Multi-target parameter adaptive non-invasive tibial nerve stimulation method and system

Abstract

This invention relates to a multi-target parameter adaptive non-invasive tibial nerve stimulation method and system, aiming to solve the problem that existing neuromodulation devices cannot provide on-demand treatment based on dynamic changes in patient symptoms. It utilizes several targeted electrode units configured on the medial malleolus and sole of the foot to apply electrical stimulation synchronously or independently, and collects induced plantar nerve signals. Within a preset time window, the peak amplitude and latency of the F-wave, meeting stability conditions, are extracted as real-time feedback signals. Based on this, an individualized target interval is established, and the dynamic deviation of the current F-wave peak amplitude relative to this interval is calculated. Using the mean square error of the dynamic deviation as a loss function, a gradient descent method is used to iteratively adjust parameters such as stimulation intensity, frequency, pulse width, and the ratio of dual-target stimulation intensity until the F-wave accumulator converges to the target interval. This adaptively adapts to dynamic changes in bladder function, achieving closed-loop adaptive stimulation of multiple targets on the tibial nerve.

Description

Technical Field

[0001] This application relates to the field of neuromodulation technology, and in particular to a multi-target parameter adaptive non-invasive tibial nerve stimulation method and system. Background Technology

[0002] Tibial nerve stimulation, as an innovative treatment for open-foot-and-mouth disease (OAB), has enriched the treatment options for OAB to some extent. However, it still faces challenges such as reliance on physician experience for manual parameter adjustment, patient travel to and from the hospital during parameter adjustment, the need to visually observe the patient's foot during treatment to monitor foot movement, and the ineffectiveness of single-target stimulation in some individuals, leading to inconvenience, imprecise efficacy, and low patient coverage.

[0003] The treatment parameters of existing neuromodulation devices are fixed once set, while symptoms such as overactive bladder fluctuate dynamically. This means that patients either suffer unnecessary stimulation when symptoms are mild or insufficient stimulation when symptoms worsen, making it impossible to achieve "on-demand treatment" that matches symptoms in real time.

[0004] Chinese patent CN115869539A discloses a method, device, and system for visualizing current parameters of percutaneous tibial nerve stimulation. The method includes: acquiring impedance characteristic parameters of the patient's skin; calculating the attenuation ratio corresponding to the impedance characteristic parameters using an artificial intelligence algorithm; determining the stimulation current parameters reaching the user's tibial nerve based on the attenuation ratio, and visually displaying the stimulation current parameters. This invention can acquire the impedance characteristics of the patient's skin, then determine the current attenuation ratio in the human body due to the impedance characteristics based on an artificial intelligence algorithm, finally calculate the actual stimulation current reaching the user's tibial nerve using the attenuation ratio, and visualize the stimulation current on a control terminal to improve the accuracy of the display and provide doctors with treatment reference.

[0005] While the aforementioned approach improves the accuracy of stimulation current parameter settings by combining impedance characteristic parameters with artificial intelligence algorithms and provides doctors with a visual treatment reference, it still only optimizes and corrects the stimulation parameters based on skin impedance in the initial stage of treatment. During subsequent continuous treatment, the device continues to stimulate according to preset constant parameters, failing to dynamically adjust based on real-time changes in the patient's bladder fullness, symptom frequency, and intensity. Furthermore, existing parameter adjustments require doctors to visually monitor the patient's feet to observe foot movements, and single-target stimulation may be ineffective for some individuals, leading to inconvenience, inaccurate efficacy, and low patient coverage. Specifically, when patients are frequently active during the day or experience increased water intake leading to symptom exacerbation, the stimulation intensity may be insufficient to effectively inhibit abnormal nerve excitation; conversely, during nighttime rest or symptom relief periods, patients may experience unnecessary overstimulation, reducing long-term efficacy.

[0006] Therefore, we propose a multi-target parameter adaptive non-invasive tibial nerve stimulation method and system. Summary of the Invention

[0007] The main purpose of this application is to provide a multi-target parameter adaptive non-invasive tibial nerve stimulation method and system, which aims to solve the problems of insufficient coverage of single-target stimulation in the prior art and the lack of individualized adaptation of fixed parameter stimulation, which makes it impossible to establish an effective real-time physiological feedback mechanism.

[0008] To achieve the above objectives, this application provides a multi-target parameter adaptive non-invasive tibial nerve stimulation method, comprising the following steps: Step S1: Apply electrical stimulation synchronously or independently to several targeted electrode units coupled to the user's medial ankle and sole. Step S2: Collect plantar nerve signals induced by the electrical stimulation, extract F-wave features within a preset time window as real-time feedback signals. The extraction of F-waves must meet preset stability conditions, including: the occurrence rate of F-waves is stable within a preset range, and / or the amplitude change rate of F-waves is lower than a preset threshold. Step S3: Establish an individualized target interval based on the F-wave characteristics and latency, and calculate its dynamic deviation relative to the individualized target interval based on the F-wave characteristics extracted at the current time. Step S4: Using the mean square error of the dynamic deviation as the loss function, the electrical stimulation parameters are iteratively adjusted using the gradient descent method until the F-Baud levy converges to the individualized target interval.

[0009] Preferably, the electrical stimulation parameters include stimulation intensity, stimulation frequency, pulse width, and the ratio of stimulation intensity at the medial malleolus and the sole of the foot as dual target points; The medial malleolus target point is located in the surface projection area of ​​the tibial nerve behind the medial malleolus, while the plantar target point electrode is located in the surface projection area of ​​the nerve branch on the medial or lateral side of the foot.

[0010] Preferably, the stability condition determination in step S2 is based on a single measurement and / or on the trend characteristics of neural excitability under long-term monitoring, specifically; The single measurement includes: collecting the signal sequence as... n is the number of sampling points, and is based on the formula Japanese style Quantify the F-wave characteristics to obtain the peak amplitude A of the current acquisition. f With incubation period L f As a key feedback feature; in, For the preset time window for the appearance of the F wave, t onset t is the starting point of the F wave. stimThe timing of stimulus application is determined by adjusting the stimulus parameter t. stim To make the peak amplitude A of the sampled peak amplitude A f It stabilizes within the preset threshold for F-wave amplitude variation; The long-term monitored neural excitability trend characteristics include: acquiring m simplified neural excitability proxy indicators x daily. i Robust estimation of the m indicators yields the characteristic value X for the day. d Furthermore, the robust estimation employs the median or other statistics that resist outliers; Calculate the characteristic value X for the day d 7-day weighted moving average and trend slope T d When the cumulative monitoring days are less than 7 days, a weighted moving average is calculated based on the actual number of days. After 7 days, a fixed 7-day window is activated. The weight of the weighted moving average decreases as the distance between the measurement date and the current date increases. The trend slope T... d By analyzing the characteristic value X of the actual number of days d Weighted moving average smoothing was performed, and least squares linear fitting was performed based on the smoothed data and the time series corresponding to the actual number of monitoring days to quantify the direction and rate of change in neural excitability. Based on the trend slope T d The sign and magnitude of the value are adjusted to set the preset threshold for F-wave amplitude change. When T... d >0 indicates an upward trend in excitability characteristics, and the upper and / or lower limits of the amplitude change threshold are adjusted upward; when T d <0 indicates a declining trend in excitability characteristics; therefore, the upper and / or lower limits of the amplitude change threshold are lowered, and the adjustment range and |T are adjusted accordingly. d | Positively correlated.

[0011] Preferably, the calculation of dynamic deviation in step S3 uses an asymmetric dead-zone control function: ; in, The peak amplitude of the F-wave extracted at the current time t. The target range for the peak amplitude of the F wave is preset based on the individualized physiological characteristics of the user, i.e., the individualized target range. The lower limit of the interval, The upper limit of the interval, The trend term representing the change in the peak amplitude of the F-wave is obtained by performing time-difference or linear fitting on the peak values ​​of the F-wave extracted from multiple consecutive stimuli. It is used to quantify the dynamic direction and rate of change in spinal cord excitability. This is the low-side feedforward compensation coefficient, used to indicate that the extracted F-wave peak amplitude is lower than the individualized target range. The high-side feedforward compensation coefficient is used to indicate that the peak amplitude of the extracted F-wave is higher than the individualized target range.

[0012] Preferably, the And it is adjusted based on the duration of the F-wave latency, specifically as follows; When the F-wave latency is prolonged, the low-side feedforward compensation coefficient is increased. The value of is used to compensate for the delay in excitatory response caused by the decrease in nerve conduction efficiency; When the F-wave latency shortens, the high-side feedforward compensation coefficient is increased. The absolute value of is used to suppress the state of over-excitation caused by increased nerve conduction velocity.

[0013] Preferably, the gradient descent method used in the iterative adjustment of the electrical stimulation parameters in step S4 is predicted by a machine learning model, and the machine learning model is a gradient boosting decision tree model, which is an additive model composed of K regression trees. ; Where z is the input feature vector, which is determined by the trend slope T d The seven-day weighted moving average, the baseline deviation of the F-wave peak amplitude relative to the individualized target interval, and the rate of change in the symptom log together constitute the total. For the k-th regression tree, The learning rate is used to control the contribution of each tree.

[0014] Preferably, the model is trained on a historical dataset by minimizing a regularized loss function: ; in, Here, l represents the optimal parameters validated for the i-th treatment in history, and l is the mean squared error loss function. This is a regularization term used to control model complexity and prevent overfitting.

[0015] Preferably, during the iterative adjustment of electrical stimulation parameters using the gradient descent method, the gradient boosting decision tree model is incrementally updated based on stochastic gradient descent. That is, after obtaining new training samples (zt, yt), the model parameters are fine-tuned in the following directions. ; Where γ is the fine-tuning learning rate, and its value is at least one-tenth of the initial training learning rate. This represents the gradient of the loss function l with respect to the input feature vector z.

[0016] Preferably, the plantar nerve signal includes at least one of electromyographic signal, nerve electrical signal and motor signal; the features extracted from the F wave include at least one of F wave amplitude, F wave latency, F wave occurrence rate, F wave duration and F wave area.

[0017] To achieve the above objectives, this application provides a system for a multi-target parameter adaptive non-invasive tibial nerve stimulation method, including a stimulation module, a signal acquisition module, and a processing and control module; The stimulation module includes a targeted electrode unit and a stimulation control unit. The targeted electrode unit contains several targeted electrode units, which are disposed at the medial malleolus and the sole of the foot, respectively corresponding to the surface projection area of ​​the tibial nerve behind the medial malleolus and the surface projection area of ​​the nerve branches on the medial or lateral side of the sole of the foot. The stimulation control unit is connected to the targeted electrode unit and is used to apply electrical stimulation to the medial malleolus and the sole of the foot according to the instructions of the processing and control module. The signal acquisition module includes a feature signal acquisition and analysis unit connected to the target electrode unit, used to acquire plantar nerve signals induced by electrical stimulation, and extract the peak amplitude and latency of the F wave as feedback signals within a preset time window; the extraction of the F wave meets stability conditions, which include: the occurrence rate of the F wave is stable within a preset range, and / or the amplitude change rate of the F wave is lower than a preset threshold. The processing and control module includes an adaptive sensing control unit, a data fusion unit, and a machine learning prediction model unit. The adaptive sensing control unit is connected to the stimulation control unit and the feature signal acquisition and analysis unit, respectively. It is used to establish an individualized target interval based on the peak amplitude and latency of the F wave, calculate the dynamic deviation of the peak amplitude of the F wave extracted at the current time relative to the individualized target interval, and use the mean square error of the dynamic deviation as the loss function to adjust the electrical stimulation parameters using the gradient descent method until the peak amplitude of the F wave converges to the individualized target interval.

[0018] The beneficial effects of the technical solution of this invention are as follows: By configuring targeted electrode units that act on the tibial nerve trunk (posterior to the medial malleolus) and its plantar branches (medial / lateral side of the sole), synergistic or independent regulation of dual target points on the medial malleolus and the sole of the foot can be achieved to simultaneously activate the main tibial nerve trunk and its plantar branches. Compared with single-point stimulation, it can more effectively mobilize the descending inhibitory pathway at the spinal cord level (reflected by the F wave) and significantly expand the coverage of neural regulation, enhancing the regulatory effect on bladder function.

[0019] By precisely setting the F-wave extraction time window after the M-wave and before the interference of voluntary muscle contraction, and introducing strict stability criteria (e.g., the F-wave occurrence rate is stable within a preset range, and the amplitude change rate is below a preset threshold), interference from electrical stimulation artifacts, background electromyography, and motor artifacts is effectively eliminated. Combined with the quantification of peak amplitude and latency from a single measurement, and the analysis of long-term monitoring trends, the purity, accuracy, and repeatability of the feedback signal are ensured. This reduces the need for frequent manual adjustments, lowers the burden on users or operators, and stabilizes the F-wave within the individualized target range. This avoids ineffective treatment due to insufficient stimulation or discomfort or even damage caused by excessive stimulation, improving treatment safety and comfort, and contributing to better long-term user compliance.

[0020] By establishing individualized F-wave target intervals and introducing an asymmetric dead zone control function with feedforward compensation to calculate dynamic deviations, the system differentiates between two physiological states: F-waves below the lower limit and above the upper limit. This allows the stimulation strategy to closely align with the user's individual differences and real-time state. By incorporating changes in F-wave latency into the adaptive adjustment of the compensation coefficient, when a prolonged latency indicates decreased neural conduction efficiency, the system actively enhances low-side compensation to promote excitatory recruitment; conversely, when a shortened latency indicates overexcitation, the system actively enhances high-side inhibition to prevent overstimulation, thereby maintaining neural excitability within the optimal physiological range throughout the treatment process.

[0021] By employing a gradient boosting decision tree model to predict gradient direction, the parameter optimization process, which incorporates trend slope, weighted moving average, baseline bias, and symptom log change rate, not only relies on current F-wave feedback but also fully absorbs historical treatment experience and long-term trend information. The model's regularization design effectively prevents overfitting, ensuring generalization ability across different user groups. Combined with the incremental update mechanism of stochastic gradient descent, the system can continuously learn during clinical use, constantly optimizing prediction accuracy as treatment data accumulates.

[0022] In summary, this invention constructs a complete closed-loop control system that organically integrates four major modules: multi-target stimulation, high-fidelity signal acquisition, intelligent feature extraction, and adaptive parameter optimization. This forms an automated closed loop from physiological state perception to treatment parameter adjustment, thereby monitoring minute changes in nerve excitability in real time, actively predicting their development trend, and completing fine-tuning of stimulation parameters without the user's awareness. This ensures that the peak amplitude of the F wave is always maintained within the individualized target range, thus achieving precise control of the tibial nerve and its spinal reflex pathway. Attached Figure Description

[0023] Figure 1 This is a front view of the target electrode unit of the multi-target parameter adaptive non-invasive tibial nerve stimulation method in one embodiment of this application; Figure 2 This is a side view of the target electrode unit of a multi-target parameter adaptive non-invasive tibial nerve stimulation method in one embodiment of this application; Figure 3 This is a schematic diagram of the adaptive algorithm layer of the multi-target parameter adaptive non-invasive tibial nerve stimulation method in one embodiment of this application; Figure 4 This is a schematic diagram of the incremental learning process of the multi-target parameter adaptive non-invasive tibial nerve stimulation method in one embodiment of this application; Figure 5 This is a block diagram of a multi-target parameter adaptive non-invasive tibial nerve stimulation system in one embodiment of this application; Figure 6 This is a structural block diagram of a terminal device in one embodiment of this application. Detailed Implementation

[0024] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0025] Furthermore, descriptions using terms such as "first" and "second" in this application are for descriptive purposes only (e.g., to distinguish identical or similar elements) and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, technical solutions from different embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If a combination of technical solutions is contradictory or impossible to implement, such a combination should be considered nonexistent and not within the scope of protection claimed in this application.

[0026] Tibial nerve stimulation, as a non-invasive neuromodulation technique, has been widely used to treat lower urinary tract dysfunction diseases such as overactive bladder (OAB), urinary incontinence, and neurogenic bladder. Its mechanism of action lies in stimulating the tibial nerve, regulating the excitability of the sacral plexus, and thereby affecting the synergistic function of the bladder detrusor muscle and urethral sphincter, improving urine storage and voiding control.

[0027] Current non-invasive tibial nerve stimulation techniques primarily employ a single-target stimulation approach, placing a single stimulating electrode only in the tibial nerve's surface projection area behind the medial malleolus. However, the tibial nerve branches into the medial and lateral plantar nerves behind the medial malleolus, respectively innervating muscles and skin sensation in different areas of the foot. Single-target stimulation struggles to simultaneously activate multiple nerve branches, resulting in insufficient coverage of neural modulation and an inability to fully mobilize descending inhibitory pathways at the spinal cord level.

[0028] On the other hand, there are significant differences in nerve conduction velocity, nerve excitability threshold, and muscle response characteristics among different patients. Current technologies mostly use stimulation modes with fixed parameters, failing to dynamically adjust according to the individual physiological characteristics of patients, resulting in inconsistent treatment effects. Some patients even experience discomfort or ineffective treatment due to inappropriate parameters.

[0029] Furthermore, existing technologies primarily rely on patients' subjective symptom feedback (such as voiding diaries and urinary urgency scores) as the basis for efficacy assessment during stimulation, and cannot obtain objective physiological indicators of the neuromuscular system in real time during treatment, thus failing to achieve closed-loop adaptive control.

[0030] Furthermore, in order to address the problems of inconvenience, inaccurate efficacy, and low population coverage caused by the reliance on doctors' experience to manually adjust parameters, the need to visually observe the patient's foot during treatment, and the ineffectiveness of single-target stimulation in some populations in existing tibial nerve stimulation techniques, a dual-target electrode that can simultaneously cover the medial malleolus and the sole of the foot can be set up, and the F-wave signal generated by stimulation can be used as objective physiological feedback.

[0031] The F-wave, specifically, is a delayed electromyographic response generated by the recurrent discharge of spinal α-motor neurons after muscle contraction induced by stimulation, and is an important indicator for assessing the excitability of spinal motor neurons. However, the F-wave signal is weak, easily interfered with by M-waves (direct muscle contraction response) and voluntary muscle contractions, and the appearance of the F-wave is random. Existing technologies lack effective methods for determining and extracting the stability of the F-wave, limiting its application in neural stimulation feedback control.

[0032] In summary, existing technologies lack a multi-target parameter adaptive non-invasive tibial nerve stimulation method capable of achieving multi-target synergistic stimulation, individualized parameter adaptation, and real-time F-wave physiological feedback closed-loop control, which includes the following steps: Step S1: Apply electrical stimulation synchronously or independently to several targeted electrode units 10 coupled to the user's medial malleolus and sole; the electrical stimulation parameters include stimulation intensity, stimulation frequency, pulse width, and the ratio of stimulation intensity at the dual target points of the medial malleolus and sole. The medial malleolus target electrode is positioned in the surface projection area of ​​the tibial nerve posterior to the medial malleolus, while the plantar target electrode is positioned in the surface projection area of ​​the medial or lateral branch of the tibial nerve on the sole of the foot. (See [link to relevant documentation]). Figures 1-2 .

[0033] Step S2: Collect the plantar nerve signal induced by the electrical stimulation, extract the peak amplitude and latency of the F wave within a preset time window as a real-time feedback signal, and the extraction of the F wave must meet the preset stability conditions. Specifically, the stability conditions include: the occurrence rate of the F wave is stable within a preset range, and / or the amplitude change rate of the F wave is lower than a preset threshold; the plantar nerve signal includes at least one of electromyographic signal, nerve electrical signal and motor signal; the features extracted from the F wave include at least one of F wave amplitude, F wave latency, F wave occurrence rate, F wave duration and F wave area.

[0034] In one embodiment, the stability condition determination in step S2 is based on a single measurement and / or on the neural excitability trend characteristics set by long-term monitoring.

[0035] Specifically, the single measurement refers to the system acquiring the F-wave signal induced by electrical stimulation in real time during a single treatment, extracting features such as F-wave amplitude and latency, and comparing them with the preset target feature range; if the current F-wave feature deviates from the target range, the algorithm automatically fine-tunes the stimulation parameters, including stimulation intensity, frequency, pulse width, and the ratio of dual-target stimulation intensity, until the F-wave feature stabilizes within the target range, ensuring that the treatment achieves the best neuromodulation effect; The aforementioned setting of neural excitability trend characteristics based on long-term monitoring refers to the systematic comprehensive analysis of the final F-wave characteristics recorded at the end of each treatment, historical efficacy data (including subjective feedback such as the patient's voiding diary synchronized through the APP, the number of urge incontinence episodes, and severity self-assessment), and the integration of the direction and amplitude of the long-term neural excitability trend line (calculated based on the time-domain characteristics of daily plantar nerve signal monitoring, such as root mean square value, average amplitude, and zero-crossing rate). The algorithm automatically learns the pattern of the patient's neural excitability fluctuating with symptoms through machine learning models (such as reinforcement learning or Bayesian optimization), and predictively adjusts the initial parameters or stimulation mode of the next treatment, so that the treatment plan can dynamically adapt to the long-term changes in the patient's symptoms, forming an intelligent closed loop of "treatment-feedback-learning-optimization".

[0036] See Figure 3 The system's decision-making is based on three layers of data input: long-term physiological trends (extracting F-wave / electromyography trend features), multi-source data fusion (such as subjective symptom reports), and intelligent decision-making and prediction (such as personalized initial parameters).

[0037] Furthermore, after each treatment, the actual efficacy data is used to feed back and update the model, enabling the system to continuously evolve and become more intelligent with use. Thus, during non-treatment periods or treatment intervals, the system periodically collects plantar nerve signals or nerve electrical signals in a low sampling rate and low power consumption mode, extracting noise-robust temporal features, including root mean square value, average amplitude, zero-crossing rate, wavelength, etc. On the other hand, during the device initialization phase or symptom stabilization period, the system establishes the patient's personal physiological baseline, calculates the average value and normal fluctuation range of each feature (such as mean ± 2 standard deviations), performs moving average processing on the monitored feature values ​​daily to generate a smooth "neural excitability trend line", and monitors the direction (rising or falling), slope changes and whether it exceeds the personal baseline range in real time. When the trend line shows a persistent abnormality (such as continuously rising above the threshold), the system can provide early warning or trigger additional treatment intervention, providing key long-term state input for the inter-session adaptive algorithm.

[0038] In one preferred embodiment, the single measurement includes: collecting a signal sequence S={s1,s2,...,s...} n}, where n is the number of sampling points, and is based on the formula Japanese style Quantify the F-wave characteristics to obtain the peak amplitude A of the current acquisition. f With incubation period L f As a key feedback feature; in, For the preset time window for the appearance of the F wave, t onset t is the starting point of the F wave. stim The timing of stimulus application is determined by adjusting the stimulus parameter t. stim To make the peak amplitude A of the sampled peak amplitude A f It stabilizes within the preset threshold for F-wave amplitude variation; The long-term monitored neural excitability trend characteristics include: acquiring m simplified neural excitability proxy indicators x daily. i Robust estimation of the m indicators yields the characteristic value X for the day. d Furthermore, the robust estimation employs the median or other statistics that resist outliers; Calculate the seven-day weighted moving average and trend slope Td of the characteristic value Xd for the day. When the cumulative monitoring days are less than 7 days, the weighted moving average is calculated based on the actual number of days. ; ; ; in, and These are the time series number and the mean of the eigenvalues, respectively.

[0039] And the trend slope T d The direction and rate of change in neural excitability were quantified, and a fixed 7-day window was activated after 7 days. The weight of the weighted moving average decreased as the distance from the measurement date to the current date increased; the trend slope T... d By analyzing the characteristic value X of the actual number of days d The result is obtained by least-squares linear fitting with the corresponding time series, which is used to quantify the direction and rate of change in neural excitability. Based on the trend slope T d The sign and magnitude of the value are adjusted to set the preset threshold for F-wave amplitude change. When T... d >0 indicates an upward trend in excitability characteristics, and the upper and / or lower limits of the amplitude change threshold are adjusted upward; when T d <0 indicates a declining trend in excitability characteristics; therefore, the upper and / or lower limits of the amplitude change threshold are lowered, and the adjustment range and |T are adjusted accordingly. d | Positively correlated.

[0040] Step S3: Establish an individualized target interval based on the F-wave peak amplitude and latency, and calculate its dynamic deviation relative to the individualized target interval based on the F-wave peak amplitude extracted at the current time; Step S4: Using the mean square error of the dynamic deviation as the loss function, the electrical stimulation parameters are iteratively adjusted using the gradient descent method until the peak amplitude of the F wave converges to the individualized target range.

[0041] In this embodiment, the stimulation target is precisely positioned on the tibial nerve trunk (posterior to the medial malleolus) and its plantar branches (medial / lateral plantar nerves). By monitoring the temporal characteristics of plantar electromyography daily (such as root mean square value and zero-crossing rate), a smooth "neural excitability trend line" is generated. The system can quantify the long-term direction and rate of change in neural excitability (trend slope Td) and dynamically adjust the allowable change threshold of F-wave amplitude accordingly. Simultaneously, the system integrates multi-source data such as patient subjective reports (e.g., voiding diaries, symptom scores) and uses machine learning models to predictively optimize the initial parameters for the next treatment. This allows the treatment plan to dynamically adapt to fluctuations in the patient's condition, thereby avoiding unnecessary overstimulation and reducing patient discomfort while ensuring efficacy.

[0042] On the other hand, the gradient descent method is used to iteratively optimize the F-wave characteristic deviation using the mean square error as the loss function. This ensures that the stimulation parameters can track and maintain neural excitability within the individualized target range in real time. Based on the predictive adjustment mechanism of long-term trends, the system can provide early warning or proactive intervention before the patient's symptoms worsen, which helps to maintain long-term stable control of symptoms and reduce the frequency of sudden events such as urge incontinence.

[0043] In addition, it supports convenient synchronization of patients' subjective feedback through mobile applications and operates in low-power mode during non-treatment periods, which greatly facilitates patients' daily use and data management and enhances patients' initiative in participating in treatment.

[0044] In one embodiment, the gradient descent method used in step S4 to iteratively adjust the electrical stimulation parameters has a gradient direction predicted by a machine learning model, and the machine learning model is a gradient boosting decision tree model, which is an additive model composed of K regression trees. ; Where z is the input feature vector, which is composed of the trend slope Td, the seven-day weighted moving average, the baseline deviation of the F-wave peak amplitude relative to the individualized target interval, and the rate of change in the symptom log. For the k-th regression tree, The learning rate is used to control the contribution of each tree.

[0045] In this embodiment, by using the additive combination of multiple regression trees, the nonlinear mapping relationship between different feature combinations and the optimal stimulation parameter adjustment can be automatically learned. Specifically, when the peak amplitude of the patient's F wave deviates from the individualized target range, the model can predict the direction of stimulation parameter adjustment that will bring the patient's state closer to the target range based on the treatment experience of similar historical cases. This avoids the defects of traditional PID control, such as difficulty in parameter tuning and response lag, and achieves prospective adjustment of electrical stimulation parameters.

[0046] In one preferred embodiment, the model is trained on a historical dataset by minimizing a regularization loss function: ; in, Here, l represents the optimal parameters validated for the i-th treatment in history, and l is the mean squared error loss function. This is a regularization term used to control model complexity and prevent overfitting.

[0047] In this embodiment, by introducing a regularized loss function, the model automatically balances fitting accuracy and model complexity during training, effectively avoiding the decline in generalization ability caused by excessive pursuit of historical data fitting.

[0048] In one more preferred embodiment, during the iterative adjustment of the electrical stimulation parameters using the gradient descent method, the gradient boosting decision tree model is incrementally updated based on stochastic gradient descent. That is, after obtaining new training samples (zt, yt), the model parameters are fine-tuned in the following directions. ; Where γ is the fine-tuning learning rate, and its value is at least one-tenth of the initial training learning rate. This represents the gradient of the loss function l with respect to the input feature vector z.

[0049] In this embodiment, by introducing the aforementioned incremental update mechanism, the model achieves an evolution from group experience to individual wisdom. In the early stages of treatment, the model mainly relies on the group patterns in historical training data to provide gradient predictions, ensuring the reliability of the basic adjustment direction. As the number of treatments increases, the model gradually captures the patient's unique neural response patterns through fine-tuning, and the fitting accuracy of its nonlinear correlation with the patient's F-wave characteristics and symptom scores gradually improves, thereby making the predicted gradient direction based on the gradient descent method in step S4 closer to the true steepest descent direction.

[0050] To more clearly illustrate the incremental learning mechanism of the machine learning model in this invention, a simplified experimental example is provided below, such as... Figure 4 As shown. Those skilled in the art should understand that this experimental example does not constitute any limitation on the scope of protection of this invention.

[0051] In the simplified scenario, the output of the prediction model F is defined by two key stimulus parameters: the intensity of the medial malleolus target stimulus. (Unit: mA) and foot target stimulation intensity (Unit: mA). For a patient's specific state feature vector Zt collected before a certain treatment, the model's prediction based on the current parameter Fold is: F(Zt;Fold)=( , = (1.5, 0.8); Furthermore, in this treatment, the system, through real-time acquisition of F-wave signals and in-session closed-loop optimization, ultimately determined the stimulation parameters that actually produced the optimal physiological response. The data pair (Zt, (1.7, 0.75)) constitutes a new sample for incremental learning, and the mean squared error function l is used to quantify the difference between the model's predicted value and the actual optimal value. For this sample, the loss is calculated as follows: ; To reduce prediction errors in similar states like Zt in the future, the gradient of the loss function with respect to the model output needs to be calculated. This gradient indicates the direction of the steepest descent: ; ; ; therefore, The physical meaning of this result is that, to reduce losses, the amount of force applied to the target area should be appropriately increased. The predicted value, and appropriately reduce the impact. The predicted value.

[0052] The system performs gradient descent updates based on a preset incremental learning rate γ=0.01. This operation essentially backpropagates gradient information to the model's internal parameters, causing the output of model F(Zt;Fold) to automatically move closer to the (1.7,0.75) direction, for example, it might become (1.504,0.799), thus achieving personalized parameter presetting based on individual treatment history.

[0053] The above experimental example demonstrates one iteration of gradient descent. Since the incremental learning rate γ is set to a small value, such as 0.01, this update is a fine-tuning of the model parameters. As a result, the predicted values ​​of the model output (1.504, 0.799) compared to (1.5, 0.8) before the update clearly point in the direction of the better parameters (1.7, 0.75) validated in this treatment. This mechanism ensures that the model can safely, stably, and continuously optimize its internal parameters based on each new treatment feedback, thereby simulating and realizing a gradual accumulation process similar to that of clinical expert experience, ultimately dynamically matching the optimal stimulus parameters for individual patients.

[0054] In one embodiment, the calculation of dynamic deviation in step S3 uses an asymmetric dead-zone control function: ; in, The peak amplitude of the F-wave extracted at the current time t. The target range for the peak amplitude of the F wave is preset based on the individualized physiological characteristics of the user, i.e., the individualized target range. The lower limit of the interval, The upper limit of the interval, The trend term representing the change in the peak amplitude of the F-wave is obtained by performing time-difference or linear fitting on the peak values ​​of the F-wave extracted from multiple consecutive stimuli. It is used to quantify the dynamic direction and rate of change in spinal cord excitability. This is the low-side feedforward compensation coefficient, used to indicate that the extracted F-wave peak amplitude is lower than the individualized target range. The high-side feedforward compensation coefficient is used to indicate that the peak amplitude of the extracted F-wave is higher than the individualized target range.

[0055] In one preferred embodiment, the And it is adjusted based on the duration of the F-wave latency, specifically as follows; When the F-wave latency is prolonged, the low-side feedforward compensation coefficient is increased. The value of is used to compensate for the delay in excitatory response caused by the decrease in nerve conduction efficiency; When the F-wave latency shortens, the high-side feedforward compensation coefficient is increased. The absolute value of is used to suppress the state of over-excitation caused by increased nerve conduction velocity.

[0056] See Figure 5 To achieve the above objectives, this application provides a system for a multi-target parameter adaptive non-invasive tibial nerve stimulation method, including a stimulation module, a signal acquisition module, and a processing and control module. The stimulation module includes a targeted electrode unit 10 and a stimulation control unit 20. The targeted electrode unit 10 includes a plurality of targeted electrode units 10, which are disposed at the medial malleolus and the sole of the foot, respectively corresponding to the surface projection area of ​​the tibial nerve behind the medial malleolus and the surface projection area of ​​the nerve branch on the medial or lateral side of the sole of the foot. The stimulation control unit 20 is connected to the targeted electrode unit 10 and is used to apply electrical stimulation to the medial malleolus and the sole of the foot according to the instructions of the processing and control module. The signal acquisition module includes a feature signal acquisition and analysis unit 30, which is connected to the target electrode unit 10. It is used to acquire plantar nerve signals induced by electrical stimulation and extract the peak amplitude and latency of the F wave as feedback signals within a preset time window. The time window is located after the M wave and / or before the interference of voluntary muscle contraction. The extraction of the F wave meets the stability conditions, which include: the occurrence rate of the F wave is stable within a preset range, and / or the amplitude change rate of the F wave is lower than a preset threshold. The processing and control module includes an adaptive sensing control unit 40, a data fusion unit 60, and a machine learning prediction model unit 50. The adaptive sensing control unit 40 is connected to the stimulation control unit 20 and the feature signal acquisition and analysis unit 30, respectively. It is used to establish an individualized target interval based on the peak amplitude and latency of the F wave, calculate the dynamic deviation of the peak amplitude of the F wave extracted at the current time relative to the individualized target interval, and use the mean square error of the dynamic deviation as the loss function to adjust the electrical stimulation parameters using the gradient descent method until the peak amplitude of the F wave converges to the individualized target interval.

[0057] Furthermore, this application embodiment also provides a terminal device, the internal structure of which can be as follows: Figure 6As shown, the terminal device includes a processor, memory, communication interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory of the terminal device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and the database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the terminal device is a dedicated dataset or data storage area in the embedded terminal, used to store data called by the computer program. The communication interface of the terminal device is used for data communication with external terminals. The input device of the terminal device is used to receive signals input from external devices. When the computer program is executed by the processor, it implements a multi-target parameter adaptive non-invasive tibial nerve stimulation method as described in the above embodiments.

[0058] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the solution of this application, and does not constitute a limitation on the terminal device to which the solution of this application is applied.

[0059] Furthermore, this application also proposes a readable storage medium comprising a computer program that, when executed by a processor, implements the steps of the multi-target parameter adaptive non-invasive tibial nerve stimulation method as described in the above embodiments. It is understood that the readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0060] Those skilled in the art will understand that implementing all or part of the processes in the multi-target parameter adaptive non-invasive tibial nerve stimulation method of the above embodiments can be accomplished by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the multi-target parameter adaptive non-invasive tibial nerve stimulation method described above. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0061] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, apparatus, article, or multi-target parameter adaptive non-invasive tibial nerve stimulation method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or multi-target parameter adaptive non-invasive tibial nerve stimulation method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or multi-target parameter adaptive non-invasive tibial nerve stimulation method that includes that element.

[0062] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A multi-target parameter adaptive non-invasive tibial nerve stimulation system, comprising a stimulation module, a signal acquisition module, and a processing and control module; Its features are, The stimulation module includes a targeted electrode unit and a stimulation control unit. The targeted electrode unit contains several targeted electrode units disposed at the medial malleolus and the sole of the foot, respectively corresponding to the surface projection area of ​​the tibial nerve behind the medial malleolus and the surface projection area of ​​the nerve branch on the medial or lateral side of the sole of the foot. The stimulation control unit is connected to the targeted electrode unit and is used to apply electrical stimulation synchronously or independently to the several targeted electrode units coupled to the user's medial malleolus and the sole of the foot according to the instructions of the processing and control module. The signal acquisition module includes a feature signal acquisition and analysis unit connected to the target electrode unit, used to acquire plantar nerve signals induced by electrical stimulation, and extract F-wave features as feedback signals within a preset time window; and the extraction of the F-wave satisfies stability conditions; the stability conditions include: the occurrence rate of the F-wave is stable within a preset range, and / or the amplitude change rate of the F-wave is lower than a preset threshold. The processing and control module includes an adaptive sensing control unit, a data fusion unit, and a machine learning prediction model unit. The adaptive sensing control unit is connected to the stimulation control unit and the feature signal acquisition and analysis unit, respectively. It is used to establish an individualized target interval based on the F-wave features, calculate the dynamic deviation of the F-wave features extracted at the current time relative to the individualized target interval, and use the mean square error of the dynamic deviation as the loss function to adjust the electrical stimulation parameters using the gradient descent method until the F-wave accretion converges to the individualized target interval.

2. The multi-target parameter adaptive non-invasive tibial nerve stimulation system according to claim 1, characterized in that, The electrical stimulation parameters include stimulation intensity, stimulation frequency, pulse width, and the ratio of stimulation intensity at the medial malleolus and the sole of the foot.

3. The multi-target parameter adaptive non-invasive tibial nerve stimulation system according to claim 1, characterized in that, The stability condition determination is based on a single measurement and / or on the neural excitability trend characteristics of an individual over a long period of monitoring. The single measurement includes: collecting a signal sequence S={s1,s2,...,s...} n }, where n is the number of sampling points, and is based on equation A f= max(S[t a :t b ])-min(S[t a :t b ]) and formula L f= t onset- t stim Quantify the F-wave characteristics to obtain the peak amplitude A of the current acquisition. f With incubation period L f As a key feedback feature; [t a :t b ] represents the preset time window for the appearance of the F wave, t onset t is the starting point of the F wave. stim The timing of stimulus application is determined by adjusting the stimulus parameter t. stim To make the peak amplitude A of the sampled peak amplitude A f It stabilizes within the preset threshold for F-wave amplitude variation; The long-term monitored neural excitability trend characteristics include: acquiring m simplified neural excitability proxy indicators x daily. i Robust estimation of the m indicators yields the characteristic value X for the day. d Furthermore, the robust estimation employs the median or other statistics that resist outliers; Calculate the characteristic value X for the day d 7-day weighted moving average and trend slope T d When the cumulative monitoring days are less than 7 days, a weighted moving average is calculated based on the actual number of days. After 7 days, a fixed 7-day window is activated. The weight of the weighted moving average decreases as the distance between the measurement date and the current date increases. The trend slope T... d By analyzing the characteristic value X of the actual number of days d Weighted moving average smoothing was performed, and least squares linear fitting was performed based on the smoothed data and the time series corresponding to the actual number of monitoring days to quantify the direction and rate of change in neural excitability. Based on the trend slope T d The sign and magnitude of the value are adjusted to set the preset threshold for F-wave amplitude change. When T... d >0 indicates an upward trend in excitability characteristics, and the upper and / or lower limits of the amplitude change threshold are adjusted upward; when T d <0 indicates a declining trend in excitability characteristics; therefore, the upper and / or lower limits of the amplitude change threshold are lowered, and the adjustment range and |T are adjusted accordingly. d | Positively correlated.

4. The multi-target parameter adaptive non-invasive tibial nerve stimulation system according to claim 1, characterized in that, The dynamic deviation is calculated using an asymmetric dead-zone control function: ； in, The peak amplitude of the F-wave extracted at the current time t. The target range for the peak amplitude of the F wave is preset based on the individualized physiological characteristics of the user, i.e., the individualized target range. The lower limit of the interval. The upper limit of the interval, The trend term representing the change in the peak amplitude of the F-wave is obtained by performing time-difference or linear fitting on the peak values ​​of the F-wave extracted from multiple consecutive stimuli. It is used to quantify the dynamic direction and rate of change in neural excitability. This is the low-side feedforward compensation coefficient, used to indicate that the extracted F-wave peak amplitude is lower than the individualized target range. The high-side feedforward compensation coefficient is used to indicate that the peak amplitude of the extracted F-wave is higher than the individualized target range.

5. The multi-target parameter adaptive non-invasive tibial nerve stimulation system according to claim 4, characterized in that, The And it is adjusted based on the duration of the F-wave latency, specifically as follows: When the F-wave latency is prolonged, the low-side feedforward compensation coefficient is increased. The value of is used to compensate for the delay in excitatory response caused by the decrease in nerve conduction efficiency; When the F-wave latency shortens, the high-side feedforward compensation coefficient is increased. The absolute value of is used to suppress the state of over-excitation caused by increased nerve conduction velocity.

6. The multi-target parameter adaptive non-invasive tibial nerve stimulation system according to claim 1, characterized in that, The gradient descent method is described in which the gradient direction is predicted by a machine learning model, and the machine learning model is a gradient boosting decision tree model, which is an additive model consisting of K regression trees. ； Where z is the input feature vector, which is determined by the trend slope T d The seven-day weighted moving average, the baseline deviation of the F-wave peak amplitude relative to the individualized target interval, and the rate of change in the symptom log together constitute the total. For the k-th regression tree, The learning rate is used to control the contribution of each tree.

7. The multi-target parameter adaptive non-invasive tibial nerve stimulation system according to claim 6, characterized in that, The model is trained on a historical dataset by minimizing the regularization loss function: ； in, Here, l represents the optimal parameters validated for the i-th treatment in history, and l is the mean squared error loss function. This is a regularization term used to control model complexity and prevent overfitting.

8. The multi-target parameter adaptive non-invasive tibial nerve stimulation system according to claim 7, characterized in that, During the iterative adjustment of electrical stimulation parameters using the gradient descent method, the gradient boosting decision tree model is incrementally updated based on stochastic gradient descent, thus obtaining new training samples (z). t ,y t The model parameters are fine-tuned in the following directions; ； Where γ is the fine-tuning learning rate, and its value is at least one-tenth of the initial training learning rate. This represents the gradient of the loss function l with respect to the input feature vector z.

9. The multi-target parameter adaptive non-invasive tibial nerve stimulation system according to claim 1, characterized in that, The plantar nerve signals include at least one of electromyographic signals, nerve electrical signals, and motor signals; the features extracted from the F wave include at least one of the following: F wave amplitude, F wave latency, F wave occurrence rate, F wave duration, and F wave area.

Images