Auxiliary scheme generation method and system for dysphagia rehabilitation

Abstract

The invention discloses an auxiliary scheme generation method and system for dysphagia rehabilitation, belongs to the technical field of medical rehabilitation intelligent evaluation, and is used for solving the problem of poor optimization of a swallowing training path. According to the method, myoelectricity, tongue pressure, throat images and acoustic signals in the swallowing process are collected through a multi-channel sensing device, a unified multi-mode swallowing behavior feature vector is constructed, obstacle stage recognition and functional grade judgment are achieved in combination with a pre-training model and historical data, and a structured evaluation result is output; matching training actions with high adaptation degree based on the knowledge graph, fusing muscle group matching degree, edge weight and function level factors to generate a personalized training path, and setting strength, frequency and rhythm parameters; in the training process, the motion and myoelectricity deviation is monitored in real time, a prompt signal is output, and the behavior completion degree is recorded; and the rehabilitation progress map is dynamically updated according to feedback, and if the deviation exceeds the limit, the path is automatically reconstructed and doctor intervention is synchronized, so that the rehabilitation efficiency and safety are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical rehabilitation intelligent assessment, and more specifically, the present invention relates to a method and system for generating an auxiliary plan for dysphagia rehabilitation. Background Art

[0002] Swallowing disorder is a widespread functional disorder among patients with common diseases such as Parkinson's disease, postoperative head and neck surgery, and senile degenerative diseases. The rehabilitation process usually relies on professional assessors to judge the degree of the disorder through observation, scale scoring, etc., and to develop rehabilitation training plans based on clinical experience. Existing rehabilitation methods mostly use standardized movement templates for training, such as effortful swallowing, tongue propulsion, and glottis closure.

[0003] Deficiencies in existing technologies: Existing dysphagia rehabilitation methods generally rely on static templates or empirical rules to match training movements during the development of training pathways. These methods lack a structured understanding and computational representation of patients' real-time assessment results, functional level labels, and muscle group restriction information. This makes it difficult to optimize training movements and configure parameters based on multidimensional adaptation factors, resulting in a lack of personalized adaptability in training pathways. This is particularly true for patients with varying impairment locations, severity levels, and varying compliance. Current methods are unable to dynamically adjust the sequencing of training movements or reconstruct movement combinations, and lack a precise matching mechanism for training intensity, frequency, and rhythm parameters. This impacts the targeted and sustained effectiveness of the rehabilitation process, limiting the advancement of intelligent and refined swallowing rehabilitation. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method and system for generating an auxiliary program for dysphagia rehabilitation, so as to solve the problem of poor optimization of the swallowing training path in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions: A method for generating an auxiliary plan for dysphagia rehabilitation comprises the following steps: Based on the multi-channel sensor equipment, myoelectric signals, tongue pressure data, laryngeal movement images and speech features during swallowing are collected to construct a multimodal swallowing behavior feature vector and uniformly map it to a standard time frame sequence. Combining a pre-trained swallowing function assessment model with historical rehabilitation data, it performs impairment stage identification and functional level determination, and outputs structured assessment results that include labels for impairment type, location, and severity. Based on the evaluation results, the relevant training action nodes in the rehabilitation knowledge graph are matched, the fitness weights of the training actions are calculated, and a ranking list of actions is generated to build a personalized training path with intensity, frequency, and rhythm parameters. During the training process, interactive monitoring is carried out synchronously. Based on real-time motion capture and electromyographic response recognition algorithms, training deviations are identified and prompt signals are output. At the same time, behavioral completion and deviation level indicators are recorded. The rehabilitation progress map is updated based on training feedback records and stage-by-stage scores. If deviation levels are continuously identified as exceeding the limit, the training path is automatically reconstructed and synchronized with the doctor's terminal to assist in strategy optimization and intervention.

[0006] In a preferred embodiment, a multi-channel sensor device collects myoelectric signals, tongue pressure data, laryngeal movement images and speech features during swallowing, constructs a multimodal swallowing behavior feature vector, and uniformly maps it to a standard time frame sequence. The specific process is as follows: The surface electromyography electrode array is used to obtain the time-series electromyographic voltage signals generated by the muscle groups located in the mandibular, laryngeal and sublingual areas during swallowing, and record them as electromyographic signal data; The tongue pressure sensor pad is used to obtain the unit pressure change value of the contact area between the tongue and the upper palate of the mouth, and the pressure duration is recorded synchronously as tongue pressure data; An infrared visual camera is used to obtain the vertical movement trajectory of the Adam's apple during swallowing, and the speed and amplitude of the movement are calculated and recorded as laryngeal movement image data. Use a close-range microphone to collect acoustic response signals during swallowing, extract the onset time, peak frequency, and duration, and record them as speech feature data; Four types of heterogeneous sensor data, namely electromyographic signal data, tongue pressure data, laryngeal movement image data, and speech feature data, are uniformly mapped to a standardized time frame structure with the start of the swallowing event as the reference point to form a swallowing behavior feature vector set.

[0007] In a preferred embodiment, a pre-trained swallowing function assessment model is combined with historical rehabilitation data to perform impairment stage identification and functional level discrimination, and output a structured assessment result containing labels of impairment type, location, and severity. The specific process is as follows: The constructed swallowing behavior feature vector is input into a pre-trained swallowing function assessment model, which includes a convolutional structure for extracting static features and a recurrent neural network structure for modeling temporal changes; An attention mechanism was used to enhance the responsiveness of key frame segments of a swallowing function assessment model, identifying feature differences during swallow initiation, laryngeal elevation, and contraction recovery. After the swallowing function assessment model processes the input swallowing behavior feature vector, it outputs a set of structured prediction results. The structured prediction results include classification probability predictions of the obstruction stage. In the output structured prediction results, the obstruction stages involved in the swallowing process are identified, including the oral preparation stage, the pharyngeal stage, and the esophageal transit stage, and the corresponding obstruction type labels are output. The prediction results of the swallowing function assessment model are matched with samples with clear functional level labels in the historical rehabilitation database, and the range of the current functional level is confirmed by similarity measurement method; Combining the recognition results of the swallowing function assessment model with the historical score mapping relationship, a structured assessment result is generated, which includes the type of obstacle, the main restricted area and the severity level label.

[0008] In a preferred embodiment, the prediction results of the swallowing function assessment model are matched with samples with clear functional level labels in the historical rehabilitation database, and the range of the current functional level is confirmed by a similarity measurement method. The specific process is as follows: A standard sample set containing functional grade labels is extracted from a historical rehabilitation database, wherein each sample in the standard sample set includes a swallowing feature vector of the same modality dimension as that of the current patient and its corresponding verified functional grade label; Performing feature normalization processing on the feature vector output by the current swallowing function evaluation model; The characteristic distance between the normalized current feature vector and each sample in the standard sample set is calculated in sequence. The characteristic distance calculation is based on the cumulative results of the numerical differences between each modality, including the difference in electromyographic response, the difference in tongue pressure peak value, the difference in laryngeal trajectory change, and the difference in acoustic duration. Generate a sample similarity ranking list based on all distance calculation results, and select at least two samples with similarity higher than a set threshold as reference samples; The distribution of corresponding functional level labels in the reference samples is statistically analyzed, and the functional level range of the current assessment object is determined based on the level label with the highest frequency, which is used as the final output severity judgment result.

[0009] In a preferred embodiment, the evaluation results are matched to the relevant training action nodes in the rehabilitation knowledge graph, the fitness weights of the training actions are calculated, and an action ranking list is generated. The specific process is as follows: Based on the structured assessment results, the obstacle type, restricted area, and severity level labels are extracted and input into the pre-built rehabilitation knowledge graph as a set of nodes to be matched; The rehabilitation knowledge graph is a heterogeneous graph structure that contains edges between training action nodes and obstacle type nodes. The edge weights represent the degree of adaptability of the action to a specific obstacle. Starting from the obstacle type node in the rehabilitation knowledge graph, a breadth-first search is performed to traverse the adjacent action nodes and extract all directly or indirectly connected candidate training action nodes; For each candidate training action node, the fitness weight between the candidate training action node and the evaluation result is calculated. The fitness weight is determined by the edge weight, the matching degree of the muscle group targeted by the action, and the fitness factor of the patient's current functional level. All candidate training action nodes are sorted from high to low according to the fitness weight, and an action sorting list is generated as the input candidate set for training path construction.

[0010] In a preferred embodiment, for each candidate training action node, the fitness weight between the candidate training action node and the evaluation result is calculated. The fitness weight is determined by the edge weight, the matching degree of the muscle group targeted by the action, and the fitness factor of the patient's current functional level. The specific process is as follows: Obtain the main restricted muscle group information from the evaluation results, compare it with the target muscle groups involved in the candidate training action nodes, and calculate the muscle group matching index. The muscle group matching index is quantified based on the overlap of the target muscle groups, with complete coverage being recorded as 1, overlap being recorded as 0.5, and no overlap being recorded as 0; Extract the edge weights between candidate training action nodes and evaluation nodes in the rehabilitation knowledge graph. The edge weights represent the effectiveness of the training action for this type of disorder in clinical experience and are preset during graph construction and normalized to the interval [0, 1]. According to the patient's functional level label in the evaluation results, the range of movement intensity allowed by the current level is extracted and compared with the training intensity label of the candidate movement. If the intensity matches, the adaptation factor is assigned to 1, otherwise the adaptation factor value is reduced according to the decreasing function; The edge weights, muscle group matching indexes, and functional level adaptation factors are normalized item by item and combined in product form to obtain the fitness weights of the candidate movements. The fitness weights are used to represent the comprehensive matching priority of the training movements under the current patient condition.

[0011] In a preferred embodiment, a personalized training path including intensity, frequency, and rhythm parameters is constructed as follows: Selecting a set of training action nodes whose fitness weights are higher than a preset threshold from the fitness weight sorted list as a candidate action set for the training path; For each training action node, extract the parameter template associated with it in the knowledge graph. The parameter template includes the recommended intensity parameter, recommended frequency parameter, and standard rhythm parameter of the training action. Match the functional level label in the evaluation result with the training level requirement of the candidate training action. If the current patient level is lower than the required level of the training action, the training action will be removed from the candidate set. According to the patient's historical compliance records and stage recovery curve, the recommended intensity parameters, recommended frequency parameters and standard rhythm parameters of the candidate training movements are adjusted to form the target execution parameters for the current individual state; According to the fitness weight order of the training action nodes, the training actions with set execution parameters are arranged in sequence and combined into a personalized training path containing at least two training action nodes. Each action node in the personalized training path contains clear intensity parameters, frequency parameters and rhythm parameters.

[0012] In a preferred embodiment, interactive monitoring is performed synchronously during the training process. Based on real-time motion capture and electromyographic response recognition algorithms, training deviations are identified and prompt signals are output. At the same time, behavioral completion and deviation level indicators are recorded. The specific process is as follows: At the beginning of the training action, the visual acquisition channel and the electromyographic signal acquisition channel are activated to obtain the action image sequence of the current training action and the corresponding electromyographic activation timing signal; Extract the displacement trajectories of key points of the throat and head based on the action image sequence, and perform a time sequence comparison with the standard action template of the training action to identify spatial deviation behaviors such as posture offset, start delay and insufficient amplitude; Analyze the onset, peak amplitude, and duration of muscle activation based on electromyographic signals and compare them with the muscle group activation target parameters preset for the training movement to identify electrophysiological deviations such as insufficient electromyographic response, delayed activation, or abnormal tension. If any of the identified spatial deviation behaviors or electrophysiological deviation behaviors exceeds the set threshold, a prompt signal output is triggered, which includes voice prompts, image prompts, and vibration prompts; After each training action is completed, the difference between the actual execution parameters of the training action and the target parameters is recorded to calculate the behavioral completion degree of the action; Based on the cumulative number of spatial deviation behaviors and electrophysiological deviation behaviors and the deviation intensity level, a deviation level index for the training action is generated and stored in the training behavior log as a reference for progress evaluation.

[0013] In a preferred embodiment, the rehabilitation progress map is updated based on the training feedback records and staged scores. If the deviation level exceeds the limit, the training path is automatically reconstructed and synchronized with the doctor's terminal to assist in strategy optimization and intervention. The specific process is as follows: After each training action is completed, the corresponding behavior completion and deviation level indicators are recorded in the training feedback record, and the feedback data sequence of the current training cycle is constructed in chronological order; At the end of the training cycle, a stage score is calculated based on the feedback data sequence. The stage score is generated based on the total number of training actions, the average completion rate of the behavior, and the deviation level distribution, and is used to reflect the quality of rehabilitation execution at the current stage. Based on the staged scores and historical score curves, the progress status of the corresponding time nodes in the rehabilitation progress map is updated. The rehabilitation progress map is a recovery trajectory structure constructed in a time series. Each node contains the stage score, action compliance label and deviation risk factor. Count the number of deviation level violations that occur during the current training cycle. If the deviation level indicators of multiple consecutive training movements exceed the preset threshold, the training path reconstruction mechanism will be triggered; In the training path reconstruction mechanism, the fitness weight sorting process is called and the fitness weights are sorted from high to low. Based on the sorting results, candidate training action nodes are re-screened and a new training action sorting list is constructed; The reconstructed training path and the current evaluation results are packaged and sent to the doctor's terminal. After receiving it, the doctor's terminal performs manual intervention, including adding or deleting training actions, adjusting parameter configurations, or modifying rhythm factors, and finally synchronizes the updates to the patient's execution end.

[0014] A system for generating an auxiliary plan for dysphagia rehabilitation, used to implement the above-mentioned method for generating an auxiliary plan for dysphagia rehabilitation, comprising: The swallowing data acquisition module is used to collect myoelectric signals, tongue pressure data, laryngeal movement images and speech features during swallowing based on multi-channel sensing equipment, construct multimodal swallowing behavior feature vectors, and uniformly map them to a standard time frame sequence; The obstacle stage identification module is used to combine the pre-trained swallowing function assessment model with historical rehabilitation data to perform obstacle stage identification and functional level judgment, and output structured assessment results including obstacle type, location and severity labels; The rehabilitation training path module is used to match relevant training action nodes in the rehabilitation knowledge graph based on the evaluation results, calculate the fitness weights of the training actions, generate a ranked list of actions, and construct a personalized training path with intensity, frequency, and rhythm parameters; The training monitoring module is used to conduct interactive monitoring during the training process. Based on real-time motion capture and electromyographic response recognition algorithms, it identifies training deviations and outputs prompt signals, while also recording behavioral completion and deviation level indicators. The rehabilitation training optimization module is used to update the rehabilitation progress map based on training feedback records and stage scores. If the deviation level exceeds the limit is continuously identified, the training path will be automatically reconstructed and synchronized with the doctor's terminal to assist in strategy optimization and intervention.

[0015] Technical effects and advantages of the present invention: The present invention uses a multi-channel sensing device to synchronously collect electromyographic signals, tongue pressure data, laryngeal movement images and acoustic response information during the swallowing process, and constructs a high-resolution multimodal swallowing behavior feature vector through unified time alignment and feature normalization processing, accurately capturing the physiological response characteristics of each key stage in the whole swallowing process; combining the pre-trained swallowing function assessment model with the historical rehabilitation database, it performs stage recognition and functional level discrimination of swallowing disorders, and outputs structured obstacle type, restricted area and severity label; on this basis, based on the knowledge graph reasoning mechanism, it matches the rehabilitation training action nodes with high adaptability, and integrates the edge The system uses weight, muscle group matching and functional level adaptation factors to build a personalized training path, and clearly configures core parameters such as intensity, frequency and rhythm to enhance the pertinence and compliance of the training plan. During the training process, it uses motion capture and electromyographic response to identify deviant behaviors, trigger real-time prompts and record the completion of the behavior and the deviation level, to achieve interactive monitoring and intelligent feedback throughout the process. Finally, the rehabilitation progress map is dynamically updated based on the training feedback. If the deviation exceeds the limit, the path reconstruction is triggered and the doctor is linked to optimize the strategy and intervene, thereby improving the scientificity, accuracy and intelligence of rehabilitation, and significantly improving rehabilitation efficiency and safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The present invention is a flowchart of a method for generating an auxiliary program for dysphagia rehabilitation.

[0017] Figure 2 This is a structural schematic diagram of a system for generating an auxiliary plan for dysphagia rehabilitation according to the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] Example 1: Figure 1 As shown, a method for generating an auxiliary program for dysphagia rehabilitation includes the following steps: Based on the multi-channel sensor equipment, myoelectric signals, tongue pressure data, laryngeal movement images and speech features during swallowing are collected to construct a multimodal swallowing behavior feature vector and uniformly map it to a standard time frame sequence. Combining a pre-trained swallowing function assessment model with historical rehabilitation data, it performs impairment stage identification and functional level determination, and outputs structured assessment results that include labels for impairment type, location, and severity. Based on the evaluation results, the relevant training action nodes in the rehabilitation knowledge graph are matched, the fitness weights of the training actions are calculated, and a ranking list of actions is generated to build a personalized training path with intensity, frequency, and rhythm parameters. During the training process, interactive monitoring is carried out synchronously. Based on real-time motion capture and electromyographic response recognition algorithms, training deviations are identified and prompt signals are output. At the same time, behavioral completion and deviation level indicators are recorded. The rehabilitation progress map is updated based on training feedback records and stage-by-stage scores. If deviation levels are continuously identified as exceeding the limit, the training path is automatically reconstructed and synchronized with the doctor's terminal to assist in strategy optimization and intervention.

[0020] Step 1: Based on the multi-channel sensing equipment, myoelectric signals, tongue pressure data, laryngeal movement images and speech features during swallowing are collected to construct a multimodal swallowing behavior feature vector and uniformly map it to a standard time frame sequence. The specific implementation is as follows: Surface electromyography (EMG) electrode arrays are placed on the patient's mandibular margin, upper larynx, and sublingual muscles. The electrodes are connected non-invasively by contacting the skin. A high-sensitivity dual-channel EMG preamplifier module is used to collect muscle contraction electrical signals during swallowing. The acquisition frequency is set to above 1kHz to ensure that rapid changes in electrical activity related to swallowing can be captured. The collected data is recorded in time series form, including the starting point, peak value, and attenuation trend of EMG activation. This data is used to reflect the initial drive of the tongue and the protection of the larynx, and is used as EMG signal data. A flexible tongue pressure sensing pad is placed on the upper palate, in contact with the back of the patient's tongue, sensing changes in contact pressure between the tongue and the roof of the mouth during swallowing. A high-precision strain gauge pressure sensing unit collects pressure changes per unit time in real time, and simultaneously records the duration of each pressure event. This allows the intensity and rhythmic characteristics of the tongue's propulsion process to be extracted and used as tongue pressure data. An infrared camera placed directly in front of the patient's neck captures the vertical movement trajectory of the Adam's apple during swallowing. A key point recognition algorithm between consecutive frames is used to extract the displacement of the Adam's apple. The movement velocity and maximum displacement amplitude are then calculated. This data captures the closing and opening process of the larynx, reflecting the coordination of the food bolus's movement through the esophageal entrance, and serves as laryngeal motion image data. A close-range microphone is placed on the side of the patient's neck, and the voice signal acquisition module captures the throat sounds associated with swallowing in real time. The captured content includes the starting point of the sound at the moment of swallowing, the peak frequency component, and the duration of the entire acoustic event. After filtering and noise reduction processing to eliminate environmental interference, the data is used to analyze the integrity and contraction rhythm of the glottal vibration and serve as voice feature data. After completing the collection of four types of heterogeneous sensor data, namely electromyographic signal data, tongue pressure data, laryngeal movement image data, and speech feature data, a standardized time axis is constructed with the swallowing initiation moment as a unified reference point, and the feature data of different modalities are mapped to a unified time frame structure. This structure aligns the original signals according to relative time, normalizes the length, and generates a fixed-length, multimodal joint vector for input into the subsequent evaluation model. The final swallowing behavior feature vector set contains the electromyographic response curve, tongue pressure peak sequence, laryngeal movement trajectory parameters and acoustic signal time-frequency characteristics, constituting a complete swallowing behavior feature expression.

[0021] Combining the pre-trained swallowing function assessment model with historical rehabilitation data, we perform impairment stage identification and functional level determination, and output structured assessment results that include labels for impairment type, location, and severity. The specific implementation is as follows: The swallowing behavior feature vector under the standard time frame structure is used as input and input into the swallowing function evaluation model. The swallowing evaluation model adopts an end-to-end neural network structure. The front end of the swallowing function evaluation model adopts a multi-channel convolution structure to extract the spatial distribution characteristics of electromyographic signals, tongue pressure changes, image trajectories and acoustic parameters respectively, and fuses the expressions of each modality through feature splicing in the middle layer. The back end of the model adopts a recurrent neural network structure (such as bidirectional gated recurrent unit Bi-GRU or LSTM) to model the changing trend of swallowing behavior in the time dimension and capture the dynamic evolution relationship between different action stages. The attention mechanism is introduced after the recurrent structure. The attention mechanism is based on The importance weight allocation strategy of time series performs weighted enhancement on the key frame segments in the input sequence. The key frame segments are delineated by labeled samples or template matching, specifically including: swallowing initiation period (corresponding to the first rise in electromyography and the appearance of tongue pressure), laryngeal elevation period (corresponding to the peak of vertical displacement of the laryngeal trajectory), and contraction recovery period (corresponding to electromyography attenuation and the retraction of the Adam's apple). The attention mechanism assigns higher weights to the feature frames in these time periods, so that the model focuses on learning the feature differences of this stage during training, thereby improving the accuracy of obstacle stage identification. The model output adopts a multi-classification structure to identify the obstacle stages in the swallowing process and output the corresponding obstacle type label. That is, after the swallowing function assessment model processes the input swallowing behavior feature vector, it outputs a set of structured prediction results. The structured prediction results include classification probability predictions of the obstruction stage. The obstruction stage involved in the swallowing process is identified in the output structured prediction results. The obstruction stage classification is based on the physiological process of swallowing movements, mainly including oral preparation stage obstruction (such as tongue propulsion weakness), pharyngeal stage obstruction (such as laryngeal elevation hysteresis) and esophageal transport stage obstruction (such as swallowing residue or delivery delay). The three types of obstruction labels are probabilistically judged through a soft classification strategy, and multi-label results can be output simultaneously to reflect mixed obstructions. The refined division of functional levels is to perform similarity matching between the feature vectors output by the current model and the standard sample set in the historical rehabilitation database. The standard sample set is constructed by rehabilitation centers or clinical institutions. Each sample has a complete functional level label (such as mild, moderate, and severe) and contains multimodal feature expressions consistent with the current model input dimension. The similarity calculation is based on measurement methods such as weighted Euclidean distance or Mahalanobis distance. The difference in electromyographic response, tongue pressure peak, laryngeal movement trajectory and acoustic characteristics are compared respectively to generate a ranked list. From this list, a reference sample set with a similarity higher than the set threshold is selected. Based on the frequency distribution of the functional level labels in the reference samples, the functional level range of the current assessment object is determined. The above recognition results are output in a structured manner to generate an evaluation report. The structured evaluation results include the following three types of fields: one is the obstacle type field, which reflects the type of identified obstacle stage; the second is the main restricted part field, which extracts the labels of the muscle groups involved in the restriction, such as the anterior tongue, pharyngeal constrictor muscles or glottis closure muscles; the third is the severity level field, which marks the functional level grading result of the current patient, which can be labeled according to a five-level or three-level system for subsequent training path generation.

[0022] Specifically, the key frame segment refers to the time segment corresponding to the three core action stages in the swallowing physiological process in the standardized time series covered by the entire swallowing behavior feature vector, including: The key frame segment of the swallowing initiation period is the time interval from the start of the mouth closure to the initial appearance of the tongue pressure peak. During this segment, the electromyographic signal shows the first wave of rise, the tongue pressure sensor detects obvious pressure changes, and is accompanied by the initial vibration of the glottis. This is the key triggering stage for the start of the swallowing movement. The key frame segment of the laryngeal elevation period corresponds to the stage of rapid rise of the Adam's apple. During this period, the laryngeal motion trajectory shows the maximum vertical displacement, the acceleration of the laryngeal key points in the visual image is significantly enhanced, and the electromyographic signal reaches the second wave peak. This is the main movement output stage during swallowing; The key frame segment of the contraction recovery period starts from the cessation of laryngeal movement and the recovery of downward movement, and ends when the electromyographic signal steadily decays. This segment is usually accompanied by muscle relaxation and rapid weakening of electrophysiological activity. It is the transition stage between the completion of the swallowing action and the preparation for the next swallowing. The start and end time points of each key frame segment can be determined by the following method: using the rate of change of the electromyographic signal, the derivative of the tongue pressure signal, the velocity threshold of the displacement of the key points of the larynx, and the energy mutation point in the acoustic signal as the basis for segmentation, presetting the segment boundary time ratio in the standard swallowing template, combining the event alignment benchmark dynamic mapping of the current collected data, and using timing alignment techniques such as dynamic time warping to find the optimal correspondence between the current sample and the standard template at the key feature points, and then calibrating the key frame segments.

[0023] The swallowing function assessment model prediction results are matched with samples with clear functional level labels in the historical rehabilitation database. The current functional level range is confirmed by similarity measurement method. More specific steps include: A standard sample set is extracted from a historical rehabilitation database. The sample set consists of a large number of historical patient data that have completed the rehabilitation process and have a clear functional level. Each standard sample contains a complete multimodal swallowing behavior feature vector, which is consistent with the feature extraction structure used for the current assessment subject, ensuring that the data dimensions and semantics are fully aligned. Each sample also has a verified functional level label. The label grading system can adopt a five-level system (such as level 1 to level 5) or a three-level system (mild, moderate, and severe), which is set according to actual clinical standards. The multimodal feature vectors output by the current swallowing function assessment model are preprocessed to ensure that their numerical range and unit scale are consistent with the standard sample set. The normalization process is set according to the physical quantity units of each modality: the maximum peak amplitude is used as the normalization reference for electromyographic signals, the upper pressure limit is used as the reference interval for tongue pressure data, the maximum displacement amplitude is used for laryngeal movement trajectories, and the acoustic features are normalized to the standard interval based on the duration and frequency values. The normalization process for each modality uses linear scaling mapping to ensure that all feature components fall into a unified numerical range, which can usually be set between [0, 1] or [-1, 1]. After normalization is completed, each sample data in the standard sample set is traversed in turn to calculate the feature distance between it and the current evaluation feature. The feature distance is a metric that comprehensively measures multimodal differences. The text description combination calculation is performed as follows: First, the difference terms of the electromyographic response are calculated, including the activation start time difference, peak amplitude difference, and duration difference. Second, the maximum pressure difference and pressure maintenance time difference in the tongue pressure data are extracted. Then, the maximum rising displacement and falling time in the motion trajectory of the key points of the larynx are compared to obtain the laryngeal trajectory difference. Finally, the start time offset, frequency bandwidth difference, and total duration difference of the acoustic features are analyzed. The multiple feature differences in each mode are weighted and accumulated to obtain the total difference of the mode. That is, the differences of all modes are added together to obtain the final multimodal feature distance. After the feature distance calculation is completed, a similarity-ranked list is generated based on the distance results of all samples. To screen out outliers, only samples with a similarity to the current evaluation object that exceeds a set threshold are retained to form a reference sample set. The set threshold can be a fixed distance value or a certain percentage of samples (such as the top 10%). In this example, at least two reference samples are retained to ensure diversity and stability in functional level judgment. The functional level labels corresponding to each sample in the reference sample set are further counted, and the frequency of label occurrence is statistically analyzed. The functional level with the highest frequency is used as the level judgment result for the current assessment object. This means that the current patient's functional level at the time of assessment is considered to be closest to the status of the sample corresponding to this type of label. If there are multiple level labels with the same frequency, a weighted voting method is introduced to weight the label votes according to the sample similarity values. The level labels corresponding to samples with higher weights have higher priority. The output of the matching process is the functional level label corresponding to the current patient. As part of the structured assessment results, it is annotated in the severity level field for personalized training path selection and parameter adjustment.

[0024] Based on the evaluation results, the relevant training action nodes in the rehabilitation knowledge graph are matched, the fitness weights of the training actions are calculated, and a ranking list of actions is generated. A personalized training path with intensity, frequency, and rhythm parameters is constructed. The specific implementation is as follows: Receive the structured assessment results output by the swallowing function assessment module and extract the three core elements contained therein: the obstacle type label, the main restricted muscle group location label, and the severity level label. These three labels constitute the set of nodes to be matched for action matching and are passed into the rehabilitation knowledge graph as input; The rehabilitation knowledge graph is modeled using a heterogeneous graph structure. The nodes in the rehabilitation knowledge graph include but are not limited to different semantic levels such as obstacle type nodes, training action nodes, muscle group nodes, and functional level nodes. The edges in the graph represent the association relationship between the nodes, especially the training adaptation edge between the obstacle type node and the training action node. Each edge is provided with an edge weight parameter. The edge weight represents the degree of adaptation of the training action to a certain obstacle type in clinical experience or literature data. It is obtained in advance by expert annotation, system learning, or empirical statistics, and the value range is usually normalized to [0, 1]. In the knowledge graph, a breadth-first search traversal operation is performed with the obstacle type node as the starting point. The maximum number of search layers is 2 hops to ensure that both directly related actions and training action nodes indirectly related through muscle group nodes are captured. During the search process, all training action nodes that meet the connectivity conditions are collected as a candidate set of training actions. For each candidate training action node, the adaptation weight between it and the current evaluation result is calculated to determine its adaptation priority in this round of training. The calculation of fitness weight is based on three core factors: The edge weight factor is to extract the associated edge weight between the current training action node and the obstacle type node, reflecting the historical effect or fitness of the action when training this type of obstacle; The muscle group matching factor is based on the target muscle group affected by the candidate training action and the restricted muscle group label extracted from the evaluation results. If the training action completely covers the target muscle group, the value is assigned to 1; if it partially covers the target muscle group, the value is 0.5; if there is no intersection, the value is 0; The functional level adaptation factor is based on the functional level label assessed by the current patient to determine whether the training movement meets the training intensity range allowed by the current level. If it is a perfect match, it is assigned a value of 1; if it exceeds the current level requirement, a decrement factor less than 1 is set according to the training level deviation, such as 0.8, 0.5 or 0; After normalizing the above three factors separately, they are combined in a multiplication manner to obtain the fitness weight corresponding to each candidate training action node, that is: the fitness weight is equal to the edge weight multiplied by the muscle group matching multiplied by the functional level adaptation factor. This combination method can effectively suppress the inadaptive movements when any factor is zero, and improve the consistent expression ability of multi-dimensional adaptability.

[0025] All candidate training action nodes are sorted from high to low according to the fitness weight to generate a sorted list of training actions. The actions in the sorted list will be used as a set of candidate actions for the subsequent construction of personalized training paths. Training action nodes with higher fitness weights will be given priority to ensure that each training task in the training path is highly matched with the patient's current state, thereby improving the efficiency and pertinence of rehabilitation training.

[0026] The specific process of building a personalized training path with intensity, frequency, and tempo parameters is as follows: Select training action nodes with weight values ​​higher than a set threshold from the fitness weight sorting list to form a training action candidate set. The threshold is set as the minimum fitness requirement value to ensure training safety and effectiveness under the current functional state. For each training action node in the training action candidate set, extract the parameter template associated with it in the preset rehabilitation knowledge graph. The parameter template includes: training intensity parameter (indicating the muscle group load level when the action is performed), training frequency parameter (indicating the recommended number of repetitions per unit cycle), and training rhythm parameter (indicating the time rhythm and control beat of each execution); Extract the functional level label output in the evaluation result and compare it with the training level requirement of the candidate training action node. If the current patient's functional level is lower than the minimum level requirement of the training action, the training action node is eliminated and the candidate set is updated; Obtain the patient's historical compliance records, including the frequency of completion of past training tasks, movement deviation levels, and training interruptions; and extract the patient's staged recovery curve, including functional score trends and deviation correction feedback results; Based on the training tolerance and adaptation trend reflected in the compliance record and recovery curve, the parameter template of each candidate training action is adaptively adjusted. The adjustment method includes: when compliance is poor or recovery is slow, reducing the training intensity parameters and rhythm parameters; when the recovery stability is good and the functional level is gradually improved, appropriately increasing the frequency and rhythm parameters to form a target execution parameter group under the individual state. The training action nodes with completed parameter settings are arranged in sequence according to the fitness weight of the training action nodes from high to low to construct a personalized training path. The personalized training path consists of at least two training action nodes, and each training action node contains determined training intensity parameters, training frequency parameters and training rhythm parameters to guide the specific implementation of the subsequent training plan.

[0027] During the training process, interactive monitoring is carried out synchronously. Based on real-time motion capture and electromyographic response recognition algorithms, training deviations are identified and prompt signals are output. At the same time, behavioral completion and deviation level indicators are recorded. The specific implementation is as follows: Before the start of the training movements in each round of personalized training path, the visual acquisition device and the electromyographic signal acquisition device are started respectively. The visual acquisition device is used to continuously capture the image sequence of the patient's neck area, laryngeal contour and head posture, and the electromyographic acquisition device is used to record the electromyographic activation response of the target muscle group during the training process. Both are synchronized and aligned with the training start time as a unified benchmark. The image data uses the key point recognition algorithm to extract the spatial displacement trajectory of the larynx and head, and compares it with the preset standard training movement template frame by frame to identify the spatial deviation of the displacement trajectory in terms of angle change, movement amplitude and speed; at the same time, the electromyographic response signal is parsed into three dimensions: the starting time point of muscle activation, the peak amplitude value and the activation duration, and is compared item by item with the muscle group activation target range preset in the training movement to identify whether there are electrophysiological deviations such as premature activation, insufficient peak or insufficient activation; The process of setting deviation identification thresholds involves triggering visual or auditory prompts if either spatial or electrophysiological deviation exceeds the set error limit, prompting the patient to adjust their posture or force output. The prompt type is tied to the corresponding deviation type, ensuring clear and instructive feedback. After each training session, the actual execution parameters, including actual movement duration, peak displacement, and electromyographic activation characteristics, are recorded and subtracted from the target parameters set in the current training session to generate a behavioral completion index for the movement. A behavioral completion value closer to 1 indicates more accurate execution. The system also counts the number of spatial and electrophysiological deviation behaviors identified during each training session, along with their magnitude levels. A weighted integral method is used to calculate a comprehensive deviation level index. For example, if two spatial deviations and one electrophysiological deviation occur at a medium intensity level, a medium deviation level index is generated. Completion and deviation level data for all training sessions are stored in real time in the training behavior log for use in rehabilitation progress assessment and subsequent pathway reconstruction.

[0028] For example, during a laryngeal lift training session, the motion capture and electromyographic signal acquisition channels were activated simultaneously. The camera recorded that the patient's Adam's apple lifted only 60% of the standard template, with a rise delay of 0.8 seconds and a dwell time of less than 0.5 seconds, significantly below the training requirements. At the same time, the peak value of the target muscle group activation signal collected by the electromyographic sensor was only 50% of the preset target, and the duration was short, showing premature attenuation. Based on this, the system identified the dual deviation behaviors of insufficient lift amplitude and insufficient muscle activation intensity. Based on the comparison results of spatial displacement and electromyographic response, the system triggers voice and interface prompts to inform the patient that the laryngeal lift amplitude and maintenance time are low, prompting them to pay attention to the quality of the movement. After the training is completed, the system compares the actual and target parameters, generates a behavior completion degree of 0.55, and evaluates the deviation level of the current movement as 2 based on the accumulated deviation items, which is written into the cycle feedback record as part of the training log.

[0029] The rehabilitation progress map is updated based on training feedback records and staged scores. If deviation levels exceed the limit are continuously identified, the training path is automatically reconstructed and synchronized with the doctor's terminal to assist in strategy optimization and intervention. The specific implementation is as follows: After each training action is completed, the behavioral completion index and deviation level index of the current action are automatically recorded and stored in chronological order to form a training feedback data sequence. This sequence covers the execution results of all training actions in the current training cycle. The behavioral completion index represents the execution quality of each training action and is usually normalized by the degree of difference between the target parameter and the actual parameter. The deviation level index is based on the frequency and amplitude of spatial deviation and electrophysiological deviation to obtain a numerical label. After the training cycle is completed, the system performs statistical analysis on the feedback data sequence within the cycle and calculates the stage score of the current stage. The stage score is generated according to the following rules: the total number of training actions in the current cycle is used as the base number, combined with the average value of the behavioral completion of all training actions, and the deviation level distribution is weighted. For example, the number of moderate deviations is multiplied by a weight of 0.5, and the severe deviation is multiplied by a weight of 1 to generate a stage execution score (the numerical range can be normalized to [0, 100]). This score reflects the overall quality level of rehabilitation task execution; The stage-by-stage scoring results will be mapped to the rehabilitation progress map. The map is a trajectory structure built based on time series. Each time node represents a comprehensive status node of a training cycle, including the stage score value, training compliance label (such as high, medium, and low), and deviation risk factor (accumulated based on the frequency and severity of the deviation level). The number of times the deviation level indicator exceeds the set threshold in the current training cycle is counted. If the deviation level of multiple consecutive training movements exceeds the preset threshold (such as three consecutive high deviation levels), the current training path is judged to be at risk of mismatch, and the training path reconstruction mechanism is automatically triggered. After triggering the reconstruction mechanism, the system re-calls the aforementioned evaluation results and knowledge graph matching process, and recalculates the fitness weights of candidate training action nodes based on the patient's current functional level label and training compliance record, giving priority to action nodes with high fitness, moderate training intensity, and good historical compliance, and generating a new ranking list of training actions. This list constitutes a new training path, and combined with the current evaluation results, a path execution parameter package (including the intensity, frequency, and rhythm parameters of each action) is generated; this path package is automatically sent to the doctor's terminal, where the doctor can intervene in the training path, such as replacing training actions, adjusting rhythm control, resetting training intensity or upper and lower limits of rhythm, etc. The reconstructed path confirmed by the doctor is synchronously updated to the patient's execution end to achieve real-time path optimization and closed-loop adjustment of rehabilitation strategies. This process ensures that training feedback and doctor intervention form a linkage mechanism to improve the adaptability and safety of training.

[0030] It should be noted that the thresholds involved in the embodiment can be determined according to specific scenarios and requirements.

[0031] The present invention uses a multi-channel sensing device to synchronously collect electromyographic signals, tongue pressure data, laryngeal movement images and acoustic response information during the swallowing process, and constructs a high-resolution multimodal swallowing behavior feature vector through unified time alignment and feature normalization processing, accurately capturing the physiological response characteristics of each key stage in the whole swallowing process; combining the pre-trained swallowing function assessment model with the historical rehabilitation database, it performs stage recognition and functional level discrimination of swallowing disorders, and outputs structured obstacle type, restricted area and severity label; on this basis, based on the knowledge graph reasoning mechanism, it matches the rehabilitation training action nodes with high adaptability, and integrates the edge The system uses weight, muscle group matching and functional level adaptation factors to build a personalized training path, and clearly configures core parameters such as intensity, frequency and rhythm to enhance the pertinence and compliance of the training plan. During the training process, it uses motion capture and electromyographic response to identify deviant behaviors, trigger real-time prompts and record the completion of the behavior and the deviation level, to achieve interactive monitoring and intelligent feedback throughout the process. Finally, the rehabilitation progress map is dynamically updated based on the training feedback. If the deviation exceeds the limit, the path reconstruction is triggered and the doctor is linked to optimize the strategy and intervene, thereby improving the scientificity, accuracy and intelligence of rehabilitation, and significantly improving rehabilitation efficiency and safety.

[0032] Example 2: A system for generating auxiliary plans for rehabilitation of dysphagia, such as Figure 2 As shown, specifically including: The swallowing data acquisition module is used to collect myoelectric signals, tongue pressure data, laryngeal movement images and speech features during swallowing based on multi-channel sensing equipment, construct multimodal swallowing behavior feature vectors, and uniformly map them to a standard time frame sequence; The obstacle stage identification module is used to combine the pre-trained swallowing function assessment model with historical rehabilitation data to perform obstacle stage identification and functional level judgment, and output structured assessment results including obstacle type, location and severity labels; The rehabilitation training path module is used to match relevant training action nodes in the rehabilitation knowledge graph based on the evaluation results, calculate the fitness weights of the training actions, generate a ranked list of actions, and construct a personalized training path with intensity, frequency, and rhythm parameters; The training monitoring module is used to conduct interactive monitoring during the training process. Based on real-time motion capture and electromyographic response recognition algorithms, it identifies training deviations and outputs prompt signals, while also recording behavioral completion and deviation level indicators. The rehabilitation training optimization module is used to update the rehabilitation progress map based on training feedback records and stage scores. If the deviation level exceeds the limit is continuously identified, the training path will be automatically reconstructed and synchronized with the doctor's terminal to assist in strategy optimization and intervention.

[0033] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in the embodiments of this application are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, ATA hard drives, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state ATA hard drive.

[0034] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0035] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0036] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0037] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0038] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0039] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A method for generating an auxiliary program for dysphagia rehabilitation, characterized in that: The steps include: Based on the multi-channel sensor equipment, myoelectric signals, tongue pressure data, laryngeal movement images and speech features during swallowing are collected to construct a multimodal swallowing behavior feature vector and uniformly map it to a standard time frame sequence. Combining a pre-trained swallowing function assessment model with historical rehabilitation data, it performs impairment stage identification and functional level determination, and outputs structured assessment results that include labels for impairment type, location, and severity. Based on the evaluation results, the relevant training action nodes in the rehabilitation knowledge graph are matched, the fitness weights of the training actions are calculated, and a ranking list of actions is generated to build a personalized training path with intensity, frequency, and rhythm parameters. During the training process, interactive monitoring is carried out synchronously. Based on real-time motion capture and electromyographic response recognition algorithms, training deviations are identified and prompt signals are output. At the same time, behavioral completion and deviation level indicators are recorded. The rehabilitation progress map is updated based on training feedback records and stage-by-stage scores. If deviation levels are continuously identified as exceeding the limit, the training path is automatically reconstructed and synchronized with the doctor's terminal to assist in strategy optimization and intervention.

2. The method for generating an auxiliary plan for dysphagia rehabilitation according to claim 1, characterized in that: Based on the multi-channel sensor equipment, myoelectric signals, tongue pressure data, laryngeal movement images and speech features during swallowing are collected. A multimodal swallowing behavior feature vector is constructed and uniformly mapped to a standard time frame sequence. The specific process is as follows: The surface electromyography electrode array is used to obtain the time-series electromyographic voltage signals generated by the muscle groups located in the mandibular, laryngeal and sublingual areas during swallowing, and record them as electromyographic signal data; The tongue pressure sensor pad is used to obtain the unit pressure change value of the contact area between the tongue and the upper palate of the mouth, and the pressure duration is recorded synchronously as tongue pressure data; An infrared visual camera is used to obtain the vertical movement trajectory of the Adam's apple during swallowing, and the speed and amplitude of the movement are calculated and recorded as laryngeal movement image data. Use a close-range microphone to collect acoustic response signals during swallowing, extract the onset time, peak frequency, and duration, and record them as speech feature data; Four types of heterogeneous sensor data, namely electromyographic signal data, tongue pressure data, laryngeal movement image data, and speech feature data, are uniformly mapped to a standardized time frame structure with the start of the swallowing event as the reference point to form a swallowing behavior feature vector set.

3. The method for generating an auxiliary plan for dysphagia rehabilitation according to claim 2, characterized in that: Combining the pre-trained swallowing function assessment model with historical rehabilitation data, we perform impairment stage identification and functional level determination, and output a structured assessment result that includes labels for impairment type, location, and severity. The specific process is as follows: The constructed swallowing behavior feature vector is input into a pre-trained swallowing function assessment model, which includes a convolutional structure for extracting static features and a recurrent neural network structure for modeling temporal changes; An attention mechanism was used to enhance the responsiveness of key frame segments of a swallowing function assessment model, identifying feature differences during swallow initiation, laryngeal elevation, and contraction recovery. After the swallowing function assessment model processes the input swallowing behavior feature vector, it outputs a set of structured prediction results. The structured prediction results include classification probability predictions of the obstruction stage. In the output structured prediction results, the obstruction stages involved in the swallowing process are identified, including the oral preparation stage, the pharyngeal stage, and the esophageal transit stage, and the corresponding obstruction type labels are output. The prediction results of the swallowing function assessment model are matched with samples with clear functional level labels in the historical rehabilitation database, and the range of the current functional level is confirmed by similarity measurement method; Combining the recognition results of the swallowing function assessment model with the historical score mapping relationship, a structured assessment result is generated, which includes the type of obstacle, the main restricted area and the severity level label.

4. The method for generating an auxiliary plan for dysphagia rehabilitation according to claim 3, characterized in that: The swallowing function assessment model prediction results are matched with samples with clear functional level labels in the historical rehabilitation database, and the current functional level range is confirmed by similarity measurement method. The specific process is as follows: A standard sample set containing functional grade labels is extracted from a historical rehabilitation database, wherein each sample in the standard sample set includes a swallowing feature vector of the same modality dimension as that of the current patient and its corresponding verified functional grade label; Performing feature normalization processing on the feature vector output by the current swallowing function evaluation model; The characteristic distance between the normalized current feature vector and each sample in the standard sample set is calculated in sequence. The characteristic distance calculation is based on the cumulative results of the numerical differences between each modality, including the difference in electromyographic response, the difference in tongue pressure peak value, the difference in laryngeal trajectory change, and the difference in acoustic duration. Generate a sample similarity ranking list based on all distance calculation results, and select at least two samples with similarity higher than a set threshold as reference samples; The distribution of corresponding functional level labels in the reference samples is statistically analyzed, and the functional level range of the current assessment object is determined based on the level label with the highest frequency, which is used as the final output severity judgment result.

5. The method for generating an auxiliary plan for dysphagia rehabilitation according to claim 4, characterized in that: According to the evaluation results, the relevant training action nodes in the rehabilitation knowledge graph are matched, the fitness weights of the training actions are calculated, and an action ranking list is generated. The specific process is as follows: Based on the structured assessment results, the obstacle type, restricted area, and severity level labels are extracted and input into the pre-built rehabilitation knowledge graph as a set of nodes to be matched; The rehabilitation knowledge graph is a heterogeneous graph structure that contains the associated edges between training action nodes and obstacle type nodes. The edge weights represent the degree of adaptability of the action to a specific obstacle. Starting from the obstacle type node in the rehabilitation knowledge graph, a breadth-first search is performed to traverse the adjacent action nodes and extract all directly or indirectly connected candidate training action nodes; For each candidate training action node, the fitness weight between the candidate training action node and the evaluation result is calculated. The fitness weight is determined by the edge weight, the matching degree of the muscle group targeted by the action, and the fitness factor of the patient's current functional level. All candidate training action nodes are sorted from high to low according to the fitness weight, and an action sorting list is generated as the input candidate set for training path construction.

6. The method for generating an auxiliary plan for dysphagia rehabilitation according to claim 5, characterized in that: For each candidate training action node, the fitness weight between the candidate training action node and the evaluation result is calculated. The fitness weight is determined by the edge weight, the matching degree of the muscle group targeted by the action, and the fitness factor of the patient's current functional level. The specific process is as follows: Obtain the main restricted muscle group information from the evaluation results, compare it with the target muscle groups involved in the candidate training action nodes, and calculate the muscle group matching index. The muscle group matching index is quantified based on the overlap of the target muscle groups, with complete coverage being recorded as 1, overlap being recorded as 0.5, and no overlap being recorded as 0; Extract the edge weights between candidate training action nodes and evaluation nodes in the rehabilitation knowledge graph. The edge weights represent the effectiveness of the training action for this type of disorder in clinical experience and are preset during graph construction and normalized to the interval [0, 1]. According to the patient's functional level label in the evaluation results, the range of movement intensity allowed by the current level is extracted and compared with the training intensity label of the candidate movement. If the intensity matches, the adaptation factor is assigned to 1, otherwise the adaptation factor value is reduced according to the decreasing function; The edge weights, muscle group matching indexes, and functional level adaptation factors are normalized item by item and combined in product form to obtain the fitness weights of the candidate movements. The fitness weights are used to represent the comprehensive matching priority of the training movements under the current patient condition.

7. The method for generating an auxiliary plan for dysphagia rehabilitation according to claim 6, characterized in that: Build a personalized training path with intensity, frequency, and cadence parameters. The specific process is as follows: Selecting a set of training action nodes whose fitness weights are higher than a preset threshold from the fitness weight sorted list as a candidate action set for the training path; For each training action node, extract the parameter template associated with it in the knowledge graph. The parameter template includes the recommended intensity parameter, recommended frequency parameter, and standard rhythm parameter of the training action. Match the functional level label in the evaluation result with the training level requirement of the candidate training action. If the current patient level is lower than the required level of the training action, the training action will be removed from the candidate set. According to the patient's historical compliance records and stage recovery curve, the recommended intensity parameters, recommended frequency parameters and standard rhythm parameters of the candidate training movements are adjusted to form the target execution parameters for the current individual state; According to the fitness weight order of the training action nodes, the training actions with set execution parameters are arranged in sequence and combined into a personalized training path containing at least two training action nodes. Each action node in the personalized training path contains clear intensity parameters, frequency parameters and rhythm parameters.

8. The method for generating an auxiliary plan for dysphagia rehabilitation according to claim 7, characterized in that: During the training process, interactive monitoring is carried out simultaneously. Based on real-time motion capture and electromyographic response recognition algorithms, training deviations are identified and prompt signals are output. At the same time, behavioral completion and deviation level indicators are recorded. The specific process is as follows: At the beginning of the training action, the visual acquisition channel and the electromyographic signal acquisition channel are activated to obtain the action image sequence of the current training action and the corresponding electromyographic activation timing signal; Extract the displacement trajectories of key points of the throat and head based on the action image sequence, and perform temporal comparison with the standard action template of the training action to identify spatial deviation behavior; Analyze the starting time, peak amplitude, and duration of muscle activation based on electromyographic signals, and compare them with the muscle group activation target parameters preset in the training movement to identify electrophysiological deviation behaviors; If any of the identified spatial deviation behaviors or electrophysiological deviation behaviors exceeds the set threshold, a prompt signal output is triggered; After each training action is completed, the difference between the actual execution parameters of the training action and the target parameters is recorded to calculate the behavioral completion degree of the training action; Based on the cumulative number of spatial deviation behaviors and electrophysiological deviation behaviors and the deviation intensity level, a deviation level index for the training action is generated and stored in the training behavior log as a reference for progress evaluation.

9. The method for generating an auxiliary plan for dysphagia rehabilitation according to claim 8, characterized in that: The rehabilitation progress map is updated based on training feedback records and staged scores. If deviation levels exceed the limit are continuously identified, the training path is automatically reconstructed and synchronized with the doctor's terminal to assist in strategy optimization and intervention. The specific process is as follows: After each training action is completed, the corresponding behavior completion and deviation level indicators are recorded in the training feedback record, and the feedback data sequence of the current training cycle is constructed in chronological order; At the end of the training cycle, a stage score is calculated based on the feedback data sequence. The stage score is generated based on the total number of training actions, the average completion rate of the behavior, and the deviation level distribution, and is used to reflect the quality of rehabilitation execution at the current stage. Based on the staged scores and historical score curves, the progress status of the corresponding time nodes in the rehabilitation progress map is updated. The rehabilitation progress map is a recovery trajectory structure constructed in a time series. Each node contains the stage score, action compliance label and deviation risk factor. Count the number of deviation levels exceeding the limit in the current training cycle. If the deviation level indicators of consecutive training actions exceed the preset threshold, the training path reconstruction mechanism will be triggered; In the training path reconstruction mechanism, the fitness weight sorting process is called and the fitness weights are sorted from high to low. The candidate training action nodes are re-screened based on the sorting and a new training action sorting list is constructed. The reconstructed training path and the current evaluation results are packaged and sent to the doctor terminal. After receiving the data, the doctor terminal performs manual intervention and synchronously updates the intervention results to the patient execution terminal.

10. A system for generating an auxiliary plan for rehabilitation of dysphagia, used to implement the method for generating an auxiliary plan for rehabilitation of dysphagia according to any one of claims 1 to 9, characterized in that: include: The swallowing data acquisition module is used to collect myoelectric signals, tongue pressure data, laryngeal movement images and speech features during swallowing based on multi-channel sensing equipment, construct multimodal swallowing behavior feature vectors, and uniformly map them to a standard time frame sequence; The obstacle stage identification module is used to combine the pre-trained swallowing function assessment model with historical rehabilitation data to perform obstacle stage identification and functional level judgment, and output structured assessment results including obstacle type, location and severity labels; The rehabilitation training path module is used to match relevant training action nodes in the rehabilitation knowledge graph based on the evaluation results, calculate the fitness weights of the training actions, generate a ranked list of actions, and construct a personalized training path with intensity, frequency, and rhythm parameters; The training monitoring module is used to conduct interactive monitoring during the training process. Based on real-time motion capture and electromyographic response recognition algorithms, it identifies training deviations and outputs prompt signals, while also recording behavioral completion and deviation level indicators. The rehabilitation training optimization module is used to update the rehabilitation progress map based on training feedback records and stage scores. If the deviation level exceeds the limit is continuously identified, the training path will be automatically reconstructed and synchronized with the doctor's terminal to assist in strategy optimization and intervention.

Images