Wireless tongue myoelectricity biofeedback instrument for dysphagia assessment and training

Abstract

The invention discloses a wireless tongue myoelectricity biofeedback instrument for dysphagia assessment and training. The wireless tongue myoelectricity biofeedback instrument comprises an intraoral module and a handheld terminal. The handheld terminal injects micro-current into the electrode in the intraoral module to stimulate tongue muscles according to an instruction of the handheld terminal, and electromyographic voltage signals, tongue body movement signals and optional pressure signals generated by tongue muscle movement are synchronously collected. The collected multi-mode signals are subjected to real-time preprocessing and dynamic impedance compensation in the intraoral module, and then are packaged and sent to the handheld terminal. And the handheld terminal performs deep analysis on the data, extracts multi-dimensional features including cross-modal fusion features, decodes a motion intention through a lightweight deep learning model, and generates a quantitative evaluation report containing five dimensions of muscle strength, endurance, speed, coordination and efficiency based on the motion intention. According to the evaluation report, a personalized gamification training scheme is automatically formulated, seamless closed loop of evaluation and training is realized, and accurate, quantitative and personalized rehabilitation of tongue muscle functions is realized.

Description

Technical Field

[0001] This application belongs to the field of rehabilitation equipment technology, specifically relating to a wireless tongue electromyography biofeedback device for swallowing disorder assessment and training. Background Technology

[0002] Dysphagia is a common and serious complication of diseases such as stroke, Parkinson's disease, and postoperative head and neck tumors, leading to aspiration, aspiration pneumonia, malnutrition, and social isolation, severely impacting patients' quality of life and prognosis. As the "starting engine" of swallowing and the main power source for bolus propulsion, accurate assessment and effective training of the tongue muscles are central to swallowing rehabilitation.

[0003] Current mainstream commercial devices, such as the Iowa Oral Observatory Instrument (IOPI), can only measure static vertical pressure between the tongue and the hard palate. However, actual swallowing is a dynamic process involving complex three-dimensional tongue movements (protrusion, retraction, elevation, lateral deviation, and curling) and temporal coordination. Snapshot-style measurements based solely on pressure cannot assess tongue muscle endurance, coordination between muscle groups, movement speed, and abnormal compensatory patterns, thus having limited correlation with improvements in clinical swallowing function.

[0004] Traditional surface electromyography (sEMG) biofeedback devices used for swallowing rehabilitation typically place electrodes on the skin of the neck under the jaw. This location collects mixed signals from multiple muscles, including the geniohyoid, digastric, and platysma muscles, resulting in a low signal-to-noise ratio and making it highly susceptible to interference from head posture, skin impedance, saliva swallowing, and electrocardiogram activity. Feedback training based on these non-specific signals cannot ensure that the patient truly activates the target tongue muscles and may even reinforce incorrect neck muscle compensatory patterns, delaying the rehabilitation process.

[0005] In current clinical practice, assessment (using VFSS, manometers, etc.) and training (using simple tools or biofeedback devices) are performed by different devices at different times, resulting in completely disconnected data. Therapists cannot dynamically adjust the treatment plan based on the immediate effects of each training session, nor can they quantify the degree of improvement in specific tongue muscle function. The rehabilitation process becomes a black box, relying on the therapist's experience and lacking precise control based on objective data.

[0006] Although some studies have attempted wireless electromyography (EMG) acquisition, most of them are laboratory prototypes with fatal flaws: (1) High latency: Wireless transmission and processing latency often exceeds 50ms, making it impossible to achieve real-time feedback synchronized with swallowing movements; (2) High power consumption: Continuous high sampling rate transmission results in short device battery life, making it impossible to support complete training courses; (3) Data packet loss: In the complex electromagnetic environment of the oral cavity and neck, signal transmission is unstable, affecting the accuracy of assessment; (4) Lack of synchronous multimodal data: Only EMG voltage signals are transmitted, which cannot be accurately time-domain aligned and fused with tongue kinematic data for analysis.

[0007] Therefore, there is an urgent need in this field for a wireless intelligent rehabilitation device that can overcome the above-mentioned defects, achieve specific acquisition of tongue muscle activity, synchronous quantification of multidimensional functions, and seamless closed-loop integration with personalized adaptive training. Summary of the Invention

[0008] To address the aforementioned problems in the existing technology, this application provides a wireless tongue electromyography biofeedback device for swallowing disorder assessment and training. The technical problem to be solved by this application is achieved through the following technical solution: In a first aspect, this application provides a wireless tongue electromyography biofeedback device for swallowing disorder assessment and training, comprising: an intraoral module and a support structure, wherein the intraoral module includes a flexible electrode array and a multilayer microsystem module; the flexible electrode array includes multiple flexible electrodes, the flexible electrode array being petal-shaped and attached to the surface of the tongue; each flexible electrode is located at the center of the petal and communicates with the multilayer microsystem module located at the center via micron-level wires; the intraoral module is encapsulated in a medical silicone shell; the intraoral module communicates with an external handheld terminal through the multilayer microsystem module; The intraoral module is used to inject microcurrents into each flexible electrode to stimulate the patient's tongue muscles according to each task instruction, and to collect electromyographic voltage signals and motion signals generated by the patient's tongue movements according to each task instruction. The electromyographic voltage signals and motion signals are preprocessed and then a quality data package is generated. The quality data package is sent to the handheld terminal, which generates an assessment report based on the quality data package, and then generates a personalized training plan based on the assessment report, so that the patient can start the personalized training plan through the handheld terminal to complete daily tongue muscle training.

[0009] Secondly, this application provides a wireless tongue electromyography biofeedback method for swallowing disorder assessment and training, which uses the wireless tongue electromyography biofeedback device for swallowing disorder assessment and training described in the first aspect to complete the feedback.

[0010] Beneficial effects: 1. This application solves the problem of mixed signals in traditional neck electromyography by directly attaching a petal-shaped flexible electrode array to the tongue; the miniaturized intraoral module is encapsulated with medical silicone, making it comfortable and imperceptible to wear, and achieving true wireless freedom of movement.

[0011] 2. This application constructs a quantitative assessment system covering five dimensions: muscle strength, endurance, speed, coordination, and efficiency by integrating multimodal signals of electromyography, movement, and stress, and calculating fusion features such as efficiency ratio and delay. This system can accurately diagnose the root cause of functional impairment (such as distinguishing between insufficient strength and low efficiency), and surpasses the limitations of single stress or electromyography assessment.

[0012] 3. This application achieves a complete closed loop, from accurate assessment to automatic generation of personalized training plans, and then to gamified execution and real-time dynamic difficulty adjustment. The reinforcement learning-based adaptive engine ensures that training is always within the patient's optimal challenge range, greatly improving the personalization, enjoyment, and efficiency of rehabilitation.

[0013] 4. The assessment report of this application integrates swallowing risk prediction and rehabilitation potential prediction, providing direct support for clinical decision-making; the entire solution takes into account both automation and therapist review, ensuring professionalism and safety.

[0014] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0015] Figure 1 This application provides an example schematic diagram of a wireless tongue electromyography biofeedback device for swallowing disorder assessment and training; Figure 2 This is an example schematic diagram of another wireless tongue electromyography biofeedback device for swallowing disorder assessment and training provided in this application; Figure 3 This is a schematic diagram of the multi-layer microsystem module provided in this application; Figure 4 This is a schematic diagram of the lightweight neural network model provided in this application. Detailed Implementation

[0016] The present application will be described in further detail below with reference to specific embodiments, but the implementation of the present application is not limited thereto.

[0017] Firstly, such as Figure 1 and Figure 2 This application provides a wireless tongue electromyography biofeedback device for swallowing disorder assessment and training, comprising: an intraoral module and a support structure. The intraoral module includes a flexible electrode array and a multilayer microsystem module. The flexible electrode array includes multiple flexible electrodes arranged in a petal shape and attached to the surface of the tongue. Each flexible electrode is located at the center of the petal and communicates with the multilayer microsystem module located at the center via a micron-level wire. The intraoral module is encapsulated in a medical silicone shell. The intraoral module communicates with an external handheld terminal through the multilayer microsystem module. The intraoral module is used to inject microcurrents into each flexible electrode to stimulate the patient's tongue muscles according to each task instruction, and to collect electromyographic voltage signals and motion signals generated by the patient's tongue movements according to each task instruction. The electromyographic voltage signals and motion signals are preprocessed and then a quality data package is generated. The quality data package is sent to the handheld terminal, which generates an assessment report based on the quality data package, and then generates a personalized training plan based on the assessment report, so that the patient can start the personalized training plan through the handheld terminal to complete daily tongue muscle training.

[0018] It is worth noting that the petal-shaped electrode array in this application is not a decorative design, but a functional structure designed to address the issues of tongue curvature fit and signal specificity. For example, in one specific implementation, the array is designed with multiple petals, each petal having a 2 mm diameter gold-plated electrode contact embedded at its end. When worn by a patient, one petal fits under the tip of the tongue, primarily collecting signals from the anterior genioglossus muscle (responsible for protraction); the left and right petals fit along the lateral edges of the tongue, collecting signals from the styloglossus muscle (responsible for lateral deviation and retraction); and the other petal fits in the middle of the tongue dorsum, collecting signals from the intralingual muscles (responsible for elevation and shaping). The leads from all the petals converge at a multi-layered microsystem module located at the center of the palatal support. 。 This module functions like a miniature controller, with a motion sensor at the top, a processing chip in the middle, and a wireless transmission module at the bottom. It is integrated into a space of only 8mm square using three-dimensional stacking and packaging technology. The entire module is encapsulated using medical-grade silicone through liquid infusion, ultimately forming a personalized mouthguard that fits snugly against the dome of the palate. Patients experience no foreign body sensation and can move freely while wearing it.

[0019] The wireless tongue electromyography biofeedback device used in this application for swallowing dysphagia assessment and training begins operation with a specific task command issued by a handheld terminal app. For example, the command is to perform maximal voluntary contraction and extension. Upon receiving the command, the multi-layer microsystem module in the intraoral module first applies a safe microcurrent stimulation (e.g., 30Hz frequency, 200µs pulse width) to the tongue muscles via flexible electrodes to induce electromyographic activity for easy acquisition, and simultaneously activates the built-in inertial sensor. When the patient performs the extension movement, the electrodes acquire weak electromyographic signals (typically in the microvolt range), while the inertial sensor records the acceleration and angular velocity of the tongue. These raw signals are immediately preprocessed within the intraoral module: the electromyographic voltage signal is amplified and power frequency interference is filtered out, the motion signal undergoes temperature compensation calibration, and the contact impedance of each electrode is automatically checked. Signal quality markers (e.g., good signal or unstable contact) are packaged together and transmitted to the handheld terminal via a low-latency wireless link. The handheld terminal's analysis software performs in-depth processing of data packets, calculating indicators such as maximum muscle strength in the forward extension direction being 85% of the norm, but neuromuscular efficiency being only 60%, and integrating these into an assessment report. Subsequently, based on the report's conclusion of low efficiency, a precise force exertion space roaming training program is automatically retrieved from the game library: requiring the patient to propel a virtual spaceship to a target with minimal electromyographic consumption. The patient can start training with this program the next day, and the game difficulty is dynamically adjusted based on real-time performance, thus completing a closed loop from precise assessment to personalized intervention.

[0020] The wireless tongue electromyography biofeedback device for swallowing disorder assessment and training described in this application also includes a pressure feedback palatal support, worn on the hard palate, which has a built-in 8×8 thin-film pressure sensor array for collecting pressure signals at various locations on the tongue.

[0021] In one specific embodiment of this application, the multilayer microsystem module includes a miniature IMU sensor, a microcontroller, and a wireless communication module. The miniature IMU sensor is composed of multiple axis accelerometers and multiple gyroscopes. The multilayer microsystem module is divided into six layers, stacked from bottom to top as a flexible antenna and sensor interface layer, an analog front-end and power management layer, a core digital processing layer, a wireless communication and radio frequency layer, a motion sensing layer, and a protection and thermal management cover layer. The miniature IMU sensor is disposed in the motion sensing layer, the microcontroller is disposed in the core digital processing layer, and the wireless communication module is disposed in the wireless communication and radio frequency layer.

[0022] The structure of the multi-layer microsystem module 12 is as follows: Figure 3As shown, a six-layer vertically stacked package is used, from bottom to top: a flexible antenna and sensor interface layer 121, an analog front-end and power management layer 122, a core digital processing layer 123, a wireless communication and RF layer 124, a motion sensing layer 125, and a protection and thermal management cover layer 126. The core digital processing layer 123 integrates a microcontroller (such as an ARM Cortex-M4 core), the motion sensing layer 125 integrates a miniature IMU sensor (including a three-axis accelerometer and a three-axis gyroscope), and the wireless communication and RF layer 124 integrates a dual-mode wireless chip supporting Wi-Fi 6 and BLE 5.2. This stacked design achieves complete signal acquisition, processing, and communication functions within a very small volume (approximately 8mm × 8mm × 3mm).

[0023] In one specific embodiment of this application, the step of collecting electromyographic voltage signals and motion signals generated by the patient moving their tongue according to each task instruction, preprocessing the electromyographic voltage signals and the motion signals, and then generating a quality data package includes: The first step involves acquiring electromyographic voltage signals through various flexible electrodes and acquiring pressure signals through the pressure feedback palate support. The second step involves collecting the triaxial acceleration and triaxial angular velocity generated by the tongue movement using the multiple axial accelerometers and multiple gyroscopes. The signal preprocessing and quality data packet generation process ensures data reliability at the source and is a core design element for providing structured input for in-depth backend analysis. This process begins with the simultaneous acquisition of multimodal signals: a flexible electrode array directly contacts the tongue's dorsal mucosa to acquire raw electromyographic voltage signals at the microvolt level; simultaneously, a pressure feedback palatal support captures the pressure distribution at the tongue-palatine contact surface; and a multi-axis inertial measurement unit integrated into the microsystem module records the triaxial acceleration and angular velocity generated by the tongue's movement in real time.

[0024] The third step involves bandpass filtering the electromyographic voltage signal acquired by each flexible electrode, calculating the contact impedance of the bandpass-filtered electromyographic voltage signal, dynamically compensating the electromyographic voltage signal based on the contact impedance to obtain a preprocessed electromyographic voltage signal, and setting a quality mark for the electromyographic voltage signal if the contact impedance exceeds the normal range. The fourth step involves performing low-pass and high-pass filtering based on the triaxial acceleration and triaxial angular velocity to obtain the preprocessed motion signal. To address the issue of electromyographic signals being highly susceptible to environmental interference, this application first applies a hardware-and-digital combined bandpass filter to the signal from each electrode channel to preserve the effective frequency band reflecting muscle activity and initially suppress noise. A key step involves periodically injecting a weak test signal into the electrodes and calculating and monitoring the contact impedance at the electrode-tissue interface in real time. This impedance value directly reflects the electrical connection quality of the signal path. Based on this, dynamic compensation is performed: if the impedance fluctuates within the normal range, the amplifier bias or gain is automatically adjusted to stabilize the signal baseline; if the impedance instantaneously jumps above a threshold (e.g., due to poor contact caused by tongue slippage), not only is a clear quality marker added to the data for that time period, but an algorithm is also initiated to interpolate adjacent channel data, striving to maximize the continuity and availability of the signal at the hardware level. Simultaneously, for motion signals from the inertial sensor, high-frequency electronic noise is removed through low-pass filtering, and the linear acceleration component reflecting actual tongue movement is separated through high-pass filtering.

[0025] The fifth step involves assembling the preprocessed electromyographic voltage signal, the preprocessed motion signal, the pressure signal, and the quality marker into a quality data package based on timestamps.

[0026] Finally, all preprocessed signal streams—compensated EMG signals, filtered kinematic data, raw pressure signals, and accompanying quality markers—are assigned a unified timestamp accurate to the millisecond level, generated by the master clock. These elements are synchronously encapsulated to form a structured quality data packet. This packet not only contains multimodal physiological information but, more importantly, embeds metadata (quality markers) regarding data reliability, laying a solid foundation for subsequent precise signal alignment, advanced denoising, feature fusion, and reliable analysis on the handheld terminal.

[0027] In one specific embodiment of this application, the handheld terminal generates an evaluation report based on the quality data packet, including: The first step is for the handheld terminal to parse the timestamp in the quality data packet and use the timestamp to align the preprocessed motion signal, preprocessed electromyographic voltage signal and pressure signal to the same time axis to obtain aligned data. The handheld terminal first parses the high-precision timestamps in the quality data packets, strictly aligning the preprocessed signal streams from different physical sensors (electromyography, inertial measurement unit, pressure) onto the same time axis. This millisecond-level time synchronization is the foundation for subsequent multimodal causal analysis and feature fusion, ensuring the authenticity and analyzability of the temporal relationship between neural drive, mechanical output, and motion performance.

[0028] The second step is to read the quality marker of the preprocessed electromyographic voltage signal, and remove the running artifacts and the electromyographic voltage signal at the estimated missing location in the preprocessed electromyographic voltage signal by using a wavelet transform and IMU reference joint denoising algorithm based on the quality marker, so as to obtain a pure electromyographic voltage signal. The handheld terminal initiates an advanced signal filtering mechanism based on the quality markers attached to the electromyography (EMG) signals. For data segments marked as being affected by motion artifacts, a joint denoising algorithm combining wavelet transform and IMU reference is invoked. This algorithm uses synchronously acquired kinematic data as a reference template to intelligently separate and remove artifact components introduced by head shaking or significant tongue sliding in the wavelet domain. Simultaneously, it utilizes spatially adjacent normal channel data to perform high-fidelity estimation of signals marked as missing due to poor contact, ultimately outputting a pure EMG voltage signal. The core advantage of this step lies in maximizing the preservation of valuable physiological data, significantly improving signal usability and the reliability of analysis.

[0029] The third step involves extracting the electromyographic features of the pure electromyographic voltage signal, the motion features of the preprocessed motion signal, the pressure features of the pressure signal, the bimodal fusion features fused from both sides, and the trimodal fusion features fused from all three. The electromyographic features, motion features, bimodal fusion features, and trimodal fusion features are then input into a lightweight deep learning model to predict the specific movement intentions and associated information of the tongue. Based on high-quality synchronous data, multi-level feature mining is performed. It not only extracts fundamental features of each modality (such as the amplitude and frequency of electromyography, the range and speed of movement, and the peak and distribution of pressure), but also generates bimodal and trimodal fusion features through innovative algorithms. For example, it calculates the "electromyography-pressure efficiency ratio" to quantify the economy of neuromuscular work, or analyzes the coupling strength of "electromyography-movement-pressure" to assess system synergy. These deep features, along with the fundamental features, are input into a lightweight deep learning model that can accurately decode the specific movement intentions the patient attempts to perform (such as extension or elevation) and output related information such as confidence levels.

[0030] In this application, the lightweight deep learning model is a core algorithm module specifically designed for real-time and accurate decoding of tongue movement intentions. Its lightweight characteristics are mainly reflected in its simplified model structure, small number of parameters, and low computational complexity, thus enabling it to run efficiently on the processor of a handheld terminal (such as a tablet computer) and meet the stringent requirements for low-latency real-time feedback during evaluation and training.

[0031] Lightweight deep learning models typically employ pruned and quantized convolutional neural networks (CNNs) or temporal convolutional networks (TCNs) architectures, rather than computationally intensive general-purpose large-scale models. Their input is not the raw signal waveform, but rather a multimodal feature vector that has been fully extracted and normalized in the front end. This vector integrates the time-frequency features of electromyography, the posture and dynamics of motion, the spatial distribution of pressure, and the most innovative cross-modal fusion features (such as efficiency ratio and latency). This strategy of using high-level features as input significantly reduces the data complexity that the model needs to learn, accelerates convergence, and improves interpretability.

[0032] The lightweight deep learning model in this application features a structure specifically optimized for tongue muscle activity patterns. (Reference) Figure 4As shown, the input to the input layer is a fused feature vector of approximately 48 dimensions. This is the "fuel" of the entire model, extracted and fused from preprocessed electromyography, motion, and stress signals by the handheld terminal. The feature selection and dimensionality reduction layer acts as an intelligent filter, automatically learning and weighting the features most effective for intent discrimination while reducing data dimensionality and minimizing subsequent computational burden. It typically consists of a fully connected layer or a one-dimensional convolutional layer, followed by a Dropout operation to prevent overfitting on small datasets. This reduces the feature dimension from 48 to approximately 32, conveying more discriminative information. The temporal-spatial joint understanding layer is the "brain" of the model, responsible for understanding feature patterns. It needs to interpret both how features change over time (e.g., how electromyography changes from resting to activated to relaxed) and the spatial (or logical) relationships between different features (e.g., how channels representing different muscles are co-activated). This layer can employ a depthwise separable convolutional network or a lightweight temporal convolutional network. Compared to standard convolution, it can efficiently extract spatiotemporal features with minimal parameters and computational cost, making it ideal for running on handheld device processors. The global context aggregation layer integrates all local feature maps (representing information at different time points) output from the previous layer to form a global feature vector that represents the entire analysis time window (e.g., 300 milliseconds). Global average pooling can be used, i.e., averaging over all time points. A lightweight attention mechanism allows the model to learn to "focus" on the most important moments within the time window (e.g., peak action moments) and assign them higher weights. The intent classification and information generation head, based on the extracted global features, simultaneously performs multiple prediction tasks and outputs the "association information" mentioned above. A multi-branch parallel structure can be used, with each branch being a small fully connected network. Branch 1 (intent graph classification) outputs a probability distribution using the Softmax activation function to determine the most likely motion intent (e.g., "reaching forward"). Branch 2 (confidence regression) outputs a scalar between 0 and 1 using the Sigmoid function, representing the model's confidence in the classification. Branch 3 (Quality Labels) outputs one or more binary labels (such as "compensated" or "fatigued"), which can be generated using the Sigmoid function or in combination with preset rules. The output layer outputs a structured data dictionary or object, which is fed into the evaluation report generation engine as the basic unit for constructing the task performance matrix.

[0033] The model was pre-trained in the cloud using a large-scale multimodal dataset of tongue muscles from healthy individuals and patients with various swallowing disorders. It learned a robust mapping relationship from complex feature combinations to specific motor intentions (such as forward extension, upward movement, and lateral deviation). In actual deployment, the model can also be fine-tuned on handheld terminals with a small amount of data to quickly adapt to the personalized physiological characteristics of different patients, thereby further improving the individualized accuracy of intention decoding.

[0034] The lightweight deep learning model presented in this application balances accuracy, real-time performance, and power consumption. It can distinguish subtle motor intentions (such as differentiating between a simple upward movement and a leftward upward movement) with extremely high confidence and output structured information including intention category, confidence score, and possible anomaly markers (such as detected compensation). This provides a reliable technical guarantee for generating accurate assessment reports and achieving immersive real-time game control, serving as a key software foundation for realizing intelligent and personalized rehabilitation functions.

[0035] The fourth step is to generate an evaluation report based on the specific motion intentions and related information corresponding to all task instructions.

[0036] The handheld terminal integrates the motor intentions decoded from all assessment tasks and their rich contextual information (such as execution quality, reaction time, and efficiency indicators). Through a rule engine and statistical model, it automatically generates a comprehensive, quantitative assessment report. The significant advantage of this process lies in achieving automatic translation from signal to intention to clinical meaning. It not only reveals whether the patient can perform the action but also analyzes in depth how to do it and the efficiency, providing personalized rehabilitation with data depth and decision-making basis far exceeding traditional methods.

[0037] In one specific embodiment of this application, the electromyographic features include time-domain features, frequency-domain features, nonlinear features, and multi-channel spatial features; the motion features include posture features, kinematic features, dynamic features, and stability features; the pressure features include intensity features, temporal features, distribution features, and dynamic features; the bimodal fusion features include electromyographic-pressure fusion features, electromyographic-motion fusion features, and pressure-motion fusion features; the trimodal fusion features include electromyographic-motion-pressure fusion features; and the associated information includes confidence scores, multi-class probability distributions, temporal information, and quality markers.

[0038] Specifically, time-domain features, such as the root mean square (RMS) value, reflect the average intensity of muscle contraction; frequency-domain features, such as the median frequency, whose decrease indicates muscle fatigue; nonlinear features, such as sample entropy, are used to assess the complexity of neural control signals; and multi-channel spatial features reveal the synergistic or compensatory patterns between different muscle groups by analyzing the correlation between signals from different electrodes. Motion features, based on inertial sensor data, describe the kinematic performance of the tongue: posture features, such as pitch angle, directly correspond to the angle at which the tongue is raised; kinematic features include displacement and peak velocity; dynamic features, such as jerk, are used to quantify the smoothness of movement; and stability features assess control precision by analyzing minute tremors during posture maintenance. Pressure features, derived from the palate sensor array, quantify mechanical output: intensity features, such as average pressure; temporal features, including pressure rise time; distribution features, by calculating the pressure center coordinates and symmetry index, describe the spatial application of force; and dynamic features focus on the fluctuation characteristics of pressure during the task.

[0039] Bimodal fusion features are generated by correlating signals from different modalities. For example, the efficiency ratio in the electromyography-pressure fusion feature is calculated by dividing the pressure peak value by the integrated electromyography value at which that pressure was generated, with units of kPa / μVs, directly measuring the economy of converting neural drive into effective force. Electromyography-motor fusion features, such as conduction delay, refer to the time difference (unit: ms) between the electromyography initiation point and the motor initiation point. Pressure-motor fusion features can calculate the similarity between the pressure center trajectory and the tongue movement trajectory. Trimodal fusion features may obtain the strongest coupling coefficient between the electromyography, pressure, and motor signals through canonical correlation analysis.

[0040] When a lightweight deep learning model outputs a predicted intent category (such as extending forward), it includes a series of related information: a confidence score (such as 0.92) to express the certainty of the prediction; a multi-class probability distribution to show the probability of all possible intents (such as 85% for extending forward and 10% for raising); temporal information including the start and end times of the action; and a quality label, which is a label determined by the model based on the input features, regarding the performance of this action, such as whether there is compensation or insufficient force.

[0041] This implementation method achieves in-depth mechanistic assessment of tongue muscle function through a multi-level, multi-modal feature system. It not only extracts basic and deep features of electromyography, movement, and stress (such as neural control complexity, muscle coordination patterns, and movement fluency), but also integrates features across modalities (such as "EMG-stress efficiency ratio" and "EMG-motor delay"), inputting them into a lightweight deep learning model to output structured relational information (confidence, temporal sequence, and quality labels). Ultimately, it generates a detailed functional profile that goes beyond simple scoring, including abnormal pattern diagnosis and execution quality analysis, thus providing direct and reliable guidance for developing precise, personalized rehabilitation plans.

[0042] In one specific embodiment of this application, generating an evaluation report based on the specific motion intentions and associated information corresponding to all task instructions includes: The first step is to reorganize the specific motion intentions and associated information corresponding to all task instructions into a task matrix; Based on the evaluation agreement, this application categorizes and reorganizes the decoded events (such as intentions and confidence and timing information) scattered at different time points according to task categories (such as strength test and endurance test) to form a structured data table. Each row represents an independent movement attempt, and the columns contain multi-dimensional attributes such as intention type, confidence, reaction time, and quality mark.

[0043] The second step is to filter out specific motion intentions and related information in the task matrix whose confidence scores are lower than the confidence threshold to obtain the reconstructed task matrix; This application applies a preset confidence threshold (e.g., 0.7) to the recombined task matrix, automatically filtering out low-quality data entries with confidence scores below this threshold. This ensures that subsequent analysis is based only on reliable, high-quality intent recognition results, thereby guaranteeing the accuracy and stability of the evaluation conclusions and obtaining a clean, high-quality recombined task matrix.

[0044] The third step is to calculate the features of the recombined task matrix in multiple dimensions to obtain multidimensional features, and then aggregate the multidimensional features into a multidimensional clinical function score. This application, on the reorganized task matrix, no longer performs simple counting and statistics, but instead performs complex multidimensional feature calculations. It analyzes the distribution and relationships of data in the matrix across time, space, and patterns, calculating high-level features such as execution accuracy, stability of reaction speed, and consistency of performance across different tasks. Then, through specific algorithmic models (such as weighted aggregation or principal component analysis), these multidimensional features are mapped and aggregated into several comprehensive functional dimension scores with clear clinical significance, such as percentage scores for five dimensions: muscle strength, endurance, coordination, speed, and efficiency.

[0045] The fourth step involves assessing the swallowing risk and predicting the rehabilitation potential based on the multidimensional clinical function scores, and generating an assessment report based on the assessment and prediction results.

[0046] This application uses the aggregated multidimensional clinical functional scores as core inputs, which are fed into two parallel analysis modules. The swallowing risk assessment module combines the patient's individual information (such as age and diagnosis) to weight and logically judge the functional scores, outputting a quantified risk level (low / medium / high) and major risk factors. The rehabilitation potential prediction module predicts the patient's potential for functional improvement and potential bottlenecks in a specific future period based on the current functional baseline, historical progress trends, and learning curve models. Finally, by integrating these assessment and prediction results, a complete and graphically rich assessment report is automatically generated, covering a functional overview, risk warnings, potential analysis, and personalized recommendations.

[0047] In one specific embodiment of this application, the step of assessing swallowing risk and predicting rehabilitation potential based on the multidimensional clinical function scores, and generating an assessment report based on the assessment and prediction results, includes: The first step is to weight the multidimensional clinical function scores to obtain the baseline risk score; This application uses pre-defined weights in the clinical knowledge base (e.g., assigning higher weights to lift endurance and retraction coordination) to perform a weighted summation of clinical function scores across multiple dimensions, such as muscle strength, endurance, and coordination, to generate an initial, quantified baseline risk score.

[0048] The second step is to use the patient's basic information as a clinical correction factor to correct the basic risk score to obtain the final score, and to determine the swallowing risk level based on the range of the final score. This application uses basic information such as the patient's age, specific diagnosis, and disease course as key clinical correction factors. For example, for elderly patients or patients with specific brainstem lesions, their baseline risk score will be adjusted upwards according to rules. The final score obtained after correction will be compared with a preset risk threshold range to objectively determine the swallowing risk level of "low," "medium," or "high."

[0049] The third step is to analyze the risk factors based on the multidimensional clinical function scores and generate a risk heatmap based on the swallowing risk level and risk factors. This application goes beyond a single level; it further analyzes the scores from each dimension to identify the specific, most contributing risk factors (such as "the risk of delayed pharyngeal initiation is increased because the levator endurance is only 50% of the norm"). Combining the risk level with these specific factors, a risk heatmap is automatically generated to visually demonstrate the severity and relative importance of each risk factor.

[0050] The fourth step is to calculate the rehabilitation potential factor using the multidimensional clinical function score, and predict the potential cycle based on the rehabilitation potential factor. This application uses current multidimensional functional scores, historical progress trends (if any), and a learning curve model trained on a large amount of data to calculate a comprehensive rehabilitation potential factor, and based on this, predicts the functional improvement goals that may be achieved in a specific future period (such as 4 weeks or 12 weeks).

[0051] The fifth step is to set the potential level based on the predicted potential cycle and identify the bottleneck factors that restrict rehabilitation. Based on the predicted improvement, this application categorizes rehabilitation potential into "high," "medium," and "low" levels. Simultaneously, by analyzing the constraints between various dimensions (for example, even with adequate muscle strength, an extremely low "efficiency" dimension severely limits overall functional improvement), it intelligently identifies key bottlenecks hindering the rehabilitation process.

[0052] The sixth step is to record the swallowing risk level, risk heat map, predicted potential cycle, potential level, and bottleneck factors in the assessment report.

[0053] This application structurally integrates all the aforementioned outputs—swallowing risk levels, visualized risk heatmaps, quantified predictive potential cycles, potential levels, and clearly identified bottleneck factors—and records them in the final assessment report. This allows the report to not only describe the current functional status but also include risk warnings and forward-looking rehabilitation plans, providing therapists with a core basis for developing precise intervention strategies.

[0054] In one specific embodiment of this application, generating a personalized training plan based on the assessment report, so that the patient can initiate the personalized training plan through the handheld terminal to complete daily tongue muscle training, includes: The first step is to analyze the risk level and potential prediction in the assessment report, and through the built-in clinical rule base, map specific functional impairments to targeted digital training games, and set personalized intensity, number of sets and progression rules for each training task to form a personalized training plan. This application provides an in-depth analysis of the core conclusions in the assessment report, particularly the swallowing risk level and rehabilitation potential prediction. By utilizing a built-in rule knowledge base constructed based on clinical guidelines and expert experience, abstract descriptions of functional impairments (such as "low efficiency dimension score") are automatically mapped to specific, targeted digital training games (e.g., "precision force control game"). Simultaneously, based on the patient's current ability baseline, personalized starting intensity (e.g., target pressure value), training volume (e.g., number of sets and repetitions), and dynamic progression rules (e.g., "increase difficulty by 5% after three consecutive successful attempts") are set for each training task, thereby automatically generating a structured, periodic training plan.

[0055] The second step is to push the personalized training plan to the therapist's review platform to complete the review and fine-tuning, thereby generating a review completion signal; This application automatically pushes the generated personalized training plan to a dedicated review platform for therapists. On this platform, therapists can comprehensively review the plan's rationale and safety, and fine-tune and optimize the selection of training tasks and parameter settings based on their extensive clinical experience. After review, the therapist confirms the plan on the platform, generating a review completion signal. This step ensures that the AI-generated plan is always under the supervision of clinical experts, combining automation efficiency with professional judgment to guarantee the safety and effectiveness of the rehabilitation plan.

[0056] The third step is to synchronize the review completion signal and the personalized training plan to the handheld terminal so that the patient can complete the daily training of the tongue muscles according to the personalized training plan.

[0057] This application, representing approval, along with the finalized personalized training plan, will be synchronized to the patient's handheld device. The patient can then clearly view daily training tasks through an application on the device and initiate and complete targeted tongue muscle training with gamified guidance. The entire process achieves a closed loop from assessment, planning, approval to execution, enabling precise, personalized, and safe rehabilitation training to be carried out efficiently on the patient's end.

[0058] Secondly, this application provides a wireless tongue electromyography biofeedback method for swallowing disorder assessment and training, which uses the wireless tongue electromyography biofeedback device for swallowing disorder assessment and training from the first aspect to complete the feedback.

[0059] Step 1: The patient wears the intraoral module on the hard palate, allowing its "petal-shaped" flexible electrode array to naturally conform to the surface of the tongue. A handheld terminal (such as a tablet) pairs with the intraoral module wirelessly (such as via Bluetooth). After the instrument is activated, a calibration procedure is executed: guiding the patient to perform tongue relaxation and extreme movements in specific directions (such as forward extension and upward elevation) to collect individualized resting baseline, maximum electromyographic reference values, and range of motion, completing personalized initialization of signal acquisition parameters. Following visual and voice guidance from the handheld terminal app, the patient performs a set of preset standardized assessment tasks. For example: Task A (Maximum Strength Test): The instruction is "Extend your tongue forward with maximum force," for 3 seconds. Task B (Endurance Test): The instruction is "Maintain tongue elevation with half the force," until exhaustion. Task C (Coordination Test): The instruction is "Follow the screen dot to move the tip of your tongue," completing a complex trajectory (such as an "∞" shape). During each task, the intraoral module simultaneously performs the following operations: Step Two: According to the task instructions, apply safe microcurrent stimulation to the tongue muscles via flexible electrodes (if needed) and acquire the raw electromyographic voltage signals. Acquire the triaxial acceleration and triaxial angular velocity motion signals of the tongue using built-in miniature IMU sensors (accelerometer and gyroscope). (If a pressure feedback palatal support is used) Simultaneously acquire the pressure distribution signal of the tongue-palatine contact surface. Perform real-time preprocessing of the raw signals within the intraoral module: bandpass filter the electromyographic signals and calculate contact impedance for dynamic compensation and quality labeling; filter the motion signals to separate linear acceleration. Package the preprocessed signals and quality labels with precise timestamps to generate a "quality data package" and send it to the handheld terminal.

[0060] Step 3: Strictly synchronize EMG, motion, and pressure signals using timestamps. Based on the quality markers of the EMG signals, advanced algorithms (such as wavelet transform combined with IMU reference) are used to remove motion artifacts and repair or eliminate low-quality data segments to obtain a clean signal. Temporal (e.g., root mean square value RMS), frequency (e.g., median frequency MF), nonlinear, and spatial features are extracted from the clean EMG signals; postural, kinematic, dynamic, and stability features are extracted from the motion signals; and intensity, temporal, distribution, and dynamic features are extracted from the pressure signals. Cross-modal features are calculated. For example, the "EMG-pressure efficiency ratio" (peak pressure / peak EMG) is calculated to quantify neuromuscular efficiency; and the "EMG-motor delay" is calculated to assess the synchronicity of nerve conduction and motor initiation. The extracted rich feature set is input into a lightweight deep learning model, which outputs the patient's specific motor intention (e.g., "extend_high intensity") and associated confidence, temporal, and quality marker information.

[0061] Step 4: Reorganize the decoded motor intentions and related information by task, filter low-confidence data, and form a reliable task performance matrix. Calculate higher-dimensional performance indicators (such as accuracy, speed, and stability) on this matrix, and aggregate them to generate clinical functional scores across five dimensions: muscle strength, endurance, speed, coordination, and efficiency. Weight the functional scores and adjust them based on patient age, diagnosis, and other clinical information to determine low, medium, and high risk levels, and analyze the main risk factors. Based on current functional level, historical trends, and the learning model, predict the functional improvement potential and bottlenecks for future cycles (e.g., 4 weeks). Automatically generate a comprehensive assessment report with graphics and text, including a five-dimensional radar chart, risk heatmap, potential prediction curve, and specific diagnostic recommendations.

[0062] Step 5: Analyze the assessment report and automatically map diagnosed problems (such as "inefficiency") to targeted digital training games (such as "Precision Force Space Roaming") using a built-in clinical rule base. Personalized initial difficulty, number of training sets, and progression rules are set for each game, forming a detailed one-week training plan. This plan is pushed to the therapist's management platform for review and necessary fine-tuning to ensure safety and professionalism. Once approved, the plan is locked. The training plan is synchronized to the patient's handheld device. The patient launches the app daily and follows the guidance for gamified training. Their tongue muscle activity controls the game progress in real time, providing multimodal biofeedback (visual, auditory, tactile). During training, the integrated reinforcement learning algorithm dynamically fine-tunes the game difficulty based on the patient's real-time performance (such as success rate and fatigue level), keeping it within the "optimal challenge range." Each training session's data is uploaded to the cloud to update the patient's personal digital twin model. This application can analyze progress periodically (e.g., weekly) and automatically iterate and optimize the training plan for the next cycle based on the latest data.

[0063] This application discloses a wireless tongue electromyography biofeedback device and method for assessing and training dysphagia, belonging to the field of medical device technology. The device includes an intraoral module and a handheld terminal. The intraoral module comprises a petal-shaped flexible electrode array conforming to the tongue surface and a highly integrated multilayer microsystem module, encapsulated in a medical silicone shell. This module can inject microcurrent stimulation into the tongue muscles according to instructions from the handheld terminal, and simultaneously collect electromyographic voltage signals, tongue movement signals, and optional pressure signals generated by tongue muscle activity. The collected multimodal signals undergo real-time preprocessing and dynamic impedance compensation within the intraoral module before being packaged and sent to the handheld terminal. The handheld terminal performs deep analysis of the data, extracting multidimensional features including cross-modal fusion features, decoding the movement intent through a lightweight deep learning model, and generating a quantitative assessment report covering five dimensions: muscle strength, endurance, speed, coordination, and efficiency. Finally, the system automatically develops a personalized gamified training program based on the assessment report, achieving a seamless closed loop between assessment and training. This application solves the technical problems of existing technologies, such as single evaluation dimensions, non-specific signal acquisition, disconnect between evaluation and training, and large latency and high power consumption in real-time wireless feedback, and realizes precise, quantitative, and personalized rehabilitation of tongue muscle function.

[0064] To verify the effectiveness and advancement of the solution described in this application, computer simulation and comparative experiments were conducted.

[0065] In a simulation environment, electrophysiological models of the main tongue muscles (genioglossus, hyoidohyoidus, etc.) and a neck compensatory muscle model were constructed. The same mixed electromyographic signals were input to both a traditional neck-attached electrode array and the proposed "petal-shaped" intraoral electrode array. Results showed that for the "tongue protrusion" movement, the signal-to-noise ratio (SNR) of the target electromyographic (genioglossus) component in the intraoral array was an average of 25.8 dB, while the SNR of the traditional neck electrode was only 12.3 dB, and its signal contained up to 35% non-target muscle activity. This indicates that the electrode array proposed in this application can achieve more specific and purer tongue muscle signal acquisition.

[0066] Data from 20 patients with mild dysphagia were simulated and assessed using a conventional commercial tongue depressor (measuring only pressure) and the instrument described in this application. The conventional device only provides a single indicator: "maximum tongue pressure." The instrument described in this application generates a multidimensional assessment report. Results showed that 7 patients had normal maximum tongue pressure (>40 kPa), but the assessment report showed significantly low scores (<50 points) in the "neuromuscular efficiency" dimension, identifying compensatory patterns. In subsequent targeted training, these 7 patients showed significant improvement in their actual swallowing safety scores (assessed via video-fluorescence swallowing imaging) after training only in "efficiency," while the control group, which only trained "strength," showed no significant improvement. This demonstrates the crucial value of the multimodal fusion assessment described in this application for revealing deep functional impairments and guiding precise rehabilitation.

[0067] In a simulated environment, 10 patients were trained on the same game for two weeks using two different algorithms. The results showed that the group using the adaptive algorithm described in this application had an average 28% higher score on the "flow experience" questionnaire during training, an average 15% higher daily training task completion rate, and after two weeks, an average 22% higher improvement in the electromyography-stress efficiency ratio compared to the fixed-difficulty group. This indicates that the personalized adaptive training engine described in this application can significantly improve patient engagement and training efficiency.

[0068] It is worth noting that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0069] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. A wireless tongue electromyography biofeedback device for assessing and training dysphagia, characterized in that, include: Intraoral module and support structure, wherein the intraoral module includes a flexible electrode array and a multilayer microsystem module; The flexible electrode array includes multiple flexible electrodes in a petal shape, which are attached to the surface of the tongue. Each flexible electrode is located at the center of the petal and communicates with a multi-layer microsystem module located at the center via micron-level wires. The intraoral module is encapsulated in a medical silicone shell. The intraoral module communicates with an external handheld terminal through the multi-layer microsystem module; The intraoral module is used to inject microcurrents into each flexible electrode to stimulate the patient's tongue muscles according to each task instruction, and to collect electromyographic voltage signals and motion signals generated by the patient moving the tongue according to each task instruction. The electromyographic voltage signals and motion signals are preprocessed and then a quality data package is generated. The quality data packet is sent to the handheld terminal, which generates an assessment report based on the quality data packet and then generates a personalized training plan based on the assessment report, so that the patient can start the personalized training plan through the handheld terminal to complete daily tongue muscle training.

2. The wireless tongue electromyography biofeedback device for swallowing disorder assessment and training according to claim 1, characterized in that, The wireless tongue electromyography biofeedback device for swallowing disorder assessment and training also includes a pressure feedback palatal support, worn on the hard palate, which has a built-in 8×8 thin-film pressure sensor array for collecting pressure signals at various locations on the tongue.

3. The wireless tongue electromyography biofeedback device for swallowing disorder assessment and training according to claim 2, characterized in that, The multi-layer microsystem module includes a miniature IMU sensor, a microcontroller, and a wireless communication module. The miniature IMU sensor consists of multiple axis accelerometers and multiple gyroscopes. The multi-layer microsystem module is divided into six layers, stacked from bottom to top: a flexible antenna and sensor interface layer, an analog front-end and power management layer, a core digital processing layer, a wireless communication and radio frequency layer, a motion sensing layer, and a protection and thermal management cover layer. The miniature IMU sensor is located in the motion sensing layer, the microcontroller is located in the core digital processing layer, and the wireless communication module is located in the wireless communication and radio frequency layer.

4. The wireless tongue electromyography biofeedback device for swallowing disorder assessment and training according to claim 3, characterized in that, The process of collecting electromyographic voltage signals and motion signals generated by the patient's tongue movements according to each task instruction, preprocessing the electromyographic voltage signals and motion signals, and then generating a quality data package includes: Electromyographic voltage signals are acquired through various flexible electrodes, and pressure signals are acquired through the pressure feedback palate support. The triaxial acceleration and triaxial angular velocity generated by the tongue movement are collected by the multiple axial accelerometers and multiple gyroscopes. The electromyographic voltage signal acquired by each flexible electrode is bandpass filtered, and the contact impedance of the bandpass filtered electromyographic voltage signal is calculated. The electromyographic voltage signal is dynamically compensated according to the contact impedance to obtain a preprocessed electromyographic voltage signal. If the contact impedance exceeds the normal range, a quality mark is set for the electromyographic voltage signal. Based on the triaxial acceleration and triaxial angular velocity, low-pass and high-pass filtering are performed to obtain the preprocessed motion signal; The preprocessed electromyographic voltage signal, the preprocessed motion signal, the pressure signal, and the quality marker are combined into a quality data package according to the timestamp.

5. The wireless tongue electromyography biofeedback device for swallowing disorder assessment and training according to claim 4, characterized in that, The handheld terminal generates an evaluation report based on the quality data packet, including: The handheld terminal parses the timestamp in the quality data packet and uses the timestamp to align the preprocessed motion signal, preprocessed electromyographic voltage signal, and pressure signal onto the same time axis to obtain aligned data; The quality marker of the preprocessed electromyographic voltage signal is read, and the running artifacts and the electromyographic voltage signal at the estimated missing location are removed from the preprocessed electromyographic voltage signal by wavelet transform and IMU reference joint denoising algorithm according to the quality marker, so as to obtain a pure electromyographic voltage signal. The electromyographic features of the pure electromyographic voltage signal, the motion features of the preprocessed motion signal, the pressure features of the pressure signal, the pairwise fused bimodal fusion features, and the trimodal fusion features are extracted; and the electromyographic features, motion features, bimodal fusion features, and trimodal fusion features are input into a lightweight deep learning model to predict the specific movement intention and associated information of the tongue. An evaluation report is generated based on the specific motion intent and associated information corresponding to all task instructions.

6. The wireless tongue electromyography biofeedback device for swallowing disorder assessment and training according to claim 5, characterized in that, The electromyographic features include time-domain features, frequency-domain features, nonlinear features, and multi-channel spatial features; the motion features include posture features, kinematic features, dynamic features, and stability features; the pressure features include intensity features, temporal features, distribution features, and dynamic features. The dual-modal fusion features include electromyographic-mechanical fusion features, electromyographic-motor fusion features, and pressure-motor fusion features; the trimodal fusion features include electromyographic-motor-pressure fusion features; the associated information includes confidence scores, multi-class probability distributions, temporal information, and quality labels.

7. The wireless tongue electromyography biofeedback device for swallowing disorder assessment and training according to claim 5, characterized in that, The process of generating an evaluation report based on the specific motion intentions and associated information corresponding to all task instructions includes: All task instructions and their corresponding specific motion intentions and associated information are reorganized into a task matrix; The reconstructed task matrix is ​​obtained by filtering out specific motion intentions and related information in the task matrix whose confidence scores are lower than the confidence threshold; The recombined task matrix is ​​calculated to obtain multidimensional features in multiple dimensions, and the multidimensional features are aggregated into a multidimensional clinical function score. The multidimensional clinical function scores are used to assess swallowing risk and predict rehabilitation potential, and an assessment report is generated based on the assessment and prediction results.

8. The wireless tongue electromyography biofeedback device for swallowing disorder assessment and training according to claim 7, characterized in that, The process of assessing swallowing risk and predicting rehabilitation potential using the multidimensional clinical function scores, and generating an assessment report based on the assessment and prediction results, includes: The baseline risk score is obtained by weighting the multidimensional clinical function scores. The patient's basic information is used as a clinical correction factor to correct the baseline risk score to obtain the final score, and the swallowing risk level is determined based on the range of the final score. Risk factors are analyzed based on the multidimensional clinical function scores, and a risk heatmap is generated based on the swallowing risk level and risk factors. The rehabilitation potential factor is calculated using the multidimensional clinical function score, and the potential cycle is predicted based on the rehabilitation potential factor. Potential levels are set based on the predicted potential cycle, and bottleneck factors that restrict recovery are identified. Record the swallowing risk level, risk heat map, predicted potential cycle, potential level, and bottleneck factors in the assessment report.

9. The wireless tongue electromyography biofeedback device for swallowing disorder assessment and training according to claim 1, characterized in that, The step of generating a personalized training plan based on the assessment report, enabling the patient to initiate the personalized training plan via the handheld terminal to complete daily tongue muscle training, includes: The risk level and potential prediction in the assessment report are analyzed, and specific functional impairments are mapped to targeted digital training games through the built-in clinical rule base. Personalized intensity, number of sets and progression rules are set for each training task to form a personalized training plan. The personalized training plan is pushed to the therapist's review platform to complete the review and fine-tuning, thereby generating a review completion signal; The review completion signal and the personalized training plan are synchronized to the handheld terminal so that the patient can complete the daily training of the tongue muscles according to the personalized training plan.

10. A wireless tongue electromyographic biofeedback method for assessing and training dysphagia, characterized in that, Feedback is provided using the wireless tongue electromyography biofeedback device for swallowing disorder assessment and training as described in any one of claims 1 to 9.

Images