Pelvic floor muscle rehabilitation accurate training and evaluation method fusing myoelectricity and posture recognition

Abstract

The invention discloses a pelvic floor muscle rehabilitation precise training and evaluation method fusing myoelectricity and posture recognition, and particularly relates to the technical field of intelligent rehabilitation, electromyographic signals and motion posture data are synchronously collected through a multi-channel sensor, and a personalized rehabilitation demand model is constructed after real-time filtering; the model can synchronously output a multi-dimensional user state vector representing the real-time physiological state of a user and a dynamic adjustment factor used for quantitatively adjusting the training intensity; training intensity and modes are automatically adjusted based on model output, and meanwhile real-time posture correction and power generation guidance are provided for the user through multi-mode interaction; recording and analyzing whole-course training data, generating an effect evaluation report and continuously optimizing a training scheme; closed-loop rehabilitation training from state sensing, intelligent decision making to accurate execution is achieved, and the problems that in traditional pelvic floor muscle rehabilitation training, the training effect is poor and evaluation is inaccurate due to the fact that individual differences are large and feedback is inaccurate are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent rehabilitation technology, and more specifically, to a method for precise training and assessment of pelvic floor muscle rehabilitation that integrates electromyography and posture recognition. Background Technology

[0002] With the continuous development of modern rehabilitation medicine, the treatment of pelvic floor muscle dysfunction has gradually received widespread attention from the medical community and society. Existing pelvic floor muscle rehabilitation training largely relies on fixed training equipment and standardized training models. While these can assist patients in muscle recovery to some extent, they have significant shortcomings in terms of individual adaptability and real-time response. Different patients exhibit significant differences in electromyographic signal characteristics, muscle strength, and postural control abilities. Traditional equipment often uses a uniform training intensity and frequency, making it difficult to achieve targeted training adjustments based on individual differences. Furthermore, the lack of comprehensive recognition and analysis of real-time physiological signals during training makes it impossible to effectively assess the patient's immediate training status and provide dynamic feedback during training, resulting in unstable and delayed rehabilitation outcomes.

[0003] In response to the above situation, existing technologies still have significant problems in the intelligent adjustment of personalized rehabilitation training. Traditional methods, after collecting electromyography and posture information, usually only perform single-dimensional analysis, failing to model the correlation between different physiological signals and thus unable to accurately depict the patient's comprehensive physiological state during training. Secondly, the generation of training programs often relies on human experience or preset templates, lacking a data-driven personalized optimization mechanism, making it difficult to achieve adaptive adjustments to training intensity, mode, and duration. Therefore, there is an urgent need for an intelligent rehabilitation training and assessment method that can integrate electromyography signals and posture recognition data to achieve dynamic adjustment and individualized optimization of the training process, thereby improving the scientific rigor and precision of pelvic floor muscle rehabilitation. Summary of the Invention

[0004] To overcome the aforementioned deficiencies in the prior art, this invention provides a precise training and assessment method for pelvic floor muscle rehabilitation that integrates electromyography and posture recognition. By constructing a personalized rehabilitation needs model that can simultaneously output user state vectors and dynamic adjustment factors, and based on this model, automatic adjustment of training parameters and multimodal real-time guidance are achieved. At the same time, a continuous optimization mechanism for the training program based on full-process data tracking is established to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Step S1 Signal Acquisition and Real-time Data Synchronization: Through the synchronized operation of the multi-channel electromyography sensor and posture sensor, the user's electromyography signals and motion posture data are acquired in real time, and noise is removed through real-time filtering; Step S2: Dynamic Data Analysis and Model Building: Using machine learning algorithms, based on real-time collected electromyography and posture data, combined with the user's historical rehabilitation data and physiological parameters, a personalized rehabilitation needs model is established. Step S3: Training Intensity and Mode Adjustment: Based on the personalized rehabilitation needs model, perform rehabilitation exercises, record the user's real-time electromyography and postural characteristics, dynamically evaluate the user's electromyography signals and postural data, output a quantified user state vector and dynamic adjustment factor, and automatically adjust the intensity, mode and duration of training. Step S4 Real-time Feedback and Training Adjustment Strategy: Based on the user state vector output by the rehabilitation needs model, specific posture correction and force exertion guidance are provided to the user in real time through voice, visual or tactile interfaces to guide the user to make self-adjustments in order to make up for the shortcomings of purely automatic adjustment in the neuromuscular re-education process. Step S5: Full Data Recording and Effect Evaluation: Record the electromyographic signals, posture data, and training effects throughout each training session and store them in the cloud.

[0006] Preferably, the process of establishing a personalized rehabilitation needs model includes: Sub-step 201 Data feature extraction: Obtain noise-removed electromyographic signals and motion posture data, extract key electromyographic waveform features through waveform feature extraction algorithm and standardize them, and extract the user's action angle and motion frequency based on the posture sensor; Sub-step 202 Feature selection and correlation analysis: Screen and analyze the extracted features; evaluate the relationship between each feature and the training effect through machine learning algorithms, and select the features that are associated with the prediction of the training effect. Sub-step 203 Model training and parameter optimization: After feature selection is completed, the rehabilitation needs model is trained; machine learning algorithms are used to learn from the collected rehabilitation data to generate a model that can predict the user's training status. Sub-step 204 Personalized training plan generation: Based on the training output results, a personalized training plan will be generated for each user; the training plan sets the training intensity, frequency and duration according to the user's physiological characteristics and historical rehabilitation data, and can also be dynamically adjusted through real-time feedback.

[0007] Preferably, the user state vector is a multi-dimensional quantized array used to comprehensively represent the user's real-time physiological state, and it includes at least the following three dimensions of evaluation indicators: Fatigue Index: The user's current level of muscle fatigue calculated based on changes in the frequency components and amplitude decay trend of electromyographic signals; Force efficiency score: The score is calculated based on the matching degree between the current electromyographic signal amplitude and the target force level, as well as the standardization of muscle activation patterns. Postural compliance: An assessment of user movement conformity calculated based on the deviation between real-time posture data and standard rehabilitation movement templates.

[0008] Preferably, the quantification formula for the dynamic adjustment factor is calculated as follows: Adjustment factor = w1 × (current EMG amplitude sliding average / user's personal baseline amplitude) + w2 × (1 - normalized root mean square error of posture angle) + w3 × (real-time fatigue coefficient calculated based on the slope of EMG amplitude decrease during this training); where the weight coefficients w1, w2, and w3 are trained using the ridge regression algorithm to learn the optimal weight combination that can predict the training effect. After each training session, the dynamic adjustment factor is automatically calculated, and the intensity of the next training cycle is scaled up or adjusted stepwise according to the preset segmented decision rules.

[0009] Preferably, the output of step S3 includes a recommended intensity level, optimal training frequency, and maximum duration of a single training session; a linear regression model based on online learning is used, with the maximum value and standard deviation of the user's historical electromyography (EMG) signal amplitude as input features, to output the recommended intensity level, and the regression coefficient is dynamically updated by the residual between the real-time EMG amplitude and the predicted value; a long short-term memory (LSTM) network model is used to analyze the time-series data of the pelvic movement trajectory within the user's continuous training cycle, predict the muscle function recovery curve, and thus output the optimal training frequency; an endurance decay model is established using an exponential decay function, and the maximum duration of a single training session is calculated by combining the decay constant determined by nonlinear curve fitting and the safety factor dynamically adjusted based on the real-time fatigue coefficient.

[0010] Preferably, step S3 includes the following: Sub-step 301: Match the user state vector output by the personalized rehabilitation needs model with the pre-set training mode knowledge base; by calculating the similarity between the user state vector and the ideal state vector of each mode in the knowledge base, select the most suitable basic training mode for the current user, and load the baseline intensity, baseline frequency and baseline duration parameters corresponding to the mode. Sub-step 302: Fine-tune the training parameters in real time based on the dynamic adjustment factor, multiply the baseline training intensity by the dynamic adjustment factor to obtain the final training intensity value; when the value of the dynamic adjustment factor is continuously lower than the preset safety threshold, automatically switch the training mode to a lower intensity mode to ensure training safety.

[0011] Preferably, it includes a proactive safety intervention mechanism that continuously monitors the fatigue index in the user's state vector. When it is predicted that the fatigue index will exceed the safety threshold in subsequent training periods, it intervenes in advance and actively reduces the training intensity.

[0012] Preferably, step S4 includes the following: Sub-step 401: Guidance requirement identification and priority ranking: Analyze the deviation between the values ​​of each dimension in the user state vector and the preset threshold to identify the problems that need to be corrected; prioritize all identified problems, and prioritize the problems with the highest security risks and the most significant impact on training results as guidance content. Sub-step 402 Personalized guidance content generation and multimodal mapping: Based on the guidance priority determined in sub-step 401, select the most matching correction instruction template from the pre-set guidance corpus; combine the abstract instruction template with the user's specific real-time data to generate concrete guidance content; assign appropriate feedback modalities to the guidance content according to the problem type and training scenario, such as using tactile vibration prompts for key posture deviations, using voice prompts for insufficient force exertion skills, and using visual charts for training status overview; Sub-step 403 Guidance effect evaluation and dynamic adjustment of feedback: After issuing guidance, continuously monitor the subsequent user state vector to evaluate the guidance effect; if the relevant indicators do not improve, the feedback intensity will be increased, for example, by superimposing voice reminders on top of visual cues; conversely, if the indicators return to normal, positive confirmation will be given and the feedback frequency will be reduced to avoid information overload.

[0013] Preferably, when generating guidance content, multiple feedback modalities are triggered simultaneously when a complex error pattern is identified, thereby accelerating the user's understanding and correction process by stimulating multiple senses in a coordinated manner. A dynamic content density adjustment mechanism is introduced during the guidance process to calculate the user's cognitive load level in real time. When it is detected that the user's state vector shows signs of confusion or fatigue after receiving multiple guidance messages in a row, non-critical guidance will be automatically paused, and only the most essential safety prompts will be retained.

[0014] Preferably, step S5 includes the following: Sub-step 501 Multi-source data aggregation and standardization preprocessing: The electromyography signal waveforms, posture sensor raw data, user state vectors and dynamic adjustment factors output in real time by the rehabilitation demand model, and all training parameter adjustment records generated during a single training session are timestamped and aggregated. The aggregated data is then standardized and cleaned to eliminate the influence of device differences and abnormal fluctuations, forming a structured training session data package. Sub-step 502 Multi-dimensional rehabilitation effect assessment and report generation: The training effect is quantitatively assessed from three dimensions: short-term dimension: analysis of the improvement curve of electromyographic amplitude stability and postural compliance in a single training session; medium-term dimension: comparison of the mean changes of key indicators in this training and recent historical training; long-term dimension: tracking the gap between the user's baseline level and rehabilitation goals, and generating an assessment report including trend charts and improvement suggestions. Sub-step 503 Personalized training program closed-loop optimization: The training effect evaluation results are correlated with the historical data of the user's state vector to identify effective training modes and parameter combinations; when it is found that a user's training effect in a specific fatigue index range is consistently better than in other ranges, the parameters of the user's rehabilitation demand model will be automatically adjusted to make it more inclined to recommend training intensity and mode near this range, so as to realize the self-evolution of the training program.

[0015] The technical effects and advantages of this invention are as follows: This invention constructs a personalized rehabilitation needs model that can synchronously output user state vectors and dynamic adjustment factors. Based on this model, it realizes automatic adjustment of training parameters and real-time user guidance, achieving closed-loop intelligent rehabilitation from data perception to decision execution. It synchronously collects and fuses multi-channel electromyographic signals and posture data, and transforms the raw data into a multi-dimensional state assessment including fatigue index, force exertion efficiency score, and posture compliance through the rehabilitation needs model. This generates dynamic adjustment factors for quantifying and adjusting training intensity, effectively solving the problems of single training modes and delayed adjustments caused by the lack of deep understanding of the user's physiological state in traditional rehabilitation equipment. It ensures that every training decision is based on multi-dimensional real-time physiological state assessment.

[0016] This invention further establishes a closed-loop iterative mechanism for continuous self-optimization of training programs through full-process data tracking and multi-dimensional effect evaluation. It records and stores electromyographic signals, posture data, user state vectors, and training parameters generated during each training session in the cloud. By comparing and analyzing short-term training indicator changes with long-term rehabilitation trends, it automatically generates an effect evaluation report containing improvement suggestions. This effectively solves the problem of traditional rehabilitation training programs relying on subjective experience and lacking long-term continuous optimization, ensuring that the entire rehabilitation process proceeds along the optimal path and significantly improving the scientific nature of rehabilitation training. Attached Figure Description

[0017] Figure 1 This is a flowchart of the pelvic floor muscle rehabilitation precision training and assessment method integrating electromyography and posture recognition of the present invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0020] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0021] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0022] See Figure 1 The present invention provides a flowchart of a precise training and assessment method for pelvic floor muscle rehabilitation that integrates electromyography and posture recognition. Figure 1 The pelvic floor muscle rehabilitation precision training and assessment method integrating electromyography and posture recognition, as shown, includes: Step S1: Signal Acquisition and Real-Time Data Synchronization By synchronously working with multi-channel electromyography (EMG) sensors and posture sensors, the user's EMG signals and motion posture data are collected in real time. Noise is removed through real-time filtering to obtain high-quality, synchronous multimodal physiological and motion data, providing an accurate and reliable data foundation for subsequent analysis. To further elaborate, during data preprocessing, a dynamic filtering technique based on "user physiological feedback" was used to reduce the impact of individual differences in the signal on the results. By adjusting the feedback of the user's physiological state before and after each training session, the filtering parameters were optimized, enabling signal processing to better adapt to the user's physical condition. This not only improved the accuracy of the signal but also reduced errors caused by differences between users, ensuring the quality and accuracy of the data. Step S2: Dynamic Data Analysis and Model Building By using machine learning algorithms, based on real-time collected electromyography and posture data, combined with users' historical rehabilitation data and physiological parameters, a personalized rehabilitation needs model is established to identify the training status and physical characteristics of different users and generate a scientific training mode. Step S3: Training Intensity and Mode Adjustment Based on a personalized rehabilitation needs model, rehabilitation exercises are performed. After recording the user's real-time electromyography and postural characteristics, the user's electromyography signals and postural data are dynamically evaluated. A quantified user state vector and dynamic adjustment factor are output, and the intensity, mode and duration of training are automatically adjusted to ensure that the training meets the user's rehabilitation needs and avoid inappropriate training intensity. Step S4: Real-time feedback and training adjustment strategy: The user state vector output by the rehabilitation needs model is used to generate and provide specific posture correction and force exertion guidance to users in real time through voice, visual or tactile interfaces, guiding users to make self-adjustments to make up for the shortcomings of purely automatic adjustment in the neuromuscular re-education process; human-computer interaction guidance is provided, based on the output of the rehabilitation needs model, to provide users with real-time and operable posture or force exertion suggestions to guide users to make self-corrections. Step S5: Full Data Recording and Effect Evaluation The electromyographic signals, posture data, and training effects of each training session are recorded and stored in the cloud. Detailed effect evaluation reports are generated through multi-dimensional data analysis. Based on this, personalized training programs are continuously optimized to ensure the long-term effectiveness of rehabilitation training.

[0023] Furthermore, the process of establishing a personalized rehabilitation needs model includes: Sub-step 201 Data feature extraction: Obtain noise-removed electromyographic signals and motion posture data, extract key electromyographic waveform features through waveform feature extraction algorithms and standardize them, such as amplitude and frequency, and extract the user's action angle and motion frequency based on the posture sensor; Sub-step 202 Feature selection and correlation analysis: Screen and analyze the extracted features; evaluate the relationship between each feature and the training effect through machine learning algorithms, and select features that are associated with the prediction of training effect; for example, the amplitude and frequency of electromyographic signals and the balance and stability in posture data will be important bases for model input; explore the intrinsic correlation between these features and identify the improvement effect of different feature combinations on training effect. In this invention, machine learning algorithms are used to evaluate the relationship between each feature and the training effect. The training effect is a multi-dimensional indicator involving feedback from multiple physiological and motor levels, such as muscle strength, postural stability, and movement accuracy. Supervised learning methods such as regression analysis, decision tree algorithms, and support vector machines are used to extract key features (such as amplitude and frequency of electromyography signals, and balance and stability in posture data) from real-time collected electromyographic signals and posture data, and to quantify the impact of each feature on different training effects. In this process, feature selection techniques (such as L1 regularization, information gain method, and recursive feature elimination method) are used to screen out the features most relevant to the training effect, thereby optimizing the model's predictive ability. Specifically, by weighting the contributions of different features, features that have a significant effect on the prediction of training effect are selected, thereby establishing a personalized training model with high predictive accuracy.

[0024] To further enhance the model's accuracy, a multi-level feature selection method based on "physiological feedback optimization" was adopted in the feature selection stage. Combined with deep learning algorithms, the high-dimensional feature combinations under different physiological states of the user were automatically identified by analyzing the hierarchical relationship between features over multiple training cycles, ensuring that the input features in each training stage best reflect the user's physiological response. Sub-step 203 Model Training and Parameter Optimization: After feature selection, the rehabilitation needs model is trained; machine learning algorithms (such as deep neural networks, support vector machines, etc.) are used to learn from the collected rehabilitation data to generate a model that can predict the user's training status; during the training process, the model parameters are continuously adjusted so that each round of training can be optimized based on the user's real-time data, ensuring that the training suggestions output by the model meet the personalized needs of each user. The explanation explains that the parameters of the adjusted model include the weights, bias terms, and learning rate in the neural network, as well as the kernel function parameters and penalty factor in the support vector machine. These parameters directly affect the model's response to input data and the accuracy of prediction results. During each training round, based on real-time user feedback data (such as electromyography signals, posture data, fatigue levels, etc.), these parameters are automatically adjusted using a progressive optimization strategy employing Bayesian optimization or grid search. For example, the learning rate is dynamically adjusted using an adaptive learning rate adjustment algorithm (such as the Adam optimizer) to flexibly adjust the model's learning speed and convergence according to the user's training status. When the user's fatigue level is high or muscle recovery speed is slow, the learning rate is reduced accordingly to slow down the model's update speed and avoid overfitting. Conversely, when the training status is good, the learning rate is increased to accelerate the model's optimization process. The updates of weights and bias terms are also dynamically adjusted according to the training results. For example, if a certain feature (such as the frequency of electromyography signals or the stability of posture) is detected to have weak predictive ability for the training effect, the model will reduce the weight of that feature to reduce its impact during the training process; conversely, it will increase the weight of that feature if the predictive ability is strong. In addition, during the model training phase, this invention adopts a progressive training optimization strategy to ensure that the model parameters are not adjusted all at once, but through multiple iterations of optimization. The feedback from each training provides a basis for adjustment for the next training, avoiding the "overfitting" problem in traditional machine learning, and gradually improving the model's adaptability to individual differences as the training cycle progresses. Sub-step 204 Personalized Training Plan Generation: Based on the training output (referring to the trained rehabilitation needs model), a personalized training plan will be generated for each user. The training plan sets the training intensity, frequency, and duration according to the user's physiological characteristics and historical rehabilitation data, and can also be dynamically adjusted through real-time feedback. For example, when the user shows signs of fatigue, the training intensity will be automatically reduced or the rest time will be increased to ensure that the training process does not exceed the user's physiological capacity. The user's training plan will be optimized regularly based on the training effect to maintain the optimal training effect. The explanation explains that the recommended intensity level is derived from electromyography (EMG) signal amplitude regression analysis (using a linear regression model, with the maximum value and standard deviation of historical EMG signal amplitude as input features, directly mapped to a preset intensity level scale, such as 1-10, and dynamically adjusted by the residual between the real-time collected EMG amplitude and the predicted value to correct the intensity recommendation). Based on the optimal training frequency predicted by posture sequence, the system analyzes the user's posture sequence data within continuous training cycles, such as the periodic changes in pelvic movement trajectory, to predict their muscle function recovery curve, thereby outputting a suggested training session for the next 24 hours. The training plan includes the specific time points and intervals, as well as the maximum duration of a single training session calculated based on historical endurance data (based on the user's historical endurance data, such as the longest time in the previous training session when the electromyographic signal amplitude maintained above 80% of the target level, an endurance decay model is established using an exponential decay function to calculate the theoretical maximum duration that the current training session will not lead to muscle failure, and multiplied by a safety factor dynamically adjusted according to the real-time fatigue coefficient, such as 0.8-1.2, as the final execution duration); the training plan is invoked by a preset decision rule engine; this engine has embedded adjustment logic based on a clinical knowledge base.

[0025] The linear regression model for intensity levels is constructed as follows: It fits historical data using the least squares method, specifically in the form: Intensity Level = β0 + β1 (maximum EMG amplitude) + β2 (standard deviation of EMG amplitude). The model parameters (β0, β1, β2) are initially obtained using population data for training. After each subsequent training session with the user, the residuals between the real-time EMG amplitude and the model's predicted values ​​are updated using an online learning algorithm (such as stochastic gradient descent). This allows for personalized fine-tuning of the model parameters, ensuring that the intensity recommendations always adapt to changes in the user's abilities.

[0026] Building the LSTM prediction model for training frequency: An LSTM network architecture with at least one hidden layer is used. Its input is normalized user pose sequence data (such as time series of pelvic tilt angles in the sagittal and coronal planes) from multiple consecutive training epochs. The network learns sequence dependencies through time-stepping and uses mean squared error as the loss function. It is trained using the backpropagation algorithm and ultimately outputs a curve representing the degree of muscle recovery, which determines the optimal time point for the next training iteration. The duration exponential decay model is constructed as follows: The endurance decay function is established as duration = T_maxe^(-λ·t), where the key parameter, the decay constant λ, is determined by nonlinear curve fitting (e.g., using the Levenberg-Marquardt algorithm) of the "electromyography amplitude maintenance duration" data from each training session. The safety factor is obtained by linear interpolation from a predefined lookup table based on the real-time calculated fatigue coefficient (e.g., the slope of the electromyography signal decline at the beginning of the training session), thereby achieving adaptive scaling of the theoretical duration.

[0027] Furthermore, the user state vector is a multi-dimensional quantized array used to comprehensively represent the user's real-time physiological state, and it includes at least the following three dimensions of evaluation metrics: Fatigue Index: The user's current level of muscle fatigue calculated based on changes in the frequency components and amplitude decay trend of electromyographic signals; Force efficiency score: The score is calculated based on the matching degree between the current electromyographic signal amplitude and the target force level, as well as the standardization of muscle activation patterns; Postural compliance: An assessment of user movement conformity calculated based on the deviation between real-time posture data and standard rehabilitation movement templates; The fatigue index is obtained by: calculating the median frequency of the electromyography (EMG) signal in real time, and analyzing the slope of the decrease of the median frequency in the most recent time window through linear regression, while calculating the attenuation rate of the EMG signal amplitude; after normalizing the slope and attenuation rate, mapping them to a fatigue index between 0 and 1 through a pre-trained support vector machine regression model, the closer the index value is to 1, the higher the degree of fatigue.

[0028] The method for obtaining the force exertion efficiency score is as follows: First, calculate the absolute difference between the current electromyographic signal amplitude and the preset target force exertion level, and normalize the difference; at the same time, analyze the electromyographic signals of the agonist and antagonist muscles and calculate their synergistic contraction rate; finally, input the normalized difference sequence within a time window and the synergistic contraction rate into a feedforward neural network, which outputs a force exertion efficiency score between 0 and 1, with a higher score indicating better force exertion efficiency.

[0029] The posture compliance is obtained by comparing the real-time collected three-dimensional pelvic posture angle sequence with the standard rehabilitation movement template and calculating the root mean square error of the two in the pitch, tilt, and rotation dimensions. At the same time, the dynamic time warping algorithm is used to calculate the overall morphological difference between the real-time posture sequence and the standard template in terms of time sequence. After normalizing all error terms and weighted summing, the result is mapped to a posture compliance score between 0 and 1, with a higher score indicating more standardized movements.

[0030] Further details include the introduction of a dynamic adjustment factor when generating personalized training plans. After each training session, the plan is fine-tuned based on the user's real-time physiological data, adjusting the intensity, frequency, and duration of the training to ensure that the training plan can respond to the user's physical changes in real time. This avoids the uniformity and rigidity of traditional training plans, allowing each user to obtain a highly personalized, flexible, and adjustable training program.

[0031] The quantitative formula for the dynamic adjustment factor is explained below: Adjustment factor = w1 × (current EMG amplitude sliding average / user's personal baseline amplitude) + w2 × (1 - normalized root mean square error of posture angle) + w3 × (real-time fatigue coefficient calculated based on the slope of EMG amplitude decline during this training); where each input variable in the formula comes from feature extraction of the original sensor signal. For example, "root mean square error of posture angle" is calculated by comparing the difference between the real-time pelvic tilt angle and the standard rehabilitation posture; the weight coefficients w1, w2, and w3 are not set manually, but are obtained through a supervised learning process: using historical training data and the corresponding "training effect score" evaluated by the rehabilitation therapist as labels, the ridge regression algorithm is used for training to learn the weight combination that can best predict the training effect; At the application level, the dynamic adjustment factor is automatically calculated after each training session, and the intensity of the next training cycle is scaled up or adjusted stepwise according to a preset piecewise decision rule. For ease of understanding, the adjustment is performed through a piecewise function decision logic: if the adjustment factor ∈ [0.8, 1.2], the original plan is maintained; if the factor < 0.8, the intensity of the next training cycle is set to the original intensity × the adjustment factor; if the factor > 1.2, a fixed step size is added to the original intensity. This design ensures that the dynamic adjustment is entirely based on objective data and automated algorithms, achieving precise and reproducible fine-tuning of the training plan.

[0032] Furthermore, step S3 includes the following: Sub-step 301: Match the user state vector output by the personalized rehabilitation needs model with the pre-set training mode knowledge base; by calculating the similarity between the user state vector and the ideal state vector of each mode in the knowledge base, select the most suitable basic training mode for the current user, and load the baseline intensity, baseline frequency and baseline duration parameters corresponding to the mode. Sub-step 302: Fine-tune the training parameters in real time based on the dynamic adjustment factor, multiply the baseline training intensity by the dynamic adjustment factor to obtain the final training intensity value; when the value of the dynamic adjustment factor is continuously lower than the preset safety threshold, automatically switch the training mode to a lower intensity mode to ensure training safety.

[0033] In a further technical solution, the personalized rehabilitation needs model is implemented through a neural network with a dual-branch output structure. One branch of the network outputs a dynamic adjustment factor through regression analysis, and the other branch generates a user state vector through a multilayer perceptron. Further technical solutions also include a proactive safety intervention mechanism that continuously monitors the fatigue index in the user's state vector. When it is predicted that the fatigue index will exceed the safety threshold in subsequent training periods, it intervenes in advance and actively reduces the training intensity.

[0034] Furthermore, step S4 includes the following: Sub-step 401: Guidance requirement identification and priority ranking: Analyze the deviation between the values ​​of each dimension in the user state vector and the preset threshold to identify the problems that need to be corrected; prioritize all identified problems, and prioritize the problems with the highest safety risks (such as a sharp drop in posture compliance) and the most significant impact on training results (such as persistently low force efficiency scores) as guidance content. Sub-step 402 Personalized guidance content generation and multimodal mapping: Based on the guidance priority determined in sub-step 401, select the most matching correction instruction template from the pre-set guidance corpus; combine the abstract instruction template with the user's specific real-time data to generate concrete guidance content, such as: "Currently, the pelvis is tilted 5 degrees to the left, please adjust slightly to the right"; according to the problem type and training scenario, assign an appropriate feedback modality to the guidance content, such as using tactile vibration prompts for key posture deviations, using voice prompts for insufficient force exertion skills, and using visual charts for training status overview; Sub-step 403: Guidance Effectiveness Evaluation and Dynamic Feedback Adjustment: After issuing guidance, continuously monitor subsequent user state vectors to evaluate the guidance effectiveness. If relevant indicators do not improve, increase the feedback intensity, for example, by overlaying voice reminders on top of visual cues; conversely, if indicators return to normal, provide positive confirmation and reduce the feedback frequency to avoid information overload.

[0035] In a further technical solution, a multimodal fusion guidance strategy is employed when generating guidance content. When a complex error pattern is identified, multiple feedback modalities are triggered simultaneously. For example, the error area is highlighted on the screen while voice instructions explain the correction direction, and the device generates directional tactile cues. This collaborative stimulation of multiple senses accelerates the user's understanding and correction process. A dynamic content density adjustment mechanism is introduced during the guidance process. This mechanism calculates the user's cognitive load level in real time. When it detects that the user's state vector shows signs of confusion or fatigue after receiving multiple consecutive guidance messages, non-critical guidance is automatically paused, retaining only the most essential safety prompts. This ensures that the user receives necessary information without being overwhelmed by information, thus improving the training experience.

[0036] Furthermore, step S5 includes the following: Sub-step 501 Multi-source data aggregation and standardization preprocessing: The electromyography signal waveforms, posture sensor raw data, user state vectors and dynamic adjustment factors output in real time by the rehabilitation demand model, and all training parameter adjustment records generated during a single training session are timestamped and aggregated. The aggregated data is then standardized and cleaned to eliminate the influence of device differences and abnormal fluctuations, forming a structured training session data package. Sub-step 502 Multi-dimensional rehabilitation effect assessment and report generation: The training effect is quantitatively assessed from three dimensions: short-term dimension: analysis of the improvement curve of electromyographic amplitude stability and postural compliance in a single training session; medium-term dimension: comparison of the mean changes of key indicators in this training and recent historical training; long-term dimension: tracking the gap between the user's baseline level and rehabilitation goals, and generating an assessment report including trend charts and improvement suggestions. Sub-step 503 Personalized training program closed-loop optimization: The training effect evaluation results are correlated with the historical data of the user's state vector to identify effective training modes and parameter combinations; when it is found that a user's training effect in a specific fatigue index range is consistently better than in other ranges, the parameters of the user's rehabilitation demand model will be automatically adjusted to make it more inclined to recommend training intensity and mode near this range, so as to realize the self-evolution of the training program.

[0037] In a further technical solution, a model optimization triggering mechanism based on statistical process control was established. When the effect evaluation indicators of multiple consecutive training sessions exceed the control limits calculated based on the user's historical data, the retraining process of the rehabilitation needs model is automatically started, and the model weights are updated using the accumulated new data to ensure that the model always adapts to the user's dynamically changing rehabilitation status.

[0038] In a further technical solution, a rehabilitation trajectory prediction algorithm is introduced. Based on the user's historical progress rate and current state vector characteristics, the remaining time and optimal path to reach the rehabilitation goal are predicted. By comparing the difference between the predicted trajectory and the expected trajectory, the training program is adjusted in advance to provide users with personalized stage goals and expectations, thereby enhancing rehabilitation confidence and compliance.

[0039] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A precise training and assessment method for pelvic floor muscle rehabilitation that integrates electromyography and posture recognition, characterized in that, include: Step S1 Signal Acquisition and Real-time Data Synchronization: Through the synchronized operation of the multi-channel electromyography sensor and posture sensor, the user's electromyography signals and motion posture data are acquired in real time, and noise is removed through real-time filtering; Step S2: Dynamic Data Analysis and Model Building: Using machine learning algorithms, based on real-time collected electromyography and posture data, combined with the user's historical rehabilitation data and physiological parameters, a personalized rehabilitation needs model is established. Step S3: Training Intensity and Mode Adjustment: Based on the personalized rehabilitation needs model, perform rehabilitation exercises, record the user's real-time electromyography and postural characteristics, dynamically evaluate the user's electromyography signals and postural data, output a quantified user state vector and dynamic adjustment factor, and automatically adjust the intensity, mode and duration of training. Step S4 Real-time Feedback and Training Adjustment Strategy: Based on the user state vector output by the rehabilitation needs model, specific posture correction and force exertion guidance are provided to the user in real time through voice, visual or tactile interfaces to guide the user to make self-adjustments; Step S5: Full Data Recording and Effect Evaluation: Record the electromyographic signals, posture data, and training effects throughout each training session and store them in the cloud.

2. The method for precise training and assessment of pelvic floor muscle rehabilitation integrating electromyography and posture recognition according to claim 1, characterized in that, The process of establishing a personalized rehabilitation needs model includes: Sub-step 201 Data feature extraction: Obtain noise-removed electromyographic signals and motion posture data, extract key electromyographic waveform features through waveform feature extraction algorithm and standardize them, and extract the user's action angle and motion frequency based on the posture sensor; Sub-step 202 Feature selection and correlation analysis: Screen and analyze the extracted features; evaluate the relationship between each feature and the training effect through machine learning algorithms, and select the features that are associated with the prediction of the training effect. Sub-step 203 Model training and parameter optimization: After feature selection is completed, the rehabilitation needs model is trained; machine learning algorithms are used to learn from the collected rehabilitation data to generate a model that can predict the user's training status. Sub-step 204 Personalized training plan generation: Based on the training output results, a personalized training plan will be generated for each user; the training plan sets the training intensity, frequency and duration according to the user's physiological characteristics and historical rehabilitation data, and can also be dynamically adjusted through real-time feedback.

3. The method for precise training and assessment of pelvic floor muscle rehabilitation integrating electromyography and posture recognition according to claim 1, characterized in that, The user state vector is a multi-dimensional quantized array used to comprehensively represent the user's real-time physiological state, and it includes at least the following three dimensions of evaluation metrics: Fatigue Index: The user's current level of muscle fatigue calculated based on changes in the frequency components and amplitude decay trend of electromyographic signals; Force efficiency score: The score is calculated based on the matching degree between the current electromyographic signal amplitude and the target force level, as well as the standardization of muscle activation patterns. Postural compliance: An assessment of user movement conformity calculated based on the deviation between real-time posture data and standard rehabilitation movement templates.

4. The method for precise training and assessment of pelvic floor muscle rehabilitation integrating electromyography and posture recognition according to claim 1, characterized in that, The quantification formula for the dynamic adjustment factor is calculated as follows: Adjustment factor = w1 × (current EMG amplitude sliding average / user's personal baseline amplitude) + w2 × (1 - normalized root mean square error of posture angle) + w3 × (real-time fatigue coefficient calculated based on the slope of EMG amplitude decrease during this training); where the weight coefficients w1, w2, and w3 are trained using the ridge regression algorithm to learn the optimal weight combination that can predict the training effect. After each training session, the dynamic adjustment factor is automatically calculated, and the intensity of the next training cycle is scaled up or adjusted stepwise according to the preset segmented decision rules.

5. The method for precise training and assessment of pelvic floor muscle rehabilitation integrating electromyography and posture recognition according to claim 1, characterized in that, The output of step S3 includes the recommended intensity level, the optimal training frequency, and the maximum duration of a single training session. A linear regression model based on online learning is used, with the maximum value and standard deviation of the user's historical electromyography (EMG) signal amplitude as input features, and the recommended intensity level is output. The regression coefficient is dynamically updated by the residual between the real-time EMG amplitude and the predicted value. A long short-term memory network model is used to analyze the time series data of the pelvic movement trajectory in the user's continuous training cycle, predict the muscle function recovery curve, and thus output the optimal training frequency. An endurance decay model is established using an exponential decay function. The maximum duration of a single training session is calculated by combining the decay constant determined through nonlinear curve fitting and the safety factor dynamically adjusted based on the real-time fatigue coefficient.

6. The method for precise training and assessment of pelvic floor muscle rehabilitation integrating electromyography and posture recognition according to claim 1, characterized in that, Step S3 includes the following: Sub-step 301: Match the user state vector output by the personalized rehabilitation needs model with the pre-set training mode knowledge base; by calculating the similarity between the user state vector and the ideal state vector of each mode in the knowledge base, select the most suitable basic training mode for the current user, and load the baseline intensity, baseline frequency and baseline duration parameters corresponding to the mode. Sub-step 302: Fine-tune the training parameters in real time based on the dynamic adjustment factor, multiply the baseline training intensity by the dynamic adjustment factor to obtain the final training intensity value; When the value of the dynamic adjustment factor remains below the preset safety threshold, the training mode will be automatically switched to a lower intensity mode to ensure training safety.

7. The method for precise training and assessment of pelvic floor muscle rehabilitation integrating electromyography and posture recognition according to claim 1, characterized in that, It includes a proactive safety intervention mechanism that continuously monitors the fatigue index in the user's state vector. When it is predicted that the fatigue index will exceed the safety threshold in subsequent training periods, it intervenes in advance and actively reduces the training intensity.

8. The method for precise training and assessment of pelvic floor muscle rehabilitation integrating electromyography and posture recognition according to claim 1, characterized in that, Step S4 includes the following: Sub-step 401: Guidance requirement identification and priority ranking: Analyze the deviation between the values ​​of each dimension in the user state vector and the preset threshold to identify the problems that need to be corrected; prioritize all identified problems, and prioritize the problems with the highest security risks and the most significant impact on training results as guidance content. Sub-step 402 Personalized guidance content generation and multimodal mapping: Based on the guidance priority determined in sub-step 401, select the most matching correction instruction template from the preset guidance corpus; By combining abstract instruction templates with specific real-time user data, concrete guidance content is generated. Based on the problem type and training scenario, appropriate feedback modalities are assigned to the guidance content, such as using tactile vibration prompts for key posture deviations, using voice prompts for insufficient force exertion skills, and using visual charts to provide an overview of the training status. Sub-step 403 Guidance effect evaluation and dynamic adjustment of feedback: After issuing guidance, continuously monitor the subsequent user state vector to evaluate the guidance effect; if the relevant indicators do not improve, the feedback intensity will be increased, for example, by superimposing voice reminders on top of visual cues; conversely, if the indicators return to normal, positive confirmation will be given and the feedback frequency will be reduced to avoid information overload.

9. The method for precise training and assessment of pelvic floor muscle rehabilitation integrating electromyography and posture recognition according to claim 1, characterized in that, When generating guidance content, if a complex error pattern is identified, multiple feedback modalities will be triggered simultaneously to accelerate the user's understanding and correction process by stimulating multiple senses in a coordinated manner. A dynamic content density adjustment mechanism is introduced during the guidance process to calculate the user's cognitive load level in real time. When it is detected that the user's state vector shows signs of confusion or fatigue after receiving multiple guidance messages in a row, non-critical guidance will be automatically paused, and only the most essential safety tips will be retained.

10. The method for precise training and assessment of pelvic floor muscle rehabilitation integrating electromyography and posture recognition according to claim 1, characterized in that, Step S5 includes the following: Sub-step 501 Multi-source data aggregation and standardization preprocessing: The electromyography signal waveforms, posture sensor raw data, user state vectors and dynamic adjustment factors output in real time by the rehabilitation demand model, and all training parameter adjustment records generated during a single training session are timestamped and aggregated. The aggregated data is then standardized and cleaned to eliminate the influence of device differences and abnormal fluctuations, forming a structured training session data package. Sub-step 502 Multi-dimensional rehabilitation effect assessment and report generation: The training effect is quantitatively assessed from three dimensions: short-term dimension: analysis of the improvement curve of electromyographic amplitude stability and postural compliance in a single training session; medium-term dimension: comparison of the mean changes of key indicators in this training and recent historical training; long-term dimension: tracking the gap between the user's baseline level and rehabilitation goals, and generating an assessment report including trend charts and improvement suggestions. Sub-step 503 Personalized training program closed-loop optimization: The training effect evaluation results are correlated with the historical data of the user's state vector to identify effective training modes and parameter combinations; when it is found that a user's training effect in a specific fatigue index range is consistently better than in other ranges, the parameters of the user's rehabilitation demand model will be automatically adjusted to make it more inclined to recommend training intensity and mode near this range, so as to realize the self-evolution of the training program.

Images