets of data structures and instructions 924 (e.g., software) embodying or utilized by any one or more of the methodologies or functions described herein. The instructions 924 may also reside, completely or at least partially, within the main memory 904, the static memory 906, and / or within the processor 902 during execution thereof by the machine 900, with the main memory 904, the static memory 906, and the processor 902 also constituting machine-readable media.

[0111] While the machine-readable medium 922 is illustrated in an example embodiment to be a single medium, the term “machine-readable medium” may include a single medium or multiple media (e.g., a centralized or distributed database, and / or associated caches and servers) that store the one or more instructions 924. The term “machine-readable medium” shall also be taken to include any tangible medium that is capable of storing, encoding, or carrying instructions for execution by the machine and that causes the machine to perform any one or more of the methodologies of the present disclosure or that is capable of storing, encoding or carrying data structures utilized by or associated with such instructions. The term “machine-readable medium” shall accordingly be taken to include, but not be limited to, solid-state memories, and optical and magnetic media. Specific examples of machine-readable media include non-volatile memory, including but not limited to, by way of example, semiconductor memory devices (e.g., electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM)) and flash memory devices; magnetic disks such as internal hard disks and removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks. Acomputer-readable storage device may be a machine-readable medium 922 that excluded transitory signals.

[0112] The instructions 924 may further be transmitted or received over a communications network 926 using a transmission medium via the network interface device 920 utilizing any one of a number of well-known transfer protocols (e.g., frame relay, internet protocol (IP), transmission control protocol (TCP), user datagram protocol (UDP), hypertext transfer protocol (HTTP), etc.). Examples of communication networks include a Local Area Network (LAN), a Wide Area Network (WAN), the Internet, mobile telephone networks, and wireless data networks (e.g., Wi-Fi, 3G, and 4G LTE / LTE-A or WiMAX networks, or the like). The term “transmission medium” shall be taken to include any intangible medium that is capable of storing, encoding, or carrying instructions for execution by the machine, and includes digital or analog communications signals or other intangible medium to facilitate communication of such software.

[0113] To better illustrate the systems, devices, and methods described herein, such as can include or use a machine learning-based model to process physiologic status information from one or multiple sources to identify a present or imminent seizure event, a non-limiting set of Example embodiments are set forth below as numerically identified Examples.

[0114] Example 1 is a method comprising: sensing physiologic status information about a patient using a sensor coupled to an implantable vagus nerve stimulation (VNS) system; applying, using a processor circuit, a machine learning-based model to the physiologic status information to detect a seizure event or to determine that a seizure event is imminent for the patient; and in response to information from the machine learning-based model indicating the seizure event was detected or is imminent, controlling a signal generator of the VNS system to provide a VNS therapy signal to the patient to treat the seizure event.

[0115] In Example 2, the subject matter of Example 1 optionally includes training the machine learning-based model to detect a seizure event for thepatient using the physiologic status information about the patient received from the sensor.

[0116] In Example 3, the subject matter of Example 2 optionally includes training the machine learning-based model to detect the seizure event includes using first vagal nerve activity information sensed from a vagus nerve of the patient.

[0117] In Example 4, the subject matter of Example 3 optionally includes training the machine learning-based model including using the first vagus nerve activity information and at least one other physiologic statusindicating feature about the patient, wherein the first vagus nerve activity information and the other physiologic status-indicating feature are concurrently received from or about the patient.

[0118] In Example 5, the subject matter of Example 4 optionally includes applying the machine learning-based model including using later-received second vagus nerve activity information from the patient.

[0119] In Example 6, the subject matter of Example 5 optionally includes applying the machine learning-based model including exclusively using the later-received second vagus nerve activity information from the patient to detect the seizure event or determine that the seizure event is imminent for the patient.

[0120] In Example 7, the subject matter of any one or more of Examples 4-6 optionally includes receiving audio and / or visual information about patient movements or vocalizations; wherein training the machine learning-based model includes using the audio and / or visual information about the patient movements or vocalizations as an input feature for the model.

[0121] In Example 8, the subject matter of any one or more of Examples 3-7 optionally includes sensing the vagus nerve activity information using one or more electrodes coupled to the implantable VNS system and disposed at or near a vagus nerve of the patient.

[0122] In Example 9, the subject matter of Example 8 optionally includes sensing other physiologic status information about the patient, wherein the other physiologic status information is other than information about vagusnerve activity. In Example 8, training the machine learning-based model to detect the seizure event can include using the vagus nerve activity information together with the other physiologic status information about the patient, wherein the machine learning-based model is configured to identify correlations between seizure occurrences and one or more features of the vagus nerve activity information and the other physiologic status information about the patient.

[0123] In Example 10, the subject matter of Example 9 optionally includes sensing the other physiologic status information about the patient using an implantable sensor.

[0124] In Example 11, the subject matter of Example 10 optionally includes applying the machine learning-based model to detect a seizure event or to determine that a seizure event is imminent for the patient including using the vagus nerve activity information and the other physiologic status information about the patient.

[0125] In Example 12, the subject matter of Example 11 optionally includes sensing vagus nerve activity timing characteristics.

[0126] In Example 13, the subject matter of any one or more of Examples 8-12 optionally includes sensing heart rate information about the patient; and wherein training the machine learning-based model to detect the seizure event includes using the vagus nerve activity information together with the heart rate information about the patient. The machine learning-based model can be configured to identify correlations between seizure occurrences and one or more features of the vagus nerve activity information and the heart rate information.

[0127] Example 14 is a method comprising: receiving training data from a patient using one or more physiologic sensors; using the training data, determining one or more training data features, including one or more features representing vagus nerve activity information from the patient; training a machine learning model to identify a seizure event for the patient using the training data features; receiving patient monitoring data from the patient using the same or other physiologic sensors; using the patientmonitoring data, determining one or more monitoring data features representing subsequent vagus nerve activity information from the patient; and applying the trained machine learning model to process the monitoring data features and provide a classification of a seizure status of the patient.

[0128] In Example 15, the subject matter of Example 14 optionally includes applying the trained machine learning model to provide the classification of the seizure status of the patient as one of a non-seizure status, a seizure event status, and an imminent seizure event status.

[0129] In Example 16, the subject matter of Example 15 optionally includes receiving a validation input indicating a validity of the of the seizure status and, in response, updating the machine learning model based on a relationship between the validation input and the monitoring data features.

[0130] In Example 17, the subject matter of any one or more of Examples 14-16 optionally includes, in response to the classification indicating an imminent or present seizure event status, providing a vagus nerve stimulation therapy to the patient.

[0131] In Example 18, the subject matter of Example 17 optionally includes updating one or more parameters of the vagus nerve stimulation therapy based on the classification of the seizure status of the patient.

[0132] In Example 19, the subject matter of any one or more of Examples 14-18 optionally includes receiving the training data from the patient using one or more physiologic sensors including receiving vagus nerve activity information from the patient.

[0133] In Example 20, the subject matter of Example 19 optionally includes determining the one or more training data features including determining a timing or intensity feature of the vagus nerve activity information.

[0134] In Example 21, the subject matter of any one or more of Examples 14-20 optionally includes receiving the training data comprising receiving vagus nerve activity information and other physiologic status information about the patient; wherein determining the one or more training data featuresincludes determining first features that represent the vagus nerve activity information and determining second features that represent the other physiologic status information about the patient; and wherein training the machine learning model includes using the first and second features.

[0135] In Example 22, the subject matter of Example 21 optionally includes receiving patient monitoring data from the patient including receiving subsequent vagus nerve activity information without receiving subsequent other physiologic status information about the patient, or receiving subsequent physiologic status information about the patient without receiving subsequent vagus nerve activity information.

[0136] Example 23 is a seizure event identification and response system, the system comprising: a sensor configured to sense physiologic status information about a patient, the sensed physiologic status information corresponding to a sensing duration; an implantable vagus nerve stimulation (VNS) system configured to provide a VNS therapy to a vagus nerve of a patient; and a processor circuit configured to: receive the physiologic status information about the patient from the sensor; apply a machine learningbased model to the physiologic status information to detect a seizure event or to determine that a seizure event is imminent for the patient. In response to information from the machine learning-based model indicating the seizure event was detected or is imminent, Example 23 can include controlling a signal generator of the VNS system to provide the VNS therapy to the vagus nerve of the patient to treat the seizure event.

[0137] In Example 24, the subject matter of Example 23 optionally includes the implantable VNS system comprising the sensor and the processor circuit.

[0138] In Example 25, the subject matter of any one or more of Examples 23-24 optionally includes an external interface device configured to receive validation information from the patient or a clinician about whether a seizure event was experienced by the patient concurrently with, or following, the sensing duration.

[0139] In Example 26, the subject matter of Example 25 optionally includes the processor circuit is configured to update the machine learningbased model based on the received validation information.

[0140] In Example 27, the subject matter of any one or more of Examples 23-26 optionally includes the processor circuit is configured to: receive a patient input indicating whether a seizure event was experienced by the patient; and update the machine learning-based model based on the patient input.

[0141] In Example 28, the subject matter of Example 27 optionally includes the implantable VNS system comprising a magnetic field sensor, and wherein the magnetic field sensor is configured to receive the patient input indicating whether the seizure event was experienced by the patient.

[0142] In Example 29, the subject matter of any one or more of Examples 23-28 optionally includes the sensor comprising one or more implantable electrodes configured to sense neural activity information from a nerve of the patient.

[0143] In Example 30, the subject matter of Example 29 optionally includes the sensor comprising a heart rate sensor configured to sense heart rate information about the patient, and wherein the processor circuit is configured to apply the machine learning-based model to process the neural activity information and the heart rate information together to detect the seizure event or to determine that the seizure event is imminent for the patient.

[0144] In Example 31, the subject matter of any one or more of Examples 29-30 optionally includes the sensor comprising an accelerometer configured to measure motion information about the patient, and wherein the processor circuit is configured to apply the machine learning-based model to process the neural activity information and the motion information together to detect the seizure event or to determine that the seizure event is imminent for the patient.

[0145] In Example 32, the subject matter of Example 31 optionally includes the accelerometer is configured to measure cardiac motioninformation about the patient, and the processor circuit is configured to use the cardiac motion information as an input to the machine learning-based model.

[0146] Example 33 is at least one machine -readable medium including instructions that, when executed by processing circuitry, cause the processing circuitry to perform operations to implement of any of any one or more of Examples 1-32.

[0147] Example 34 is an apparatus comprising means to implement any one or more of Examples 1-32.

[0148] Example 35 is a system to implement any one or more of Examples 1-32.

[0149] Example 36 is a method to implement any one or more of Examples 1-32.

[0150] Each of these non-limiting examples can stand on its own, or can be combined in various permutations or combinations with one or more of the other examples.

[0151] The above detailed description includes references to the accompanying drawings, which form a part of the detailed description. The drawings show, by way of illustration, specific embodiments in which the invention can be practiced. These embodiments are also referred to generally as “examples.” Such examples can include elements in addition to those shown or described. However, the present inventors also contemplate examples in which only those elements shown or described are provided. Moreover, the present inventors also contemplate examples using any combination or permutation of those elements shown or described (or one or more aspects thereof), either with respect to a particular example (or one or more aspects thereof), or with respect to other examples (or one or more aspects thereof) shown or described herein.

[0152] In the event of inconsistent usages between this document and any documents so incorporated by reference, the usage in this document controls.

[0153] In this document, the terms “a” or “an” are used, as is common in patent documents, to include one or more than one, independent of any other instances or usages of “at least one” or “one or more.” In this document, the term “or” is used to refer to a nonexclusive or, such that “A or B” includes “A but not B,” “B but not A,” and “A and B,” unless otherwise indicated. In this document, the terms “including” and “in which” are used as the plain- English equivalents of the respective terms “comprising” and “wherein.” Also, in the following aspects, the terms “including” and “comprising” are open-ended, that is, a system, device, article, composition, formulation, or process that includes elements in addition to those listed after such a term in a claim are still deemed to fall within the scope of that claim. Moreover, in the following aspects, the terms “first,” “second,” and “third,” etc., are used merely as labels, and are not intended to impose numerical requirements on their objects.

[0154] Method examples described herein can be machine or computer- implemented at least in part. Some examples can include a computer- readable medium or machine-readable medium encoded with instructions operable to configure an electronic device to perform methods as described in the above examples. An implementation of such methods can include code, such as microcode, assembly language code, a higher-level language code, or the like. Such code can include computer readable instructions for performing various methods. The code may form portions of computer program products. Such instructions can be read and executed by one or more processors to enable performance of operations comprising a method, for example. The instructions are in any suitable form, such as but not limited to source code, compiled code, interpreted code, executable code, static code, dynamic code, and the like.

[0155] Further, in an example, the code can be tangibly stored on one or more volatile, non-transitory, or non-volatile tangible computer-readable media, such as during execution or at other times. Examples of these tangible computer-readable media can include, but are not limited to, hard disks, removable magnetic disks, removable optical disks (e.g., compactdisks and digital video disks), magnetic cassettes, memory cards or sticks, random access memories (RAMs), read only memories (ROMs), and the like.

[0156] The above description is intended to be illustrative, and not restrictive. For example, the above-described examples (or one or more aspects thereof) may be used in combination with each other. Other embodiments can be used, such as by one of ordinary skill in the art upon reviewing the above description. The Abstract is provided to allow the reader to quickly ascertain the nature of the technical disclosure. It is submitted with the understanding that it will not be used to interpret or limit the scope or meaning of the claims. Also, in the above Detailed Description, various features may be grouped together to streamline the disclosure. This should not be interpreted as intending that an unclaimed disclosed feature is essential to any claim. Rather, inventive subject matter may lie in less than all features of a particular disclosed embodiment. Thus, the following aspects are hereby incorporated into the Detailed Description as examples or embodiments, with each aspect standing on its own as a separate embodiment, and it is contemplated that such embodiments can be combined with each other in various combinations or permutations.

Claims

CLAIMSWhat is claimed is:

1. A method comprising: sensing physiologic status information about a patient using a sensor coupled to an implantable vagus nerve stimulation (VNS) system; applying, using a processor circuit, a machine learning-based model to the physiologic status information to detect a seizure event or to determine that a seizure event is imminent for the patient; and in response to information from the machine learning-based model indicating the seizure event was detected or is imminent, controlling a signal generator of the VNS system to provide a VNS therapy signal to the patient to treat the seizure event.

2. The method of claim 1, comprising training the machine learning-based model to detect a seizure event for the patient using the physiologic status information about the patient received from the sensor.

3. The method of claim 2, wherein training the machine learning-based model to detect the seizure event includes using first vagal nerve activity information sensed from a vagus nerve of the patient.

4. The method of claim 3, wherein training the machine learning-based model includes using the first vagus nerve activity information and at least one other physiologic status-indicating feature about the patient, wherein the first vagus nerve activity information and the other physiologic statusindicating feature are concurrently received from or about the patient.

5. The method of claim 4, wherein applying the machine learning-based model includes using later-received second vagus nerve activity information from the patient.

6. The method of claim 5, wherein applying the machine learning-based model includes exclusively using the later-received second vagus nerve activity information from the patient to detect the seizure event or determine that the seizure event is imminent for the patient.

7. The method of claim 4, comprising receiving audio and / or visual information about patient movements or vocalizations; wherein training the machine learning-based model includes using the audio and / or visual information about the patient movements or vocalizations as an input feature for the model.

8. The method of claim 3, comprising sensing the vagus nerve activity information using one or more electrodes coupled to the implantable VNS system and disposed at or near a vagus nerve of the patient.

9. The method of claim 8, comprising sensing other physiologic status information about the patient, wherein the other physiologic status information is other than information about vagus nerve activity; and wherein training the machine learning-based model to detect the seizure event includes using the vagus nerve activity information together with the other physiologic status information about the patient, wherein the machine learning-based model is configured to identify correlations between seizure occurrences and one or more features of the vagus nerve activity information and the other physiologic status information about the patient.

10. The method of claim 9, comprising sensing the other physiologic status information about the patient using an implantable sensor.

11. The method of claim 10, wherein applying the machine learning-based model to detect a seizure event or to determine that a seizure event is imminent for the patient includes using the vagus nerve activity information and the other physiologic status information about the patient.

12. The method of claim 11, wherein sensing the vagus nerve activity information includes sensing vagus nerve activity timing characteristics.

13. The method of claim 8, comprising sensing heart rate information about the patient; and wherein training the machine learning-based model to detect the seizure event includes using the vagus nerve activity information together with the heart rate information about the patient, wherein the machine learning-based model is configured to identify correlations between seizure occurrences and one or more features of the vagus nerve activity information and the heart rate information.

14. A method comprising: receiving training data from a patient using one or more physiologic sensors; using the training data, determining one or more training data features, including one or more features representing vagus nerve activity information from the patient; training a machine learning model to identify a seizure event for the patient using the training data features; receiving patient monitoring data from the patient using the same or other physiologic sensors; using the patient monitoring data, determining one or more monitoring data features representing subsequent vagus nerve activity information from the patient; and applying the trained machine learning model to process the monitoring data features and provide a classification of a seizure status of the patient.

15. The method of claim 14, wherein applying the trained machine learning model provides the classification of the seizure status of the patient as one of a non-seizure status, a seizure event status, and an imminent seizure event status.

16. The method of claim 15, comprising receiving a validation input indicating a validity of the of the seizure status and, in response, updating the machine learning model based on a relationship between the validation input and the monitoring data features.

17. The method of claim 14, wherein in response to the classification indicating an imminent or present seizure event status, the method comprises providing a vagus nerve stimulation therapy to the patient.

18. The method of claim 17, comprising updating one or more parameters of the vagus nerve stimulation therapy based on the classification of the seizure status of the patient.

19. The method of claim 14, wherein receiving the training data from the patient using one or more physiologic sensors includes receiving vagus nerve activity information from the patient.

20. The method of claim 19, wherein determining the one or more training data features includes determining a timing or intensity feature of the vagus nerve activity information.

21. The method of claim 14, wherein receiving the training data comprises receiving vagus nerve activity information and other physiologic status information about the patient; wherein determining the one or more training data features includes determining first features that represent the vagus nerve activity information and determining second features that represent the other physiologic status information about the patient; wherein training the machine learning model includes using the first and second features.

22. The method of claim 21, wherein receiving patient monitoring data from the patient includes receiving subsequent vagus nerve activity information without receiving subsequent other physiologic status information about thepatient, or receiving subsequent physiologic status information about the patient without receiving subsequent vagus nerve activity information.

23. A seizure event identification and response system, the system comprising: a sensor configured to sense physiologic status information about a patient, the sensed physiologic status information corresponding to a sensing duration; an implantable vagus nerve stimulation (VNS) system configured to provide a VNS therapy to a vagus nerve of a patient; and a processor circuit configured to: receive the physiologic status information about the patient from the sensor; apply a machine learning-based model to the physiologic status information to detect a seizure event or to determine that a seizure event is imminent for the patient; and in response to information from the machine learning-based model indicating the seizure event was detected or is imminent, controlling a signal generator of the VNS system to provide the VNS therapy to the vagus nerve of the patient to treat the seizure event.

24. The system of claim 23, wherein the implantable VNS system comprises the sensor and the processor circuit.

25. The system of claim 23, comprising an external interface device configured to receive validation information from the patient or a clinician about whether a seizure event was experienced by the patient concurrently with, or following, the sensing duration.

26. The system of claim 25, wherein the processor circuit is configured to update the machine learning-based model based on the received validation information.

21. The system of claim 23, wherein the processor circuit is configured to: receive a patient input indicating whether a seizure event was experienced by the patient; and update the machine learning-based model based on the patient input.

28. The system of claim 27, wherein the implantable VNS system comprises a magnetic field sensor, and wherein the magnetic field sensor is configured to receive the patient input indicating whether the seizure event was experienced by the patient.

29. The system of claim 23, wherein the sensor comprises one or more implantable electrodes configured to sense neural activity information from a nerve of the patient.

30. The system of claim 29, wherein the sensor comprises a heart rate sensor configured to sense heart rate information about the patient, and wherein the processor circuit is configured to apply the machine learning-based model to process the neural activity information and the heart rate information together to detect the seizure event or to determine that the seizure event is imminent for the patient.

31. The system of claim 29, wherein the sensor comprises an accelerometer configured to measure motion information about the patient, and wherein the processor circuit is configured to apply the machine learning-based model to process the neural activity information and the motion information together to detect the seizure event or to determine that the seizure event is imminent for the patient.

32. The system of claim 31, wherein the accelerometer is configured to measure cardiac motion information about the patient, and wherein the processor circuit is configured to use the cardiac motion information as an input to the machine learning-based model.

Images