Post-stroke dysphagia rehabilitation recognition system, feedback stimulation equipment and storage medium

Abstract

The invention belongs to the technical field of brain-computer interfaces, and particularly relates to a post-stroke dysphagia rehabilitation recognition system, feedback stimulation equipment and a storage medium. The system comprises a signal acquisition module, an electroencephalogram signal preprocessing module, a neural network feature extraction module, a statistical feature extraction module and a swallowing intention recognition module, and can recognize whether a patient has a swallowing intention or not through a Mama deep learning model based on space-time statistical features. The invention also constructs brain-computer interface equipment for rehabilitation of dysphagia after stroke in combination with electroencephalogram signal acquisition equipment and electrical stimulation equipment. The technical scheme of the invention has the advantages of being accurate in recognition, good in rehabilitation treatment effect and capable of forming center-periphery closed loop feedback, and has a very good application prospect.

Description

Post-stroke dysphagia rehabilitation identification system, feedback stimulation device and storage medium Technical Field

[0001] This invention belongs to the field of brain-computer interface technology, specifically relating to a post-stroke dysphagia rehabilitation recognition system, a feedback stimulation device, and a storage medium. Background Technology

[0002] Dysphagia is a very common complication after stroke, and it has long been a global clinical challenge. Current treatments for post-stroke dysphagia mainly include postural adjustment, feeding training, balloon dilation, and electrical stimulation therapy. Among these, electrical stimulation therapy uses electrical currents to stimulate swallowing-related muscles and nerves, bypassing the damaged central nervous system control, directly activating peripheral muscle contractions, enhancing muscle strength and coordination, and simultaneously promoting the recovery of nerve tissue.

[0003] While these treatments can improve patients' swallowing function to some extent, they generally have a slow onset of action and require significant human and material resources to implement. For example, low-frequency electrical stimulation therapy requires specialized doctors to spend considerable time conducting a comprehensive evaluation of the patient before determining the location and timing of the electrical stimulation.

[0004] Furthermore, existing devices and technologies for swallowing disorder rehabilitation on the market mainly suffer from the following problems: Lack of personalization: Most devices cannot accurately identify the individual characteristics of different patients and cannot provide the most suitable treatment plan based on individual differences. Some traditional rehabilitation devices only provide fixed patterns of electrical stimulation, failing to consider the specificity of patients' electroencephalogram (EEG) signals, making it difficult to meet diverse treatment needs.

[0005] Limited signal processing precision: The devices are not precise enough in monitoring patients' physiological signals and cannot effectively extract key information closely related to swallowing function. Some devices struggle to accurately distinguish between valid signals and noise when processing EEG signals, leading to inaccurate judgment of the patient's swallowing intention and affecting treatment outcomes.

[0006] Lack of closed-loop control: Current technology has failed to fully establish an effective central nervous system-peripheral nervous system-muscle control closed loop, and cannot adjust treatment strategies in real time based on the patient's neural signal feedback, which greatly reduces the pertinence and effectiveness of rehabilitation training.

[0007] Inadequate intelligent monitoring and assessment: The lack of a comprehensive intelligent monitoring and assessment system prevents real-time tracking of patient rehabilitation progress and timely optimization of training effects. Many devices fail to comprehensively collect and analyze patient data during treatment, making it difficult to accurately assess treatment effectiveness and thus hindering timely adjustments to treatment parameters to achieve optimal rehabilitation results. Summary of the Invention

[0008] To address the problems of existing technologies, this invention provides a post-stroke dysphagia rehabilitation identification system, a feedback stimulation device, and a storage medium.

[0009] A post-stroke dysphagia rehabilitation and identification system includes: a signal acquisition module configured to acquire electroencephalogram (EEG) signals via a brain-computer interface device; an EEG signal preprocessing module configured to preprocess the EEG signals in real time to reduce interference signals; a neural network feature extraction module configured to extract neural network features from the preprocessed EEG signals using a convolutional neural network; a statistical feature extraction module configured to extract statistical features from the preprocessed EEG signals by solving for the upper bound, lower bound, and mean of the signals; and a swallowing intention recognition module configured to input the neural network features and statistical features into a Mamba deep learning model in real time to identify whether the patient has a swallowing intention.

[0010] Preferably, the acquisition points of the EEG signal include: 8-11 cm lateral to Cz point, 2-4 cm anterior to Cz point, on both sides of the occipital protuberance with Oz point as a reference, below Oz point, C3, C4, Fz, F3, F4, T3, T4, Pz, P3, P4, Oz, O1, O2; and / or, the sampling rate of the EEG signal is 256 Hz-1024 Hz; the structure of the Mamba deep learning model includes: 1) a first two-dimensional convolutional layer, a first batch of normalized layers, a first depthwise separable two-dimensional convolutional layer, a second batch of normalized layers, a second two-dimensional convolutional layer, a first average pooling layer, a third two-dimensional convolutional layer, a third batch of normalized layers, a second depthwise separable two-dimensional convolutional layer, an exponential linear unit, a second average pooling layer, a first splicing layer, a fourth two-dimensional convolutional layer, a Mamba module, an RNN network, a flattening layer, and a softmax layer, all connected in sequence. The function and output layer; 2) also includes a global statistical feature branch, which includes: simultaneously introducing the output of the first average pooling layer into three statistical branches, taking the minimum, average and maximum values ​​in the spatial dimension respectively, and passing the three vectors obtained through the second concatenation layer and the third average pooling layer in sequence; the output of the third average pooling layer and the output of the second average pooling layer are concatenated through the first concatenation layer.

[0011] Preferably, the preprocessing includes: performing 0.1-60Hz bandpass filtering, 50Hz notch filtering, and removing baseline drift.

[0012] Preferably, the formula for calculating the statistical characteristics is: in, For statistical characteristics, The EEG signal after GAN data enhancement. It is to find the upper bound of the signal. It is to find the lower bound of the signal. This involves calculating the weighted mean of the signal; the formula for extracting and updating the neural network features is as follows: It is the current state quantity. Let A be the state variables at time step t-1, A be the state transition matrix, and B be the influence of the control variables on the state variables. yes and raw EEG signals Projected onto a higher-dimensional vector.

[0013] Preferably, it further includes: an EEG signal data generation module, configured to generate EEG signal training data enhanced and augmented by an adversarial network; and a training parameter update module, configured to train the optimal parameters of the Mamba deep learning model.

[0014] Preferably, the training parameter update module combines historical usage data and doctor feedback to periodically optimize and update the deep learning model.

[0015] Preferably, it also includes a feedback stimulation module configured to send a control signal to the rehabilitation device when the patient’s swallowing intention is detected.

[0016] The present invention also provides a post-stroke dysphagia rehabilitation feedback stimulation device, comprising: an electroencephalogram (EEG) signal acquisition device configured to acquire the patient's EEG signals; a control unit configured to integrate the above-mentioned post-stroke dysphagia rehabilitation recognition system and control the electrical stimulation potential and intensity of the electrical stimulation device; and an electrical stimulation device configured to receive the control signal from the control unit and apply electrical stimulation to the patient.

[0017] Preferably, the electrical stimulation device applies electrical stimulation in the following mode: during the oral preparation phase, the electrical stimulation is focused on the buccinator muscle; during the pharyngeal phase, the electrical stimulation is focused on the pharyngeal constrictor muscle, pharyngeal levator muscle, mylohyoid muscle, and thyrohyoid muscle.

[0018] The present invention also provides a computer-readable storage medium having stored thereon a computer program for implementing the above-described post-stroke dysphagia rehabilitation system.

[0019] This invention provides a brain-computer interface rehabilitation device for post-stroke dysphagia that applies deep learning-driven personalized EEG biomarker recognition. Its working principle is as follows: A deep learning model identifies EEG signal features associated with swallowing function, extracts personalized EEG biomarkers, and determines in real-time whether the patient has swallowing intent. Based on the prediction results of the deep learning model, it controls the location and pattern of electrical stimulation applied to the patient's swallowing muscles, thereby effectively improving post-stroke dysphagia. The technical solution of this invention achieves the following beneficial technical effects: 1. Addressing the problem of insufficient personalization in existing dysphagia rehabilitation devices, this invention extracts personalized EEG biomarkers from the patient's EEG signals and controls the electrical stimulation signal based on the prediction results of the deep learning model, enabling the control of the electrical stimulation signal according to the individual differences of each patient.

[0020] 2. This invention optimizes the data preprocessing method and the model training data processing method, and improves the deep learning model algorithm, so as to make the signal processing more accurate and the judgment of the patient's swallowing intention more accurate, which is conducive to improving the rehabilitation treatment effect.

[0021] 3. This invention improves the feature extraction network structure of the model by designing a statistical feature extraction branch, and inputs two types of features, namely neural network features and statistical features, into the Mamba deep learning model. By using the two types of features together to identify swallowing intention, it has better recognition accuracy.

[0022] 4. This invention constructs a rehabilitation system with a closed loop of central nervous system-peripheral nerve-muscle control, which has better therapeutic effects.

[0023] 5. The rehabilitation device of this invention can detect the patient's electroencephalogram (EEG) signals in real time and determine whether the patient has a swallowing intention. Simultaneously, by comprehensively collecting and analyzing patient data, this invention can optimize the training model and adjust treatment parameters in a timely manner to achieve the best rehabilitation effect.

[0024] Obviously, based on the above description of the present invention, and according to common technical knowledge and conventional methods in the field, various other modifications, substitutions or alterations can be made without departing from the basic technical concept of the present invention.

[0025] The following detailed embodiments further illustrate the above-described content of the present invention. However, this should not be construed as limiting the scope of the present invention to the following examples. All technologies implemented based on the above-described content of the present invention fall within the scope of the present invention. Attached Figure Description

[0026] Figure 1 is a schematic diagram of the data processing flow of the Mamba deep learning model in Example 1, which involves acquiring signals, processing signals, determining intent, and performing control.

[0027] Figure 2 is a schematic diagram of the specific design of the model in Example 1. The feature extraction network structure of the deep learning model is improved by designing a statistical feature extraction branch, as shown in the dashed box in the figure.

[0028] Figure 3 is a schematic diagram of the data processing flow during the training and updating process of the Mamba deep learning model in Example 1. Detailed Implementation

[0029] It should be noted that the algorithms for data acquisition, transmission, storage and processing steps not specifically described in the embodiments, as well as the hardware structures and circuit connections not specifically described, can all be implemented using content already disclosed in the prior art.

[0030] Example 1: Post-Stroke Swallowing Dysphagia Rehabilitation and Recognition System This example provides a system for real-time determination of whether a post-stroke dysphagia patient has a swallowing intention. It includes: a signal acquisition module configured to acquire electroencephalogram (EEG) signals via a brain-computer interface device; an EEG signal preprocessing module configured to preprocess EEG signals in real-time to reduce interference signals; a neural network feature extraction module configured to extract neural network features from the preprocessed EEG signals using a convolutional neural network; a statistical feature extraction module configured to extract statistical features from the preprocessed EEG signals by solving for the upper bound, lower bound, and mean of the signal; and a swallowing intention recognition module. The module is configured to input the neural network features and statistical features into the Mamba deep learning model in real time to identify whether the patient has a swallowing intention; the training parameter update module is configured to train the optimal parameters of the Mamba deep learning model; in addition, the training module can combine historical usage data and doctor feedback to periodically optimize the deep learning model (as shown in Figure 3); the EEG signal data generation module is configured to enhance and augment the EEG signals in the training data by generating a Generative Adversarial Network (GAN); the feedback stimulation module is configured to send a control signal to the rehabilitation device when the patient's swallowing intention is detected.

[0031] In the signal acquisition and preprocessing stage, the sampling rate was first set to 1000Hz to ensure sufficient frequency of EEG signal data could be acquired. Next, a 0.1-60Hz bandpass filter was applied; this filtering range effectively preserves the effective frequency bands in the EEG signal while removing low-frequency and high-frequency noise interference. Simultaneously, a 50Hz notch filter was used to suppress power line interference, further improving signal quality. Finally, baseline drift was removed to ensure the stability of the acquired EEG signal baseline.

[0032] As shown in Figure 1, in this embodiment, combining neural network features with statistical features to learn the temporal features of EEG signals can not only better capture the temporal patterns of EEG signals under different individuals or different experimental conditions, but also improve the interpretability of the model, accelerate the convergence speed of the model, and thus more robustly capture the temporal changes of EEG signals, thereby improving the learning effect of EEG signal temporal features.

[0033] In the specific data processing workflow, it is assumed that the acquired raw EEG signal is Where t represents the time series, It is a vector sequence that changes over time, with each time step t corresponding to a vector dimension of d.

[0034] First, during the training phase, the signal set after GAN data augmentation is... Dividing the data into time series segments, each time step inputs an EEG vector containing swallowing-related features. Train the Mamba deep learning model based on statistical features.

[0035] After the EEG vector is input into the network, the statistical features are first calculated: in, It is to find the upper bound of the signal. It is to find the lower bound of the signal. It calculates the weighted average of the signals, where the weights are learned by the network. In the model, the inputs are... and And projected to higher dimensions .

[0036] The neural network features are then obtained, and the expressions for calculating and updating these features are as follows: in, It is the state quantity at time step t. Let A be the state variables at time step t-1, A be the state transition matrix, and B be the influence of the control variables on the state variables. The output at the current time step is... in, This is the system's output at time step t, i.e., the result of the swallowing intention judgment. C represents the influence of the current state variable on the output. D represents the influence of the current control variable on the output, which is a learnable vector that provides additional linear correction to the output. The final output result... Projecting back to a lower dimension is denoted as . It records the temporal information of past swallowing-related EEG signals, thereby capturing the dynamic changes of EEG signals before, during, and after swallowing.

[0037] The output layer receives the output of the hidden layer at the last time step T. This is mapped to the final output dimension through a fully connected layer. Let the weight matrix of the output layer be... , dimension , where o is the output dimension (binary classification prediction of whether swallowing disorder exists, o = 2), and the bias vector is If the dimension is 0, then the formula for calculating the output y is: in The function transforms the output into a probability distribution, such that the output y represents the probability of belonging to each category (having swallowing intention or not). A swallowing probability threshold can be determined through training; exceeding the threshold indicates a swallowing disorder.

[0038] During training, the gradient is calculated using the backpropagation through-time (BPTT) algorithm. Let the true swallowing intention label be yi. The cross-entropy loss function L is used to measure the difference between the predicted result yi and the true label yi. The difference between them is expressed by the formula: The BPTT algorithm is used to backpropagate the error of the output layer to the hidden and input layers at each time step, calculating the contribution of each network parameter to the error. Specifically, for the weight matrix... , , and bias vector , They are then calculated using the chain rule to determine their gradients with respect to the loss function (L). The gradient is calculated as follows: in, Calculated based on the derivative formula of the cross-entropy loss function, according to Calculate the derivative of a function using the formula. .

[0039] As shown in Figure 2, the Mamba deep learning model includes a Mamba module and an RNN network. Specifically, the structure of the Mamba deep learning model includes: 1) a first two-dimensional convolutional layer, a first batch of normalization layers, a first depthwise separable two-dimensional convolutional layer, a second batch of normalization layers, a second two-dimensional convolutional layer, a first average pooling layer, a third two-dimensional convolutional layer, a third batch of normalization layers, a second depthwise separable two-dimensional convolutional layer, an exponential linear unit (ELU), a second average pooling layer, a first concatenation layer, a fourth two-dimensional convolutional layer, a Mamba module, an RNN network, a flattening layer, a softmax function, and an output layer, all connected in sequence.

[0040] 2) It also includes a global statistical feature branch, which includes: simultaneously introducing the output of the first average pooling layer into three statistical branches, taking the minimum (MIN), average (MEAN), and maximum (MAX) values ​​in the spatial dimension respectively, and passing the three vectors obtained through the second concatenation layer and the third average pooling layer in sequence; the output of the third average pooling layer and the output of the second average pooling layer are concatenated through the first concatenation layer.

[0041] The specific parameters for each layer or module are as follows, where the parameters [x, y] correspond to the parameters of the first and second dimensions, respectively: Based on the calculated gradient information, stochastic gradient descent (SGD) or other optimization algorithms are used to adjust the RNN network parameters to minimize the loss function (L). In each iteration, the formula for updating the parameters is as follows ( ): in The learning rate controls the step size of parameter updates. Through multiple iterations of training, the RNN gradually learns the EEG signal patterns related to swallowing, improving recognition accuracy.

[0042] Through the aforementioned data flow and training mechanism, EEG signals can be effectively processed to achieve accurate identification and prediction of swallowing intentions.

[0043] Example 2: Post-stroke Swallowing Dysfunction Rehabilitation Feedback Stimulation Device. This example provides a post-stroke swallowing dysphagia rehabilitation brain-computer interface device based on Example 1. It includes: an EEG signal acquisition device configured to acquire the patient's EEG signals; a control unit configured to integrate the post-stroke swallowing dysphagia rehabilitation recognition system of Example 1 and control the electrical stimulation potential and intensity of the electrical stimulation device; and an electrical stimulation device configured to receive the control signal from the control unit and apply electrical stimulation to the patient.

[0044] In this rehabilitation device, the EEG signal acquisition equipment uses a 10-20 acquisition system to focus on acquiring EEG signals from the brain regions associated with motor imagery. This system can accurately locate brain regions related to motor imagery, ensuring that the acquired signals are targeted and effective. In practical application, this embodiment uses the ANT Neuro EEG acquisition equipment, which has high accuracy and stability in EEG signal acquisition and can meet the requirements of this device. Acquisition points: 8-11cm lateral to Cz point, 2-4cm anterior to it; with Oz point as a reference, at appropriate distances to the sides and below the occipital protuberance; C3, C4, Fz, F3, F4, T3, T4, Pz, P3, P4, Oz, O1, O2.

[0045] In the signal acquisition and preprocessing stage, the sampling rate was first set to 1000Hz to ensure sufficient frequency of EEG signal data could be acquired. Next, a 0.1-60Hz bandpass filter was applied; this filtering range effectively preserves the effective frequency bands in the EEG signal while removing low-frequency and high-frequency noise interference. Simultaneously, a 50Hz notch filter was used to suppress power line interference, further improving signal quality. Finally, baseline drift was removed to ensure the stability of the acquired EEG signal baseline.

[0046] In this rehabilitation device, the electrical stimulation equipment has a dedicated interface. After the deep learning model analyzes and processes the electroencephalogram (EEG) signals, accurately identifies EEG signal patterns related to swallowing function, and determines the swallowing intention, it generates corresponding control signals. These control signals are transmitted from the control unit containing the deep learning model to the electrical stimulation device via transmission lines, adjusting the location and intensity of the electrical stimulation based on prior assessments and predictions. The electrical stimulation device is designed for different stages of the swallowing process, especially the oral preparation phase, the pharyngeal phase, and subsequent stages.

[0047] During the oral preparation phase, which is the stage when food is chewed in the mouth to form a bolus, the operation of the electrical stimulation device focuses on the cheek muscles, causing depolarization of the muscle fibers, thereby triggering muscle contraction.

[0048] During the pharyngeal phase, which is the complex process of transporting a food bolus from the pharynx to the esophagus via the swallowing reflex, the operation of the electrical stimulation device primarily targets the pharyngeal constrictor muscles, pharyngeal levator muscles, mylohyoid muscles, and thyrohyoid muscles. During stimulation, electrical pulses stimulate the nerve endings of the pharyngeal levator muscles, triggering muscle contraction, thereby elevating the larynx and pharynx. This, in conjunction with the contraction of the pharyngeal constrictor muscles, ensures the smooth progress of swallowing.

[0049] As can be seen from the above embodiments, this invention provides a brain-computer interface rehabilitation system and device that applies deep learning-driven personalized EEG biomarker recognition to post-stroke dysphagia. Given the high incidence of post-stroke dysphagia and the problems of slow onset of action, reliance on human and material resources, and failure to establish an effective control loop in existing treatment programs, this device first uses a deep learning model to identify EEG signal features associated with swallowing function to extract personalized EEG biomarkers, determining whether the patient has a swallowing intention. These biomarkers are then applied to the brain-computer interface device to convert them into electrical stimulation applied to the patient's swallowing muscles. This device combines self-supervised learning and generative adversarial networks to enhance model capabilities, using recurrent neural networks to track changes in EEG signals in real time and mine personalized patterns, thereby accurately determining swallowing intentions for different patients and providing a basis for individualized intervention. Its function is to treat post-stroke dysphagia by providing brain-computer interface-based swallowing training stimulation through the combination of deep learning algorithms and real-time EEG signal monitoring. It also features an intelligent monitoring and evaluation system to track the rehabilitation process in real time and optimize training effects. Therefore, this invention has excellent application prospects in the rehabilitation treatment of post-stroke dysphagia.

Claims

1. A rehabilitation and identification system for post-stroke dysphagia, characterized in that, include: The signal acquisition module is configured to acquire electroencephalogram (EEG) signals via a brain-computer interface device. The EEG signal preprocessing module is configured to preprocess EEG signals in real time to reduce interference signals; The neural network feature extraction module is configured to extract neural network features from the preprocessed EEG signals through a convolutional neural network. The statistical feature extraction module is configured to extract statistical features from the preprocessed EEG signal by solving for the upper bound, lower bound, and mean of the signal; the swallowing intention recognition module is configured to input the neural network features and statistical features into the Mamba deep learning model in real time to identify whether the patient has a swallowing intention. The structure of the Mamba deep learning model includes: 1) a first two-dimensional convolutional layer, a first batch of normalization layers, a first depthwise separable two-dimensional convolutional layer, a second batch of normalization layers, a second two-dimensional convolutional layer, a first average pooling layer, a third two-dimensional convolutional layer, a third batch of normalization layers, a second depthwise separable two-dimensional convolutional layer, an exponential linear unit, a second average pooling layer, a first concatenation layer, a fourth two-dimensional convolutional layer, a Mamba module, an RNN network, a flattening layer, a softmax function, and an output layer; 2) a global statistical feature branch, which includes: simultaneously introducing the output of the first average pooling layer into three statistical branches, taking the minimum, average, and maximum values ​​in the spatial dimension respectively, and passing the resulting three vectors sequentially through the second concatenation layer and the third average pooling layer; the output of the third average pooling layer and the output of the second average pooling layer are concatenated through the first concatenation layer.

2. The post-stroke dysphagia rehabilitation and identification system according to claim 1, characterized in that: The acquisition points for the EEG signals include: 8-11 cm lateral to Cz point, 2-4 cm anterior to Cz point, both sides of the occipital protuberance with Oz point as a reference, below Oz point, C3, C4, Fz, F3, F4, T3, T4, Pz, P3, P4, Oz, O1, and O2; and / or, the sampling rate of the EEG signals is 256 Hz-1024 Hz; and / or, the preprocessing includes: 0.1-60 Hz bandpass filtering, 50 Hz notch filtering, and baseline drift removal.

3. The post-stroke dysphagia rehabilitation and identification system according to claim 1, characterized in that: The formula for calculating the statistical characteristics is as follows: in, For statistical characteristics, The EEG signal after GAN data enhancement. It is to find the upper bound of the signal. It is to find the lower bound of the signal. This involves calculating the weighted mean of the signal; the formula for extracting and updating the neural network features is as follows: It is the current state quantity. Let A be the state variables at time step t-1, A be the state transition matrix, and B be the influence of the control variables on the state variables. yes and raw EEG signals Projected onto a high-dimensional vector.

4. The post-stroke dysphagia rehabilitation and identification system according to claim 1, characterized in that: Also includes: The EEG signal data generation module is configured to generate EEG signal training data that enhances and augments adversarial networks. The training parameter update module is configured to train the optimal parameters for the Mamba deep learning model.

5. The post-stroke dysphagia rehabilitation and identification system according to the model of claim 4, characterized in that: The training parameter update module combines historical usage data and doctor feedback to regularly optimize and update the deep learning model.

6. The post-stroke dysphagia rehabilitation and identification system according to claim 1, characterized in that: Also includes: The feedback stimulation module is configured to send a control signal to the rehabilitation device when it detects that the patient has the intention to swallow.

7. A feedback stimulation device for post-stroke dysphagia rehabilitation, characterized in that, include: The electroencephalogram (EEG) signal acquisition device is configured to acquire the patient's EEG signals; The control unit is configured to integrate the post-stroke dysphagia rehabilitation identification system according to any one of claims 1-6, and to control the electrical stimulation potential and intensity of the electrical stimulation device; the electrical stimulation device is configured to receive the control signal from the control unit and apply electrical stimulation to the patient.

8. The post-stroke dysphagia rehabilitation feedback stimulation device according to claim 7, characterized in that: The electrical stimulation device applies electrical stimulation in the following pattern: during the oral preparation phase, the electrical stimulation is focused on the buccinator muscle; during the pharyngeal phase, the electrical stimulation is focused on the pharyngeal constrictor muscle, pharyngeal levator muscle, mylohyoid muscle, and thyrohyoid muscle.

9. A computer-readable storage medium, characterized in that, It stores a computer program for implementing the post-stroke dysphagia rehabilitation system according to any one of claims 1-6.

Images