Optimization method of dysphagia training based on multimodal stimulation

Abstract

The present invention relates to an optimization method for swallowing disorder training involving multimodal stimulation, and the following technical scheme is adopted, including the following steps: collecting basic physiological data of patients including age, gender, and medical history, and real-time swallowing data including throat muscle activity, electrophysiological signals, and swallowing pressure; using multimodal data fusion technology to fuse the S1.1 real-time swallowing data to generate a personalized swallowing function evaluation report; and based on the evaluation results, generating a personalized swallowing disorder training program through an intelligent algorithm, which includes a combination of multiple stimulation methods such as touch, electrical stimulation, light stimulation, and sound stimulation; constructing a real-time data acquisition and feedback system, and using sensors to monitor the multimodal data of patients including muscle activity, sound signals, facial expressions, and breathing frequency in real time during the training process; using the dynamic time warping DTW algorithm, the real-time monitoring data is compared with the expected indicators in the training program, and the intensity and mode of the stimulation are adjusted in real time.

Description

Technical Field

[0001] The present invention relates to an optimization method for dysphagia training, in particular to an optimization method for dysphagia training with multi-modal stimulation. Background Art

[0002] China's invention patent 2022105223246, a VR-based swallowing disorder training method and its equipment, takes VR (virtual reality) technology and mechanical assisted training methods as examples. This technology guides patients to perform swallowing training through VR equipment and uses swallowing recognition equipment to monitor physiological changes during swallowing. Although this technology is unique in visual guidance and action assistance, it still has obvious shortcomings in some key issues.

[0003] First, the VR-based swallowing disorder training method relies more on visual feedback and physical guidance of the robotic arm. This training method is effective in specific scenarios, but it ignores the synergistic effect of multimodal stimulation. The swallowing process involves complex multisensory responses, such as touch, muscle contraction, electrophysiological signals, etc. It is difficult to cover all the sensory needs of patients during the swallowing process by simply relying on VR technology and mechanical assistance for swallowing training. For example, the triggering of the swallowing reflex depends not only on visual perception, but more on the response of the neuromuscular system. Although traditional VR training equipment can guide patients to complete the swallowing action, it lacks in biofeedback and multisensory stimulation integration of the swallowing process. For patients with swallowing disorders, a single mode of training cannot stimulate their nervous system to fully participate, thereby limiting the development of neural plasticity and failing to effectively promote the recovery of swallowing reflexes. Secondly, although VR technology can provide visual guidance and virtual scenes, its ability to handle the personalized needs of patients with swallowing disorders is insufficient. Each patient's condition, course of disease, type and degree of swallowing disorders are different, and a single VR-guided training cannot be adaptively adjusted according to the patient's real-time status. Although the public technical documents propose to monitor the swallowing status of patients, such as mouth opening, swallowing, choking, etc., through swallowing recognition equipment, no specific mechanism is proposed for in-depth analysis of real-time physiological feedback data and adjustment of stimulation patterns. The swallowing disorder training of patients requires not only the guidance of external actions, but also real-time monitoring of their muscle activity, breathing rate, heart rate and other physiological signals, and personalized real-time adjustment based on these feedback data. However, the existing VR training methods fail to achieve multimodal integration and processing capabilities of patients' real-time feedback signals, which limits the accuracy of training and the flexibility of personalized adjustments.

[0004] Furthermore, VR devices rely more on visual feedback and action guidance, but lack the coordinated use of multimodal stimulation such as touch and electrical stimulation. Studies have shown that the recovery of swallowing function requires not only visual guidance, but also multi-sensory coordinated stimulation to activate the patient's neural network. For example, tactile stimulation, electrical stimulation and sound stimulation can work together at a specific time point to enhance the activation effect of the swallowing reflex. However, the VR-based training method cannot fully utilize the synergistic effect of these multi-sensory stimulations, resulting in a relatively single training effect. In contrast, the multimodal stimulation synergistic training method can systematically stimulate the patient's swallowing reflex and improve neural plasticity by integrating multiple modes such as touch, electrical stimulation, light stimulation and sound stimulation, thereby accelerating the rehabilitation process. The limitation of VR technology is that it lacks the integration of multimodal stimulation and cannot fully utilize the response of the nervous system to multi-sensory input, resulting in its training effect being far inferior to the multimodal stimulation training method. Summary of the invention

[0005] The purpose of the present invention is to provide an optimized method for dysphagia training based on multimodal stimulation, thereby solving some of the drawbacks pointed out in the background art.

[0006] The present invention solves the above-mentioned technical problems by adopting the following technical solution, which includes the following steps:

[0007] S1. Generation of personalized training programs for multimodal stimulation:

[0008] S1.1. Collect the patient's basic physiological data including age, gender, medical history and real-time swallowing data including throat muscle activity, electrophysiological signals, and swallowing pressure;

[0009] S1.2, using multimodal data fusion technology, the real-time swallowing data of S1.1 is fused and processed to generate a personalized swallowing function assessment report;

[0010] S1.3, and based on the evaluation results, generate a personalized dysphagia training program through an intelligent algorithm, which includes a combination of multiple stimulation methods such as tactile, electrical stimulation, light stimulation and sound stimulation;

[0011] S2. Synchronous optimization and real-time feedback of multi-mode stimulation:

[0012] S2.1. Build a real-time data collection and feedback system, using sensors to monitor the multimodal data of patients during training, including muscle activity, sound signals, facial expressions, and breathing rate;

[0013] S2.2, through the dynamic time warping (DTW) algorithm, the real-time monitoring data is compared with the expected indicators in the training plan, and the intensity and mode of stimulation are adjusted in real time;

[0014] S2.3. Use machine learning algorithms to analyze patients’ responses to different stimulation patterns and optimize training programs.

[0015] S3. Synergistic enhancement strategy of multimodal stimulation:

[0016] S3.1. Introduce a synergistic enhancement algorithm for multimodal stimulation to analyze the synergistic effects of different stimulation modes including tactile and electrical stimulation, light stimulation and sound stimulation;

[0017] S3.2. Use a multi-objective optimization algorithm to dynamically adjust the order, frequency, and intensity of synergistic stimulation based on individual patient characteristics and training feedback;

[0018] S4, Integration of Intelligent Adaptive Swallowing Training System:

[0019] S4.1. Integrate personalized training programs, real-time feedback systems, and synergistic enhancement strategies into the intelligent adaptive swallowing training system; adaptively adjust training programs and stimulation patterns based on patients’ real-time feedback and training progress.

[0020] Furthermore, the method for adjusting the intensity and mode of stimulation in real time includes: using a DTW algorithm to deal with the temporal variability of the patient during swallowing to process the time offset, calculating the difference between the real-time data and the ideal training target to adjust the training plan; the calculation formula of the path difference is as follows:

[0021]

[0022] Among them, D DTW (i, j) represents the minimum difference path between the patient's real-time monitoring data and the ideal training index; f(x k ) and g(y i-k ) represent the real-time physiological data of the patient's throat muscle activity and the expected training index, respectively, while the time series function reflects the deviation between the patient's physiological signals at different time points and the ideal target;

[0023] Where p is the power of the norm, which is used to adjust the measurement method of different data errors; γ k is the alignment adjustment factor, which is used to balance the alignment cost between different time periods; |ij| q is the weight term for processing time offset, controlling the influence of time interval, where q is an important factor for adjusting time difference; integral It means that the deviation of the time series is accumulated over the entire time range T, and the differences in each time period are weighted.

[0024] Further, the method for adjusting the intensity and mode of stimulation in real time includes: when the DTW algorithm detects a deviation between the patient's real-time data and the training expected index, adjusting the intensity and mode of stimulation according to the calculated difference value;

[0025] An adaptive adjustment mechanism is introduced to make the stimulation intensity respond to the patient's immediate feedback; by analyzing the size of the difference value, it is decided whether to increase or decrease the stimulation intensity; the adaptive adjustment is achieved through the formula:

[0026]

[0027] Where S(t) represents the intensity of stimulation applied to the patient, which is adjusted as time t changes; α is the initial stimulation intensity, which defines the starting value of the stimulation; the exponential decay part Controls the intensity of the stimulus, where D DTW (t) is the difference between the patient’s real-time physiological data and the expected target, which is calculated by the DTW algorithm and represents the actual deviation at each time point;

[0028] R(t) is the patient's immediate physiological response, representing the patient's actual reaction state at each moment. DTW The high-order power p of (t)-R(t) is used to adjust the sensitivity of the deviation; δ is the adjustment rate, which controls the speed and amplitude of the response adjustment; the periodic adjustment term Introducing the sine wave function To simulate the natural fluctuations in the patient's physiological cycle, κ is the adjustment amplitude factor.

[0029] Furthermore, the method for adjusting the intensity and mode of stimulation in real time includes: introducing a closed-loop feedback mechanism; the feedback data after each training is input into a machine learning model, and the subsequent training parameters are optimized by analyzing the feedback information; the closed-loop feedback mechanism analyzes and optimizes the long-term training data by weighted averaging; the long-term training effect is measured by an optimization formula:

[0030]

[0031] Among them, E opt It is a measure of the overall optimization effect, obtained by weighted averaging the long-term training effect; is the time integration operation, which is used to calculate the average performance within the time T; λ is the weight coefficient of the stimulus intensity S(t), which reflects the sensitivity to the stimulus intensity. Controls the adaptability to different stimulus intensities, p is used to adjust the weighted processing of large and small stimuli; β is the error control factor, which is used to balance the tolerance to errors, |D DTW (t)| qIt is a weighted treatment of the real-time error by raising the difference to the power of q.

[0032] Furthermore, the method for dynamically adjusting the order, frequency and intensity of the synergistic stimulation includes: introducing an adaptive multimodal neural network architecture, the network architecture is composed of multiple layers of neurons, the input layer receives physiological data and training feedback signals from multiple sensors, the middle layer adjusts the weights of various stimulation modes in real time through an adaptive weight mechanism, and the output layer gives a stimulation order, frequency and intensity combination; the multimodal neural network optimization process is represented by the function:

[0033]

[0034] Among them, P opt represents the stimulus scheme combination, and represents the stimulus combination found within time T;

[0035] S 1 (t),S 2 (t),…,S n (t) is the intensity function of different stimulation modes including electrical stimulation, light stimulation and tactile stimulation at time t; α 1 ,α 2 ,…,α n is the weight coefficient of different stimulation modes, indicating the relative importance of each mode in the whole training process; integral Calculate the weighted intensity of each stimulation pattern over the entire time interval.

[0036] Furthermore, the method for dynamically adjusting the order, frequency and intensity of synergistic stimulation includes: establishing an individualized response model of the patient through a deep learning model, predicting the physiological response under different stimulation combinations according to the specific response of each patient; and combining a reinforcement learning algorithm; reinforcement learning optimizes the stimulation strategy through a feedback loop mechanism, evaluates the current training effect after each training, and adjusts the stimulation order and frequency of the next training through the value function of reinforcement learning; the optimization process is represented by the value function of reinforcement learning:

[0037]

[0038] Where V(t) represents the long-term value of different training stimulus combinations at time t; S k (t) is the intensity function of the kth stimulation pattern at time t, indicating the performance of the combination of different stimulation patterns in training; r(S k (t)) is the immediate reward function, which reflects the effect of the stimulation mode and the immediate response effect of the patient during training; C(S k (t)) is the cost function, which represents the negative impact or discomfort of the stimulation pattern on the patient; γ kis a discount factor used to balance short-term and long-term benefits.

[0039] Furthermore, the method for dynamically adjusting the sequence, frequency and intensity of the synergistic stimulation includes: real-time monitoring of the patient's biosignal data including electromyographic signals, heart rate, respiratory rate, and throat muscle activity, and the biosignal data is pre-processed and sent to the neural network model as an input feedback signal; dynamically adjusting the stimulation parameters according to the feedback signal; and the total utility of the biofeedback is measured by the feedback integral function:

[0040]

[0041] Where F(t) is the total feedback utility of the patient within time T, which represents the effect of adjusting the stimulation mode according to the biofeedback data; R 1 (t),R 2 (t),…,R n (t) represents different types of biofeedback signals including electromyographic signals, heart rate, and respiratory rate; β 1 ,β 2 ,…,β n is the importance weight of each biofeedback signal, indicating the relative importance of different biosignals in the overall feedback; integral It means that the feedback in each time period is accumulated and calculated, which is used to optimize the biofeedback data at different training moments.

[0042] Furthermore, the method for dynamically adjusting the order, frequency and intensity of synergistic stimulation includes: optimizing the training effect based on a benefit-cost trade-off model, making real-time adjustments according to the ratio of benefit to cost in each training, and evaluating the training effect through a long-term optimization formula:

[0043]

[0044] Among them, E opt It is the long-term optimization effect, which means the overall effect after multiple trainings; P opt (t) is the stimulation combination function, which indicates the combination of stimulation modes selected at each time point; D(t) is the degree of discomfort or cost function of the patient during the training process, which indicates the negative effect of the stimulation mode on the patient; λ and μ are the weight coefficients of benefit and cost, which are used to balance the therapeutic effect of stimulation and the physiological tolerance of the patient; p and q are adjustment parameters, which control the sensitivity to different degrees of stimulation effect and cost; integral Used to calculate the average effect over a long period of time.

[0045] The present invention has the following beneficial effects:

[0046] The present invention can generate a personalized dysphagia training program by comprehensively analyzing the individual characteristics and physiological data of the patient. Different stimulation modes (such as electrical stimulation, light stimulation, tactile stimulation and sound stimulation) are combined and adjusted according to the patient's real-time feedback, ensuring that the training program is highly personalized and can meet the specific needs of different patients, significantly improving the effectiveness of the training.

[0047] The introduction of the dynamic time warping (DTW) algorithm and real-time biofeedback system can monitor the patient's physiological signals (such as electromyographic signals, heart rate, respiratory rate, etc.) during the training process and adjust the intensity, frequency and sequence of stimulation according to real-time data. This real-time adjustment mechanism ensures that the training program can respond to the patient's immediate reaction at any time, improving the accuracy and effect of the training.

[0048] Through multimodal neural networks and synergistic enhancement algorithms, the present invention can comprehensively analyze the synergistic effects between different stimulation modes and optimize the combination of multimodal stimulation. This allows multiple stimulation modes to interact with each other, producing a stronger training effect than a single stimulation mode, and promoting the rapid recovery of the patient's swallowing function. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of the optimization method of dysphagia training based on multimodal stimulation of the present invention.

[0050] Figure 2 This is a flow chart of the method for adjusting the intensity and mode of stimulation in real time according to the present invention.

[0051] Figure 3 The present invention is a flow chart of the method for dynamically adjusting the order, frequency and intensity of co-stimulation.

[0052] Figure 4 A in the figure is a photograph of a human half pharynx treated with Sihler stain (a whole nerve staining technique) of the present invention, showing the pharyngeal plexus formed by the pharyngeal branches of the X and IX nerves, innervating the upper, middle and lower pharyngeal constrictors (i.e., SPC, MPC and IPC) and the cricopharyngeus muscle (CP); 1, pharyngeal branch of the X nerve (Ph-X); 2, pharyngeal branch of the IX nerve (Ph-IX); IX-L, lingual branch of the IX nerve; ESLN, external superior laryngeal nerve; ISLN, internal superior laryngeal nerve; UE, upper esophagus; Figure 4 B and C in the figure are longitudinal section micrographs of the cervical X nerve trunk (B) and Ph-X (C) of a PD patient with dysphagia in the present invention; the sections were immunostained with PAS; a large number of PAS immunoreactive axons (darkly stained lines and dots) were present in both the X nerve and the Ph-X nerve; Figure 4D in the figure is a cross-section of the IPC muscle of a PD patient with dysphagia in the present invention, stained with the monoclonal antibody NOQ7-5-4D against type I muscle fibers (dark staining); wherein, a large number of small atrophic muscle fibers are present in the IPC muscle; original magnification: (BD) 200 times.

[0053] Figure 5 A in the figure shows the human larynx (side view) and its innervated nerves according to the present invention; Figure 5 B in the figure shows the position of the intrinsic laryngeal muscles of the present invention (posterior view); wherein, the cervical X nerve gives rise to the superior laryngeal nerve (SLN) and the recurrent laryngeal nerve (RLN); the SLN is further divided into internal (ISLN) and external (ESLN) branches, innervating the mucosa and the cricothyroid muscle (CT), respectively; the RLN innervates the remaining muscles (i.e., the posterior cricoarytenoid muscle, PCA; the thyroarytenoid muscle, TA; the lateral cricoarytenoid muscle, LCA; the interarytenoid muscle, IA); E, epiglottis; H, hyoid bone; TC, thyroid cartilage; Figure 5 C and D are Sihler-stained human laryngeal mucosa of the present invention, showing the distribution and branching pattern of ISLN and sensory nerve endings on the epiglottic laryngeal surface (C) and postcricoid region (D) mucosa.

[0054] Figure 6 The human tongue and its anatomical relationship to the larynx, pharynx, and other structures are shown for the present invention. CP, cricopharyngeus; CT, cricothyroid; GG, genioglossus; HG, hyoglossus; IL, inferior longitudinal muscle; IPC, inferior pharyngeal constrictor; M, mandible; MPC, middle pharyngeal constrictor; SG, styloglossus; SPC, superior pharyngeal constrictor; TC, thyroid cartilage; TH, thyrohyoid muscle; UE, upper esophagus.

[0055] Figure 7 This is a photo of a human tongue treated with Sihler stain according to the present invention, showing the branches and distribution pattern of the tongue nerves. Among them, the twelfth nerve is divided into two parts, the outer side (green circle) and the inner side (purple circle) in the posterior part of the tongue to supply the tongue muscles. The lingual branch of the ninth nerve (IX-L) sends out multiple secondary branches to supply the mucosa covering the posterior third of the tongue and the lingual papilla (black dots). The lingual nerve (LN) sends out a bundle of branches to supply the mucosa covering the anterior two-thirds of the tongue. There are many communication branches between the LN and the twelfth nerve. DETAILED DESCRIPTION

[0056] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0057] The first step involves the generation of personalized training plans for multimodal stimulation:

[0058] Combined with Figure 1The specific steps are: S1.1 Collect the patient's basic physiological data, including age, gender, medical history and other information, and collect the patient's swallowing data in real time. These data include throat muscle activity, electrophysiological signals and swallowing pressure. Through real-time monitoring of these data, the system can accurately capture the patient's specific physiological manifestations during the swallowing process. Next, use multimodal data fusion technology to fuse these real-time swallowing data. The role of multimodal data fusion technology is to integrate data from different physiological sensors. These data come from different dimensions, times or formats. Through fusion processing, they can more comprehensively and accurately reflect the patient's actual swallowing function status.

[0059] The fused data will be used to generate a personalized swallowing function assessment report, which is based on a multi-dimensional analysis of the data and can accurately assess the patient's current degree and characteristics of swallowing disorders. Based on the generated assessment results, the system further generates a personalized swallowing disorder training program through intelligent algorithms (such as machine learning or optimization algorithms). This training program integrates a combination of multiple stimulation methods, including multimodal stimulation such as tactile stimulation, electrical stimulation, light stimulation, and sound stimulation. Tactile stimulation can enhance the patient's perception of swallowing through skin or muscle contact feedback, electrical stimulation activates the swallowing reflex by directly stimulating the throat or related muscles, and light stimulation and sound stimulation use sensory responses to assist patients in completing swallowing movements.

[0060] S2 involves the synchronous optimization and real-time feedback of multimodal stimulation. First, a real-time data acquisition and feedback system is constructed, which uses a variety of sensors to monitor the patient's physiological and behavioral data in real time during the training process. These monitoring data include multimodal data such as muscle activity, sound signals, facial expressions, and respiratory rate. Muscle activity data is mainly used to track the contraction and relaxation of the patient's throat and swallowing-related muscles. Sound signals can capture the patient's swallowing behavior during the phonation process. Facial expressions can reflect the patient's emotions and tension during the swallowing process, and changes in respiratory rate can help judge the patient's breathing coordination and swallowing synchronization. Through the real-time collection of these data, the system can comprehensively monitor the patient's dynamic physiological state during training. Next, the dynamic time warping DTW algorithm is used to process these real-time monitored data.

[0061] The core function of the DTW algorithm is to compare the patient's actual response data with the expected indicators in the training program. Since the swallowing reaction time of patients in training is different, the DTW algorithm can perform nonlinear matching of signals of different time lengths or rhythms to ensure that the system can accurately capture the time difference between the actual physiological response and the expected target. The advantage of this algorithm is that it can handle the nonlinear time changes in the patient's swallowing response process, thereby achieving accurate comparison and matching. Through this comparison, the system can adjust the intensity and mode of stimulation in real time to ensure that the stimulation program of each training can be optimized according to the patient's immediate response. If it is found that the patient's muscle activity or other physiological parameters deviate greatly from the expected target, the system will immediately increase the stimulation intensity. Conversely, if the patient's response is close to expectations, the stimulation intensity will be reduced or the stimulation mode will be adjusted to prevent overstimulation.

[0062] S3 involves a synergistic enhancement strategy for multimodal stimulation. The core of this strategy is to analyze the synergistic effects between multiple stimulation modes by introducing a synergistic enhancement algorithm for multimodal stimulation. Different stimulation modes include tactile stimulation, electrical stimulation, light stimulation, and sound stimulation. The synergistic effect of these modes can improve patients' neural responses and motor control during swallowing training. Tactile stimulation provides sensory feedback through physical contact with the skin or muscles, while electrical stimulation directly stimulates the patient's throat or related muscle groups through electric current to activate the swallowing reflex; light stimulation uses light signals to stimulate the senses to enhance the patient's ability to respond to the external environment, and sound stimulation mobilizes the patient's attention and swallowing response through the auditory system.

[0063] When these different stimulation modes are used in combination, the synergy between them produces a stronger training effect than when used alone. Through the synergistic enhancement algorithm, the system is able to quantify the effects of these combinations and analyze which combinations of stimulation modes can produce the best synergistic effects, thereby improving the efficiency and effectiveness of training. In order to achieve the optimal training plan for the individual characteristics of patients, the system further introduces a multi-objective optimization algorithm. The multi-objective optimization algorithm can handle multiple conflicting goals at the same time, such as stimulation sequence, frequency, and intensity, and dynamically adjust these stimulation parameters based on the individual characteristics of the patient (such as medical history, age, muscle response, etc.) and real-time training feedback.

[0064] Specifically, the system first collects the individual characteristics of the patient and the real-time response during training, and then inputs this data into a multi-objective optimization algorithm, which explores the best combination of stimulation parameters in multiple dimensions to ensure the optimal application of the synergistic stimulation mode. Each patient's response is different, so the system can continuously learn and adjust the order, frequency and intensity of the stimulation mode through intelligent optimization methods, so that the synergistic effect of the stimulation is adapted to the patient's actual response.

[0065] S4 involves the integration of an intelligent adaptive swallowing training system. The core of this step is to integrate personalized training programs, real-time feedback systems, and synergistic enhancement strategies into an intelligent adaptive swallowing training system, thereby providing patients with a comprehensive, real-time adjusted personalized treatment plan. First, the personalized training program is generated based on the patient's individual characteristics (such as age, gender, medical history, muscle response, etc.) and swallowing function assessment, which ensures that each patient can receive customized training content that meets their specific needs.

[0066] The real-time feedback system continuously collects the patient's physiological data (such as muscle activity, respiratory rate, heart rate, etc.) during the training process through sensors and biomonitoring equipment, and compares the real-time data with the expected indicators through the dynamic time warping (DTW) algorithm. Through this comparison, the system can identify the deviation between the patient's actual response and the ideal training goal, and adjust the intensity, mode and frequency of the stimulation in real time. The synergistic enhancement strategy is to apply a combination of multiple stimulation modes (such as tactile, electrical stimulation, light stimulation, and sound stimulation) to training. The system analyzes the synergistic effects of different stimulation modes through a multi-objective optimization algorithm to find the best stimulation combination.

[0067] The intelligent adaptive swallowing training system integrates these functions and adaptively adjusts the training program based on the patient's real-time feedback and training progress. This means that the system can not only monitor the patient's response during training, but also dynamically adjust the stimulation mode and intensity according to the patient's status. For example, when the system detects that the patient's swallowing response is insufficient, it will increase the intensity of electrical stimulation or change the frequency of tactile stimulation; on the contrary, if the system detects that the patient reacts too strongly or is not adapted to a certain stimulation mode, it will reduce the stimulation intensity or switch to another stimulation mode.

[0068] Embodiment 1:

[0069] Combined with Figure 2 A patient named Xiao Li is a stroke rehabilitation patient who suffers from dysphagia, which causes difficulty in eating and swallowing. The therapist designed a dysphagia training program for him based on multimodal stimulation, which includes a combination of electrical stimulation, tactile stimulation, sound stimulation, and light stimulation. Since Xiao Li's swallowing reaction is slow and his reaction fluctuates in timing during different training periods, the system uses the dynamic time warping (DTW) algorithm to process these timing variations and adjust the stimulation intensity and mode of the training program in real time.

[0070] During each training session, Xiao Li's throat muscle activity is monitored in real time by sensors, and the physiological data obtained include electromyography (EMG) signals, which reflect the activity intensity of the throat muscles. At the same time, the system records the ideal training target value, which is the standard signal of muscle activity during normal swallowing. In order to capture Xiao Li's swallowing response at different time points, the system records 100 data points per second to generate a time series f(x k ), where f(x k ) represents the muscle activity data at time point k.

[0071] The system uses the DTW algorithm to process the difference between Xiao Li's real-time monitoring data and the ideal training target, using the calculation formula for the path difference:

[0072]

[0073] In this formula, f(x k ) is the muscle activity data of Xiao Li at time point k, g(y i-k ) is the standard signal of the ideal training target at the same time point. By comparing f(x k ) and g(y i-k ), the system can calculate the error between the patient's actual performance and expected performance.

[0074] Here p is the power of the norm, which is used to adjust the measurement method of different data errors. For example, if p = 2, it means that the system measures the error with Euclidean distance, and if p = 1, it means that Manhattan distance is used. This depends on the trainer's requirements for error sensitivity. Usually, the value of p ranges from 1 to 3.

[0075] γ k is the alignment adjustment factor, which determines how the system balances the cost of time alignment. It can range from 0.1 to 1.0. If γ k The larger the value, the stricter the system is in time alignment, while smaller values ​​allow for larger time deviations. |ij| q A weight term used to handle time offset, where q is the adjustment factor for time difference, and the value is generally between 1 and 2. If the time series offset is too large, the system will increase the error weight according to this term to reflect the impact of the time misalignment of the swallowing process on the training results.

[0076] For example, in a certain training session, Xiao Li at time point t 1 The muscle activity at the position lags behind the expected value by 200 milliseconds (i.e., the time difference |ij| = 0.2 seconds). k =0.5, q=1.5, p=2. Substituting into the formula:

[0077]

[0078] The calculated minimum difference path D DTW (i,j) is 0.18 (assuming that there are appropriate matching points in the training protocol), which means that Xiao Li's swallowing response has a small timing difference compared to the target, but there is a slight lag at some time points.

[0079] According to system feedback, if the difference path D DTW If (i,j) exceeds 0.2, the system will automatically adjust the training parameters. In Xiao Li's case, because his swallowing reaction lags by 200 milliseconds, the system adjusts the intensity of electrical stimulation in real time, such as increasing the current intensity from 5mA to 8mA, while shortening the time interval of electrical stimulation to help Xiao Li activate throat muscle reactions faster. At this time, the frequency of tactile stimulation will also increase accordingly, for example, from 60 times per minute to 80 times per minute, to ensure that the stimulation mode adapts to the actual needs of the patient.

[0080] During Xiao Li's swallowing training, the system detected a large difference between his swallowing target and the expected swallowing target in a certain training session through the DTW algorithm. In order to better help Xiao Li improve his swallowing response, the system decided to adjust the stimulation intensity and mode based on the calculated difference value. In this training, the system introduced an adaptive adjustment mechanism so that the stimulation intensity could dynamically respond to Xiao Li's immediate feedback to ensure the adaptability of the training.

[0081] Assume that during a certain training session, the difference between Xiao Li’s muscle activity data and the expected target is D DTW (t) = 0.3, which means that his swallowing reaction is obviously delayed. Through the sensor, the system also captures Xiao Li's immediate physiological response R(t) = 0.25 in real time, showing that his actual swallowing muscle activity is 5 units weaker than the expected index. At this time, the system needs to adjust the stimulation through an adaptive mechanism.

[0082] The adaptive adjustment formula of the system is:

[0083]

[0084] In this formula, S(t) represents the stimulus intensity adjusted as time t changes. The system first analyzes the difference D DTW (t)-R(t)=0.3-0.25=0.05, which is a small deviation, indicating that although Xiao Li's swallowing reaction is delayed, it is close to the expected goal. We substitute the specific data into the formula.

[0085] Initial stimulation intensity α=5 mA: The initial electrical stimulation intensity was set to 5 mA.

[0086] Exponential decay part: The system is based on the difference value D DTWThe exponential decay part is calculated by the difference between (t) and R(t). Set p = 2 to use the squared error metric. The integral part is calculated as:

[0087]

[0088] If the adjustment rate δ is set to 0.5, the exponential decay part is:

[0089] exp(-0.5·0.0025t)=exp(-0.00125t)

[0090] This part gradually decays as time t increases, so that the stimulation intensity is reduced when the patient's response improves, avoiding overstimulation.

[0091] During the training process, the system also took into account Xiao Li’s natural physiological cycle and introduced periodic adjustment items to simulate the natural fluctuations of the patient’s physiological state:

[0092]

[0093] Assuming the adjustment amplitude factor κ = 1.5 and the period T = 10 seconds, the periodic adjustment item is:

[0094]

[0095] Through this sine wave function, the system can adapt to Xiao Li's physiological changes and fine-tune the stimulation. The function of this item is to help the system make reasonable dynamic adjustments to the intensity of electrical stimulation when the patient's physiological state undergoes periodic changes.

[0096] Assume that Xiao Li's current training time is t = 5 seconds, substitute it into the above formula to calculate:

[0097]

[0098] The calculation results in:

[0099] S(5)=5·exp(-0.00625)+1.5·2

[0100] Finally, S(5)=4.968mA+3≈7.968mA.

[0101] Through this calculation process, the system adjusted the electrical stimulation intensity from the initial 5 mA to 7.968 mA at the 5-second moment, which gave Xiao Li's muscle activity a stronger stimulation and helped him correct the delayed swallowing reaction in time.

[0102] Following Xiao Li's dysphagia training example, after each training session, the real-time physiological data collected by the sensor, including muscle activity, electrophysiological signals, and respiratory rate, are fed back to the system through a closed-loop feedback mechanism. The system inputs this data into the machine learning model, and by analyzing these feedback data, the system can continuously and adaptively optimize the training process. The core of the closed-loop feedback mechanism is to evaluate the overall effect of the system training by weighted averaging the long-term training data, and then optimize the intensity, frequency, and pattern of the stimulation parameters.

[0103] In Xiao Li's training, the system continuously optimizes the stimulation intensity of each training session through a long-term feedback loop, adjusting in real time to suit his physiological state. In order to measure the effect of long-term training, the system uses the following optimization formula for calculation:

[0104]

[0105] In this formula, E opt It represents the long-term optimization effect of the overall training of the system. After each training, the system will summarize and optimize the performance during the training process. Specifically, each symbol in the formula has a clear physical meaning:

[0106] is the time integral term, which is used to calculate the average performance within time T, indicating the accumulation of feedback at each time point. λ is the weight coefficient of the stimulus intensity S(t), which usually ranges from 0.1 to 1 and indicates the sensitivity of the system to the stimulus intensity. A larger λ value means that the system is more sensitive to changes in stimulus intensity.

[0107] It is a weighted term for stimulus intensity, reflecting the system's adaptability to large and small stimulus intensities. Let p = 2, which means that the system pays more attention to large stimulus changes and controls the adaptability of stimulus intensity. β is the error control factor, which usually ranges from 0.1 to 0.9 and is used to balance the system's tolerance for errors. A larger β value indicates that the system is more sensitive to errors and will adjust quickly to reduce errors.

[0108] |D DTW (t)| q It is a weighted processing of the real-time error, reflecting the difference between the patient's physiological response and the expected target. q is the power of the adjustment time error, and its value range is usually 1 to 3, reflecting the different sensitivities of the system to the error.

[0109] In Xiao Li's training process, assuming that the stimulus intensity S(t) is 8mA, the difference D detected by the system DTW (t) is 0.2, and Xiao Li's swallowing reaction is slightly delayed. Substitute these data into the formula, set λ = 0.7, β = 0.5, p = 2, q = 1.5, and calculate:

[0110] First, the weighting term for stimulus intensity is calculated:

[0111]

[0112] Calculate the error weighting term:

[0113] |D DTW (t)| q =|0.2| 1.5 =0.2 1.5 ≈0.089

[0114] Substitute into the formula to calculate the long-term optimization effect E opt :

[0115]

[0116] After calculation, the long-term optimization effect E opt ≈1.9365. This shows that Xiao Li maintained a good balance between training intensity and error. The system optimized the stimulation intensity based on his immediate feedback and continued to reduce the error.

[0117] Embodiment 2:

[0118] Combined with Figure 3 Taking Xiao Li's swallowing disorder training as an example, the system introduced an adaptive multimodal neural network architecture in the subsequent stage of his rehabilitation training to dynamically adjust the order, frequency and intensity of collaborative stimulation. The neural network architecture consists of multiple layers of neurons. The input layer receives real-time monitored physiological data (such as electromyographic signals, electrophysiological signals, respiratory rate, etc.) and training feedback signals from sensors. Through the adaptive weight mechanism of the middle layer, the system can dynamically adjust the weights of various stimulation modes, and the output layer gives the optimal combination of stimulation order, frequency and intensity based on these weights.

[0119] During Xiao Li's training, the stimulation patterns included a combination of electrical stimulation, light stimulation, tactile stimulation, and sound stimulation. Each stimulation pattern has a different intensity function, which represents the intensity of the stimulation at each time point. The goal of the neural network optimization is to find the most effective stimulation combination for Xiao Li throughout the training process.

[0120] Neural network optimization is represented by the following function:

[0121]

[0122] In this formula, P opt S represents the optimal stimulation combination found within time T, ensuring that each stimulation mode has the best effect in training. 1 (t),S 2 (t),…,Sn (t) are the intensity functions of electrical stimulation, light stimulation, tactile stimulation and sound stimulation at different time points, representing the intensity they exert on Xiao Li at different time points.

[0123] α 1 ,α 2 ,…,α n is the weight coefficient of different stimulation modes, which is used to control the relative importance of different stimulation modes in the whole training process. The weight value range is usually between 0.1 and 1, indicating the degree of dependence of the system on a certain stimulation mode. A higher weight means that the mode occupies a more important position in the optimization process. The system can calculate the weighted intensity of each stimulation pattern over the entire time interval, which can fully reflect Xiao Li's overall response to different stimuli during training.

[0124] The physiological data monitored by the sensor include Xiao Li’s electromyographic signals and breathing rate at different time points during the training process. Assume that during a training session, the system detects electrical stimulation S 1 (t) The intensity at time t is 7 mA, and the light stimulus S 2 (t) is 5 Lux, tactile stimulation S 3 The intensity of (t) is 4N (Newton). These data are fed into the input layer of the neural network, and the adaptive weight mechanism in the middle layer continuously adjusts the relative weights of these stimulation patterns based on the feedback.

[0125] Assuming that in Xiao Li's current training process, the effect of electrical stimulation on him is significantly better than other modes, the system will increase the weight of electrical stimulation accordingly. 1 = 0.8, and set the weights of light stimulation and tactile stimulation to α 2 =0.4 and α 3 =0.5, reflecting the contribution of these stimulation patterns to Xiao Li’s recovery.

[0126] The system determines the optimal stimulation combination by calculating the weighted intensity of the stimulation pattern within the entire time interval T. Assuming that Xiao Li's training time is 10 minutes, that is, T = 600 seconds, the system substitutes the actual data for calculation:

[0127]

[0128] Calculate the weighted intensity of each stimulation pattern:

[0129] Electrical stimulation: 0.8 7 = 5.6

[0130] Light stimulation: 0.4·5=2

[0131] Tactile stimulation: 0.5·4=2

[0132] The total weighted intensity is:

[0133]

[0134] The integral calculation result is:

[0135] P opt =9.6·600=5760

[0136] This value P opt =5760 represents the optimal combination intensity of each stimulation mode during Xiao Li's training. Through this value, the system can determine whether the current stimulation combination is effective for Xiao Li's rehabilitation.

[0137] In the subsequent training process, the system will continuously adjust these weight coefficients according to Xiao Li's actual response (such as muscle activity, breathing rate, etc.). For example, if the effect of light stimulation gradually increases, the system will increase the weight of light stimulation α 2 It is raised to 0.6, and the weights of other stimuli are reduced accordingly, thereby dynamically adjusting the stimulation order, frequency and intensity combination.

[0138] In the previous training process, the system adjusted the weights of electrical stimulation, light stimulation, and tactile stimulation through a multimodal neural network. Now, a deep learning model is used to establish Xiao Li's individualized response model to predict his physiological responses under different stimulation combinations. This model is based on data collected from multiple training sessions and can accurately predict Xiao Li's muscle activity, breathing rate, and other physiological responses under specific stimulation combinations. Combined with the reinforcement learning algorithm, it optimizes future stimulation strategies through continuous learning and feedback.

[0139] The system first uses a deep learning algorithm to build an individualized response model for Xiao Li. This model is based on the data accumulated during the training process (such as electromyographic signals, breathing rate, heart rate, etc.), and trains a neural network to predict Xiao Li’s response to each combination of stimuli. For example, electrical stimulation S 1 (t), light stimulation S 2 (t) and tactile stimulation S 3 The intensity data of (t) is used as input, and the system predicts the physiological response under these stimulation combinations through deep learning algorithms. Assume that the model predicts that when the electrical stimulation is 7mA, the light stimulation is 5Lux, and the tactile stimulation is 4N, Xiao Li's electromyographic signal intensity increases to the expected level, which means that the current stimulation combination is effective.

[0140] After each training session, the system combines the reinforcement learning algorithm to evaluate Xiao Li's immediate training effect. Reinforcement learning optimizes the stimulation strategy through a feedback loop mechanism. That is, after each training session, the system evaluates the current stimulation combination based on Xiao Li's feedback data and optimizes the training plan by adjusting the next stimulation sequence and frequency. The reinforcement learning process can be represented by the value function V(t):

[0141]

[0142] In this formula, V(t) represents the long-term value of different stimulus combinations at time t.

[0143] S k (t) is the intensity function of the kth stimulation pattern, for example, electrical stimulation S 1 (t), light stimulation S 2 (t) and other intensities at different time points. k (t)) is the immediate reward function, which indicates the immediate effect of the stimulation mode on Xiao Li during training, such as Xiao Li's muscle activity, swallowing reaction, etc. The value of the immediate reward function is usually related to the degree of fit with the training goal. Assuming that the immediate reward r(S) for electrical stimulation of 7 mA is 1 (t)) = 0.8, the immediate reward r(S 2 (t))=0.6.

[0144] C(S k (t)) is the cost function, which reflects the negative impact or discomfort caused by the stimulation mode on Xiao Li. For example, if the electrical stimulation is too strong, it may cause discomfort. Assume that C(S 1 (t))=0.1, the tactile stimulation intensity is relatively high and may make the patient feel uncomfortable. Assuming C(S 3 (t)) = 0.2. k It is a discount factor used to balance short-term benefits and long-term benefits, and its value usually ranges from 0.8 to 1. Setting γ = 0.9 indicates that the system pays more attention to long-term benefits.

[0145] Assume that in Xiao Li's current training, the system applies electrical stimulation, light stimulation, and tactile stimulation at the same time. According to the current training situation, enter the following parameters:

[0146] Electrical stimulation 1 (t) = 7mA, light stimulation S 2 (t) = 5 Lux, tactile stimulation S 3 (t) = 4N

[0147] Reward function: r(S 1 (t)) = 0.8, r(S 2 (t)) = 0.6, r(S 3 (t))=0.7

[0148] Cost function: C(S 1 (t))=0.1、C(S 2 (t))=0.05、C(S 3 (t))=0.2

[0149] Discount factor: γ = 0.9

[0150] Substitute the reinforcement learning value function for calculation:

[0151] V(t)=max(0.9 1 [0.8-0.1]+0.9 2 ·[0.6-0.05]+0.9 3 [0.7-0.2])

[0152] Calculate the value of each item:

[0153] Electrical stimulation: 0.9 0.7 = 0.63

[0154] Light stimulation: 0.81 0.55 = 0.4455

[0155] Tactile stimulation: 0.729 0.5 = 0.3645

[0156] The total value function V(t) is:

[0157] V(t)=0.63+0.4455+0.3645=1.44

[0158] This result V(t) = 1.44 represents the long-term value of the current training stimulation combination. Through the feedback loop of reinforcement learning, the system can evaluate the effect and cost of the current stimulation combination after each training, and adjust the next stimulation order and frequency according to the calculated long-term value. For example, if the value of electrical stimulation is high and the cost of tactile stimulation is high, the system will increase the frequency or intensity of electrical stimulation in the next training, while reducing the intensity of tactile stimulation to optimize the overall effect.

[0159] In subsequent training, biofeedback monitoring was further strengthened, and real-time collection of various biological signals including Xiao Li's electromyography, heart rate, respiratory rate, and throat muscle activity was carried out. Through real-time monitoring of these biological signals, the system can fully grasp Xiao Li's physiological state during training. During each training session, these data will be preprocessed and input into the neural network model as feedback signals. The system dynamically adjusts the stimulation parameters based on these feedback signals to ensure that the stimulation mode can adapt to Xiao Li's real-time physiological state. The total utility of biofeedback is measured by the following feedback integral function:

[0160]

[0161] During the training, Xiao Li's electromyographic signal R 1 (t), heart rate R 2 (t), respiratory rate R 3 (t) and throat muscle activity R 4(t) is collected in real time. Through the sensor system, these signals are monitored 100 times per second to generate time series data. In order to ensure the stability and accuracy of the data, the system pre-processes these signals, including filtering, denoising, and standardization, so that each biological signal can accurately reflect Xiao Li's physiological state.

[0162] The system calculates the total utility F(t) of biofeedback through the feedback integral function and dynamically adjusts the stimulation parameters according to the biosignal feedback. Assuming the training time T = 600 seconds, the system monitors Xiao Li's biosignals and accumulates feedback data during the entire training. Assuming that in the current training, the electromyographic signal R 1 (t), heart rate R 2 (t), respiratory rate R 3 (t) and throat muscle activity R 4 The real-time data of (t) is as follows:

[0163] EMG 1 (t) = 0.7 (indicates the intensity of muscle activity)

[0164] Heart rate 2 (t) = 75 bpm (heart beats per minute)

[0165] Respiratory rate 3 (t) = 18 times / min

[0166] Throat muscle activity 4 (t) = 0.6 (reflects the contraction degree of throat muscles)

[0167] The system sets different importance weights β for each biological signal 1 ,β 2 ,…,β n , adjust these weights according to the contribution of the signal to the training effect. Assume:

[0168] The weight of the electromyographic signal β 1 =0.5 (Muscle activity is particularly important for the recovery of swallowing function)

[0169] Heart rate weight β 2 =0.2

[0170] The weight of breathing rate β 3 =0.1

[0171] The weight of throat muscle activity β 4 =0.4

[0172] After substituting these data and weights, the feedback utility F(t) is calculated:

[0173]

[0174] First calculate the weighted value of each signal:

[0175] EMG signal: 0.5 0.7 = 0.35

[0176] Heart rate: 0.2 75 = 15

[0177] Respiratory rate: 0.1·18=1.8

[0178] Throat muscle activity: 0.4 0.6 = 0.24

[0179] The sum is:

[0180]

[0181] The integral result is:

[0182] F(t)=17.39·600=10434

[0183] This value F(t) = 10434 represents the total utility of Xiao Li's biofeedback data within 600 seconds. The system uses this feedback utility value to evaluate whether the current training is effective for Xiao Li's physiological response.

[0184] Through the feedback utility F(t), the system can identify the impact of different biological signals on the training effect. If the feedback utility of the electromyographic signal is found to be high, indicating that Xiao Li's muscle activity has increased in the current training, the system may maintain or increase the intensity of electrical stimulation. For example, if the feedback utility f(t) shows that the weight of the throat muscle activity signal is low, indicating that the recovery of this part is not ideal, the system may increase the intensity or frequency of tactile stimulation to help activate the relevant muscles. By dynamically adjusting these stimulation modes, the system ensures that the training can adapt to Xiao Li's physiological needs in real time.

[0185] In the following training, the system will continue to monitor and accumulate Xiao Li's biofeedback data. After each training session, the result of the feedback integral function f(t) will be stored and used to optimize the next training program. If long-term fluctuations in heart rate and respiratory rate are detected, the system may adjust the order of electrical stimulation or reduce excessive stimulation to avoid discomfort to Xiao Li.

[0186] Then, a dynamic optimization method based on the benefit-cost trade-off model was introduced to balance the therapeutic effect of stimulation with Xiao Li's physiological tolerance. The system makes real-time adjustments according to the benefit-cost ratio during each training process, and evaluates the overall effect of the training through a long-term optimization formula. The purpose of this method is to ensure that the rehabilitation training for dysphagia is effective while minimizing the discomfort or negative impact of stimulation on the patient.

[0187] The system uses the following long-term optimization formula to measure the training effect:

[0188]

[0189] E opt Represents the long-term optimization effect, which is the overall effect calculated after multiple trainings. opt (t) is the stimulation combination function, which indicates the combination of stimulation modes selected at each time point, such as the combination of electrical stimulation, light stimulation, and tactile stimulation. D(t) is the degree of discomfort or cost function of the patient during the training process, which indicates the negative effects of the stimulation mode on the patient, which may be pain, fatigue, etc. caused by excessive stimulation.

[0190] λ and μ are the weight coefficients of benefit and cost respectively. · ranges from 0.5 to 1.5, which is used to indicate the system's attention to the treatment effect, while μ ranges from 0.1 to 1, which is used to balance the negative effects of stimulation. p and q are adjustment parameters for controlling the sensitivity of stimulation effect and cost respectively. Usually, p ranges from 1 to 3, and q ranges from 1 to 2. Integration Used to calculate the average effect within the training time T.

[0191] During Xiao Li's training, the system continuously adjusted the intensity combination of electrical stimulation, light stimulation, and tactile stimulation. Assume that in one training session, the intensity of electrical stimulation S 1 (t) = 6 mA, the intensity of light stimulation is S 2 (t) = 4 Lux, the intensity of tactile stimulation is S 3 (t) = 3N. The system inputs these data into the stimulus combination function P opt (t), and calculate the overall stimulus benefit.

[0192] Assume that the discomfort D(t) during training is 0.2, which means that Xiao Li feels some discomfort (such as muscle tension or slight pain) during training. The system sets the benefit weight λ=1.2, the cost weight μ=0.5, and the adjustment parameters p=2 and q=1.5.

[0193] Substitute into the formula for calculation, first calculate the stimulus combination P opt The weighted value of (t):

[0194]

[0195] Then, calculate the inappropriate cost term:

[0196] D(t) q =0.2 1.5 =0.089

[0197] Finally, substitute these values ​​into the optimization formula to calculate the long-term optimization effect E opt :

[0198]

[0199] Simplified calculation:

[0200]

[0201] The integral calculation result is:

[0202] E opt =4.2815·T

[0203] This result E opt This means that during long-term training, Xiao Li's training effect has reached a good balance. The system weighs the benefits and costs to ensure that while ensuring the effectiveness of the training, Xiao Li's discomfort during the training is minimized.

[0204] During the training process, the system will adjust the order, frequency and intensity of the stimulation combination according to the benefit-cost ratio monitored in real time. If Xiao Li's discomfort increases at a certain point in time (for example, D(t) rises to 0.4), the system will automatically reduce the intensity of electrical stimulation or reduce the frequency of tactile stimulation to alleviate the discomfort. At the same time, the system will keep the intensity of light stimulation unchanged, because it does not significantly increase Xiao Li's negative feelings while improving the training benefits. For example, if in the next training, the intensity of electrical stimulation is adjusted to 5mA and the intensity of tactile stimulation is reduced to 2N, after substituting into the formula, the system will recalculate the optimization effect and continuously optimize according to the new training situation.

[0205] After many training sessions, the system has accumulated a large amount of data about Xiao Li's physiological reactions, and can better predict Xiao Li's reactions and feelings under different stimulation combinations. The data after each training session is stored and used to further optimize the training program. Through the feedback loop mechanism, the system gradually finds the best training program that can effectively promote Xiao Li's swallowing function recovery without causing him too much discomfort. As the training progresses, the system may reduce the frequency of certain stimulations or increase the intensity of stimulation according to Xiao Li's recovery progress. For example, light stimulation may be increased to 6Lux and electrical stimulation maintained at 5mA to ensure that Xiao Li's training progresses smoothly and effectively.

[0206] Embodiment 3:

[0207] Swallowing is a complex neural reflex and muscle coordination process involving multiple levels of cerebral cortex, brainstem and ganglion control, including precise control of tongue, pharyngeal, laryngeal and cricopharyngeal muscles. Dysphagia is usually caused by neurological diseases (such as stroke, Parkinson's disease, cerebral palsy, etc.) or head and neck injuries, resulting in poor or complete loss of neuromuscular coordination, and patients are unable to effectively complete swallowing movements.

[0208] Combination Figure 4 AD in the figure shows the innervation of the human pharyngeal plexus (SPC, MPC, IPC and CP) by the X (vagus nerve) and IX (glossopharyngeal nerve) nerves through Sihler staining; the vagus nerve branches in the larynx and their innervation of the laryngeal muscles, especially the distribution of the recurrent laryngeal nerve (RLN) and the superior laryngeal nerve (SLN); the anatomical structure of the tongue muscles and their nerve branches, especially the lingual nerve (LN), the hypoglossal nerve (XII) and the lingual branch of the glossopharyngeal nerve (IX), which play a key role in the motor control of the tongue muscles. The in-depth analysis of these neuromuscular structures provides a theoretical basis for the optimization of multimodal stimulation for dysphagia training.

[0209] Combined with the above neuroanatomical information, the swallowing disorder training method proposed in the present invention utilizes multimodal stimulation such as electrical stimulation, light stimulation, tactile stimulation and sound stimulation, combined with an intelligent feedback system, and gradually restores the patient's swallowing function through a personalized dynamic adjustment strategy. The following are the steps and explanations of the specific training program.

[0210] Figure 4 Figure A shows the innervation of the pharynx, specifically the pharyngeal branches of the X and IX nerves, which innervate the superior pharyngeal constrictor (SPC), middle pharyngeal constrictor (MPC), inferior pharyngeal constrictor (IPC), and cricopharyngeus (CP), respectively. These pharyngeal muscles are responsible for pushing the food bolus down into the esophagus during swallowing and closing the airway to prevent food from entering the trachea. Figure 4 B and Figure 4 Figure C shows the pathological state of the X nerve and Ph-X nerve in Parkinson's disease patients during dysphagia. It was found that PAS immunoreactivity was found in the nerve fibers of these patients, indicating pathological changes in the nerve fibers. Figure 4 D in the figure further demonstrates the atrophy of the inferior pharyngeal constrictor (IPC) in patients with Parkinson's disease. This muscle fiber atrophy can seriously affect swallowing function.

[0211] In this training program, electrical stimulation is used to activate these pharyngeal nerves, especially the innervated branches of the Ph-X nerve and the X nerve, thereby improving the contractile strength of the pharyngeal constrictor and cricopharyngeal muscles. The present invention recommends the use of low-frequency electrical stimulation (about 30Hz-60Hz) to directly act on the Ph-X nerve branch area through subcutaneous electrodes to activate the superior pharyngeal constrictor (SPC) and the middle pharyngeal constrictor (MPC) to enhance their contractile ability. During the electrical stimulation process, the patient's biological signals (such as electromyographic signals, activity signals of throat muscles, etc.) will be monitored in real time by sensors and sent to the neural feedback system. The changes in these signals are analyzed in real time by the dynamic time warping (DTW) algorithm and compared with the expected training goals. The system adjusts the intensity and frequency of electrical stimulation according to the neural response.

[0212] For a Parkinson's patient whose pharyngeal constrictor muscles had already experienced slight atrophy, moderate-intensity electrical stimulation was used in the initial training phase, with a current of 5mA and a frequency of 50Hz for 10 minutes. The biofeedback system monitored the electromyographic activity and found that the contraction amplitude of SPC and MPC was lower than the expected target (the difference was about 20%). The system automatically adjusted the current to 7mA and increased the stimulation frequency to 60Hz in real time. After 30 minutes of training, the activity of the patient's pharyngeal constrictor muscles returned to near normal levels, and the strength of muscle contraction increased by 15%. This result shows that electrical stimulation can effectively promote the recovery of pharyngeal constrictor muscles through real-time feedback and dynamic adjustment.

[0213] Figure 5 Figure A shows the branch structures of the superior laryngeal nerve (SLN) and recurrent laryngeal nerve (RLN) of the vagus nerve. The internal and external branches of the SLN innervate the mucosa and cricothyroid muscle of the larynx, respectively, while the RLN innervates the remaining laryngeal muscles (such as the interarytenoid muscle and thyroarytenoid muscle). Figure 5 C and Figure 5 D in the figure shows the distribution of nerve endings in the laryngeal mucosa, especially the distribution of ISLN and sensory nerve endings in the epiglottis and postcricoid region. The dense innervation of these areas suggests that they play an important role in the swallowing reflex and laryngeal protection mechanism.

[0214] The training program of the present invention realizes precise control of laryngeal muscles by electrical stimulation of SLN and RLN branches, combined with light stimulation and sound stimulation. Electrical stimulation acts on the cricothyroid muscle (CT) and interarytenoid muscle (IA) through non-invasive electrodes to enhance the contraction force of laryngeal muscles and prevent food from entering the trachea and causing choking and coughing. Light stimulation stimulates the sensory nerve endings of the larynx through LED light sources of specific frequencies to stimulate the swallowing reflex. Sound stimulation stimulates the laryngeal sensory plexus through sounds of specific frequencies to enhance the patient's perception of the swallowing action.

[0215] A patient with post-stroke dysphagia, during training, the system electrically stimulated the RLN and SLN at the same time, with a current of 6mA, light stimulation of blue light with a wavelength of 480nm, and a sound stimulation frequency of 2000Hz. The biofeedback sensor detected that the patient's laryngeal closure function was weak, and the airway closure was delayed during swallowing, resulting in mild choking. Based on the feedback, the system increased the electrical stimulation intensity of the RLN to 8mA and extended the time of sound stimulation. After 15 minutes of adjustment training, the patient's choking symptoms were significantly reduced and the laryngeal closure function was improved.

[0216] Figure 6 , 7 The anatomical relationship between the tongue and the laryngeal and pharyngeal structures is demonstrated, especially the branch distribution of the hypoglossal nerve (XII nerve) and the lingual nerve (LN), which innervate the hyoglossus (HG), styloglossus (SG) and inferior longitudinal (IL) muscles respectively. Figure 7The innervation pattern of the tongue is shown, with the key roles of the lingual nerve and the XII nerve in tongue muscle movement, making the tongue crucial for food propulsion during swallowing.

[0217] The present invention combines electrical stimulation and tactile feedback to activate the tongue muscles and restore the tongue muscles' ability to control and push food. Through electrical stimulation of the hypoglossal nerve, the system enhances the coordinated contraction function of the tongue muscles, and combines tactile stimulation and biofeedback to allow patients to perceive swallowing movements through multiple senses simultaneously during training. The feedback system monitors the movement state of the tongue muscles and dynamically adjusts the intensity of electrical stimulation according to training feedback. For a post-stroke patient with tongue muscle weakness, the initial electrical stimulation intensity was 4mA, 40Hz, and lasted for 10 minutes. Tactile feedback is attached to the patient's tongue through a sensor to record its movement trajectory. After training, it was found that the contraction amplitude of the tongue muscles was lower than expected (the difference was 25%), and the system automatically increased the current to 6mA and increased the training time to 15 minutes. After 5 days of training, the tongue muscle strength and food pushing ability were significantly improved, and the patient showed obvious progress in swallowing solid food.

[0218] The present invention can accurately design a personalized dysphagia rehabilitation training program based on multimodal stimulation. The method systematically restores the patient's pharyngeal, larynx, and tongue neuromuscular coordination function through real-time feedback, dynamic adjustment, and a combination of multimodal stimulation (electricity, light, touch, and sound).

Claims

1. An optimization method for dysphagia training based on multimodal stimulation, characterized in that The following steps are involved: S1. Generation of personalized training programs for multimodal stimulation: S1.

1. Collect the patient's basic physiological data including age, gender, medical history and real-time swallowing data including throat muscle activity, electrophysiological signals, and swallowing pressure; S1.2, using multimodal data fusion technology, the real-time swallowing data of S1.1 is fused and processed to generate a personalized swallowing function assessment report; S1.3, and based on the evaluation results, generate a personalized dysphagia training program through an intelligent algorithm, which includes a combination of multiple stimulation methods such as tactile, electrical stimulation, light stimulation and sound stimulation; S2. Synchronous optimization and real-time feedback of multi-mode stimulation: S2.

1. Build a real-time data collection and feedback system, using sensors to monitor the multimodal data of patients during training, including muscle activity, sound signals, facial expressions, and breathing rate; S2.2, through the dynamic time warping (DTW) algorithm, the real-time monitoring data is compared with the expected indicators in the training plan, and the intensity and mode of stimulation are adjusted in real time; S2.

3. Use machine learning algorithms to analyze patients’ responses to different stimulation patterns and optimize training programs. S3. Synergistic enhancement strategy of multimodal stimulation: S3.

1. Introduce a synergistic enhancement algorithm for multimodal stimulation to analyze the synergistic effects of different stimulation modes including tactile and electrical stimulation, light stimulation and sound stimulation; S3.

2. Use a multi-objective optimization algorithm to dynamically adjust the order, frequency, and intensity of synergistic stimulation based on individual patient characteristics and training feedback; S4, Integration of Intelligent Adaptive Swallowing Training System: S4.

1. Integrate personalized training programs, real-time feedback systems, and collaborative enhancement strategies into intelligent adaptive swallowing training systems; Adaptively adjust training plans and stimulation modes based on patients’ real-time feedback and training progress; The method for adjusting the intensity and mode of stimulation in real time includes: using the DTW algorithm to deal with the temporal variability of the patient during swallowing to process the time offset, calculating the difference between the real-time data and the ideal training target to adjust the training plan; the calculation formula of the path difference is as follows: Among them, D DTW (i, j) represents the minimum difference path between the patient's real-time monitoring data and the ideal training index; k represents the kth stimulation mode, n represents the total number of stimulation modes; f(x k ) and g(y i-k ) represent the real-time physiological data of the patient's throat muscle activity and the expected training index, respectively, and are functions of the time series, reflecting the deviation between the patient's physiological signals at different time points and the ideal target; Among them, p a is the power of the norm, which is used to adjust the measurement method of different data errors; γ k is the alignment adjustment factor, which is used to balance the alignment costs between different time periods; is the weight term for processing time offset, controlling the influence of time interval, where q a It is an important factor in adjusting time difference; integral It means that the deviation of the time series is accumulated within the entire time range T, and the differences in each time period are weighted; The method for adjusting the intensity and mode of stimulation in real time includes: when the DTW algorithm detects a deviation between the patient's real-time data and the expected training index, adjusting the intensity and mode of stimulation according to the calculated difference value; introducing an adaptive adjustment mechanism to make the stimulation intensity respond to the patient's immediate feedback; and determining whether to increase or decrease the stimulation intensity by analyzing the size of the difference value; The method for adjusting the intensity and mode of stimulation in real time includes: introducing a closed-loop feedback mechanism; the feedback data after each training is input into a machine learning model, and the subsequent training parameters are optimized by analyzing the feedback information; the closed-loop feedback mechanism analyzes and optimizes the long-term training data by weighted averaging; the long-term training effect is measured by an optimization formula: Among them, E opt-train It is a measure of the overall optimization effect, obtained by weighted averaging the long-term training effect; is the time integration operation, which is used to calculate the average performance within the time range T; a is the weight coefficient of stimulus intensity S(t), reflecting the sensitivity to stimulus intensity. Controls the adaptation to different stimulus intensities, p b It is used to adjust the weighted processing of large and small stimuli; β is the error control factor, which is used to balance the tolerance to errors. is a weighted processing of the real-time error, by q b Power operation.

2. The optimization method for dysphagia training based on multimodal stimulation according to claim 1, characterized in that The method for dynamically adjusting the order, frequency and intensity of collaborative stimulation includes: introducing an adaptive multimodal neural network architecture, wherein the network architecture is composed of multiple layers of neurons, the input layer receives physiological data and training feedback signals from multiple sensors, the middle layer adjusts the weights of various stimulation modes in real time through an adaptive weight mechanism, and the output layer gives a combination of stimulation order, frequency and intensity.

3. The optimization method for dysphagia training based on multimodal stimulation according to claim 2, characterized in that The method for dynamically adjusting the order, frequency and intensity of synergistic stimulation includes: establishing an individualized response model of the patient through a deep learning model, predicting the physiological response under different stimulation combinations according to the specific response of each patient; and combining a reinforcement learning algorithm; reinforcement learning optimizes the stimulation strategy through a feedback loop mechanism, evaluates the current training effect after each training, and adjusts the stimulation order and frequency of the next training through the value function of reinforcement learning; the optimization process is represented by the value function of reinforcement learning: Where V(t) represents the long-term value of different training stimulus combinations at time t; S k (t) is the intensity function of the kth stimulation pattern at time t, indicating the performance of the combination of different stimulation patterns in training; r(S k (t)) is the immediate reward function, which reflects the effect of the stimulation mode and the immediate response effect of the patient during training; C(S k (t)) is the cost function, which represents the negative impact or discomfort of the stimulation pattern on the patient; γ k is a discount factor used to balance short-term and long-term benefits.

4. The optimization method for dysphagia training based on multimodal stimulation according to claim 3, characterized in that The method for dynamically adjusting the sequence, frequency and intensity of synergistic stimulation includes: real-time monitoring of the patient's biological signal data including electromyographic signals, heart rate, respiratory rate, and throat muscle activity, and sending the biological signal data into the neural network model as input feedback signals after preprocessing; and dynamically adjusting the stimulation parameters according to the feedback signals.

5. The optimization method for dysphagia training based on multimodal stimulation according to claim 4, characterized in that The method for dynamically adjusting the order, frequency and intensity of synergistic stimulation includes: optimizing the training effect based on a benefit-cost trade-off model, making real-time adjustments according to the ratio of benefit to cost in each training, and evaluating the training effect through a long-term optimization formula: Among them, E opt-all It is the long-term optimization effect, which means the overall effect after multiple trainings; P opt (t) is the stimulus combination function, which represents the stimulus pattern combination selected at each time point; C(S k (t)) is the patient's discomfort level or cost function during the training process, indicating the negative effect of the stimulation mode on the patient; λ b and μ are the weight coefficients of benefit and cost, which are used to balance the therapeutic effect of stimulation and the physiological tolerance of the patient; p c and q c To adjust the parameters, control the sensitivity to different levels of stimulation effects and costs; integral Used to calculate the average effect over a long period of time.

Images