An evaluation method for determining the mechanical properties of the upper airway muscle group based on signal characteristics

Abstract

The application provides an upper respiratory tract muscle group mechanics characteristic determination evaluation method based on signal characteristics. First, the application collects the force signal of the upper respiratory tract muscle group through a sensor. After the signal is amplified, filtered and digitized, key characteristic variables including maximum muscle strength, muscle strength stability, total work, fatigue index and overall test time are generated. To improve evaluation accuracy, the application introduces a multi-level machine learning model, including a basic TabTransformer structure, a contrast learning framework and a meta-learning technology, to realize deep analysis of complex feature interaction. During model training, contrast learning of unlabeled data is used to enhance feature representation capability; a MAML framework is used for rapid adaptation of a small amount of labeled data, and a multi-task learning strategy is used to consider the comprehensive goal of disease classification and efficacy prediction. Finally, the performance indicators of the model on the validation set are significantly improved, and efficient evaluation and rapid deployment can be realized in actual application.

Description

Technical Field

[0001] This invention belongs to the field of mechanical property measurement and evaluation, specifically a method for measuring and evaluating the mechanical properties of upper airway muscle groups based on signal characteristics. Background Technology

[0002] Assessments of upper airway function generally focus on respiratory status, but often remain at the level of structural observation and indirect monitoring, making it difficult to accurately quantify muscle group characteristics. In traditional methods, test results are frequently affected by subjective factors or external environmental interference, making it difficult to provide precise biomechanical data. These methods remain limited in assessing upper airway muscle tone and fatigue. Furthermore, conventional testing methods struggle to capture changes in tongue dynamics, resulting in insufficient acquisition of intraoral biomechanical information and weakening assessment accuracy.

[0003] For muscles such as the genioglossus, which maintain the shape of the oropharyngeal region, a lack of sustained tension can easily lead to airway collapse and airflow obstruction. Current assessments rely on surface signals and fail to represent biomechanical levels. However, most current analyses focus only on nerve innervation and basic muscle activity, lacking precise quantification of whether upper airway muscles possess sufficient strength and endurance. Once contractile capacity or fatigue characteristics cannot be accurately determined, timely intervention or training becomes difficult, allowing potential risks to persist. Current methods often employ subjective scales or simplified biomechanical measurements, which struggle to accommodate long-term dynamic monitoring. Due to a lack of assessment dimensions, the understanding of the interaction between nerves and muscle strength remains insufficient. Previous studies attempting to measure the biomechanical characteristics of upper airway muscles have found that some individuals exhibit lower total work done in tongue extension movements and significantly shorter endurance time for tongue elevation movements, with slower fatigue recovery. These individuals often demonstrate persistent muscle function decline, indicating that relying solely on structural or simplified testing is insufficient to promptly capture key fatigue points, thus introducing uncertainty into training and intervention strategies. Therefore, more precise quantitative methods are urgently needed to identify potential functional impairments and guide individualized treatment to ensure effectiveness.

[0004] Given the understanding that upper airway muscle dysfunction (such as weakness and fatigue) can lead to airway instability, a comprehensive assessment protocol combining biomechanical data and individual characteristics is urgently needed. Only by accurately understanding the functional status of these muscle groups can we help select suitable candidates for training interventions and provide clearer guidance for subsequent treatment strategies. Current methods lack targeted quantitative and classification standards, making it difficult to perfectly align assessment results with clinical needs. Furthermore, the lack of automated and intelligent applications hinders in-depth analysis of multidimensional signals. Summary of the Invention

[0005] This invention addresses the shortcomings of existing assessment methods, such as the inability to comprehensively quantify the contractile mechanical properties of upper airway muscles and the difficulty in identifying muscle fatigue changes. It proposes a measurement method based on high-precision mechanical signal acquisition and multi-model fusion assessment. By acquiring muscle force signals of upper airway muscles under different movements using force sensors, key features such as maximum muscle strength, fatigue index, and work output are extracted. Combined with the TabTransformer model and self-supervised contrastive learning, meta-learning (MAML), and multi-task learning strategies, a comprehensive analysis and rapid assessment of muscle function status is performed. This method can learn general features using unlabeled data and quickly adapt to new tasks under limited sample conditions, thereby achieving accurate quantitative assessment of upper airway muscle dysfunction and screening patients with upper airway muscle dysfunction. This method has significant clinical value for customizing individualized and precise treatment plans for OSA patients.

[0006] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0007] A method for assessing and evaluating the mechanical properties of upper airway muscle groups based on signal characteristics, comprising the following steps:

[0008] 1) Collect the electrical and mechanical signals of the upper respiratory tract muscles under specific movements using tension and compression sensors, and then perform amplification, filtering, and analog-to-digital conversion on them in sequence;

[0009] 2) Extract signal features from the processed electrical impulse signals;

[0010] 3) The mechanical properties of the upper respiratory tract muscles are evaluated based on signal characteristics and individual parameters.

[0011] Step 2) includes the following steps:

[0012] 2.1) Extract the maximum muscle group strength F max A maximum force test is performed every minute, and the maximum force value of each training session is recorded. The maximum value F is then taken. max :

[0013] F max =max{f peak,1 ,f peak,2 ,…,f peak,i}

[0014] Among them, f peak,i The peak value of the i-th maximum force attempt;

[0015] 2.2) Extracting muscle strength stability F cons :

[0016]

[0017] in, σ is the average force value over a certain time period. f This represents the standard deviation of the force values ​​over that time period.

[0018] 2.3) Extract the total work done W:

[0019]

[0020] Where, f(t) n ) represents the electrical impulse signal of the nth segment, Δt represents the test interval, and W represents the energy consumption during the entire test process;

[0021] 2.4) Extract the overall test time T;

[0022] 2.5) Extract the fatigue index TI, and record the work done in each maximum force test conducted once per minute to form a point set of test time and discrete work data. Using linear fitting:

[0023] W(t k )≈β0+β1t k

[0024] Where β0 is the intercept in linear regression, β1 is the slope in linear regression, and t k The fatigue index TI = β1 corresponds to the time point corresponding to the kth work measurement.

[0025] Step 3) includes the following steps:

[0026] 3.1) The TabTransformer model is used to process the signal features into high-level feature representations, which serve as the global representation vector;

[0027] 3.2) Using a self-supervised contrastive learning method, a discriminative general representation is learned from unlabeled high-level feature representation data;

[0028] 3.3) Using the MAML framework, train an initial model parameter that can quickly adapt to new tasks on different tasks, and use the model parameter to train the TabTransformer model;

[0029] 3.4) Construct a multi-task learning strategy to fine-tune the trained TabTransformer model;

[0030] 3.5) Using signal features and individual parameters as input, the adjusted model is used to evaluate the mechanical properties of the upper respiratory tract muscles.

[0031] Step 3.1) includes the following steps:

[0032] 3.1.1) The numerical signal features are mapped to a high-dimensional vector space through the embedding layer, and a feature embedding sequence is formed.

[0033] 3.1.2) Input the feature embedding sequence into the TabTransformer, and obtain the feature representation Z through a multi-layer stacked self-attention and feedforward network. L ;

[0034] 3.1.3) Represent the feature Z L The pooling layer forms a global representation vector z, which is then used as the input of the high-level feature representation to the subsequent decision layer.

[0035] Step 3.2) includes the following steps:

[0036] Using the SimCLR framework, each unlabeled subject's data is treated as a sample, and two different augmentation strategies are employed to generate sample pairs X. i and

[0037] Use TabTransformer as the encoder f θ (·), mapping sample pairs to latent representation z = f θ (X);

[0038] Constructing a contrastive loss function By minimizing Learn discriminative feature representations from unlabeled data.

[0039] The contrast loss function for:

[0040]

[0041] in, Let τ be the cosine similarity and τ be the temperature coefficient.

[0042] Step 3.3) includes the following steps:

[0043] 3.3.1) Construct different tasks in different scenarios, and train the model based on these tasks;

[0044] 3.3.2) Extracting tasks from the task distribution The TabTransformer model parameters θ are obtained by performing one or more gradient descent operations on this task to obtain the adapted parameters θ. i ;

[0045] 3.3.3) Using the adapted parameter θ i The loss is calculated on the validation set of this task, and the accumulated validation loss from multiple tasks is used to update the initial parameter θ. The updated parameter θ* for:

[0046]

[0047] in, Indicates the relationship between θ and the task The parameters updated in the inner loop.

[0048] Step 3.4) includes the following steps:

[0049] The primary task is to determine whether the pathological mechanism is upper airway muscle dysfunction. Simultaneously, auxiliary classification and regression tasks are constructed, with all tasks sharing input features and using theta... * As initial parameters for the model, multiple prediction results are obtained, and a joint loss function is used. Train the model.

[0050] The joint loss function for:

[0051]

[0052] in, The cross-entropy loss is used for the main task. To assist in the cross-entropy of classification tasks, To assist in calculating the mean squared error of the regression task, λ c , λ a and λ r All are weights.

[0053] The present invention has the following beneficial effects and advantages:

[0054] 1. This invention combines force sensors with feature extraction to comprehensively capture the contractile mechanical information of the upper airway muscle groups, overcoming the data loss and misjudgment caused by previous single measurement methods.

[0055] 2. This invention integrates self-supervised contrastive learning and meta-learning techniques, which can make full use of unlabeled data and quickly adapt to new scenarios under conditions of a small number of labeled samples, ensuring the stability and accuracy of the evaluation results.

[0056] 3. This invention employs a multi-task learning strategy to simultaneously predict multiple indicators such as muscle group functional status and fatigue level, helping to quickly identify functional abnormalities and provide targeted guidance for subsequent interventions. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the structure and flow of the training evaluation guidance system of the present invention;

[0058] Figure 2 This is a diagram illustrating the muscle strength-time relationship during training, as presented in this invention.

[0059] Figure 3 This is a schematic diagram of the training and evaluation method of the present invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0061] like Figure 1 The diagram shown illustrates the structural flow of the training and evaluation guidance system of this invention, specifically including:

[0062] Signal flow 1: Signals from the upper respiratory tract muscle groups are acquired via a sensor. Specifically, a high-precision S-shaped tension / compression sensor is placed directly in front of the trainee's face. This sensor is made of S-shaped stainless steel, with a pressure sensing range of 0–20 N, an accuracy of 0.5%, and a sampling rate of 20 Hz. The measurement principle is based on the trainee extending their tongue. The increased pressure exerted by the tongue on the sensor's effective area leads to a decrease in the sensor's output resistance. The circuitry in the signal conditioning module converts this resistance change into a voltage change. The tension / compression sensor is sealed with a 1 mm food-grade silicone sleeve for waterproofing. Before use, a calibration experiment is conducted to determine the relationship between the sensor and the voltage change.

[0063] Signal flow 2: The signal processing unit amplifies and filters the received voltage, and then uses an analog-to-digital converter circuit for digital sampling. The digitized muscle voltage force signal is stored in the system's memory chip for convenient digital processing and signal playback.

[0064] Signal processing step 3 involves feature extraction of the acquired electrical muscle force signals, as shown in Table 1. The features of the electrical muscle force signals include the maximum muscle group force F. max Muscle strength stability F cons The fatigue index TI, total work done W, and overall test time T are described below:

[0065] 1) Maximum muscle group strength F max Maximum muscle strength reflects the peak force level generated by a subject when activating the muscles of the upper respiratory tract in a short period of time. If F max The value was significantly lower than the standard value for a specific population, indicating that the subject's muscle strength generation ability was insufficient. At the start of the test, the trainee performed 3-5 consecutive maximum tongue extensions, each lasting 3 seconds, with a 30-second interval between each extension. The maximum force value was recorded for each repetition, and the highest value was recorded.

[0066] F max =max{f peak,1 ,f peak,2 ,…,f peak,i}

[0067] Among them, f peak,iThe peak value of the i-th maximum force attempt.

[0068] 2) Muscle strength stability F cons Muscle strength stability is used to evaluate the stability of upper respiratory tract muscle control throughout the training process. Specifically, it uses test data (excluding maximum force tests conducted once per minute) and assesses force output variability using the coefficient of variation of the data.

[0069]

[0070] in, σ is the average force value over this time period. f F represents the standard deviation of the force values ​​over that time period. cons The smaller the value, the more stable the force is.

[0071] 3) Total work done W. Total work done is the cumulative value of the electrical muscle force signal over time during training.

[0072]

[0073] Where, f(t) n ) represents the electrical force signal of the nth segment, Δt represents the test interval, and W reflects the "energy" consumption during the entire test process, which can characterize the trainee's overall muscle strength output ability and persistence.

[0074] 4) Overall testing time T. Training duration T set The training duration can be set beforehand; however, due to personal factors or fatigue, the test may end prematurely, resulting in a shorter test time than the set training duration. Therefore, the actual test time should be used as the reference. When the actual test time is T≤T set If the time limit is reached, it indicates that the trainee has not reached the expected training time limit, which can be used to judge that the endurance is insufficient.

[0075] 5) Fatigue Index (TI). The work done in each maximum force test, conducted once per minute, is recorded to form a point set of test time and discrete work data. Using linear fitting:

[0076] W(t k )≈β0+β1t k

[0077] The fatigue index TI = β1. When TI < 0, it indicates that the amount of work done decreases over time, meaning that muscle fatigue worsens. The steeper the slope, the more severe the fatigue.

[0078] Table 1

[0079] Trainee training characteristics Units or categories Remark <![CDATA[Maximum muscle group strength F max > N The measurement results at the start of the test shall prevail. <![CDATA[Muscle strength stability F cons > Dimensionless Muscle strength fluctuations during the test Fatigue Index TI Dimensionless Evaluation of muscle strength decline during the test Total work W N·s Evaluate the overall performance of test takers Test time T Second Judging the length of the test by the tester

[0080] Signal flow 4 calculates and assesses the trainee's muscle function level based on the electromyographic signal characteristics and trainee personal parameters from signal flow 3. As shown in Table 2, the trainee's personal parameters include the trainee's age, gender, AHI index, BMI index, and neck circumference.

[0081] Table 2

[0082]

[0083] like Figure 3 The diagram shown illustrates the training and evaluation method of this invention, which specifically includes:

[0084] Step 1, Preparation before the assessment. Subjects are required to avoid excessive oral activities (such as playing wind instruments) for 30 minutes before the assessment and to remain relaxed.

[0085] Step 2, upper respiratory tract muscle group maximum strength test. The relationship between training strength and time is as follows: Figure 2 As shown, to begin this assessment training, the trainee, following the trainer's instructions, opens their mouth and bites the connecting rod of the pressure sensor, ensuring the distance between the tongue and the pressure sensor remains constant. The trainee then extends their tongue with maximum force to press against the pressure sensor in front, repeating this 3-5 times. Each movement lasts 3 seconds, with a 30-second interval between each movement. The maximum muscle group strength F is then calculated. max .

[0086] Step 3: Upper respiratory tract muscle strength maintenance and fatigue detection. In subsequent training, the training duration can be set beforehand, typically 30 minutes. The training includes two parts: endurance maintenance testing and periodic retesting of maximum muscle strength. The specific procedure for endurance maintenance testing is as follows: Set a moderate-intensity target pressure window, with a pressure window range of [F...]. MAX *50%, F MAX *70%], participants were asked to maintain tongue extension force within this range for 3-5 seconds. To help trainees focus, a fun game was used, with the stress window duration set as a dynamic value, ranging from 3-5 seconds; the rest time was also set as a dynamic value, ranging from 10-20 seconds. During the training period, maximum muscle strength was periodically retested every minute, within a range of [F]. MAX *80%, F MAX The pressure window is set at 100% and lasts for 5 seconds. If some data from the pressure test falls within the pressure window, the trainee is considered not to be fatigued and can continue testing. If the results of two consecutive maximum force tests are both below the pressure window, the trainee is considered to be fatigued and the training session should be stopped.

[0087] Step 4, Feature Extraction. After the muscle strength detection in Step 2 and the fatigue detection in Step 3, signal denoising, baseline correction, calibration, and regularization are performed to extract the electrical muscle force signal features. These features are then combined with the trainee's personal information to obtain complete features. The features are then normalized to reduce the differences in feature scale caused by physiological differences between individuals.

[0088] Step 5: Intelligent analysis and model iteration.

[0089] 1) Basic Model. The features extracted above resemble tabular data. Therefore, this invention combines the TabTransformer structure for feature extraction, effectively utilizing the multi-head self-attention mechanism in the Transformer. For the input features, the numerical features are mapped to a high-dimensional vector space through the Embedding layer, forming a feature embedding sequence. The feature embedding sequence is input into the TabTransformer. Through a multi-layered stacked self-attention and feedforward network (FNN) module, the TabTransformer can automatically capture the complex interaction relationships between features, such as F... max The nonlinear relationship with TI. The final representation Z is obtained. L The pooling layer forms a global representation vector z, which is then used as input to the subsequent decision layer as a high-level feature representation.

[0090] 2) Considering real-world clinical scenarios, while clinical experts can provide information on the pathological mechanisms of OSA patients, labeling all data is costly and time-consuming. Therefore, most of the collected data is unlabeled. To better utilize this data, we introduce a self-supervised contrastive learning method to learn discriminative general representations from unlabeled data. Using the SimCLR framework, each unlabeled subject's data is treated as a sample. First, positive sample pairs are constructed. X is generated using two different enhancement strategies. i and (For example, adding a small amount of additive noise to TI and F) cons Introducing small-range random scaling, etc., both should be mapped to similar feature representation spaces by the model. Using TabTransformer as the encoder f... θ (·), mapping the input feature vector to the latent representation Z = f θ (X). The contrastive loss function is defined as:

[0091]

[0092] in, Let be the cosine similarity, and τ be the temperature coefficient. By minimizing... The model learns discriminative feature representations on unlabeled data: data from the same subject (or similar mechanical patterns) are mapped to similar vector spaces, while feature representations from different subjects are scattered. This provides a more robust initial representation space for subsequent tasks.

[0093] 3) To further mitigate the limitations of limited data and corresponding labels in real-world clinical settings, this invention introduces the MAML framework. By training an initial model with parameters that can quickly adapt to new tasks across a series of tasks, the model can rapidly achieve high performance on target tasks with limited labeled data. In this invention, different scenarios (such as different breathing postures, different training loads, different time periods, etc.) are constructed using patients determined by clinical experts to create different "tasks," which are then used as the basis for model training. Tasks are extracted from the task distribution. Perform one or more gradient descent operations on the model parameters θ on this task to obtain the adapted parameters θ. i The loss is calculated on the validation set of this task using the adapted parameters, and the validation loss from multiple tasks is accumulated to update the initial parameters θ.

[0094]

[0095] in Indicates the relationship between θ and the task The parameters updated in the inner loop. After meta-training, θ * It possesses the ability to quickly adapt to new related tasks. When we apply θ... * When it comes to the target task, only a small amount of labeled data is needed, and good performance can be achieved through a few gradient updates.

[0096] 4) Multi-task fine-tuning. To improve the model's reliability and generalization performance for tasks, this invention constructs a multi-task learning strategy. The primary task is to determine whether the pathological mechanism is upper airway muscle dysfunction, while also considering auxiliary classification tasks (determining the degree of muscle weakness in patients: mild, moderate, and severe) and auxiliary regression tasks (predicting the trajectory of signal feature changes after intervention training). These tasks share input features and use θ obtained from meta-learning. * As initial parameters for the model, multiple predictions are output simultaneously, and the model is trained using a joint loss function.

[0097]

[0098] in, The cross-entropy loss is the primary task. Cross-entropy is used to assist in classification tasks; The mean squared error is used to assist in the regression task.

[0099] 5) Model Deployment and Practical Application. For a patient not yet receiving treatment, their complete characteristics can be input into the model to obtain multi-task assessment results, assisting clinicians in evaluation and providing guidance for upper airway muscle training programs. Over time and with the accumulation of data, the patient's upper airway muscle function may change as treatment progresses, and the model can be updated periodically. For patients currently receiving treatment, unlabeled data (mostly daily measurements) and a few clearly labeled key time points (such as a clinician follow-up on day N after treatment begins) can be collected. This data can be used to assess the effectiveness of model training and also serve as new semi-supervised data to update the model.

Claims

1. A method for measuring and evaluating the mechanical properties of upper airway muscle groups based on signal characteristics, characterized in that, Includes the following steps: 1) Acquire electrical and mechanical signals of the upper respiratory tract muscles under specific movements using tension and compression sensors, and then amplify, filter, and convert them from analog to digital in sequence; 2) Extract signal features from the processed electrical muscle force signal; 3) Assess the biomechanical properties of the upper respiratory tract muscles based on signal characteristics and individual parameters; Step 2) includes the following steps: 2.1) Extracting the maximum muscle strength of the muscle group A maximum force test is performed every minute, and the maximum force value for each training session is recorded, with the highest value being recorded. : in, The peak value of the i-th maximum force attempt; 2.2) Extracting muscle strength stability : in, The average force value over a certain period of time. This represents the standard deviation of the force values ​​over that time period. 2.3) Extract the total work done : in, This represents the electrical force signal of the nth segment. For the test interval, This refers to the energy consumption during the entire testing process. 2.4) Extract the overall test time ; 2.5) Extracting the fatigue index The work done in each maximum force test, conducted once per minute, is recorded to form a point set of test time and discrete work data. Using linear fitting: in, The intercept in linear regression. The slope in linear regression. The fatigue index at the time point corresponding to the kth work measurement. ; Step 3) includes the following steps: 3.1) The TabTransformer model is used to process the signal features into high-level feature representations, which serve as the global representation vector; 3.2) Using a self-supervised contrastive learning method, a discriminative general representation is learned from unlabeled high-level feature representation data; 3.3) Using the MAML framework, train an initial model parameter that can quickly adapt to new tasks on different tasks, and use the model parameter to train the TabTransformer model; 3.4) Construct a multi-task learning strategy to fine-tune the trained TabTransformer model; 3.5) Using signal features and individual parameters as input, the adjusted model is used to evaluate the mechanical properties of the upper respiratory tract muscles.

2. The method for measuring and evaluating the mechanical properties of upper airway muscle groups based on signal characteristics according to claim 1, characterized in that, Step 3.1) Includes the following steps: 3.1.1) The numerical signal features are mapped to a high-dimensional vector space through the embedding layer, and a feature embedding sequence is formed. 3.1.2) Input the feature embedding sequence into the TabTransformer, and obtain the feature representation through a multi-layer stacked self-attention and feedforward network. ; 3.1.3) Representing features A global representation vector is formed through the pooling layer. This information is then used as a high-level feature representation input to subsequent decision-making layers.

3. The method for measuring and evaluating the mechanical properties of upper airway muscle groups based on signal characteristics according to claim 1, characterized in that, Step 3.2) includes the following steps: Using the SimCLR framework, each unlabeled subject's data is treated as a sample, and sample pairs are generated using two different augmentation strategies. and ; Use TabTransformer as the encoder Mapping sample pairs to latent representations ; Constructing a contrastive loss function By minimizing This allows the learning of discriminative feature representations from unlabeled data.

4. The method for measuring and evaluating the mechanical properties of upper airway muscle groups based on signal characteristics according to claim 3, characterized in that, The contrast loss function for: in, For cosine similarity, This is the temperature coefficient.

5. The method for measuring and evaluating the mechanical properties of upper airway muscle groups based on signal characteristics according to claim 1, characterized in that, Step 3.3) includes the following steps: 3.3.1) Construct different tasks in different scenarios, and train the model based on these tasks; 3.3.2) Extracting tasks from the task distribution For TabTransformer model parameters Perform one or more gradient descent iterations on this task to obtain the adapted parameters. ; 3.3.3) Use the adapted parameters Calculate the loss on the validation set for this task, and update the initial parameters by accumulating the validation losses from multiple tasks. Updated parameters for: in, Indicates to In the mission The parameters updated in the inner loop.

6. The method for measuring and evaluating the mechanical properties of upper airway muscle groups based on signal characteristics according to claim 1, characterized in that, Step 3.4) includes the following steps: The primary task is to determine whether the pathological mechanism is upper airway muscle dysfunction. Simultaneously, auxiliary classification and regression tasks are constructed, with all tasks sharing input features. As initial parameters for the model, multiple prediction results are obtained, and a joint loss function is used. Train the model.

7. The method for measuring and evaluating the mechanical properties of upper airway muscle groups based on signal characteristics according to claim 6, characterized in that, The joint loss function for: in, The cross-entropy loss is used for the main task. To assist in the cross-entropy of classification tasks, To assist in the mean squared error of the regression task, , and All are weights.

Images