Method for measuring and evaluating mechanical characteristics of upper airway muscle group based on signal characteristics

Abstract

The invention provides an upper respiratory tract muscle group mechanical property measuring and evaluating method based on signal characteristics. Firstly, a force signal of an upper respiratory tract muscle group is collected through a sensor; after the signals are subjected to voltage amplification filtering and digital processing, key characteristic variables including the maximum muscle force, the muscle force stability, the total acting amount, the fatigue index and the overall testing time are generated. In order to improve the evaluation accuracy, a multi-level machine learning model is introduced, and the multi-level machine learning model comprises a basic TabTransform structure, a contrast learning framework and a meta-learning technology so as to realize deep analysis of a feature complex interaction relationship. In the model training process, through comparative learning of unlabeled data, the feature representation capability is enhanced; an MAML framework is adopted to quickly adapt to a small amount of labeled data, and a comprehensive target of disease grading and curative effect prediction is taken into account by utilizing a multi-task learning strategy. Finally, the performance index of the model on the verification set is remarkably improved, and efficient evaluation and rapid deployment can be achieved in practical application.

Description

Technical Field

[0001] The invention belongs to the field of mechanical property measurement and evaluation, and specifically is a method for measuring and evaluating the mechanical properties of upper airway muscles based on signal characteristics. Background Art

[0002] The evaluation of upper airway function generally focuses on respiratory status, but it mostly stays at the level of structural observation and indirect monitoring, and it is difficult to accurately quantify muscle group characteristics. In traditional methods, the test results are often interfered by subjective factors or external environment, and it is difficult to provide accurate mechanical data. This type of method is still limited in the evaluation of upper airway muscle tension and fatigue. In addition, conventional testing methods are difficult to capture changes in tongue dynamics, resulting in insufficient acquisition of intraoral mechanical information, which weakens the evaluation accuracy.

[0003] For muscles such as the genioglossus that can maintain the shape of the oropharyngeal region, if there is a lack of continuous tension support, channel collapse and airflow obstruction are prone to occur. Existing assessments rely on surface signals and are difficult to present mechanical levels. However, most current analyses only focus on neural innervation and basic muscle activity, and lack accurate quantification of whether the upper airway muscles have sufficient strength and endurance. Once the contraction capacity or fatigue characteristics cannot be accurately determined, it is difficult to intervene or train in time, so that potential risks continue to exist. Currently, subjective scales or simple mechanical measurements are mostly used, but it is difficult to take into account long-term dynamic monitoring. Due to the lack of evaluation dimensions, the understanding of the interactive relationship between nerves and muscle strength is still insufficient. Existing studies have attempted to measure the biomechanical properties of the upper airway muscles and found that the total work of tongue extension movement in some individuals is relatively low, and the endurance time of tongue lifting movement is also significantly shortened, and fatigue recovery is slower. Such subjects often show a continuous decline in muscle function, indicating that it is difficult to capture key fatigue points in time by relying solely on structure or simple detection, thereby bringing uncertainty to the formulation of training and intervention strategies. For this reason, more accurate quantitative methods are urgently needed to identify potential functional disorders and guide individualized treatment to ensure the effectiveness of protection.

[0004] Based on the recognition that upper airway muscle disorders (such as weak contraction and easy fatigue) may cause channel instability, there is an urgent need for a comprehensive evaluation program that combines mechanical data and individual characteristics. Only by accurately understanding the functional status of the muscle groups can we assist in screening suitable training intervention populations and provide clearer directions for subsequent treatment strategies. Existing methods still lack targeted quantification and classification standards, making it difficult for evaluation results to perfectly match clinical needs. At the same time, the lack of practical applications of automation and intelligence also hinders the in-depth analysis of multidimensional signals. Summary of the invention

[0005] In view of the problems that the existing evaluation methods cannot fully quantify the mechanical properties of upper airway muscle contraction and are difficult to identify changes in muscle fatigue, the present invention proposes a measurement method based on high-precision mechanical signal acquisition and multi-model fusion evaluation. The muscle force signals of the upper airway muscles under different actions are obtained by force sensors, and key features such as maximum muscle force, fatigue index, and work done are extracted. The TabTransformer model is combined with self-supervised comparative learning, meta-learning (MAML) and multi-task learning strategies to comprehensively analyze and quickly determine the functional state of the muscle group. This method can not only use unlabeled data to learn general features, but also quickly adapt to new tasks under the condition of few samples, thereby realizing accurate quantitative evaluation of upper airway muscle dysfunction and screening out patients with upper airway muscle dysfunction. The above method research customizes individualized and accurate treatment plans for OSA patients, which has important clinical value.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

[0007] A method for measuring and evaluating the mechanical properties of upper airway muscles based on signal characteristics comprises the following steps:

[0008] 1) The electromyographic pressure signals of the upper respiratory tract muscles under specific actions are collected through the tension pressure sensor, and then amplified, filtered, and analog-to-digital converted in sequence;

[0009] 2) extracting signal features from the processed electromyographic pressure signal;

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

[0011] The step 2) comprises the following steps:

[0012] 2.1) Extract the maximum muscle force F max , perform a maximum force test every minute, record the maximum force value of each training, and take the maximum value F max :

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

[0014] Among them, f peak,i is the peak value of the ith maximum force attempt;

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

[0016]

[0017] in, is the average force value in a certain period of time, σ f is the standard deviation of the force value during this period;

[0018] 2.3) Extract the total work W:

[0019]

[0020] Among them, f(t n ) is the electromyographic stress signal of the nth segment, Δt is the test interval, and W is 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 maximum force test work done once a minute to form a point set of test time and discrete work data Using linear fitting:

[0023] W(t k )≈β0+β1t k

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

[0025] The step 3) comprises the following steps:

[0026] 3.1) Using the TabTransformer model, the signal features are processed into high-level feature representations as global representation vectors;

[0027] 3.2) Learning discriminative general representations from unlabeled high-level feature representation data through self-supervised contrastive learning;

[0028] 3.3) Using the MAML framework, we 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) Build a multi-task learning strategy to fine-tune the trained TabTransformer model;

[0030] 3.5) The signal characteristics and individual parameters are used as input and the adjusted model is used to evaluate the mechanical properties of the upper airway muscles.

[0031] The step 3.1) comprises the following steps:

[0032] 3.1.1) The numerical signal features are mapped to a high-dimensional vector space through the Embedding layer and form a feature embedding sequence;

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

[0034] 3.1.3) Represent the feature Z L Through the pooling layer, a global representation vector z is formed and input into the subsequent decision layer as a high-level feature representation.

[0035] The step 3.2) comprises the following steps:

[0036] Using the SimCLR framework, the data of each unlabeled subject is taken as a sample, and two different enhancement strategies are used to generate sample pairs X i and

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

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

[0039] The contrast loss function for:

[0040]

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

[0042] The step 3.3) comprises the following steps:

[0043] 3.3.1) Construct different tasks in different scenarios and use them as the basis for training models;

[0044] 3.3.2) Extracting tasks from the task distribution Perform one or more gradient descents on the TabTransformer model parameters θ on this task to obtain the adapted parameters θ i ;

[0045] 3.3.3) Use the adapted parameters θ i The loss is calculated on the validation set of the task, and the validation losses of multiple tasks are accumulated to update the initial parameters θ. The updated parameters θ* for:

[0046]

[0047] in, Indicates the task θ Parameters after the inner loop update.

[0048] The step 3.4) comprises the following steps:

[0049] The main task is to determine whether the pathological mechanism is upper airway muscle dysfunction. At the same time, auxiliary classification tasks and auxiliary regression tasks are constructed. All tasks share input features and use θ * As the initial parameters of the model, multiple prediction results are obtained and the joint loss function is used Train the model.

[0050] The joint loss function for:

[0051]

[0052] in, is the cross entropy loss of the main task, is the cross entropy of the auxiliary classification task, is the mean square error of the auxiliary regression task, λ c , a and λ r All are weights.

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

[0054] 1. The present invention combines force sensors with feature extraction to comprehensively capture the contraction mechanics information of the upper airway muscles, thus overcoming the data loss and misjudgment caused by the previous single measurement method.

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

[0056] 3. The present invention adopts a multi-task learning strategy to simultaneously predict multi-dimensional indicators such as muscle functional status and fatigue level, helping to quickly identify functional abnormalities and provide targeted guidance for subsequent interventions. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0058] Figure 2 This is a muscle strength-time relationship diagram for training of the present invention;

[0059] Figure 3 It is a schematic diagram of the training evaluation method process of the present invention. DETAILED DESCRIPTION

[0060] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0061] like Figure 1 The figure is a schematic diagram of the structure flow of the training evaluation guidance system of the present invention, which specifically includes:

[0062] Signal flow 1, collects the upper respiratory tract muscle group signal through the sensor, and the specific implementation method is to place a high-precision S-shaped tension pressure sensor in front of the trainee's face. The tension pressure sensor is made of S-shaped stainless steel, with a pressure sensing range of 0-20N, an accuracy of 0.5%, and a sampling rate of 20Hz. The measurement principle is that the trainee extends the tongue, and the pressure applied by the tongue on the effective area of ​​the sensor increases, the sensor output resistance decreases, and the resistance value change is converted into a voltage value change through the circuit in the signal conditioning module. The tension pressure sensor is sealed with a 1mm edible silicone seal for waterproofing. Before use, the relationship between the sensor and the voltage change is calibrated using a calibration experiment.

[0063] Signal flow 2: The signal processing unit amplifies and filters the received voltage, and then uses the analog-to-digital conversion circuit to perform digital sampling processing. The digitized electromyographic pressure signal is stored in the system's storage chip to facilitate digital processing and signal playback.

[0064] Signal process 3, feature extraction of the collected electromyographic stress signal, as shown in Table 1, the electromyographic stress signal features include the maximum muscle group force F max , muscle strength stability F cons , fatigue index TI, total work W, and overall test time T. The characteristics are described as follows:

[0065] 1) Maximum muscle strength F max The maximum muscle group strength is that this feature can reflect the peak force level produced by the subject when the upper respiratory muscles are activated for a short time. max It is significantly lower than the standard value of a specific population, indicating that the subject's muscle force generation capacity is insufficient. When the test begins, the trainer will perform 3-5 consecutive tongue extensions with maximum force, each action lasting 3 seconds, with an interval of 30 seconds, and record the maximum force value of each time, and take the maximum value;

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

[0067] Among them, f peak,iis the peak value of the ith maximum force attempt.

[0068] 2) Muscle strength stability F cons Muscle strength stability is used to evaluate the stability of the upper respiratory tract muscle group's muscle control during the entire training process. The test process data (excluding the maximum force test once a minute) is used to evaluate the force output fluctuation using the coefficient of variation of the data:

[0069]

[0070] in, is the average force value during this period, σ f is the standard deviation of the force value during this period. cons The smaller the value, the more stable the force is.

[0071] 3) Total work W. The total work is the accumulated value of the electromyographic stress signal over time during the training process.

[0072]

[0073] Among them, f(t n ) is the electromyographic stress signal of the nth segment, Δt is the test interval, and W reflects the "energy" consumption during the entire test process, which can characterize the overall muscle output capacity and persistence of the trainee.

[0074] 4) Overall test time T. Training time T set It can be set before training. However, due to personal factors or fatigue, the test ends early, making the test time less than the set training time. Therefore, it is necessary to follow the actual test time. When the actual test time is T≤T set , it means 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. Record the maximum force test work done once a minute to form a point set of test time and discrete work data Using linear fitting:

[0076] W(t k )≈β0+β1t k

[0077] Among them, fatigue index TI = β1. When TI < 0, it means that the amount of work done decreases over time, that is, muscle fatigue increases. The steeper the slope, the more severe the fatigue.

[0078] Table 1

[0079] Trainer Training Features Unit or Category Remark <![CDATA[Maximum muscle group strength F max > N The measurement result at the beginning of the test shall prevail <![CDATA[Muscle strength stability F cons > Dimensionless Muscle strength fluctuations during the test Fatigue Index TI Dimensionless Evaluate the decrease in muscle strength during the test Total work W N·s Evaluate the overall completion strength of the test taker Test time T Second Determine how long the tester takes to test

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

[0081] Table 2

[0082]

[0083] like Figure 3 FIG. 1 is a schematic diagram of the training evaluation method process of the present invention, which specifically includes:

[0084] Step 1: Preparation before the test. Subjects were asked to avoid excessive oral activities (such as playing musical instruments) within 30 minutes before the test and to keep their body and mind relaxed.

[0085] Step 2: Upper respiratory muscle maximum strength test. The relationship between training strength and time is as follows: Figure 2 As shown, start this assessment training. First, according to the guidance of the trainer, the trainee opens his mouth and bites the connecting rod of the pressure sensor, and ensures that the distance between the tongue and the pressure sensor remains unchanged. Use the maximum force to press the force sensor in front of the tongue. Do it 3-5 times in a row, each action lasts for 3s, with an interval of 30s, and calculate the maximum muscle group force F max .

[0086] Step 3: Upper respiratory muscle strength maintenance and fatigue detection. In subsequent training, the training duration can be set before training, and the general setting time is 30 minutes. The specific training includes two parts: continuous strength maintenance test and maximum muscle strength regular retest. The specific process of continuous strength maintenance test is: set a moderate intensity target pressure window, and the pressure window range is [F MAX *50%, F MAX *70%], requiring the subjects to maintain the tongue extension force in this range for 3-5 seconds. In order to allow the trainees to concentrate, the game is set to a dynamic value for the duration of the pressure window, ranging from 3 to 5 seconds, and the rest time is a dynamic value, ranging from 10 to 20 seconds. During the training time, the maximum muscle strength is regularly retested every 1 minute, and a range of [F MAX *80%, F MAX *100%], the duration is 5 seconds. If some of the data in the pressure value test is within the pressure window, it is considered that the trainee is not in a fatigue state and can continue the test. If the completion results of two consecutive maximum force tests are both lower than the pressure window, it is considered that the trainee has entered a fatigue state and is prompted to terminate the training.

[0087] Step 4, feature extraction. After the muscle strength test in step 2 and the fatigue test in step 3, the electromyographic stress signal features can be extracted after signal denoising, baseline correction, calibration and regularization, and the complete features can be obtained by combining the trainer's personal information. The features are normalized to reduce the feature scale differences caused by physiological differences between different individuals.

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

[0089] 1) Basic model. The features extracted above are similar to tabular data. Therefore, the present invention combines the TabTransformer structure for feature extraction and effectively utilizes 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 and form a feature embedding sequence. The feature embedding sequence is input into the TabTransformer. Through the multi-layer stacked self-attention and feedforward network (FNN) modules, the TabTransformer can automatically capture the complex interactive relationships between features, such as F max The nonlinear relationship between TI and the final representation Z L Through the pooling layer, a global representation vector z is formed, which is used as a high-level feature representation input into the subsequent decision layer.

[0090] 2) Considering the real clinical scenario, although clinical experts can provide the pathological mechanism of OSA patients, it is costly and time-consuming to label all the data. Therefore, most of the collected data are unlabeled data. In order to make better use of these data, we introduce a self-supervised contrastive learning method to learn a discriminative general representation from unlabeled data. Using the SimCLR framework, the data of each unlabeled subject is regarded as a sample. First, construct a positive sample pair. Generate X through two different enhancement strategies i and (For example, adding a small additive noise to TI and F cons Introducing small random scaling, etc.), these two should be mapped to similar feature representation spaces by the model. Use TabTransformer as the encoder f θ (·), maps the input feature vector to the latent representation Z = f θ (X). The contrast loss function is defined as:

[0091]

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

[0093] 3) In order to further reduce the situation where data and corresponding labels are limited in actual clinical scenarios. The present invention introduces the MAML framework, which enables the model to quickly obtain higher performance in target tasks with only a small amount of labeled data by training an initial model parameter that can quickly adapt to new tasks on a series of tasks. In the present invention, different scenarios (such as different breathing posture conditions, different load training conditions, different time periods, etc.) are constructed based on patients determined by clinical experts to construct different "tasks" and train the model based on this. Extract tasks from task distribution Perform one or more gradient descents on the model parameters θ on this task to obtain the adapted parameters θ i The loss is calculated on the validation set of the task using the adapted parameters, and the validation losses of multiple tasks are accumulated to update the initial parameters θ.

[0094]

[0095] in Indicates the task θ The parameters updated by the inner loop above. After meta-training, θ * Have the ability to quickly adapt to relevant new 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. In order to improve the reliability and generalization performance of the model for tasks, the present invention constructs a multi-task learning strategy. The main task is to determine whether the pathological mechanism is upper airway muscle dysfunction, while considering auxiliary classification tasks (determining the degree of muscle weakness in patients: mild, moderate, severe) and auxiliary regression tasks (predicting the trajectory of signal feature changes after intervention training). These tasks share input features and use the θ obtained from meta-learning. * As the initial parameters of the model, multiple predictions are output simultaneously and trained using a joint loss function.

[0097]

[0098] in, is the cross entropy loss for the main task; is the cross entropy for the auxiliary classification task; is the mean square error of the auxiliary regression task.

[0099] 5) Model deployment and practical application. For a patient who has not yet received treatment, the complete features are input into the model, and multi-task evaluation results can be obtained to assist clinicians in evaluation and provide guidance for upper airway muscle training programs. With the passage of time and the accumulation of data, the patient's upper airway muscle function may change with the progress of treatment, and the model can be updated regularly. For patients who are receiving treatment, unlabeled data (mostly daily measurements) and very few clearly marked key time points (such as a clinician follow-up visit on the Nth day after the start of treatment) can be collected. On the one hand, the model can be used to evaluate the effectiveness of training, and on the other hand, it can be used as new semi-supervised data to update the model.

Claims

1. A method for measuring and evaluating the mechanical properties of upper airway muscles based on signal characteristics, characterized in that: The following steps are involved: 1) The electromyographic pressure signals of the upper respiratory tract muscles under specific actions are collected through the tension pressure sensor, and amplified, filtered, and analog-to-digital converted in sequence; 2) extracting signal features from the processed electromyographic pressure signal; 3) Evaluate the mechanical properties of the upper respiratory tract muscles based on signal characteristics and individual parameters.

2. A method for measuring and evaluating the mechanical properties of upper airway muscles based on signal characteristics according to claim 1, characterized in that: The step 2) comprises the following steps: 2.1) Extract the maximum muscle force F max , perform a maximum force test every minute, record the maximum force value of each training, and take the maximum value F max : F max =max{f peak,1 ,f peak,2 ,…,f peak,i } Among them, f peak,i is the peak value of the ith maximum force attempt; 2.2) Extracting muscle force stability F cons : in, is the average force value in a certain period of time, σ f is the standard deviation of the force value during this period; 2.3) Extract the total work W: Among them, f(t n ) is the electromyographic stress signal of the nth segment, Δt is the test interval, and W is the energy consumption during the entire test process; 2.4) Extract the overall test time T; 2.5) Extract the fatigue index TI and record the maximum force test work done once a minute to form a point set of test time and discrete work data Using linear fitting: W(t k )≈β0+β1t k Among them, β0 is the intercept in linear regression, β1 is the slope in linear regression, and t k At the time point corresponding to the kth work measurement, the fatigue index TI = β1.

3. The method for measuring and evaluating the mechanical properties of upper airway muscles based on signal characteristics according to claim 1, characterized in that: The step 3) comprises the following steps: 3.1) Using the TabTransformer model, the signal features are processed into high-level feature representations as global representation vectors; 3.2) Learning discriminative general representations from unlabeled high-level feature representation data through self-supervised contrastive learning; 3.3) Using the MAML framework, we 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) Build a multi-task learning strategy to fine-tune the trained TabTransformer model; 3.5) The signal characteristics and individual parameters are used as input and the adjusted model is used to evaluate the mechanical properties of the upper airway muscles.

4. The method for measuring and evaluating the mechanical properties of upper airway muscles based on signal characteristics according to claim 3, characterized in that: Step 3.1) The following steps are involved: 3.1.1) The numerical signal features are mapped to a high-dimensional vector space through the Embedding layer and form a feature embedding sequence; 3.1.2) Input the feature embedding sequence into TabTransformer, and obtain the feature representation Z through multi-layer stacked self-attention and feedforward networks L ; 3.1.3) Represent the feature Z L Through the pooling layer, a global representation vector z is formed and input into the subsequent decision layer as a high-level feature representation.

5. The method for measuring and evaluating the mechanical properties of upper airway muscles based on signal characteristics according to claim 3, characterized in that: The step 3.2) comprises the following steps: Using the SimCLR framework, the data of each unlabeled subject is taken as a sample, and two different enhancement strategies are used to generate sample pairs X i and Use TabTransformer as encoder f θ (·), mapping sample pairs to potential representation z = f θ (X); Constructing the contrastive loss function By minimizing Learning discriminative feature representations from unlabeled data.

6. The method for measuring and evaluating the mechanical properties of upper airway muscles based on signal characteristics according to claim 5, characterized in that: The contrast loss function for: in, is the cosine similarity, and τ is the temperature coefficient.

7. The method for measuring and evaluating the mechanical properties of upper airway muscles based on signal characteristics according to claim 3, characterized in that: The step 3.3) comprises the following steps: 3.3.1) Construct different tasks in different scenarios and use them as the basis for training models; 3.3.2) Extracting tasks from the task distribution Perform one or more gradient descents on the TabTransformer model parameters θ on this task to obtain the adapted parameters θ i ; 3.3.3) Use the adapted parameters θ i The loss is calculated on the validation set of the task, and the validation losses of multiple tasks are accumulated to update the initial parameters θ. The updated parameters θ * for: in, Indicates the task θ Parameters after the inner loop update.

8. The method for measuring and evaluating the mechanical properties of upper airway muscles based on signal characteristics according to claim 3, characterized in that: The step 3.4) comprises the following steps: The main task is to determine whether the pathological mechanism is upper airway muscle dysfunction. At the same time, auxiliary classification tasks and auxiliary regression tasks are constructed. All tasks share input features and use θ * As the initial parameters of the model, multiple prediction results are obtained and the joint loss function is used Train the model.

9. The method for measuring and evaluating the mechanical properties of upper airway muscles based on signal characteristics according to claim 8, characterized in that: The joint loss function for: in, is the cross entropy loss of the main task, is the cross entropy of the auxiliary classification task, is the mean square error of the auxiliary regression task, λ c , a and λ r All are weights.

Images