WO2024119066A1 - Réseau neuronal profond pour la neuromodulation en boucle fermée de la vessie - Google Patents

Réseau neuronal profond pour la neuromodulation en boucle fermée de la vessie Download PDF

Info

Publication number
WO2024119066A1
WO2024119066A1 PCT/US2023/082067 US2023082067W WO2024119066A1 WO 2024119066 A1 WO2024119066 A1 WO 2024119066A1 US 2023082067 W US2023082067 W US 2023082067W WO 2024119066 A1 WO2024119066 A1 WO 2024119066A1
Authority
WO
WIPO (PCT)
Prior art keywords
event
bladder
vesical pressure
pressure data
vesical
Prior art date
Application number
PCT/US2023/082067
Other languages
English (en)
Inventor
Vikram ABBARAJU
Steven MAJERUS
Original Assignee
Case Western Reserve University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Case Western Reserve University filed Critical Case Western Reserve University
Publication of WO2024119066A1 publication Critical patent/WO2024119066A1/fr

Links

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/20—Measuring for diagnostic purposes; Identification of persons for measuring urological functions restricted to the evaluation of the urinary system
    • A61B5/202—Assessing bladder functions, e.g. incontinence assessment
    • A61B5/205—Determining bladder or urethral pressure
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61N—ELECTROTHERAPY; MAGNETOTHERAPY; RADIATION THERAPY; ULTRASOUND THERAPY
    • A61N1/00—Electrotherapy; Circuits therefor
    • A61N1/18—Applying electric currents by contact electrodes
    • A61N1/32—Applying electric currents by contact electrodes alternating or intermittent currents
    • A61N1/36—Applying electric currents by contact electrodes alternating or intermittent currents for stimulation
    • A61N1/36007—Applying electric currents by contact electrodes alternating or intermittent currents for stimulation of urogenital or gastrointestinal organs, e.g. for incontinence control
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61N—ELECTROTHERAPY; MAGNETOTHERAPY; RADIATION THERAPY; ULTRASOUND THERAPY
    • A61N1/00—Electrotherapy; Circuits therefor
    • A61N1/18—Applying electric currents by contact electrodes
    • A61N1/32—Applying electric currents by contact electrodes alternating or intermittent currents
    • A61N1/36—Applying electric currents by contact electrodes alternating or intermittent currents for stimulation
    • A61N1/3605—Implantable neurostimulators for stimulating central or peripheral nerve system
    • A61N1/3606—Implantable neurostimulators for stimulating central or peripheral nerve system adapted for a particular treatment

Definitions

  • the present disclosure relates generally to urinary tract dysfunction, and more specifically, to systems and methods that can use a deep neural network for closed-loop bladder neuromodulation to stop urinary incontinence.
  • Urinary incontinence due to injury or a dysfunction, is a condition that significantly reduces a patient’s quality of life.
  • Neuromodulation has emerged as a means of restoring bladder function in individuals with urinary incontinence.
  • bladder pressure measurements need to be quickly and accurately classified into bladder events for real time control of bladder functions.
  • Common methods of analyzing urodynamic study (UDS) tracings to classify bladder events are prone to errors and are subjective.
  • Common systems for bladder pressure measurement are prone to significant artifacts and unable to sufficiently weed out the artifacts for good real-time bladder control via neuromodulation.
  • the present disclosure can include a system that can include a sensing device adapted to be positioned within a patient’s bladder to sense vesical pressure and a signal processing device.
  • the sensing device can sample the vesical pressure over a time period to provide vesical pressure data samples.
  • the signal processing device can include a memory storing instructions and a processor to execute the instructions to classify the sensed vesical pressure into a bladder event.
  • the instructions can include at least the following.
  • the buffer can include a plurality of vesical pressure data samples collected over a previous time period.
  • the vector of probabilities can include a probability of an abdominal event, a probability of a detrusor overactivity event, a probability of no event, and a probability of a voiding bladder contraction event.
  • the present disclosure can include a method for classifying a bladder event, including the following steps performed by a signal processing device including a processor.
  • a signal processing device including a processor.
  • the sensing device can be adapted to be positioned within a patient’s bladder to sense vesical pressure and the vesical pressure can be sampled over a time period to provide vesical pressure data samples.
  • Applying a multi-level discrete wavelet transform to the plurality of vesical pressure data samples to output approximation coefficient arrays and detail coefficient arrays.
  • labeling the event based on the vector of probabilities wherein the label can include one of the abdominal event, the detrusor overactivity event, no event, and the voiding bladder contraction event.
  • FIG. 1 is a block diagram showing a system for controlling bladder function
  • FIG. 2 is a block diagram showing details of the sensing device and the signal processing device from the system of FIG. 1 ;
  • FIG. 3 is a block diagram showing further details of the signal processing device
  • FIG. 4 is a general flow diagram illustrating the classification process followed by the signal processing device
  • FIG. 5 is an example logic flow diagram illustrating the classification process followed by the signal processing device
  • FIG. 6 is a logic flow diagram illustrating the discrete wavelet transform
  • FIG. 7 is a logic flow diagram illustrating the feature extraction
  • FIG. 8 is a diagram illustrating the neural network
  • FIG. 9 is a process flow diagram illustrating a method of classifying bladder events from bladder pressure
  • FIG. 10 is a process flow diagram illustrating a method for treating urinary tract dysfunction
  • FIGS. 11-17 are illustrations representing the system tested in a first experiment.
  • FIGS. 18-22 are graphical representations of results and/or processes of a second experiment. Detailed Description
  • the term “bladder” can refer to a hollow organ, which may be generally triangle-shaped, that stores urine from the kidneys before disposal by urination.
  • the bladder’s walls relax and expand to store urine and contract and flatten to empty urine through the urethra. Normal urination can be based on a contraction of the bladder walls and opening of the urethra triggered by a neural signal when a subject intends to urinate.
  • Urinary incontinence can refer to loss of bladder control, often resulting in involuntary urination, caused by one or more nerve and/or muscle failures in a patient’s urinary system due to dysfunction and/or injury. Involuntary urination can be due to a non-neural event, like coughing, laughing, change in posture, or the like where urination is not intended. Urinary incontinence can include urge incontinence, stress incontinence, overflow incontinence, and the like. The terms “urinary incontinence” and “overactive bladder” can be used interchangeably.
  • bladder function can refer to the normal workings of the bladder, and the entire urinary system (e.g., voiding, holding it, control of voiding speed, etc.).
  • Loss of bladder function can refer to one or more aspects of incontinence that change any aspect of normal functioning of the urinary system.
  • neuromodulation can refer to electrical stimulation of one or more nerves (such as the genital nerves, sphincter control nerves, or the like) to control bladder functions (e.g., voiding).
  • nerves such as the genital nerves, sphincter control nerves, or the like
  • control bladder functions e.g., voiding
  • closed-loop can refer to an automatic control system in which an operation (e.g., controlling bladder function via neuromodulation) is regulated by feedback (e.g., vesical pressure of the bladder).
  • an operation e.g., controlling bladder function via neuromodulation
  • feedback e.g., vesical pressure of the bladder
  • the term “deep neural network” can refer to a machine learning technique that can process data in complex ways by employing sophisticated mathematical modeling.
  • the deep neural network can have one or more hidden layers of nodes between the input and output layer.
  • the nodes can operate like parts of the human brain and, thus, can be referred to as neurons, synapses, weights, biases, and functions. If the output of any individual node is above the specified threshold value, that node is activated, sending data to the next layer of the network. Otherwise, no data is passed along to the next layer of the network.
  • the term “deep neural network” can be used interchangeably with “neural network” herein.
  • the term "patient” can refer to any warm-blooded organism including, but not limited to, a human being, a pig, a rat, a mouse, a dog, a cat, a goat, a sheep, a horse, a monkey, an ape, a rabbit, a cow, etc.
  • the terms “subject” and “patient” can be used interchangeably herein.
  • Urinary dysfunction including urinary incontinence, can develop naturally as patients age or as a consequence of injuries, such as spinal cord injury, trauma from giving birth, or the like, and can lead to a significant reduction in patients’ quality of life.
  • urinary dysfunction can be studied using one or more urodynamics study (UDS), which can be conducted with one or two catheters (or other bladder pressure sensing devices) for recording and can be static or ambulatory.
  • UDS can provide qualitative and quantitative analyses of bladder relaxation and contractile behavior during bladder filling and voiding.
  • Results of UDS can be input to a closed loop bladder neuromodulation system that can provide neuromodulation of one or more nerves, for example one or more genital nerves (e.g., the genital branch of the genitofemoral nerve).
  • genital nerves e.g., the genital branch of the genitofemoral nerve
  • interpretation of UDS data can be subjective and variable due, at least in part, to the presence of artifacts and technician-related errors, as well as to a lack of standardization.
  • an automated method for detecting and classifying bladder events can standardize clinical urodynamic interpretations. Standardized clinical interpretations can lead to standardization of feedback used as inputs to closed-loop neuromodulation systems, which can be used to treat urinary incontinence and thereby increase patients’ quality of life.
  • An aspect of the present disclosure relates to a system 100 (FIG. 1 ) that can employ closed loop neuromodulation of at least one nerve to control bladder function and treat urinary incontinence.
  • Incontinence may be caused by nerves exciting the bladder, relaxing the urethra, and/or exciting the external urethral sphincter.
  • neuromodulation of the genital nerve or, more generally, the genitofemoral nerve
  • the system 100 can detect an urge from the bladder from measured bladder vesical pressure, employ a machine learning framework, including a deep neural network, to classify the urge as a bladder event, and based on the classification, deliver a neural stimulation to control the bladder function related to the urge.
  • a machine learning framework including a deep neural network
  • the system 100 includes a sensing device 1 10 connected wirelessly and/or with a wired connection to a signal processing device 112 and a neuromodulation device 114 connected to the signal processing device 112 via a wired and/or wireless connection.
  • the sensing device 110 can be positioned within the bladder to detect bladder vesical pressure.
  • the signal processing device 1 12 can include a memory and processor (not shown in FIG. 1 ) for classifying a bladder vesical pressure sample corresponding to a change in pressure at a given time into a bladder event.
  • the neuromodulation device 114 can include a signal generator 116 and at least one electrode 118 for generating and applying a stimulation signal in response to the event classified by the signal processing device 1 12.
  • FIG. 1 shows an example of a real time, closed loop system 100 for controlling bladder function based on the classification of a bladder event at a time.
  • the system 100 can include the sensing device 110 that can include at least a pressure sensor (not shown in FIG. 1 , may include additional or alternative sensors) and can be positioned within a patient’s bladder.
  • the sensing device 110 can be free floating in the patient’s bladder, implanted in a bladder wall, or can be embodied in a catheter.
  • the sensing device 110 can sense bladder vesical pressure samples at a constant rate. For example, the sensing device 110 can detect bladder vesical pressure samples of 0.8 seconds at a rate of 10 Hz.
  • the sensing device 110 can be wirelessly connected to the signal processing device 112 (e.g., with WIFI, Blutetooth®, RF signaling, etc.) (as shown). Alternatively, the sensing device 110 and the signal processing device 112 can be connected, at least in part, with a wire. [0037]
  • the signal processing device 112 can receive the bladder vesical pressure samples from the sensing device 110 and can transform and compute the bladder vesical pressure samples to determine bladder event classification labels in real time (e.g., within about 5 seconds of receiving a given sample, 1 second of receiving a given sample, 0.5 seconds of receiving a given sample, or the like).
  • the bladder event classification labels can include an abdominal event, a detrusor overactivity event, no event, an artifact, and a voiding bladder contraction event (the only intended bladder voiding event).
  • the signal processing device 1 12 can communicate with the neuromodulation device 114 over a wired and/or wireless (e.g., WIFI, Blutetooth®, RF signaling, etc.) connection to send an output of a bladder event classification as an input to a neuromodulation device 114.
  • the neuromodulation device 114 can receive an input indicating the labeled event and, based on the input, the neuromodulation device 114 can deliver a stimulation to at least one nerve.
  • the neuromodulation device 114 can include a signal generator 116 (e.g., an electrical signal generator) in communication with the signal processing device 1 12 that can configure and generate a stimulation (e.g., electrical waveform) in response to the labeled event.
  • a signal generator 116 e.g., an electrical signal generator
  • the signal processing device 1 12 can signal the neuromodulation device 114 to deliver the stimulation to prevent voiding when the bladder event label corresponds to the voiding contraction event.
  • the signal generator 116 of the neuromodulation device 114 can, for example, configure one or more parameters of the stimulation such as frequency, amplitude, power, intensity, duration, pulsed vs continuous, AC/DC, electrode being used, or the like.
  • the neuromodulation device 114 can also include one or more electrodes 118 that can be in electrical contact with the at least one nerve (e.g., a genital branch of the genitofemoral nerve) and can apply the stimulation to the at least one nerve.
  • the one or more electrodes 118 can be implanted in direct contact with the at least one nerve.
  • the one or more electrodes 118 can be near but not in contact with the at least one nerve.
  • the one or more electrodes 118 can also be separate device(s) in communication (e.g., via one or more leads) with the signal generator 116.
  • FIG. 2 shows communication between the sensing device 110 and the signal processing device 1 12 in greater detail.
  • the sensing device 110 can be positioned within a patient’s bladder (e.g., free floating, implanted, catheter, etc.) and can sense pressure within the bladder, such as bladder vesical pressure.
  • the bladder vesical pressure can be sampled over a time period to provide vesical pressure data samples (e.g., samples of 0.8 seconds in length at a rate of 10Hz for 1 hour, 5 hours, 12 hours, 24 hours, or the like).
  • the sensing device 110 can include at least a pressure sensor 120, a battery 122, and a wireless transceiver 124.
  • the sensing device 110 can include other components not shown, such as circuitry, a housing, positioning and/or implantation mechanisms, or the like.
  • the pressure sensor 120 can detect the bladder vesical pressure within the patient’s bladder.
  • the pressure sensor 120 can be implemented as electrical resistors, piezoelectric components, resistors implemented on a MEMS device, or the like.
  • the pressure sensor can include a circuit to account for offset and/or drift cancellation.
  • the battery 122 can provide power to the pressure sensor 120 and/or the wireless transceiver 124.
  • the battery 122 can be, for instance, rechargeable.
  • the wireless transceiver 124 can transmit a signal indicating the vesical pressure within the patient’s bladder at a given rate (e.g., matching the detection rate).
  • the wireless transceiver 124 can be in wireless communication with the wireless transceiver 130 of the signal processing device 1 12.
  • the sensing device 110 can also include a biocompatible housing (not shown) that can encapsulate at least one of the pressure sensor 120, the battery 122, and the wireless transceiver 124.
  • the signal processing device 112 can include a non-transitory memory 126 that can store instructions and a processor 128 that can execute the instructions.
  • the signal processing device 112 can also include the wireless transceiver 130 that can receive the signal indicating the bladder vesical pressure from the wireless transceiver 124 of the sensing device 110.
  • the signal processing device 112 can also include components such as circuitry, a battery or power source, a user interface, and/or a display (not shown).
  • the signal processing device 112 can be a dedicated device, a computer, a smartphone, a tablet, or the like.
  • FIG. 3 shows the signal processing device 112 and the instructions that can be executed by the processor 128.
  • the signal processing device 112 can receive 132 a vesical pressure data sample (shown as PVES) at a time from the sensing device 110 via the wireless transceivers 124 and 130.
  • the bladder vesical pressure data sample can be input 134 chronologically into a buffer comprising a plurality of vesical pressure data samples collected over a previous time period.
  • the current bladder vesical pressure data sample can become a part of the plurality of bladder vesical pressure data samples. If the buffer is not full, then the input 134 can be repeat for the number of bladder vesical pressure data samples needed to fill the buffer.
  • the signal processing device 112 can apply a multi-level discrete wavelet transform 136 to the plurality of bladder vesical pressure data samples to output approximation coefficient arrays and detail coefficient arrays for each level of the multi-level discrete wavelet transform.
  • a feature extraction application 138 of the signal processing device 1 12 can extract a plurality of statistical features using a portion of the vesical pressure data samples (e.g., the most recent X number of samples), the approximation coefficient arrays, and the detail coefficient arrays as inputs.
  • the signal processing device 112 can then apply a deep neural network 140 to the plurality of statistical features extracted by 138 to compute a vector of probabilities.
  • the vector of probabilities can include a probability of an abdominal event, a probability of a detrusor overactivity event, a probability of no event, and a probability of a voiding bladder contraction event.
  • a labeler 142 can label the event based on the vector of probabilities as one of the abdominal event, the detrusor overactivity event, no event, and the voiding bladder contraction event.
  • the labeled event and/or something related to the labeled event indicative of the labeled event can be output at 144 to the neuromodulation device 114 (in other words, the output 144 can be used as an input to the neural modulation device; it should be understood that the neuromodulation device need not take action for every labeled event and/or at least a portion of the labeled events may not be output to the neuromodulation device).
  • the output can alternatively and/or additionally be sent to a display (not shown), memory 126, an external device related to a medical professional and/or researcher, or the like.
  • FIG. 4 shows an example diagram 400 of the classification of a bladder event based on a vesical pressure data sample (PVES), assuming the buffer is already full to limit N.
  • the newest vesical pressure data sample can be received and input into the buffer.
  • the buffer can be a first in first out buffer and the oldest vesical pressure data sample can be removed to make room for the newest sample.
  • the buffer can output the M most recent vesical pressure data samples. For example, the 8 most recent samples.
  • the plurality of vesical pressure data samples stored in the buffer (including the most recent) can be input into the multi-level discrete wavelet transform, which can have X levels.
  • the multi-level discrete wavelet transform can then output X number of approximation and detail coefficient arrays, each (e.g., [cAi], [cDi],...[cAx], [cDx]).
  • the approximation and detail coefficient arrays can decrease in size by half for each successive level of the multi-level discrete wavelet transform.
  • the M most recent vesical pressure data samples and the approximation and detail coefficient arrays can be input into a feature extractor where Y number of statistical features can be extracted. For example, 16 features can be extracted by compression, ratio, and/or taking the last value of one or more of the approximation and detail coefficient arrays.
  • the Y number of statistical features can be input into the neural network.
  • the neural network can compute the probability of each event classification as a vector, where each value of the vector represents a probability of a given event.
  • the probabilities vector can be filtered, for example by a majority vote filter.
  • the class probability vector (filtered or unfiltered can then be used to determine an event classification label.
  • the class event can be output and sent to the neuromodulation device, for example.
  • FIG. 5 shows another example 500 of the logic process of the signal processing device 1 12.
  • the vesical pressure data sample can be received from the signal processing device and sent to the buffer.
  • the buffer can for example, have a limit N of 512 samples.
  • the signal processing device 112 can query if the buffer is full, and if the buffer is not full can receive vesical pressure data samples until the buffer is full. If the buffer is full then the signal processing device can run, in order, the multi-level wavelet transform, the feature extraction, the neural network (where the oldest 8 samples can be discarded), and the majority vote filter to transform the vesical pressure data sample(s) into an event classification label and output the event classification label.
  • FIG. 6 shows an example 600 of the multi-level discrete wavelet transform being applied by the signal processing device 112 to the plurality of vesical pressure data samples from the buffer (including the received vesical pressure data sample at the time) to output approximation coefficient arrays and detail coefficient arrays.
  • the multi-level discrete wavelet transform can not only extract frequency information that is embedded in an input signal (e.g., the plurality of vesical pressure data samples), similar to a fast Fourier transform (FFT) and short-time Fourier Transform (STFT) but can also localize the frequency content in time with high resolution.
  • FFT fast Fourier transform
  • STFT short-time Fourier Transform
  • the approximation coefficient arrays can represent the low-frequency content of the plurality of vesical pressure data samples and the detail coefficient arrays can represent the high-frequency content of the plurality of vesical pressure data samples.
  • Each level of the multi-level discrete wavelet transform can output an approximation coefficient array and a detail coefficient array. For example, if the multi-level discrete wavelet transform is five levels, then the five-level discrete wavelet transform can output five approximation coefficient arrays and five detail coefficient arrays.
  • the multi-level discrete wavelet transform can low pass filter and high- pass filter an input and down-samples each of the results by a factor of two for each level.
  • the input for the first level can be the plurality of vesical pressure data samples.
  • the inputs for the rest of the levels can be the approximation coefficient array from the previous level.
  • the down sampling can be by a factor of two to reduce storage usage such that each set of coefficient arrays is half the size of the previous set (e.g., 256, 128, 64, 32, 16).
  • the multilevel discrete wavelet transform can be a lifting transform that can allow the approximation and detail coefficient arrays to be computed in-place for increased storage efficiency and decreased computational complexity.
  • the lifting transform can be a Daubechies 4 discrete wavelet transform.
  • the Daubechies 4 discrete wavelet transform can first split the input array (e.g., the plurality of vesical pressure data samples or the previous level’s approximation coefficient array) into even and odd-indexed components, then perform a series of update and prediction operations that can represent high- and low-pass filtering, respectively.
  • FIG. 7 shows an example 700 of the feature extraction by the signal processing device 1 12.
  • One or more statistical features can be extracted by the signal processing device via the feature extraction application.
  • the approximation coefficient arrays, the detail coefficient arrays, and a portion of the vesical pressure data samples can be input into the feature extraction application.
  • a portion of the most recent (chronologically) vesical pressure data samples such as the most recent 5, 8, or 10 samples, can be input.
  • Extracting the statistical features can include mapping a number of most recently obtained vesical pressure data samples, for example 8, from the plurality of vesical pressure data samples, the approximation coefficients, and the detail coefficients into individual statistical inputs for the neural network.
  • the number of most recently obtained vesical pressure data samples can be compressed into a single value to be a first statistical feature of the neural network. For example, by determining the mean. Additionally, two statistical features can be extracted from each of the levels of the detail and approximation coefficient arrays, by compressing (taking the mean) the larger arrays (if the event is considered from the last 8 samples, then the last four values from the first level arrays, and the last two values from the second level arrays) and not compressing the smaller arrays but simply using the last value of each array of the third, fourth, and fifth level as features. Another statistical feature can be extracted from each of the sets of approximation and detail coefficient arrays based on a ratio of each of the levels of the approximation coefficient arrays over the detail coefficient arrays. For example, 16 statistical features can be extracted.
  • Each of the extracted statistical features can be input into the deep neural network example 800 shown in FIG. 8.
  • the deep neural network can be applied to the plurality of statistical features to compute a vector of probabilities.
  • the vector of probabilities can include values representing a probability of an abdominal event, a probability of a detrusor overactivity event, a probability of no event, and a probability of a voiding bladder contraction event.
  • the deep neural network can include an input layer, an output layer and fully connected first and second layers.
  • the input and output layers can also be fully connected to the first and second layers, respectively.
  • Each of the first and second layers can include, for instance, 100 neurons.
  • Each of the neurons of the first and second layers can be activated by a common function such as a Rectified Linear Unit (ReLu) function.
  • ReLu Rectified Linear Unit
  • the second layer can also undergo a soft max activation to generate the vector of probabilities.
  • the deep neural network is pre-trained.
  • the deep neural network can be trained using 12,000 0.8 second vesical pressure data samples manually annotated by a clinical urologist.
  • the bladder event can then be labeled based on the vector of probabilities.
  • the label can be one of the abdominal event, the detrusor overactivity event, no event, and the voiding bladder contraction event based on the most likely probability.
  • the output vector of probabilities can be further filtered by filtering the vector of probabilities through a majority-vote filter before labeling the event.
  • the vector of probabilities can be determined to an initial event label and the filter can determine a final event label by storing the initial event label in a buffer of the most recent 12 initial event labels and then determining the mode as the final event label.
  • a labeled event can then be output to a neuromodulation device (alternatively, the labeled event may be output to the neuromodulation device if it is classified as a voiding bladder contraction event) or to memory storage and/or a display device for research purposes, and the entire classification method restarted for the next vesical pressure data sample.
  • Another aspect of the present disclosure can include methods 900 (FIG. 9) for classifying a bladder event using a machine learning framework (e.g., a deep neural network) and methods 1000 (FIG. 10) for automatically controlling bladder function (e.g., for controlling/stopping incontinence) through neuromodulation using the classified event.
  • a machine learning framework e.g., a deep neural network
  • methods 1000 for automatically controlling bladder function (e.g., for controlling/stopping incontinence) through neuromodulation using the classified event.
  • These methods 900, 1000 can significantly improve the realtime closed loop control of incontinence.
  • the methods can utilize a system (shown in FIGS. 1 -8) to measure a bladder pressure of a bladder at a time, determine an event classification for the bladder pressure, and to modulate at least one nerve to control at least one function of the bladder.
  • At least one step of the methods 900, 1000 can be executed by at least one component that includes at least a processor.
  • the sensing device e.g., sensing device 110
  • the signal processing device e.g., signal processing device 112
  • neuromodulation device 1 14 can configure and apply stimulation to the patient to treat incontinence based on an input related to the classification.
  • the methods 900, 1000 are shown and described as being executed serially; however, it is to be understood and appreciated that the present disclosure is not limited by the illustrated order as some steps could occur in different orders and/or concurrently with other steps shown and described herein. Moreover, not all illustrated aspects may be required to implement the methods 900, 1000, nor are the methods 900, 1000 necessarily limited to the illustrated aspects.
  • FIG. 9 illustrate is a method 900 for classifying a bladder event using vesical pressure data samples and a deep neural network.
  • the method can, for example, classify a new bladder event every about 1 second (or less).
  • a vesical pressure data sample taken at a time can be received by a signal processing device from a sensing device (can be transmitted as a signal indicating the vesical pressure at the given time).
  • the vesical pressure data sample can be, for example 0.8 seconds worth of vesical pressure data measured by the sensing device at a rate of 10 Hz.
  • the sensing device can be adapted to be positioned within a patient’s bladder (e.g., free floating and/or implanted into a bladder wall) to sense vesical pressure. It should be noted that only a single sensing device can be used for the sensing of vesical pressure, rather than one inner (in bladder) and one outer (strapped to back or abdomen) device.
  • the vesical pressure can be sampled over a time period to provide vesical pressure data samples for a window of time.
  • the sensing device can include a pressure sensor adapted to detect the vesical pressure within the patient’s bladder, a wireless transceiver adapted to transmit a signal indicating the vesical pressure within the patient’s bladder; and a battery adapted to provide power to the pressure sensor. At least one of the pressure sensor, the wireless transceiver, and the battery can be encapsulated in a biocompatible housing.
  • the signal processing device can include a wireless transceiver adapted to receive the signal indicating the vesical pressure from the wireless transceiver of the sensing device.
  • the vesical pressure data sample can be input by the signal processing device into a buffer that can include a plurality of vesical pressure data samples collected over a previous time period.
  • the buffer can be a fist in first out buffer having an upper data limit (e.g., can only store X amount of samples before the oldest sample is removed).
  • the first in first out buffer can have an upper data limit of 512 samples.
  • the upper data limit can be determined based on a balance of better time-frequency localization of the input with the quality of outputs from the wavelet transform (see step 906).
  • the upper data limit Once the upper data limit is reached (e.g., the system is initialized) the newest sample can be stored in the buffer and the oldest can be removed for each time.
  • a multi-level discrete wavelet transform can be applied by the signal processing device to the plurality of vesical pressure data samples from the buffer (including the received vesical pressure data sample at the time) to output approximation coefficient arrays and detail coefficient arrays.
  • the multi-level discrete wavelet transform can not only extract frequency information that is embedded in an input signal (e.g., the plurality of vesical pressure data samples), similar to a fast Fourier transform (FFT) and short-time Fourier Transform (STFT) but can also localize the frequency content in time with high resolution.
  • FFT fast Fourier transform
  • STFT short-time Fourier Transform
  • the approximation coefficient arrays can represent the low-frequency content of the plurality of vesical pressure data samples and the detail coefficient arrays can represent the high-frequency content of the plurality of vesical pressure data samples.
  • Each level of the multi-level discrete wavelet transform can output an approximation coefficient array and a detail coefficient array. For example, if the multi-level discrete wavelet transform is five levels, then the five-level discrete wavelet transform can output five approximation coefficient arrays and five detail coefficient arrays.
  • the multi-level discrete wavelet transform can low pass filter and high- pass filter an input and down-samples each of the results by a factor of two for each level.
  • the input for the first level can be the plurality of vesical pressure data samples.
  • the inputs for the rest of the levels can be the approximation coefficient array from the previous level.
  • the down sampling can be by a factor of two to reduce storage usage such that each set of coefficient arrays is half the size of the previous set (e.g., 256, 128, 64, 32, 16).
  • the multilevel discrete wavelet transform can be a lifting transform that can allow the approximation and detail coefficient arrays to be computed in-place for increased storage efficiency and decreased computational complexity.
  • the lifting transform can be a Daubechies 4 discrete wavelet transform.
  • the Daubechies 4 discrete wavelet transform can first split the input array (e.g., the plurality of vesical pressure data samples or the previous level’s approximation coefficient array) into even and odd-indexed components, then perform a series of update and prediction operations that can represent high- and low-pass filtering, respectively.
  • a plurality of statistical features can be extracted by the signal processing device via a feature extraction application.
  • the approximation coefficient arrays, the detail coefficient arrays, and a portion of the vesical pressure data samples can be input into the feature extraction application.
  • a portion of the most recent (chronologically) vesical pressure data samples such as the most recent 5, 8, 10, or more samples, can be input.
  • Extracting the statistical features can include mapping a number of most recently obtained vesical pressure data samples, for example 8, from the plurality of vesical pressure data samples, the approximation coefficients, and the detail coefficients into individual statistical inputs for the neural network.
  • the number of most recently obtained vesical pressure data samples can be compressed into a single value to be a first statistical feature of the neural network. For example, by determining the mean. Additionally, two statistical features can be extracted from each of the levels of the detail and approximation coefficient arrays, by compressing (taking the mean) the larger arrays (if the event is considered from the last 8 samples, then the last four values from the first level arrays, and the last two values from the second level arrays) and not compressing the smaller arrays but simply using the last value of each array of the third, fourth, and fifth level as features. Another statistical feature can be extracted from each of the sets of approximation and detail coefficient arrays based on a ratio of each of the levels of the approximation coefficient arrays over the detail coefficient arrays. For example, 16 statistical features can be extracted.
  • Each of the extracted statistical features can be input into the deep neural network.
  • the deep neural network can be applied to the plurality of statistical features to compute a vector of probabilities.
  • the vector of probabilities can include values representing a probability of an abdominal event, a probability of a detrusor overactivity event, a probability of no event, and a probability of a voiding bladder contraction event.
  • the deep neural network can include an input layer, an output layer and fully connected first and second layers.
  • the input and output layers can also be fully connected to the first and second layers, respectively.
  • Each of the first and second layers can include, for instance, 100 neurons.
  • Each of the neurons of the first and second layers can be activated by a common function such as a Rectified Linear Unit (ReLu) function.
  • ReLu Rectified Linear Unit
  • the second layer can also undergo a soft max activation to generate the vector of probabilities.
  • the deep neural network can be pre-trained. For example, the deep neural network can be trained using 12,000 0.8 second vesical pressure data samples manually annotated by a clinical urologist. However. The deep neural network can be re-trained with data from the patient and validation.
  • the event can be labeled based on the vector of probabilities.
  • the label can be one of the abdominal event, the detrusor overactivity event, no event, and the voiding bladder contraction event based on the most likely probability.
  • the output vector of probabilities can be further filtered by filtering the vector of probabilities through a majority-vote filter before labeling the event.
  • the vector of probabilities can be determined to an initial event label and the filter can determine a final event label by storing the initial event label in a buffer of the most recent 12 initial event labels and then determining the mode as the final event label.
  • a labeled even can then be output to a neuromodulation device or to memory storage and/or a display device for research purposes, and the entire classification method restarted for the next vesical pressure data sample.
  • Method 1000 of FIG. 10 gives an example of the classifier of the signal processing device outputting classified labels (or a portion of the classified labels) as an input to a neuromodulation device for closed loop control of bladder functions.
  • vesical pressure data samples can be received by a signal processing device from a sensing device as described in greater detail above.
  • the vesical pressure data samples can be classified as a bladder event by the signal processing device, as described in greater detail above and then at 1006, the output of the bladder event classification can be output as one of the abdominal event, the detrusor overactivity event, no event, and the voiding bladder contraction event.
  • the output bladder event classification can be sent to a neuromodulation device.
  • the stimulation can be configured by the neuromodulation device in response to the output.
  • the configuring can include choosing one or more parameters of an electrical stimulation waveform (e.g., amplitude, frequency, power, intensity, pulse or continuous, timing, etc.) and/or where to apply the stimulation (e.g., which electrode(s) to use).
  • the stimulation can be applied to at least one nerve of the patient, such as a genital nerve to control one or more functions of the bladder and/or urinary system to control incontinence.
  • the stimulation can trigger voiding if the bladder is full or indicate do not void if the bladder is not full or the situation does not allow for voiding.
  • ABEC Autonomous Bladder Event Classifier
  • ABEC a deep neural network-based classifier which automatically labels a bladder event from vesical pressure data in real time.
  • ABEC applies the five-level discrete wavelet transform (DWT) and computes 16 statistical features from the resulting sets of approximation and detail coefficients. These features serve as the input to a pretrained, fully-connected deep neural network which outputs an initial event label. This label is finally passed through a majority-vote filter to obtain the final event label.
  • ABEC runs on a GPU hardware system which enables acceleration of certain computational blocks to speed up inference time. After collecting 0.8 seconds worth of vesical pressure data, ABEC has an average inference time of 0.11 seconds, resulting in a new bladder event prediction every -0.91 seconds.
  • FIG. 12 A high-level overview of ABEC is shown in FIG. 12
  • N vesical pressure samples must be received and stored into a first-in first-out (FIFO) PVES buffer, where N denotes the minimum number of vesical pressure samples required to perform the five-level DWT.
  • N 512. Therefore, at a data sample rate of 10 Hz, ABEC initialization takes -51.2 seconds.
  • ABEC makes a new prediction for every 0.8 seconds (or 8 samples) of newly collected PVES data. This allows more statistical features to be computed from a longer interval of data while still maintaining a high level of time precision during real-time inference.
  • the DWT extracts the frequency information that is embedded in an input signal.
  • the DWT can also localize this frequency content in time with high resolution.
  • the DWT outputs two sets of “coefficients:” approximation coefficients representing the low-frequency content in the input signal, and detail coefficients representing the high frequency content in the input signal. These two outputs are obtained by low-pass filtering and high-pass filtering the input, respectively, followed by down-sampling each of the results by a factor of two to reduce storage usage (in accordance with the Nyquist theorem).
  • the approximation coefficients from this first “level” of the DWT serves as the input for the next level of the DWT, which outputs the second-level approximation and detail coefficients.
  • This process can be repeatedly applied to obtain more sets of approximation and detail coefficients, the size of each set of coefficients being half the size of the set from the previous level.
  • the full structure of the DWT is shown in FIG. 13, element A.
  • a useful version of the DWT is the lifting transform.
  • the lifting scheme allows the approximation and detail coefficients to be computed in-place, resulting in increased storage efficiency and decreased computational complexity.
  • the lifting scheme For the Daubechies 4 DWT, which is used by ABEC, the lifting scheme first splits the input array into its even and odd-indexed components. The lifting scheme then performs a series of update and predicts operations that represent high-pass and low-pass filtering, respectively. Just like the regular DWT, the lifting DWT can be repeatedly applied to obtain more sets of coefficients.
  • the structure of the Daubechies 4 lifting DWT is shown in FIG. 13, Element B.
  • Each set of approximation and detail coefficients is stored in an array and contains half the number of elements as the set from the preceding stage because of down-sampling.
  • the table below shows the number of elements in each of the coefficient vectors, where cAi and cDi represent the approximation coefficients and detail coefficients from the i th DWT level, respectively, as shown in Table 1 :
  • ABEC feature extraction stage maps values from the PVES buffer and the DWT outputs into individual statistical inputs for the neural network. While all 512 PVES samples (51 .2 seconds of data) are used to perform the DWT, the neural network only treats the last 8 PVES samples (or 0.8 seconds) in the FIFO buffer as the event that is to be classified. ABEC thus compresses the last 8 PVES values from the buffer into a single value by computing the mean of these 8 values, which is the first feature for the neural network.
  • ABEC After computing the approximation feature and detail feature from each level of the DWT, resulting in 10 features, ABEC also computes the ratio between these features from each level: Therefore, including the mean PVES cD CD2 value, the feature extraction stage computes a total of 16 features corresponding to the last 8 samples of PVES data, which is the event that is to be classified. This full process is shown in FIG. 14.
  • the neural network contains 2 fully-connected layers with 100 neurons that are each activated by the ReLU function.
  • the second hidden layer is also passed through the SoftMax function to generate the output layer of probabilities.
  • the full neural network architecture is shown visually in FIG. 15.
  • the deep neural network classifier was trained on a PC in MATLAB using 12,000 0.8- second PVES events. Each event was manually annotated with the assistance of a clinical urologist. The data was extracted from 1 16 anonymized UDS tracings from patients with either overactive bladder or neurogenic urinary incontinence. Cross-validation on the dataset showed an accuracy of 78%.
  • the output of the neural network must be further processed.
  • the output of the neural network is a vector of 4 values, with each value representing the probability that the current 0.8-second PVES sample belongs to ABD, DO, NONE or VOID, respectively. Therefore, the index representing the maximum probability, or the arg(max) of the output, represents the initial event label.
  • This label passes through a majority-vote filter which helps decide the final event label.
  • the filter simply stores the initial label into a buffer of 12 labels.
  • the output of the filter is the mode of this buffer, and the oldest value in the buffer is then discarded.
  • the full structure of the majority-vote filter is shown in FIG. 16. [0081 ] Test Setup
  • the test setup is shown in FIG. 17.
  • ABEC has been implemented on an NVIDIA Jetson Nano development system to test its efficacy.
  • the Jetson Nano contains 128 NVIDIA CUDA cores that are capable of parallel computing, all on a low-power edge computing system suitable for real-time inference.
  • Pre-recorded PVES data stored on a .csv file was read by the Jetson Nano running ABEC.
  • a 0.1 second delay was added after streaming a PVES value.
  • the trained neural network parameters e.g., weights, biases, and feature scaler values
  • the hardware performed ABEC, and the execution time and accuracy was measured.
  • Preliminary testing showed an execution time of 0.91 seconds and accuracies ranging from 75-80% depending on the UDS trace.
  • UDS tracings sampled at 10 Hz were obtained from 34 human subjects with overactive bladder or neurogenic urinary incontinence through urodynamic testing conducted at the Louis Stokes Cleveland Department of Veteran Affairs Medical Center and the Cleveland Clinic. Anonymized data were harvested from previous studies and were IRB exempt.
  • Each tracing included P ES, PABD, and PDET, volume and flow information and was analyzed by a blinded urologist to annotate abdominal events (ABD), voiding contractions (VOID), and detrusor overactivity (DO), see FIG. 18.
  • noisy PVES segments typically found towards the start or end of a tracing and caused by catheter motion, were discarded from each tracing.
  • DO events and voiding contractions were usually represented as low- frequency rises and falls, while abdominal artifacts contained high-frequency spikes in pressure. While the fast Fourier transform (FFT) and short-time Fourier transform (STFT) are typically used to determine the frequency content of a time-varying signal, they suffer from the trade-off between time and frequency resolution.
  • the discrete wavelet transform (DWT) was therefore used for improved temporal localization of frequency content in non-stationary signals. Previously, the Daubechies 4 wavelet was shown to be useful for extracting time-frequency localized information from PVES data.
  • Each PVES tracing was split into nonoverlapping segments of 0.8 seconds in length (8 PVES samples), and each PVES segment was then labeled with one of four possible class labels, depending on which event was present in the segment: ‘ABD’, ‘VOID,’ ‘DO’ or ‘NONE’ (if no event was present).
  • Feature selection was applied to reduce the training time and overfitting of the classifiers.
  • the most relevant features from the feature matrix were identified using the relief F algorithm, which assigns weights to features using a nearest-neighbors approach.
  • Relief F rewards features with large differences for observations of different classes and penalizes features with large differences for observations of the same class.
  • This technique identified the m highest-ranked features, where m denotes the number of features chosen for each classifier architecture.
  • the initial hyperparameters for the ANN and SVM-RBF classifiers were chosen based on prior work in physiological signal classification using DWT-based features.
  • the k value for the KNN classifier was chosen based on initial testing with the 7,861 x 55 feature matrix, which showed that using a single neighbor resulted in strong classification accuracy.
  • Each classifier was trained and evaluated using 5-fold cross validation.
  • FIG. 21 shows a flow of bladder event classification framework with underlying steps from PVES input to event class output.
  • ReliefF was applied to select the 12 most relevant features for the KNN and SVM-RBF classifiers; testing showed that there was no increase in classification accuracy when more than 12 features were used for these two classifiers.
  • setting y 0.94 improved performance of the SVM-RBF classifier.
  • adding a second hidden layer increasing the number of neurons in each hidden layer to 100, and including all 55 features substantially increased classification accuracy.
  • the final architectures for all three classifiers that were tested are shown in Table 1 .
  • FIG. 22 shows a summary of prediction results using five-fold cross validation of all three classifiers.
  • the confusion matrices (top row) show the true positives, false positives, true negatives, and false negatives per class for each classifier.
  • the ROC curves (bottom row) indicate the trade-off between sensitivity and specificity per class for each classifier.

Landscapes

  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Surgery (AREA)
  • Public Health (AREA)
  • Pathology (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Medical Informatics (AREA)
  • Molecular Biology (AREA)
  • Physiology (AREA)
  • Animal Behavior & Ethology (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Veterinary Medicine (AREA)
  • Urology & Nephrology (AREA)
  • Evolutionary Computation (AREA)
  • Fuzzy Systems (AREA)
  • Mathematical Physics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Psychiatry (AREA)
  • Signal Processing (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)

Abstract

Des événements dans la vessie peuvent être classés et délivrés à un système de neuromodulation pour commander au moins une fonction de la vessie. Un dispositif de détection dans une vessie peut détecter des événements de pression vésicale, qui peuvent être classés par un dispositif de traitement de signal. Un événement de l'échantillon de données de pression vésicale peut être entré dans un tampon avec des échantillons antérieurs, une transformée en ondelettes discrète à niveaux multiples peut être appliquée aux échantillons pour délivrer en sortie une approximation et des réseaux de coefficients de détail. Des caractéristiques statistiques peuvent être extraites à l'aide des réseaux de coefficients d'approximation et de détail et d'une partie des échantillons. Un réseau neuronal peut calculer un vecteur de probabilités à partir des caractéristiques statistiques et l'événement peut être marqué sur la base du vecteur de probabilités. L'événement peut être un événement abdominal, un événement d'hyperactivité du détrusor, aucun événement ou un événement contractile de la vessie lors de la miction, qui peut être entré dans le système de neuromodulation.
PCT/US2023/082067 2022-12-01 2023-12-01 Réseau neuronal profond pour la neuromodulation en boucle fermée de la vessie WO2024119066A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US202263429180P 2022-12-01 2022-12-01
US63/429,180 2022-12-01

Publications (1)

Publication Number Publication Date
WO2024119066A1 true WO2024119066A1 (fr) 2024-06-06

Family

ID=89619297

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2023/082067 WO2024119066A1 (fr) 2022-12-01 2023-12-01 Réseau neuronal profond pour la neuromodulation en boucle fermée de la vessie

Country Status (1)

Country Link
WO (1) WO2024119066A1 (fr)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190223775A1 (en) * 2015-04-29 2019-07-25 The Cleveland Clinic Foundation Bladder event detection for diagnosis of urinary incontinence or treatment of lower urinary tract dysfunction
CN111053549A (zh) * 2019-12-23 2020-04-24 威海北洋电气集团股份有限公司 一种智能生物信号异常检测方法及系统
US20210374500A1 (en) * 2020-05-28 2021-12-02 Hitachi, Ltd. Method for reproducibility of deep learning classifiers using ensembles
US20220280089A1 (en) * 2021-03-08 2022-09-08 Regents Of The University Of Minnesota Bladder or bowel dysfunction assessment systems and methods
CN115064257A (zh) * 2022-04-15 2022-09-16 武汉闻道复兴智能科技有限责任公司 逼尿肌过度活动诊断模型的建立方法、装置、介质和设备
WO2022232992A1 (fr) * 2021-05-06 2022-11-10 Huawei Technologies Co., Ltd. Système et procédé permettant de déterminer un risque d'accident vasculaire cérébral pour une personne

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190223775A1 (en) * 2015-04-29 2019-07-25 The Cleveland Clinic Foundation Bladder event detection for diagnosis of urinary incontinence or treatment of lower urinary tract dysfunction
CN111053549A (zh) * 2019-12-23 2020-04-24 威海北洋电气集团股份有限公司 一种智能生物信号异常检测方法及系统
US20210374500A1 (en) * 2020-05-28 2021-12-02 Hitachi, Ltd. Method for reproducibility of deep learning classifiers using ensembles
US20220280089A1 (en) * 2021-03-08 2022-09-08 Regents Of The University Of Minnesota Bladder or bowel dysfunction assessment systems and methods
WO2022232992A1 (fr) * 2021-05-06 2022-11-10 Huawei Technologies Co., Ltd. Système et procédé permettant de déterminer un risque d'accident vasculaire cérébral pour une personne
CN115064257A (zh) * 2022-04-15 2022-09-16 武汉闻道复兴智能科技有限责任公司 逼尿肌过度活动诊断模型的建立方法、装置、介质和设备

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
KARAM ROBERT: "EVENT DETECTION ALGORITHM FOR SINGLE-SENSOR BLADDER PRESSURE DATA", 1 January 2016 (2016-01-01), XP055980479, Retrieved from the Internet <URL:https://etd.ohiolink.edu/apexprod/rws_etd/send_file/send?accession=case1441297369&disposition=inline> [retrieved on 20221111] *
MAJERUS STEVE J A ET AL: "Feasibility of Real-Time Conditional Sacral Neuromodulation Using Wireless Bladder Pressure Sensor", IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING, IEEE, USA, vol. 29, 4 October 2021 (2021-10-04), pages 2067 - 2075, XP011882890, ISSN: 1534-4320, [retrieved on 20211013], DOI: 10.1109/TNSRE.2021.3117518 *

Similar Documents

Publication Publication Date Title
US11419533B2 (en) Bladder event detection for diagnosis of urinary incontinence or treatment of lower urinary tract dysfunction
Chowdhury et al. Time-frequency analysis, denoising, compression, segmentation, and classification of PCG signals
AU2020349425A1 (en) Systems and methods for seizure prediction and detection
US20200060604A1 (en) Systems and methods of automatic cough identification
Cherian et al. Theoretical and methodological analysis of EEG based seizure detection and prediction: An exhaustive review
Azimi et al. Machine learning-based automatic detection of central sleep apnea events from a pressure sensitive mat
EP3100138B1 (fr) Système d&#39;interface cerveau-ordinateur
Milani et al. Abnormal heart sound classification using phonocardiography signals
Abbasi et al. Selecting Statistical Characteristics of Brain Signals to Detect Epileptic Seizures using Discrete Wavelet Transform and Perceptron Neural Network.
KR20190111570A (ko) 확률모델 및 기계학습과 함께 특징추출을 이용한 뇌전도 신호의 멀티 주파수 대역 계수 기반 간질 발작파 검출 시스템
Zhang et al. Competition convolutional neural network for sleep stage classification
Sharathappriyaa et al. Auto-encoder based automated epilepsy diagnosis
Taşkıran et al. A deep learning based decision support system for diagnosis of Temporomandibular joint disorder
Qin et al. Patient-specific seizure prediction with scalp EEG using convolutional neural network and extreme learning machine
Bhosale et al. Classification of EMG signals using wavelet transform and hybrid classifier for parkinson’s disease detection
WO2024119066A1 (fr) Réseau neuronal profond pour la neuromodulation en boucle fermée de la vessie
Liu et al. Automatic focal EEG identification based on deep reinforcement learning
Sutandi et al. Detection of traumatic brain injury using single channel electroencephalogram in mice
Majerus et al. Real-Time Wavelet Processing and Classifier Algorithms Enabling Single-Channel Diagnosis of Lower Urinary Tract Dysfunction
Tufan et al. Non-invasive diagnosis of stress urinary incontinence sub types using wavelet analysis, shannon entropy and principal component analysis
Abbaraju et al. Machine learning for automated bladder event classification from single-channel vesical pressure recordings
Hussein Scalp and intracranial EEG quantitative analysis: robust detection and prediction of epileptic seizures
Enabling et al. OH, USA
Navea et al. Classification of wavelet-denoised musical tone stimulated EEG signals using artificial neural networks
Ortiz et al. Learning from few subjects with large amounts of voice monitoring data