A control method and system for an adaptive electrical pulse generator

Abstract

The purpose of the embodiments of this disclosure is to provide a control method and system for an adaptive electrical pulse generator. The method involves: collecting inflammation-related data from a user; extracting the user's biometric vector using a feature extraction module based on the inflammation-related data; determining the user's disease course classification using a disease course classification module based on the biometric vector; and then determining corresponding electrical pulse parameters using a parameter generation module based on the biometric vector and the disease course classification. According to the electrical pulse parameters, the adaptive electrical pulse generator applies a corresponding broadband electrical pulse stimulation scheme to the user at the electrical stimulation application site. This disclosure, through data fusion and closed-loop optimization, achieves precise personalization and dynamic adaptability in parameter generation, promoting the leap from static templates to intelligent dynamic modes in electrical pulse parameter configuration schemes.

Description

Technical Field

[0001] The embodiments of this disclosure relate to the fields of electrical pulse devices and artificial intelligence, and in particular to a control technique for an adaptive electrical pulse generator. Background Technology

[0002] In the field of bioelectric stimulation, existing devices generally employ preset parameter modes to achieve basic electrical pulse output functions. For example, they achieve basic electrical stimulation functions through preset frequency ranges and fixed waveform parameters. These parameter settings are mainly based on general electrical pulse parameters formed through clinical experience. Such devices typically provide a limited range of adjustable parameters, with their frequency output range mostly concentrated in the low-frequency band. Furthermore, they lack multi-dimensional sensing capabilities of deep tissue states, leading to significant deviations between parameter adjustments and actual physiological states.

[0003] Furthermore, existing electrical stimulation devices generally face technical limitations in clinical applications due to insufficient parameter adaptability. Because core parameters such as current intensity and frequency are mostly set in static preset modes, it is difficult to dynamically adjust them based on real-time physiological feedback obtained during physiotherapy, resulting in a significant deviation between the actual output parameters and the patient's immediate state.

[0004] These technical limitations collectively make it difficult for existing devices to achieve precise and personalized bioelectric intervention when dealing with complex pathological conditions such as aseptic inflammation of the musculoskeletal system. Summary of the Invention

[0005] The purpose of the various embodiments disclosed herein is to provide a control method and system for an adaptive electrical pulse generator.

[0006] According to one aspect of this disclosure, a control method for an adaptive electrical pulse generator is provided, wherein the method includes the following steps:

[0007] Collect users' inflammation-related data, including the site of electrical stimulation application, user baseline data, and user physiological data;

[0008] Based on the inflammation-related data, the user's biometric vector is extracted by the feature extraction module. The user's disease course classification is determined by the disease course classification module based on the biometric vector. Then, the corresponding electrical pulse parameters are determined by the parameter generation module based on the biometric vector and the disease course classification.

[0009] Specifically, the feature extraction module, the disease course classification module, and the parameter generation module are trained using a case sample set.

[0010] The case sample set includes labeled source domain samples and unlabeled target domain samples. The source domain samples include the patient's inflammation-related data, corresponding disease course labels, and electrical impulse parameters.

[0011] The source domain samples are used to perform initial training on the feature extraction module and the disease course classification module.

[0012] The inflammation-related data of the source domain samples and the target domain samples are input into the feature extraction module, and the feature extraction module is adjusted by minimizing the distribution difference of the output biological feature vectors of the two domains.

[0013] Inflammation-related data from the source domain samples and the target domain samples with domain labels are used to perform adversarial training on the feature extraction module and the domain classification module, while the source domain samples are used to optimize the feature extraction module and the disease course classification module.

[0014] The parameters of the feature extraction module are fixed, and the disease course classification module and the parameter generation module are jointly trained based on the biological feature vector of the source domain sample output by the feature extraction module.

[0015] According to the electrical pulse parameters, the adaptive electrical pulse generator applies a corresponding broadband electrical pulse stimulation scheme to the user at the electrical stimulation application site.

[0016] According to one aspect of this disclosure, a control system for an adaptive electrical pulse generator is also provided, comprising an adaptive electrical pulse generator, user equipment, and network equipment; wherein the adaptive electrical pulse generator is connected to electrode plates, and the electrode plates are arranged with a plurality of sensors; wherein the adaptive electrical pulse generator is configured to perform the following operations:

[0017] The sensor collects inflammation-related data from the user, including the user's physiological data.

[0018] According to the electrical pulse parameters, a corresponding broadband electrical pulse stimulation scheme is applied to the user at the electrical stimulation application site via the electrode pads;

[0019] The user equipment is configured to perform the following operations:

[0020] Collect inflammation-related data from the user, including at least the site of electrical stimulation application and the user's basic data;

[0021] The network device is configured to perform the following operations:

[0022] Based on the inflammation-related data, the user's biometric vector is extracted by the feature extraction module. The user's disease course classification is determined by the disease course classification module based on the biometric vector. Then, the corresponding electrical pulse parameters are determined by the parameter generation module based on the biometric vector and the disease course classification.

[0023] Specifically, the feature extraction module, the disease course classification module, and the parameter generation module are trained using a case sample set.

[0024] The case sample set includes labeled source domain samples and unlabeled target domain samples. The source domain samples include the patient's inflammation-related data, corresponding disease course labels, and electrical impulse parameters.

[0025] The source domain samples are used to perform initial training on the feature extraction module and the disease course classification module.

[0026] The inflammation-related data of the source domain samples and the target domain samples are input into the feature extraction module, and the feature extraction module is adjusted by minimizing the distribution difference of the output biological feature vectors of the two domains.

[0027] Inflammation-related data from the source domain samples and the target domain samples with domain labels are used to perform adversarial training on the feature extraction module and the domain classification module, while the source domain samples are used to optimize the feature extraction module and the disease course classification module.

[0028] The parameters of the feature extraction module are fixed, and the disease course classification module and the parameter generation module are jointly trained based on the biological feature vector of the source domain sample output by the feature extraction module.

[0029] The disclosed solution begins with the comprehensive acquisition of multimodal data. Microsensors integrated into electrode pads acquire bioelectrical impedance, surface temperature, and electromyographic signals at the treatment site, while simultaneously recording user baseline information and pain feedback. For example, during knee joint physiotherapy, the sensors continuously capture local impedance change curves, temperature gradients, and muscle activity spectra, combining these with the user's age and body mass index to construct a multidimensional data input.

[0030] The feature extraction module transforms the raw signals into a unified biometric vector, capturing cross-modal correlations. For example, periodic fluctuations in impedance, the slope of temperature rise, and the energy distribution of electromyographic signals are encoded into a 128-dimensional vector, characterizing inflammatory activity and tissue state. This module eliminates data distribution biases between different user groups through cross-domain feature alignment; for instance, it normalizes metabolic rate differences in elderly patients into a standard inflammation index, ensuring compatibility between new user data and the historical case database.

[0031] The disease course classification module determines the current disease stage based on feature vectors and uses adversarial training to eliminate interfering factors such as device acquisition bias. When a patient's features simultaneously match acute and chronic phase patterns, the module outputs a probability distribution and selects the dominant category; for example, if the acute phase probability is 65%, it locks in a mid-frequency analgesia regimen, rather than a compromise parameter. The classification results and feature vectors are input into the parameter generation module to drive dynamic decision-making.

[0032] The parameter generation module employs a multi-objective optimization strategy, referencing not only biometric vectors but also disease course classification confidence for multi-objective optimization. When there is uncertainty regarding the classification results, the system automatically selects the conservative parameter combination with the lowest risk. This mechanism enables the system to maintain treatment safety even when dealing with complex cases, such as generating progressively changing parameter schemes for patients with both acute and chronic characteristics, rather than mechanically applying fixed templates.

[0033] The adaptive electrical pulse generator executes the generation parameters, and its wide frequency range coverage supports flexible switching from low-frequency neuromodulation to deep tissue intervention. Furthermore, during physiotherapy, real-time physiological data is fed back to the system, forming a closed loop of "acquisition-decision-execution-optimization." For example, when a user's impedance stabilizes in the later stages of treatment, the device automatically reduces the frequency and extends the duty cycle, maintaining therapeutic efficacy while improving tolerance.

[0034] This disclosed solution achieves precise personalization and dynamic adaptability in parameter generation through data fusion and closed-loop optimization. Cross-domain feature alignment ensures the safe transfer of historical physiotherapy experience to new users, adversarial training mechanisms eliminate non-pathological interference, and joint optimization strategies establish a complete link from data to decision-making. The final output electrical pulse parameters not only conform to clinical pathological logic but also evolve in real time according to individual responses, driving the configuration of electrical pulse parameters from static templates to intelligent dynamic modes.

[0035] Furthermore, this disclosure employs a data-driven feedback optimization mechanism to achieve continuous learning and optimization of personalized electrical pulse parameters. For example, for patients in the chronic phase, the parameter generation module prioritizes mid-frequency continuous waves and introduces periodic intensity modulation; if an impedance drop is detected during physiotherapy indicating an acute attack, the reinforcement learning model seamlessly switches to the mid-frequency pulse mode. The reinforcement learning mechanism allows for autonomous exploration within safe thresholds; for example, after discovering that a user responds best to a specific duty cycle square wave, that user is included in the personalized strategy library. Data feedback after each physiotherapy session is used to improve the accuracy of the relevant models, making the control system of this disclosure not merely a static parameter configuration tool, but a continuously optimizing intelligent system. Attached Figure Description

[0036] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0037] Figure 1 A flowchart is shown for a control method for an adaptive electrical pulse generator according to an embodiment of the present disclosure.

[0038] The same or similar reference numerals in the accompanying drawings represent the same or similar parts. Detailed Implementation

[0039] The specific embodiments of this disclosure will be further described below with reference to the accompanying drawings.

[0040] Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments of this disclosure are described as apparatuses represented by block diagrams and processes or methods represented by flowcharts. Although the flowcharts depict the operation processes of the various embodiments of this disclosure as sequential processes, many of the operations may be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations may be rearranged. The processes of the various embodiments of this disclosure may be terminated when their operations are completed, but may also include additional steps not shown in the flowcharts. The processes of the various embodiments of this disclosure may correspond to methods, functions, procedures, subroutines, subroutines, etc.

[0041] The methods illustrated in the flowcharts and the apparatuses illustrated in the block diagrams discussed below can be implemented in hardware, software, firmware, middleware, microcode, hardware description languages, or any combination thereof. When implemented in software, firmware, middleware, or microcode, program code or code segments that perform the necessary tasks can be stored in a machine or a computer-readable medium such as a storage medium. One or more processors can perform the necessary tasks.

[0042] Similarly, it will also understand any flowchart, state transition diagram, and the like, representing various processes that can be adequately described as program code stored in a computer-readable medium and thus executed by a computer device or processor, whether or not such computer device or processor is explicitly shown.

[0043] In this document, the term "storage medium" can refer to one or more devices for storing data, including read-only memory (ROM), random access memory (RAM), magnetic RAM, core memory, disk storage media, optical storage media, flash memory devices, and / or other machine-readable media for storing information. The term "computer-readable medium" may include, but is not limited to, portable or fixed storage devices, optical storage devices, and various other media capable of storing and / or containing instructions and / or data.

[0044] A code segment can represent a procedure, function, subroutine, program, routine, subroutine, module, software package, class, or any combination of instructions, data structures, or program descriptions. A code segment can be coupled to another code segment or hardware circuit by passing and / or receiving information, data, arguments, parameters, or stored content. Information, arguments, parameters, data, etc., can be passed, forwarded, or transmitted via any suitable means, including storage sharing, message passing, token passing, network transmission, etc.

[0045] In this context, "computer device" refers to an electronic device that can perform predetermined processing procedures such as numerical calculations and / or logical calculations by running predetermined programs or instructions. It may include at least a processor and a memory, wherein the predetermined processing procedures are performed by the processor executing program instructions pre-stored in the memory, or by hardware such as ASIC, FPGA, DSP, etc., or by a combination of the above.

[0046] The term "computer device" as used above is generally embodied in the form of a general-purpose computer device, whose components may include, but are not limited to, one or more processors or processing units and system memory. System memory may include computer-readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The "computer device" may further include other removable / non-removable, volatile / non-volatile computer-readable storage media. The memory may include at least one computer program product having a set (e.g., at least one) of program modules configured to perform the functions and / or methods of the embodiments of this disclosure. The processor executes various functional applications and data processing by running programs stored in the memory.

[0047] For example, a computer program is stored in the memory for performing various functions and processes of the various embodiments of the present disclosure, and when the processor executes the corresponding computer program, the various embodiments of the present disclosure are implemented.

[0048] Typically, computer devices can be user devices or network devices, or even a combination of both. User devices include, but are not limited to, personal computers (PCs), laptops, and mobile terminals; mobile terminals include, but are not limited to, smartphones and tablets. Network devices include, but are not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing, which is a type of distributed computing consisting of a super virtual computer composed of a group of loosely coupled computers. The computer devices can operate independently to implement the embodiments of this disclosure, or they can connect to a network and implement the embodiments of this disclosure through interaction with other computer devices in the network. The networks in which the computer devices reside include, but are not limited to, the Internet, wide area networks (WANs), metropolitan area networks (MANs), local area networks (LANs), and VPN networks.

[0049] It should be noted that the user equipment, network equipment, and network mentioned are merely examples. Other existing or future computing devices or networks that are applicable to the embodiments of this disclosure should also be included within the scope of protection of this disclosure and are incorporated herein by reference.

[0050] The specific structural and functional details disclosed herein are merely representative and are intended to describe exemplary embodiments of this disclosure. However, the various embodiments of this disclosure can be implemented in many alternative forms and should not be construed as being limited solely to the embodiments set forth herein.

[0051] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0052] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments. Unless the context clearly indicates otherwise, the singular forms “a” and “an” as used herein are also intended to include the plural. It should also be understood that the terms “comprising” and / or “including” as used herein specify the presence of the stated features, integers, steps, operations, units, and / or components, without excluding the presence or addition of one or more other features, integers, steps, operations, units, components, and / or combinations thereof.

[0053] It should also be mentioned that in some alternative implementations, the functions / actions mentioned may occur in a different order than those shown in the figures. For example, depending on the functions / actions involved, the two figures shown successively may actually be executed substantially simultaneously or sometimes in reverse order.

[0054] In traditional physical therapy for aseptic inflammation, electrical stimulation is one of the main treatment methods. Electrical stimulation relies on specialized equipment and professional personnel to adjust the electrical pulse parameters. These parameters include: pulse type (monophase, biphase, modulated pulse, etc.), pulse frequency (1-9000Hz), pulse width (45-200μs), duty cycle (1:2, 1:3, 1:5, etc.), temperature (0-42℃), intensity (low, medium, high), duration (10-20min), and frequency (1-2 times daily).

[0055] The following section uses electrical pulse stimulation protocols for different stages of knee osteoarthritis as examples to introduce the adjustment of electrical pulse parameters.

[0056]

[0057] Current modulation instructions:

[0058] Taking the postoperative subacute period (2-6 weeks after knee surgery) as an example,

[0059] Pulse type: Biphasic pulse or modulated pulse (gradually increasing electrical stimulation to promote the functional recovery of muscles and joints)

[0060] Frequency: 20-50Hz (low frequency, which helps promote local blood circulation and enhance muscle function)

[0061] Pulse width: 300μs (Longer pulse width stimulation is suitable for promoting blood circulation in deep tissues and soft tissue repair)

[0062] Duty cycle: 50%-70% (A higher duty cycle enhances the therapeutic effect and promotes muscle function recovery)

[0063] Duration: 15-20 minutes (the treatment time may be extended to accelerate recovery)

[0064] Frequency of treatment: Once a day, gradually increasing to twice a day (the frequency of treatment will be gradually increased according to the progress of recovery).

[0065] Electrode heating: Heating (38-42℃). Thermotherapy can further enhance blood circulation and promote the recovery of soft tissues and joints.

[0066] Intensity: Medium to strong (the current intensity is gradually increased according to the user's comfort and tolerance to activate muscles and promote recovery).

[0067] As can be seen from the above, traditional physical therapy programs for aseptic inflammation statically classify the course of the disease and set the course according to a fixed time cycle. This cannot truly match the actual disease progress of each patient. Therefore, it is also difficult to personalize the electrical pulse stimulation program set according to a fixed course to meet the real-time physiological needs of the patient.

[0068] Several embodiments of this disclosure provide a control system for an adaptive electrical pulse generator, which can dynamically determine the user's current disease stage and determine a personalized electrical stimulation plan for the user. The control system includes an adaptive electrical pulse generator, an electrical pulse control app deployed on the user's device, and multiple artificial intelligence models deployed on network devices (cloud).

[0069] In some embodiments, the adaptive electrical pulse generator and the mobile app can collect inflammation-related data from the user. The adaptive electrical pulse generator can collect physiological data from the user via sensors (e.g., body surface temperature, degree of edema, inflammatory activity, muscle fatigue and neural control level, skin resistance, etc.). Simultaneously, the mobile app can collect data on the site of electrical stimulation application (e.g., the site of inflammation), basic user data, and physiological data (e.g., pain level assessment score, VAS) by receiving user input.

[0070] Cloud-based artificial intelligence models, for example, can determine the current stage of a user's illness and the corresponding electrical pulse parameters (such as pulse type, pulse frequency, temperature, physiotherapy intensity, physiotherapy duration, number of physiotherapy sessions, waveform, duty cycle, etc.) based on a user's inflammation-related data and a pre-trained model.

[0071] The adaptive electrical pulse generator applies a corresponding wideband electrical stimulation (1-9000Hz) scheme to the user at the site of inflammation according to these electrical pulse parameters, and continuously monitors the user's real-time feedback data during the electrical pulse stimulation process.

[0072] The mobile app uses reinforcement learning algorithms to adjust the electrical pulse parameters for the current user in real time based on the real-time feedback data from the adaptive electrical pulse generator.

[0073] When the use of the adaptive electrical pulse generator ends, the cloud-based artificial intelligence model, such as the prediction model, updates the user's individual physiological state evolution prediction and generates personalized suggestions based on the user's historical physiological state data and electrical pulse parameters.

[0074] Here, the adaptive electrical pulse generator is a portable pulse generator capable of providing precise physical electrical stimulation to specific inflammatory sites (such as the knee, shoulder, and lower back). This portable device offers an intelligent, convenient, and personalized electrical pulse stimulation solution suitable for home use, and boasts excellent adaptability, safety, and therapeutic effects, meeting the treatment needs of diverse users.

[0075] This portable device is small and lightweight, resembling a portable power bank in appearance, making it easy to carry and use. Designed for comfortable handling and ease of operation, it is moderately sized and equipped with Type-C and USB charging ports, supporting convenient use for extended periods. The casing is made of lightweight, durable, and impact-resistant plastic or alloy materials, ensuring its durability and protection in various usage environments.

[0076] The portable device is equipped with electrode pads made of self-adhesive material, which can be repeatedly attached and remain stable without easily falling off. Users simply attach the electrode pads to the area requiring electrical stimulation and activate the electrical pulse stimulation program via the portable device.

[0077] The electrode pads are made of flexible material and designed to conform to various joints and muscles of the human body. The electrode pads can be replaced as needed and are securely attached to the application site using adhesive or adjustable fixing devices. The shape and size of the electrode pads can be designed for different areas, such as the shoulders, knees, and waist, to ensure coverage of the target area and provide optimal electrical stimulation. The surface of the electrode pads uses conductive carbon paste or conductive silver paste, which effectively conducts low-frequency current and reduces skin irritation, ensuring safety and comfort.

[0078] This portable pulse generator has a built-in modulated electrical stimulation module that can provide low- and mid-frequency electrical pulses covering 1-9000Hz. Electrodes are applied to specific areas of the user's body (such as the knee, shoulder, and lower back) to regulate muscle contraction, reduce inflammation, and improve blood circulation. The device adjusts the electrical stimulation program in real time by controlling the intensity and frequency of the current, providing personalized electrical pulse output based on different disease stages, locations, and individual user differences.

[0079] The electrode pads integrate three types of micro-sensors to collect the user's physiological data. Four sets of bioelectrical impedance electrodes located at the edge of the electrode pads accurately distinguish the degree of edema between the superficial and deep tissues by alternately sending safe AC signals of 1kHz and 100kHz. A high-precision infrared temperature probe is embedded in the non-conductive area to monitor surface temperature fluctuations at the treatment site in real time, reflecting the level of inflammation. The central electrical stimulation application area automatically switches to surface electromyography (sEMG) acquisition mode during the electrical stimulation interval, reusing the conductive layer as the signal acquisition interface and using a differential amplifier circuit to capture microcurrent signals during muscle contraction to assess muscle fatigue and neural control levels. All sensors are hidden within a 0.8 mm thick medical-grade silicone substrate, ensuring both data acquisition accuracy and maintaining the electrode pads' flexible and conformable characteristics. Users only need to attach the pads to activate the system for seamless, intelligent monitoring.

[0080] These sensors continuously monitor data during the electrical pulse output process, such as skin temperature, skin resistance, and sEMG signals, and feed this data back to a mobile app in real time for users and doctors to view. This real-time feedback allows the control system to automatically adjust the current intensity and duration based on user feedback, ensuring the accuracy and effectiveness of the electrical pulse stimulation. It also automatically detects and alerts to abnormal situations (such as excessive current or excessively high skin temperature). Users simply select the appropriate area for electrical pulse stimulation, and the system automatically starts and suggests suitable duration, frequency, and current intensity.

[0081] In some embodiments, the electrical pulse generator may also be configured with an edge computing module, which extracts feature vectors from real-time data collected by the sensor by running a lightweight model, so that the mobile APP or cloud model can directly perform relevant calculations quickly based on the feature vectors.

[0082] This portable electrical pulse generator is also equipped with a communication module, supporting data transmission with user devices and the cloud. The portable device can also have a built-in rechargeable lithium battery, supporting extended continuous use. Depending on the intensity of the therapy, the battery can operate for several hours. The portable device features a USB charging port or wireless charging capability for convenient charging anytime, anywhere.

[0083] This portable device is designed for home use. Users simply wear the electrode pads and activate the device to receive electrical pulse stimulation to specific areas, eliminating the need for frequent medical visits and medication dependence. Furthermore, the portable device does not interfere with daily activities, allowing users to carry it during work, leisure, and other occasions for continuous use.

[0084] The electrical pulse control app, deployed on the user's device, features a user-friendly interface. Users can input basic personal data (such as age and BMI), view electrical pulse parameters such as duration, current intensity, and frequency in real time, and adjust parameter settings. The app is available in two versions: a doctor's version and a user version. The user version receives user input and displays parameters during the therapy process in real time. The doctor's version allows doctors to view the user's therapy records and modify electrical pulse parameters.

[0085] All user data during the use of the portable electrical pulse generator will be recorded and uploaded to the cloud for optimizing subsequent electrical pulse stimulation protocols or providing feedback on the user's treatment effects (physiological state evolution). An AI model deployed in the cloud can analyze user historical data, automatically recommend the most suitable electrical pulse stimulation protocol, and continuously optimize it.

[0086] See Figure 1 The flowchart illustrates a control method for an adaptive electrical pulse generator according to an embodiment of the present disclosure.

[0087] like Figure 1 As shown, in step S1, the control system collects the user's inflammation-related data, including the site of electrical stimulation application, the user's basic data, and the user's physiological data; in step S2, the control system extracts the user's biometric vector through a feature extraction module based on the inflammation-related data, determines the user's disease course classification through a disease course classification module based on the biometric vector, and then determines the corresponding electrical pulse parameters through a parameter generation module based on the biometric vector and the disease course classification; in step S3, the control system applies a corresponding broadband electrical pulse stimulation scheme to the user at the site of electrical stimulation application using the adaptive electrical pulse generator according to the electrical pulse parameters.

[0088] Specifically, in step S1, the control system collects the user's inflammation-related data, including the site of electrical stimulation application, the user's basic data, and the user's physiological data.

[0089] The user first attaches the flexible electrode pads to the area where electrical stimulation is applied (such as the front of the knee joint), and after starting the control system, the portable pulse generator automatically enters the data acquisition mode.

[0090] The portable pulse generator collects user physiological data through sensors built into the electrode pads. For example, it measures the initial tissue impedance (e.g., 380Ω) using a bioelectrical impedance sensor and simultaneously detects the surface temperature of the affected area (32.0℃) using a miniature infrared probe. Users select the application site for electrical stimulation, fill in basic user data (age, height, and weight, with BMI calculated from height and weight) via a mobile app, and receive feedback on their physiological data (e.g., current pain level, Visual Analogue Scale (VAS) from 0 to 10).

[0091] Here, the portable pulse generator can complete the acquisition of all physiological signal baseline values ​​within 30 seconds and transmit the data to the cloud server in encrypted form.

[0092] According to one example, the control system can also access the user's electronic physiotherapy records. For instance, a mobile app might record data such as the patient's historical baseline information, the location of inflammation, the course of the disease, and the electrical pulse parameters used during each stage of physiotherapy.

[0093] In step S2, the control system extracts a biometric vector from the user's inflammation-related data through the feature extraction module, determines the user's disease course classification through the disease course classification module based on the extracted biometric vector, and then determines the corresponding electrical pulse parameters through the parameter generation module based on the extracted biometric vector and the determined disease course classification.

[0094] According to one example, the feature extraction module generates a 128-dimensional biometric vector based on the site of electrical stimulation application, user age, BMI, VAS, and temperature, impedance, and sEMG data collected by various sensors. The disease course classification module predicts the user's disease course classification based on this biometric vector, such as [chronic phase 72%, postoperative acute phase 25%, postoperative subacute phase 2%, postoperative recovery phase 1%]. The parameter generation module generates corresponding electrical pulse parameters based on the biometric vector and the disease course classification "chronic phase", such as frequency 4000Hz, intensity 3.8mA, waveform square wave, duty cycle 30%, and duration 28 minutes.

[0095] Here, the feature extraction module, disease course classification module, and parameter generation module are jointly trained using a case sample set.

[0096] The case sample set includes two types of sample data: source domain case sample data and target domain case sample data. Source domain case sample data consists of labeled, complete case data, including a large amount of complete historical case data, such as multimodal inflammation-related data, as well as disease course labels and electrical impulse parameters. Target domain case sample data consists of unlabeled, raw patient data, including a small amount of inflammation-related data from other patients, such as treatment sites, user baseline data, and user physiological data.

[0097] The process of jointly training the feature extraction module, disease course classification module, and parameter generation module based on the case sample set is as follows:

[0098] 1. Initial training of the feature extraction module and the disease course classification module

[0099] The complete data of the source domain case samples (multimodal inflammation-related data + disease course labels + electrical impulse parameters) is input into the feature extraction module. The feature extraction module generates a biometric vector based on the inflammation-related data and outputs it to the disease course classification module for supervised training. The disease course classification module performs disease course classification prediction based on the biometric vector generated by the feature extraction module.

[0100] In one example, the feature extraction module receives raw inflammatory data (such as sensor data like temperature, impedance, and electromyography signals, as well as basic information like age and BMI) and automatically extracts hidden biomedical features using a multi-layer neural network. The disease course classification module simultaneously receives the extracted feature vectors and learns to map them to disease course labels (acute phase / chronic phase, etc.). For example, it discovers that "synchronous changes in body temperature fluctuations and decreased impedance" are potential markers of the acute phase.

[0101] When the disease course classification module makes an incorrect prediction, backpropagation will adjust the parameters of both the feature extraction module and the disease course classification module, forcing the feature extraction module to generate features that are easier to classify.

[0102] According to one example, when the disease course classification module makes an incorrect prediction, such as classifying the acute phase as the chronic phase, the loss function will backpropagate, forcing the feature extraction module to adjust its parameters: 1) enhance features that are sensitive to the acute phase (such as the rapid impedance drop pattern); 2) weaken irrelevant features (such as noise introduced by ambient temperature fluctuations).

[0103] During the initial training phase, the feature extraction module uses end-to-end supervised learning to compress the raw data into high-information-density biometric vectors. These vectors retain key biomarkers related to disease progression, filter out noise caused by individual differences, and establish a unified feature representation benchmark across patients. Deep learning-based neural networks encode complex patterns invisible to the naked eye into classifiable features through multi-layered nonlinear transformations.

[0104] Here, the feature extraction module, for example, employs a transformer structure and is trained to output a 128-dimensional biometric vector based on the user's inflammation-related data. The disease course classification module, for example, can be a classification network employing a multi-head attention mechanism, such as dividing the 128-dimensional input vector into four groups (32 dimensions per group), independently calculating the attention weights for each group, and outputting a four-dimensional probability vector P = [p1, p2, p3, p4] (corresponding to four disease stages).

[0105] This yields a preliminarily trained feature extraction module (capable of extracting key features related to disease progression) and a preliminarily usable disease progression classification module (capable of roughly determining the disease stage based on features). At this point, the trained feature extraction module establishes basic feature representation capabilities, ensuring a reasonable starting point for subsequent optimization steps.

[0106] 2. Optimized training of the feature extraction module

[0107] Inflammation-related data from both the source and target domain case samples are input into the feature extraction module, generating two batches of biofeedback vectors. The distribution difference between the two batches of biofeedback vectors is calculated, and the feature extraction module is adjusted by minimizing this difference.

[0108] In one example, inflammation-related data from both the source and target domains are simultaneously input into the feature extraction module, generating two batches of biofeedback vectors. The distributional differences between the two batches of biofeedback vectors (e.g., statistical mean / variance differences) are calculated, and the parameters of the feature extraction module are adjusted using an optimization algorithm to reduce these differences, achieving feature alignment between the source and target domains. For instance, if the impedance of patients in the target domain is generally lower than that in the source domain, the feature extraction module will learn to eliminate this systematic bias.

[0109] This results in a feature extraction module capable of generating "cross-domain universal" features (feature distributions in the source and target domains are similar). The optimized feature extraction module can eliminate feature biases caused by device and population differences, improving the model's adaptability to new patients.

[0110] In this optimized training step, the training of the disease course classification module is retained to prevent damage to the existing classification capabilities of the disease course classification module.

[0111] 3. Adversarial training between the feature extraction module and the domain classification module

[0112] Inflammation-related data from source domain case samples (each with its own domain label, source domain = 0 / target domain = 1) and inflammation-related data from target domain case samples are input into the feature extraction module to generate biofeature vectors. The generated biofeature vectors are then input into the domain classification module to determine whether the input biofeature vector originates from the source domain or the target domain.

[0113] In adversarial training, the learning objective of the feature extractor is to extract features strongly correlated with the patient's physiological state (such as inflammation activity and edema level) from the raw data (inflammation-related data) while ignoring statistical biases between groups (such as a higher average BMI among patients in the source domain) and focusing on universally applicable biomarkers (such as impedance-temperature correlation patterns). The adversarial objective of the domain classification module is to attempt to identify the data source from the input feature vector (such as determining whether the feature comes from the source or target domain). Through a gradient reversal layer, the gradient direction received by the feature extractor is opposite to that of the domain classifier: when the domain classification module successfully distinguishes the feature source, the feature extraction module is penalized and its parameters are adjusted; when the domain classification module cannot distinguish the feature source, the feature extraction module receives positive feedback.

[0114] By inverting the gradient of the domain classification module during backpropagation, the feature extraction module is forced to generate features that the domain classification module cannot distinguish from the source domain. Ultimately, the accuracy of the domain classification module's output approaches 50% of random guessing (indicating that the features are sufficiently aligned across domains). For example, the "impedance-temperature correlation pattern" of patients in the source domain is generalized to a universal pattern that the target domain can also understand. By obfuscating the domain classification module (making it unable to distinguish between the source and target domains), device-dependent implicit feature encoding can also be indirectly eliminated.

[0115] This yields a feature extraction module capable of generating "domain-invariant" features (features from which the patient's domain cannot be determined), thus completing the training of the feature extraction module. Domain adversarial training is then used to enhance feature generalization, ensuring that the feature extraction module eliminates implicit distributional differences between the source and target domains (which may include a mixture of factors such as device, patient population, and environment), while preserving universally applicable inflammation-related biomarkers across patients.

[0116] For example, the source and target domain data differ due to device differences. The source domain data comes from an older device (sampling rate 1kHz, impedance detection error ±5Ω). The target domain data comes from a newer device (sampling rate 10kHz, impedance error ±1Ω). After adversarial training, the feature extraction module automatically normalizes the impedance values ​​to relative rates of change (e.g., "impedance decreases by 10% / minute"), rather than relying on absolute dimensions. The sampling rate difference between the old and new devices is transformed into unified time-frequency features (e.g., dominant frequency of fluctuations, energy spectral entropy).

[0117] For example, there are population differences between patient data in the source domain and patient data in the target domain. Patients in the source domain are mostly young and middle-aged adults (BMI 22-26, with severe inflammatory response). Patients in the target domain are elderly (BMI 28-32, with a slow inflammatory response). After adversarial training, the feature extraction module suppresses BMI-related features (such as baseline impedance) and enhances dynamic indicators of inflammatory activity (such as temperature-impedance curve fit). The metabolic rate characteristics specific to the elderly are encoded into standardized inflammatory indices comparable to those of young and middle-aged adults.

[0118] In the adversarial training step between the feature extraction module and the domain classification module, the biometric vector generated by the feature extraction module based on inflammation-related data from the source domain case samples can also be input into the disease course classification module to further optimize disease course classification (using the true labels of the source domain data). At this point, the optimization objective of the disease course classification module (maximizing disease course prediction accuracy) and the objective of the domain classification module (minimizing domain classification accuracy) achieve a dynamic balance through parameter updates in the feature extraction module. When the domain classification module attempts to better distinguish between the source and target domains, the gradient received by the feature extraction module forces it to weaken domain-related features; when the domain classification module cannot distinguish the domain source, the feature extraction module maintains its current parameter update direction. The error in the source domain data from the disease course classification module drives the feature extraction module to strengthen inflammation-related features. Each round of parameter adjustments in the feature extraction module is simultaneously affected by both the disease course classification error and the degree of domain confusion, forming a dual-objective optimization.

[0119] 4. Joint optimization of the disease course classification module and parameter generation module

[0120] With the parameters of the feature extraction module fixed, the disease course classification module and the parameter generation module share the source-domain aligned biofeature vector output by the feature extraction module. The disease course classification module and the parameter generation module are jointly trained based on the source-domain aligned biofeature vector, as well as the disease course labels and electrical impulse parameters in the corresponding case samples.

[0121] The source domain aligned biofeature vector (reflecting the patient's physiological state) output by the feature extraction module and the disease course classification (providing prior knowledge of the disease course stage) output by the disease course classification module are input into the parameter generation module. Through regression task learning, an electrical pulse setting matching the historical physiotherapy parameters (electrical pulse parameters in the source domain) is generated.

[0122] As an example, the parameter generation module employs a multi-branch conditional generation architecture to generate safe and controllable electrical pulse parameters based on individualized patient characteristics and disease stage. For instance, the input layer receives a 128-dimensional domain-invariant feature vector and a 4-dimensional disease stage probability vector, and achieves personalized adaptation of electrical pulse parameters through a gating mechanism of feature concatenation and subparameter regression branches (such as pulse frequency branch, stimulation intensity branch, etc.).

[0123] By combining loss functions to simultaneously constrain disease course classification accuracy and parameter generation error, a collaborative cross-task optimization mechanism is constructed. For example, when the disease course classification module misclassifies a patient's chronic phase as postoperative acute phase, the parameter generation module generates intermediate-frequency pulse parameters based on the incorrect classification, resulting in a significant deviation between the output parameters and the patient's actual required intermediate-frequency physiotherapy plan for the chronic phase. In this case, the disease course classification error and parameter regression error have a cumulative effect in the loss function. The backpropagation process simultaneously injects classification logic correction gradients into the disease course classification module and transmits parameter mapping rule adjustment signals to the parameter generation module. This dual-path error correction mechanism forces the disease course classification module to improve its sensitivity to chronic phase biomarkers, while guiding the parameter generation module to establish parameter error tolerance strategies under classification uncertainty. For example, when the classification confidence is below a threshold, a cross-disease course-safe parameter template is preferentially selected. Through loss coupling, the two modules form a closed-loop optimization of "classification guiding parameter generation, and parameter feedback reinforcing classification," ensuring that even under the constraint of a fixed feature space, the system can continuously improve overall decision accuracy through task collaboration and avoid systematic bias caused by overfitting of a single task.

[0124] This leads to a disease classification module that can dynamically determine the current disease stage based on the patient's medical characteristics, and a parameter generation module that can generate personalized electrical pulse parameters based on the patient's medical characteristics and disease stage. This establishes a complete decision-making chain from "physiological data → disease stage assessment → physiotherapy plan," ensuring that the generation of electrical pulse parameters conforms to medical logic.

[0125] At this point, the training of the feature extraction module, disease course classification module, and parameter generation module can be completed.

[0126] In some embodiments, the feature extraction module, disease course classification module, and parameter generation module can be further fine-tuned through the following fifth stage.

[0127] 5. Self-training optimization of the feature extraction module, disease course classification module, and parameter generation module.

[0128] Inflammation-related data (unlabeled) from the target domain is input into the feature extraction module to generate a biofeature vector. This vector is then used by the disease course classification module to predict the corresponding disease course classification. The biofeature vector and disease course classification are then input into the parameter generation module to obtain electrical pulse parameters. The disease course classification predicted by the disease course classification module and the electrical pulse parameters generated by the parameter generation module are used as "pseudo-labels" for the corresponding target domain data.

[0129] From these results, high-confidence predictions (e.g., samples with a prediction probability > 90%) are selected and added to the training dataset as temporary samples. The entire model (including the feature extraction module, disease course classification module, and parameter generation module) is then fine-tuned using the expanded training dataset (source domain samples + temporary samples) to enhance its adaptability to target domain-specific patterns. For example, if a new type of patient is found to be more sensitive to low-frequency stimuli, the parameter generation strategy is adjusted.

[0130] During the fine-tuning phase, the feature extraction module, disease course classification module, and parameter generation module are adjusted synchronously, but are limited by the domain-invariant feature base trained earlier, ensuring that the optimization direction does not deviate from cross-domain consistency.

[0131] Therefore, the model as a whole has the ability to optimize decision-making for patients in the target domain, and can overcome the limitations of source domain data, continuously improving the accuracy of personalized electrical pulse parameters by using actual data from the target domain.

[0132] Through a multi-stage collaborative training mechanism, the model achieves a fundamental innovation in parameter generation logic. Driven by cross-domain feature alignment technology, the feature extraction module can map multi-source heterogeneous sensor data into a unified high-dimensional feature vector, overcoming the limitations of traditional devices that rely on manually set parameter combinations. For users using the electrical pulse generation device for the first time, even if there is a systematic offset between their sensor data distribution and the training source domain, the feature space optimized through domain adversarial training can still accurately capture the correlation pattern of impedance-temperature-electromyography signals, generating electrical pulse parameters adapted to individual biometrics. For example, when a user's electrical stimulation application site is detected to exhibit mid-frequency impedance fluctuations coupled with low-frequency resonance of electromyography signals, the model automatically matches a wide-band modulation strategy, improving mid-frequency energy penetration efficiency while maintaining surface nerve comfort. For users who have used the device before, electrical pulse parameters can also be generated based on real-time biosignals during each therapy session, making the therapy more consistent with the user's current physiological state. Furthermore, for users who have used the device before, the model not only acquires real-time biosignals but also retrieves the user's historical feature vectors and parameter response records, analyzing the trajectory of physiological state changes through a time-series model. For example, during the third physical therapy session for a user with chronic inflammation, the model detected that the impedance fluctuation amplitude was 35% lower than the previous two sessions. Combining this with historical data showing an increasing trend in mid-frequency energy absorption efficiency, the model automatically adjusted the pulse waveform duty cycle and introduced intermittent mid-frequency components.

[0133] Multimodal data fusion architecture endows models with decision-making dimensions that transcend conventional parameter preset schemes. By analyzing the implicit correlations between sensor data, the model can autonomously discover efficient parameter combinations that are difficult to cover with human experience. For example, in some user groups, the model-generated combination of intermittent mid-frequency pulses (7500Hz) superimposed with low-frequency background waveforms (50Hz) increases electrode energy density by 2.1 times compared to traditional single-frequency output, while maintaining the same level of ergonomic comfort. This innovative parameter generation strategy stems from the model's in-depth analysis of impedance-frequency response curves in cross-domain cases, rather than relying on a preset rule base.

[0134] Cross-device compatibility of the model is achieved through feature-level abstraction. Differences in the raw signals acquired by different electrode models are transformed into standardized biometric vectors during the feature extraction stage, ensuring consistency in the parameter generation module's decisions. When a new electrode model is connected to the device, the output parameters remain stable after online feature alignment using a small number of samples. This feature significantly reduces the model reconstruction cost during hardware iterations, enabling the model to continuously evolve across generations of devices.

[0135] Compared to the independent signal acquisition, parameter calculation, and output control modules in traditional devices, this model constructs an end-to-end optimization pathway from feature representation to parameter generation, and possesses collaborative optimization capabilities. This architectural innovation enables the model to autonomously explore efficient regions in the parameter space, such as discovering the nonlinear response characteristics of specific user groups in the 3000-4500Hz frequency band, and generating corresponding parameters accordingly. This data-driven parameter discovery mechanism fundamentally changes the development paradigm of electrostimulation devices.

[0136] In some embodiments, before outputting physiotherapy parameters, the parameter generation module can also search for similar cases in the case database based on the disease course classification results output by the disease course classification module, and adjust the generated electrical pulse parameters according to the distribution of physiotherapy parameters of these similar cases.

[0137] For example, if the parameter generation module calculates a stimulation current of 22A, but the maximum safe intensity used by users with the same disease course in the first 50 similar cases in the case database is 20A, then the pulse current of the current user will also be reduced to 20A.

[0138] The distribution of physical therapy parameters in these similar cases can also be understood as physical boundary constraints, and the parameter generation module needs to adjust the generated electrical pulse parameters according to these constraints.

[0139] In some embodiments, before outputting the electrical pulse parameters, the system also performs boundary corrections on the relevant electrical pulse parameters according to physiological safety rules. For example, for elderly users (>65 years old), even if the recommended intensity for similar cases is 15mA, the system will automatically reduce it to 12mA based on the age attenuation coefficient; if abnormal skin sensitivity is detected (initial impedance <100Ω), the electrode heating function is disabled. The final output parameters inherit both group treatment experience and are dynamically calibrated through individual physiological characteristics, achieving a balance between efficacy and safety.

[0140] According to another example, historical user feedback data can also be prioritized for fine-tuning parameters. For instance, if a patient has a history of being overly sensitive to a 50Hz frequency, the parameters for this electrical pulse will be shifted to the 60Hz frequency band.

[0141] Subsequently, in step S3, the control system applies a corresponding wideband electrical pulse stimulation scheme to the user at the site of electrical stimulation according to the determined electrical pulse parameters.

[0142] The portable pulse generator generates corresponding electrical pulse stimulation schemes according to the electrical pulse parameters determined in step S2. It can cover the low and mid frequency bands of 1-9000Hz, and its wide frequency domain modulation capability can accurately adapt to the treatment needs of different tissue depths.

[0143] The current is evenly distributed to the treatment area through the conductive layer of the electrode pads, with the intensity gradually increasing from 0mA to the target value (e.g., 12mA) to avoid sudden stimulation. The impedance sensor built into the electrode pads monitors the uniformity of the current distribution in real time. If a local impedance abnormality is detected (e.g., a sudden increase of 20% in impedance in a certain area), the electrode polarity is automatically fine-tuned to optimize the electric field coverage.

[0144] Infrared sensors continuously monitor the temperature of the affected area. If the temperature exceeds 39°C or rises by more than 1°C within 5 minutes, the physiotherapy is immediately suspended and the cooling protocol is activated.

[0145] When the impedance is less than 100Ω, the electrode may detach or the skin may be damaged, and the current output will be cut off within 0.2 seconds.

[0146] According to one example, the APP interface synchronously displays the current electrical pulse parameters, such as waveform / intensity / frequency / remaining duration, and uses color gradients to indicate the progress of the physiotherapy, such as red → green representing a decrease in inflammatory activity.

[0147] In some embodiments, Figure 1 The process shown may also include step S4 (not shown). In step S4, the control system adjusts the electrical pulse parameters output by the portable pulse generator in real time based on real-time feedback data from the user during the current electrical pulse stimulation process.

[0148] Because the portable pulse generator can output electrical pulses covering the low and mid frequency bands of 1-9000Hz, its wide frequency domain modulation capability can precisely adapt to the physiotherapy needs of different tissue depths. For example, the low frequency band (1-1000Hz) targets superficial neuromuscular modulation, while the mid frequency band (1000-9000Hz) penetrates to deep inflammatory lesions. Combined with real-time feedback data, the optimal frequency band can be dynamically selected to achieve comprehensive bioelectric intervention from the epidermis to the joint capsule.

[0149] During each physiotherapy session, the control system deeply integrates multimodal sensor data and user feedback, and adjusts electrical pulse parameters in real time through reinforcement learning algorithms to dynamically optimize the therapeutic effect.

[0150] As an example, a lightweight model is run on a portable device via edge computing to extract feature vectors from sensors, ensuring that the inference latency of the reinforcement learning model deployed in the mobile app is less than 50 milliseconds, meeting real-time requirements.

[0151] As an example, a reinforcement learning model deployed in a mobile app is used to optimize electrical impulse parameters for the user in real time on a local machine. The reinforcement learning model can then be iterated over uniformly in the cloud, allowing the mobile app's model to be updated based on these cloud-based iterations.

[0152] When a user initiates electro-pulse therapy, the portable device first continuously collects real-time data from the area receiving the electrical stimulation via sensors on the electrode pads. This data includes physiological indicators such as skin temperature, impedance changes, and muscle electrical signal intensity. Simultaneously, it records the parameters of the currently output electrical pulses (such as frequency, intensity, and waveform). This raw data undergoes preliminary cleaning and standardization by the edge computing module to remove abnormal noise interference and unify the signals from different sensors into a time-aligned digital sequence. For example, if a user suddenly experiences a 0.5°C increase in local temperature during therapy, the system immediately detects this change and marks it as a critical event.

[0153] The processed data is input into a pre-trained reinforcement learning model, which transforms it into a "feature vector" representing the current physiotherapy state. This vector contains implicit indicators such as inflammation activity, tissue response efficiency, and user tolerance. Based on this state feature and combined with historical physiotherapy records (such as the user's best parameter combination from the past three physiotherapy sessions), the model selects the next action within a preset safe parameter range: maintaining the current parameters, fine-tuning the frequency or intensity, switching waveform modes, etc. For example, when a continuous decrease in impedance value is detected and the user reports an increase in pain score, the model may decide to increase the frequency from 4000Hz to 6000Hz while reducing the intensity by 5% to balance comfort. For new users, the model can generate 2-3 sets of candidate parameters (such as 4200Hz / 4mA, 3800Hz / 3mA) within a safe range, prioritizing the option with the highest expected therapeutic benefit.

[0154] After implementing the new pulse parameters, the system continues to collect physiological feedback data over the next 5 minutes, calculating the "therapeutic gain" (such as the rate of decrease in the edema index) and the "side effect cost" (such as the degree of skin redness) of this adjustment, and generating a reward signal based on the combined results. For example, if a parameter adjustment reduces the pain score by 2 points but causes mild muscle twitching, the reward signal will weigh these two outcomes.

[0155] The reward signal triggers the model parameter update mechanism, and the reinforcement learning algorithm dynamically adjusts its decision-making strategy based on the therapeutic effect. If a certain parameter combination receives high rewards three times in a row (e.g., "5000Hz square wave + moderate intensity" is generally effective in patients with chronic conditions), the model will increase its selection priority; conversely, parameter combinations that frequently cause discomfort will be suppressed. The updated model will prioritize calling the optimized policy library in the next decision, forming a continuous optimization loop of "execution → observation → learning".

[0156] The control system disclosed herein achieves dynamic matching between electrical stimulation parameters and the user's physiological state through reinforcement learning. Its core value lies in its real-time personalized adaptation capability. Traditional devices rely on fixed programs or manual parameter adjustments, making it difficult to cope with the nonlinear changes in tissue response during physiotherapy. In contrast, the control system disclosed herein can autonomously explore the optimal parameter range in each physiotherapy session. For example, if an arthritis patient experiences a sudden change in impedance due to inflammation recurrence during physiotherapy, the system automatically completes three parameter iterations within 10 minutes, ultimately locking in an effective 8000Hz mid-frequency short pulse mode, improving efficacy by 40% compared to traditional preset programs.

[0157] Furthermore, the system autonomously uncovers cross-population patterns that are difficult to summarize manually through massive amounts of individual data. For example, for obese users with a BMI > 30, the model found that the combination of "low-frequency carrier wave + intermittent mid-frequency pulse" can effectively penetrate the fat layer. In addition, users' manual intervention behaviors (such as actively reducing intensity) are transformed into safety constraint rules, forming a personalized risk barrier. After a user manually reduced the intensity three times, the system automatically lowered the upper limit of the intensity by 15%, reducing the incidence of adverse reactions to subsequent physiotherapy to zero. This intelligent dynamic balancing mechanism not only improves efficacy and safety but also promotes a paradigm shift in electrostimulation therapy from "programmed presets" to "autonomous evolution."

[0158] In some embodiments, when the use of the portable pulse generator ends, the control system updates the user's individual physiological state evolution prediction and generates personalized suggestions for the user based on the user's historical physiotherapy data (such as physiological state data and electrical pulse parameters).

[0159] After the physiotherapy session, the control system builds a dynamic efficacy prediction (i.e., physiological state evolution prediction) model based on the user's historical data to achieve full-cycle health management.

[0160] First, multi-source treatment data from users is integrated, such as historical disease progression sequences (the evolution path of disease progression labels over the past 30 days, such as "acute phase → subacute phase → chronic phase"), physiological state response curves (daily average improvement rate of impedance, weekly decrease in pain score, and slope of muscle function recovery based on sEMG), and external environmental factors (seasonal temperature changes and user exercise intensity logs). After standardization and time alignment, this data is input into a pre-trained Transformer time series prediction model.

[0161] The predictive model captures long-term dependencies through a self-attention mechanism, such as identifying individualized physical therapy response patterns where "resistance improvement lags behind pain relief," and then outputs two core predictions:

[0162] 1) Risk of inflammation recurrence: Calculate the probability of recurrence in the next two weeks (e.g., 58%). If the probability exceeds 50%, the system will automatically recommend preventive intervention measures, such as increasing the frequency of physiotherapy from 3 times a week to 5 times a week, and adding low-frequency maintenance stimulation (10Hz, 200μs).

[0163] 2) Treatment saturation warning: Analyze the trend of the decline in the repair rate (e.g., the resistance improvement rate drops from 2% / day to 0.5% / day in the past week) and dynamically adjust the treatment plan. For example, it may indicate "The current repair has entered a plateau period. It is recommended to switch to every other day physical therapy and combine it with hot compresses".

[0164] The model's predictions can be visualized through the user's app. For example, a recovery progress bar can be presented to the user, allowing for a horizontal comparison between the expected recovery curve and the actual progress, with key milestones marked, such as "Knee flexion angle reaches 90°, requiring 5 more physiotherapy sessions." Environmental sensitivity alerts can also be provided to users, such as generating warnings based on external data, like "A cold wave is coming in the next week, increasing the risk of arthritis recurrence by 15%, it is recommended to increase the duration of home physiotherapy."

[0165] According to one example, after the end of this physiotherapy session, the control system synchronizes the user's historical physiotherapy data to the doctor; and receives the doctor's modifications to the user's medical history and / or the user's electrical pulse parameters.

[0166] The user's historical physiotherapy data will also be synchronized to the doctor's side. The doctor can query the user's physiotherapy data through the doctor version of the electrical pulse control APP, and modify the control system to the user's determined disease course or electrical pulse parameters, so that the doctor can remotely guide the patient's inflammation physiotherapy.

[0167] It should be noted that the embodiments of this disclosure can be implemented in software and / or a combination of software and hardware, for example, using an application-specific integrated circuit (ASIC), a general-purpose computer, or any other similar hardware device. In one embodiment, the software program of the embodiments of this disclosure can be executed by a processor to implement the steps or functions described above. Similarly, the software program (including associated data structures) of the embodiments of this disclosure can be stored in a computer-readable recording medium, such as RAM memory, a magnetic or optical drive, a floppy disk, or similar devices. Furthermore, some steps or functions of the embodiments of this disclosure can be implemented in hardware, for example, as circuitry that cooperates with a processor to perform the various steps or functions.

[0168] Furthermore, at least a portion of the embodiments of this disclosure can be applied as computer program products, such as computer program instructions, which, when executed by a computing device, can invoke or provide methods and / or technical solutions according to the embodiments of this disclosure through the operation of the computing device. The program instructions that invoke / provide the methods of the embodiments of this disclosure may be stored in a fixed or removable recording medium, and / or transmitted via a data stream in a broadcast or other signal carrying medium, and / or stored in the working memory of a computing device operating according to the program instructions.

[0169] It will be apparent to those skilled in the art that the embodiments of this disclosure are not limited to the details of the exemplary embodiments described above, and that the embodiments of this disclosure can be implemented in other specific forms without departing from the spirit or essential characteristics of the embodiments of this disclosure. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the embodiments of this disclosure is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be encompassed within the embodiments of this disclosure. No reference numerals in the claims should be construed as limiting the scope of the claims. Furthermore, it is apparent that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in the system claims may also be implemented by a single unit or device in software or hardware. The terms "first," "second," etc., are used to denote names and do not indicate any particular order.

Claims

1. A control system for an adaptive electrical pulse generator, comprising an adaptive electrical pulse generator, user equipment, and network equipment; wherein, The adaptive electrical pulse generator is connected to electrode plates, and the electrode plates are equipped with multiple sensors; wherein, the adaptive electrical pulse generator is configured to perform the following operations: The sensor collects inflammation-related data from the user, including the user's physiological data. According to the electrical pulse parameters, the corresponding broadband electrical pulse stimulation scheme is applied to the user at the electrical stimulation application site through the electrode pads; The user equipment is configured to perform the following operations: Collect inflammation-related data from the user, including at least the site of electrical stimulation application and the user's basic data; The network device is configured to perform the following operations: Based on the inflammation-related data, the user's biometric vector is extracted by the feature extraction module. The user's disease course classification is determined by the disease course classification module based on the biometric vector. Then, the corresponding electrical pulse parameters are determined by the parameter generation module based on the biometric vector and the disease course classification. Specifically, the feature extraction module, the disease course classification module, and the parameter generation module are trained using a case sample set. The case sample set includes labeled source domain samples and unlabeled target domain samples. The source domain samples include the patient's inflammation-related data, corresponding disease course labels, and electrical impulse parameters. The source domain samples are used to initially train the feature extraction module and the disease course classification module. When the disease course classification module makes a prediction error, backpropagation will adjust the parameters of both the feature extraction module and the disease course classification module simultaneously. The inflammation-related data of the source domain samples and the target domain samples are input into the feature extraction module, and the feature extraction module is adjusted by minimizing the distribution difference of the output biological feature vectors of the two domains. Inflammation-related data of the source domain samples and the target domain samples with domain labels are used to perform adversarial training on the feature extraction module and the domain classification module. By reversing the gradient of the domain classification module during backpropagation, the feature extraction module is forced to generate features that the domain classification module cannot distinguish from the source. At the same time, the source domain samples are used to optimize the feature extraction module and the disease course classification module. The parameters of the feature extraction module are fixed, and the disease course classification module and the parameter generation module are jointly trained based on the biological feature vector of the source domain sample output by the feature extraction module.

2. The system according to claim 1, wherein training the feature extraction module, the disease course classification module, and the parameter generation module using a case sample set further includes: The high-confidence prediction results generated based on the target domain samples are added to the case sample set as temporary samples, and the feature extraction module, the disease course classification module, and the parameter generation module are adjusted using the updated case sample set.

3. The system according to claim 1, wherein, The user equipment is also configured to perform the following operations: The electrical pulse parameters are adjusted in real time based on the user's real-time feedback data during the current electrical pulse stimulation process.

4. The system according to claim 1, wherein, The network device is also configured to perform the following operations: When the use of the adaptive electrical pulse generator ends, the user's individual physiological state evolution prediction is updated based on the user's historical physiotherapy data, and personalized suggestions are generated for the user.

5. The system according to claim 1, wherein, Before outputting the electrical pulse parameters to the adaptive electrical pulse generator, the network device is also configured to perform the following operations: The parameter generation module searches for similar cases in the case database according to the disease course classification determined by the disease course classification module, and then adjusts the generated electrical pulse parameters according to the distribution of electrical pulse parameters of the similar cases.

6. The system according to claim 1 or 4, wherein, The network device is also configured to perform the following operations: The parameter generation module performs boundary correction on the relevant electrical pulse parameters according to physiological safety rules.

7. The system according to claim 1, wherein, The user equipment is also configured to perform the following operations: After the use of the adaptive electrical pulse generator is completed, the user's historical physiotherapy data will be synchronized to the doctor's dedicated channel; Receive doctor's classification of the user's disease course and / or modification of the user's electrical pulse parameters.

Images