Muscle relaxation prediction method and system
The method and system address the delays and usability issues of existing muscle relaxation monitoring by using throat video data and vital signs to predict muscle relaxation levels, enhancing feedback speed and clinical applicability.
Patent Information
- Application Number
- CN202510450745.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-10
- Publication Date
- 2025-07-15
AI Technical Summary
The existing muscle relaxation monitors have high feedback hysteresis on the laryngeal muscle relaxation at the hand measurement position, low penetration rate, high equipment requirements, and limited clinical application.
By collecting throat video data and vital sign data in real time, extracting spatiotemporal and vital sign characteristics in the throat, combining patient static information, and using superimposed fitting of muscle relaxation prediction models of multiple regression decision trees to achieve evaluation and prediction of muscle relaxation degree.
The short feedback delay time of laryngeal muscle relaxation is achieved, which is convenient for clinical operation, and improves the accuracy and popularity of muscle relaxation monitoring.
Smart Images
Figure CN120304832A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of medical technologies, and more specifically, to a muscle relaxation prediction method and system. Background Art
[0002] Currently, for muscle relaxation monitoring during general anesthesia, the method mainly used is to monitor the adductor pollicis muscle of the hand with a muscle relaxation detector. The monitoring end of a common muscle relaxation detector includes three connectors, a positive electrode end, a negative electrode end, and an acceleration sensor end. The installation method is as follows. The patient lies with the palm up and is loosely fixed to the side of the body with a cloth towel (it is required that the upper limb is slightly separated from the trunk). The negative electrode (black) is pasted 1 cm proximal to the wrist crease, on the radial side of the flexor carpi ulnaris muscle. The positive electrode (red) is pasted 2 - 3 cm proximal to the negative electrode (black), ensuring that the electrical stimulation crosses the ulnar nerve. The acceleration sensor is fixed to the palmar surface of the thumb with adhesive tape. The monitoring site should be avoided being pressed by the surgical drape to ensure that the movement of the monitoring site is not hindered. Muscle relaxation monitoring is carried out by setting the required measurement mode, current intensity, and interval time.
[0003] However, the current measurement position of the muscle relaxation monitor is on the hand, which has a certain lag in the feedback of laryngeal muscle relaxation. Moreover, the current muscle relaxation monitor has a low penetration rate, high equipment requirements, and limited clinical applications. Summary of the Invention
[0004] This application provides a muscle relaxation prediction method and system. By extracting the spatio-temporal features of the larynx from the laryngeal video data, combining with the vital sign features and static information of the patient, the evaluation and prediction of the intraoperative muscle relaxation degree are realized, the muscle relaxation feedback delay time is short, and it is convenient for clinical operation.
[0005] This application provides a muscle relaxation prediction method, including:
[0006] During the surgical process, the laryngeal video data and vital sign data of the patient are collected in real time;
[0007] The spatio-temporal features of the larynx and the vital sign features are respectively extracted from the laryngeal video data and the vital sign data;
[0008] The spatio-temporal features of the larynx, the vital sign features, and the static information of the patient are used as input data and input into the muscle relaxation prediction model to obtain the muscle relaxation prediction result.
[0009] Preferably, extracting the spatio-temporal features of the larynx from the laryngeal video data includes:
[0010] Segmenting the laryngeal video data in the time domain into several sub-segments with a preset duration;
[0011] Extracting the corresponding spatio-temporal features from each sub-segment.
[0012] Preferably, the spatio-temporal features include laryngeal deformation features and the features of the tension and texture changes of the laryngeal muscles.
[0013] Preferably, extracting the laryngeal deformation features includes:
[0014] Capturing the inter-frame correlation within each sub-fragment;
[0015] Obtaining the deformation range of the glottis from full closure to opening and then back to closure based on the inter-frame correlation, thereby obtaining the "peak-valley" feature;
[0016] Calculating the local motion vectors of the upper and lower edges of the glottis based on the inter-frame correlation to obtain the jitter or amplitude feature of the glottis.
[0017] Preferably, the output value of the muscle relaxation prediction model is fitted by superimposing multiple regression decision trees, and the muscle relaxation prediction result is obtained by using the muscle relaxation prediction model, including:
[0018] Repeatedly execute the following steps until the number of operation rounds reaches the preset number of rounds or the error reaches the preset error:
[0019] In each round of operation, calculate the residual between the output value of the previous round and the true muscle relaxation value;
[0020] Use the input data as the input and the residual as the target to train a decision tree to obtain the output value of the leaf nodes in this round of operation;
[0021] Calculate the output value of this round of operation based on the output value of the leaf nodes and the learning rate;
[0022] Add the decision tree of this round and the output value of this round of operation to the model for the next round of iterative operation;
[0023] Sum up all the decision trees based on the learning rate to obtain the predicted value.
[0024] Preferably, extracting the features of the tension and texture changes of the laryngeal muscles includes:
[0025] Extract the edge features of the sub-fragment and perform robustness processing on abnormal lighting and angle changes to obtain the processed sub-fragment;
[0026] Identify the "collapse" or "offset" parts in the processed sub-fragment;
[0027] Obtain the features of the tension and texture changes of the laryngeal muscles based on the "collapse" or "offset" parts.
[0028] This application also provides a muscle relaxation prediction system, including a data acquisition device, a feature extraction module, and a prediction module;
[0029] The data acquisition device is used to collect the laryngeal video data and vital sign data of the patient in real time during the operation;
[0030] The feature extraction module is used to extract laryngeal spatio-temporal features and vital sign features from laryngeal video data and vital sign data respectively;
[0031] The prediction module is used to input the laryngeal spatio-temporal features, vital sign features and patient static information as input data into the muscle relaxation prediction model to obtain the muscle relaxation prediction result.
[0032] Preferably, the feature extraction module includes a video segmentation module and a spatio-temporal feature extraction module;
[0033] The video segmentation module is used to segment the laryngeal video data in the time domain into several sub-fragments with a preset duration;
[0034] The spatio-temporal feature extraction module is used to extract corresponding spatio-temporal features from each sub-fragment.
[0035] Preferably, the spatio-temporal features include laryngeal deformation features and the change features of the tension and texture of laryngeal muscles.
[0036] Preferably, the spatio-temporal feature extraction module includes a capture module, a first feature acquisition module and a second feature acquisition module;
[0037] The capture module is used to capture the inter-frame correlation degree inside each sub-fragment;
[0038] The first feature acquisition module is used to obtain the deformation range of the glottis from full closure to opening and then back to closure based on the inter-frame correlation degree, so as to obtain the "peak-valley" feature;
[0039] The second feature acquisition module is used to calculate the local motion vectors of the upper and lower edges of the glottis based on the inter-frame correlation degree to obtain the jitter or amplitude feature of the glottis.
[0040] Preferably, the spatio-temporal feature extraction module includes a preprocessing module, an identification module and a third feature acquisition module;
[0041] The preprocessing module is used to extract edge features from the sub-fragments and perform robustness processing on abnormal illumination and angle changes to obtain the processed sub-fragments;
[0042] The identification module is used to identify the "collapse" or "offset" parts in the processed sub-fragments;
[0043] The third feature acquisition module is used to obtain the change features of the tension and texture of laryngeal muscles based on the "collapse" or "offset" parts.
[0044] Preferably, the prediction module includes a residual calculation module, a decision tree training module, a round output value calculation module, an iteration module and a prediction value calculation module;
[0045] The residual calculation module is used to calculate the residual between the output value of the previous round and the true muscle relaxation value in each round of training;
[0046] The decision tree training module is used to train a decision tree with the input data as the input and the residual as the target, and obtain the output value of the leaf nodes of this round of operation;
[0047] The round output value calculation module is used to calculate the output value of this round of training based on the output value of the leaf nodes and the learning rate;
[0048] The iteration module is used to add the decision tree of this round and the output value of this round of operation to the model for the next round of iterative operation;
[0049] The predicted value calculation module is used to sum all the decision trees based on the learning rate to obtain the predicted value.
[0050] Through the following detailed description of the exemplary embodiments of the present application with reference to the accompanying drawings, other features and advantages of the present application will become clear. Brief Description of the Drawings
[0051] The drawings incorporated in and constituting a part of this specification illustrate embodiments of the present application and, together with the description, serve to explain the principles of the present application.
[0052] Figure 1 It is a flowchart of the muscle relaxation prediction method provided by the present application;
[0053] Figure 2 It is a structural diagram of an embodiment of the muscle relaxation prediction system provided by the present application. Detailed Description of the Specific Embodiment
[0054] Now, various exemplary embodiments of the present application will be described in detail with reference to the accompanying drawings. It should be noted that: Unless otherwise specifically stated, the relative arrangements of components and steps, numerical expressions and values set forth in these embodiments do not limit the scope of the present application.
[0055] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way a limitation on the present application or its application or use.
[0056] Techniques, methods, and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, the techniques, methods, and devices should be considered as part of the specification.
[0057] In all the examples shown and discussed here, any specific values should be construed as merely exemplary and not as limitations. Therefore, other examples of the exemplary embodiments may have different values.
[0058] The present application provides a muscle relaxation prediction method and system. By extracting the spatio-temporal features of the larynx from laryngeal video data and combining the vital sign features and static information of the patient, the assessment and prediction of the intraoperative muscle relaxation degree are realized, the muscle relaxation feedback delay time is short, and it is convenient for clinical operation.
[0059] As Figure 1 shown, the muscle relaxation prediction method provided by the present application includes:
[0060] S110: During the operation, the laryngeal video data and vital sign data of the patient are collected in real time.
[0061] During the operation, the laryngeal video data (in MP4 format) can be recorded by the central camera inside the body of the visual laryngeal mask. The video recording duration for each patient is a preset duration T1, and the continuous video of the entire larynx and glottis area of the patient from the placement of the laryngeal mask before the operation to the removal of the laryngeal mask at the end of the operation is recorded, including all the characteristics of the laryngeal and glottic movements under different muscle relaxation degrees of the patient.
[0062] The vital sign data of the patient during the operation (including non-invasive blood pressure, heart rate, pulse oxygen saturation, bispectral index of electroencephalogram, muscle relaxation monitoring, etc.) are recorded by the monitor, reflecting the changes in vital signs under different muscle relaxation degrees of the patient. At the end of the operation, the data is exported from the monitor through the data cable.
[0063] S120: Extract the laryngeal spatio-temporal features and vital sign features from the laryngeal video data and vital sign data respectively.
[0064] It can be understood that before extracting the data, the data also needs to be cleaned. In the laryngeal video data, some video data are blurred, have abnormal perspectives, recording interruptions, etc. Therefore, during data cleaning, the abnormal laryngeal video data are marked, and other data associated with the abnormal laryngeal video data (including vital sign data) are excluded and not used for subsequent model training.
[0065] In the laryngeal video, the larynx and glottis will show different degrees of opening and closing and jitter with the respiratory cycle, the patient's voluntary movements, and the change of anesthetic drug concentration, reflecting the difference in the depth of muscle relaxation. As an embodiment, a video understanding model based on Transformer is used to extract the laryngeal spatio-temporal features from the cleaned laryngeal video data to obtain the sequence data that can be used by the muscle relaxation prediction model. The TimeSFormer model is a deep learning model mainly used to process sequence data. The video understanding model constructs the feature representation of the video through the self-attention mechanism of space and time, can automatically identify the spatio-temporal features related to muscle relaxation, and "encode" them into the input features available for the muscle relaxation prediction model.
[0066] As an embodiment, extracting the laryngeal spatio-temporal features from the laryngeal video data (such as using the muscle relaxation prediction model) includes:
[0067] S1201: Segment the laryngeal video data in the time domain into several sub - segments with a preset duration (e.g., one respiratory cycle or n seconds).
[0068] S1202: Extract the corresponding spatio - temporal features from each sub - segment.
[0069] As an embodiment, the spatio - temporal features include laryngeal deformation features and the features of the tension and texture changes of the laryngeal muscles.
[0070] When extracting the laryngeal deformation features, use a video understanding model based on Transformer to perform temporal encoding on each sub - segment, and extract the feature vectors of the deformation of the laryngeal muscle groups or the glottis. Among them, the maximum displacement amount (such as the glottis opening degree) and the jitter frequency (the pixel difference or the feature point trajectory between frames) in the respiratory cycle can reflect a certain degree of muscle relaxation.
[0071] As an embodiment, extracting the laryngeal deformation features includes:
[0072] P1: Use the spatial / temporal self - attention mechanism to capture the inter - frame correlation within each sub - segment.
[0073] P2: Obtain the deformation range of the glottis from full closure to opening and then back to closure based on the inter - frame correlation, so as to obtain the "peak - valley" features.
[0074] P3: Calculate the local motion vectors of the upper and lower edges of the glottis based on the inter - frame correlation to obtain the jitter or amplitude features of the glottis.
[0075] When the effect of the muscle relaxant drug increases, the tension of the laryngeal soft tissue weakens, and "collapse" or "offset" features may appear in the video frames. Therefore, as an embodiment, extracting the features of the tension and texture changes of the laryngeal muscles includes:
[0076] Q1: Perform edge feature extraction on the sub - segment and perform robustness processing on abnormal illumination and angle changes to obtain the processed sub - segment.
[0077] Q2: Identify the "collapse" or "offset" parts in the processed sub - segment.
[0078] Q3: Obtain the features of the tension and texture changes of the laryngeal muscles based on the "collapse" or "offset" parts.
[0079] In vital sign data (such as indicators like heart rate, blood pressure, bispectral index BIS, etc.), certain fluctuations are positively or negatively correlated with the degree of muscle relaxation. When extracting vital sign features from vital sign data, from the monitor and preoperative visit data (such as bispectral index BIS, tetralogy of Fallot TOF, non-invasive blood pressure NIBP, blood oxygen saturation SpO2, etc.), according to clinical prior knowledge and statistical analysis, indicators with a relatively high degree of association with the degree of muscle relaxation are selected (such as the association between BIS and skeletal muscle tension, TOF ratio and block degree, etc.).
[0080] As an example, a machine learning model is used to extract vital sign features from vital sign data. Before pre-training the machine learning model, the vital sign features are differentiated or normalized, and the data is cleaned (including screening out abnormal data). Given the differences in the baseline values of different patients, the original vital signs are normalized or differentiated (such as taking "current value - baseline value") to highlight the true change part after muscle relaxant intervention. Abnormal data (such as extreme jumps or physiologically unreasonable points that may be caused by equipment failures or temporary interferences) are interpolated or removed to ensure the stability of subsequent model training.
[0081] S130: Input the laryngeal spatio-temporal features, vital sign features, and patient static information as input data into the muscle relaxation prediction model to obtain the muscle relaxation prediction result.
[0082] Among them, patient static information includes personal information and basic disease information, such as gender, age, height, weight, preoperative assessment of surgical anesthesia, etc., which are obtained through the case system and / or preoperative visit of the patient. Optionally, the vital sign data and patient static information can be combined into a csv table. After blurring the privacy data, it is integrated with the laryngeal video data into a patient information database.
[0083] It should be noted that during the training phase of the muscle relaxation prediction model, when obtaining the dataset, the laryngeal spatio-temporal features, vital sign features, and patient static information are first preprocessed to obtain fused data as the dataset. Specifically, the preprocessing includes reducing the redundant dimensions in the laryngeal spatio-temporal features, vital sign features, and patient static information through PCA (Principal Component Analysis) or AutoEncoder, retaining the most important muscle relaxation difference information, and then performing temporal alignment so that the laryngeal spatio-temporal features, vital sign features, and patient static information in the same time segment are data-matched to ensure the consistency of model training. The obtained dataset is randomly divided into a training set (e.g., 80%) and a test set (e.g., 20%) according to the patient number. The training set data is used for the training of the muscle relaxation prediction model, and the test set data is used for the evaluation of the muscle relaxation prediction model. For the test set data, the corresponding muscle relaxation value is predicted using the model, and the evaluation metrics are ROC (Receiver Operating Characteristic Curve) and AUC (Area Under Curve).
[0084] It should be noted that during the prediction phase, the spatio-temporal features, vital sign features, and patient static information are extracted according to the data dimensions obtained after dimensionality reduction in the training phase to form the input data of the muscle relaxation prediction model.
[0085] As an example, the muscle relaxation prediction model uses the XGBoost (eXtreme Gradient Boosting) algorithm, which is an efficient gradient boosting library. It is based on the gradient boosting decision tree (GBDT) framework and optimizes the loss function by gradually adding weak learners to improve the model performance.
[0086] As an example, XGBoost fits the output value of the muscle relaxation prediction model by stacking multiple regression decision trees. The muscle relaxation prediction results obtained using the muscle relaxation prediction model include:
[0087] S1301: Construction of the input feature vector:
[0088] The input features include laryngeal spatio-temporal features, vital sign features, and patient static information. First, these three types of data need to be integrated to construct an input feature vector as the input data of the muscle relaxation prediction model.
[0089] Specifically, for each sub-segment, the spatio-temporal feature is defined as V t ∈ R d The vital sign feature is defined as S t ∈R m The patient static information is defined as P t ∈ R p, first, synchronize and align the spatio-temporal features and vital sign features, then splice or broadcast these three data into the features at the corresponding time, and finally obtain the input feature vector X t =[ V t， S t， P t , as the input data.
[0090] S1302: Iterative operation
[0091] First, perform initialization: Let the initial predicted output y (0) be the global mean, and set the iterative increment.
[0092] Secondly, perform iterative operation. In each round of operation (taking the k-th round as an example), calculate the residual r = y - y ^(k-1) between the output value y ^(k-1) of the previous round and the true muscle relaxation value y. Then, using the input data X t as the input and the residual r as the target, train a decision tree T k , and obtain the output value of the leaf node of this round of operation. Based on the output value T k (X t ) and the learning rate η, calculate the output value y ^(k ) of this round of operation: y ^(k ) = y ^(k-1) + η·T k (X t ). Add the decision tree of this round and the output value of this round of operation to the model for the next round of iterative operation. It should be noted that to prevent overfitting, XGBoost penalizes the complexity Ω(Tk) of each tree and searches for the optimal feature split point through greedy splitting.
[0093] Repeat the above steps, continuously iterate the decision tree until the preset number of rounds or error convergence is reached.
[0094] S1303: Calculation of predicted value
[0095] Sum all the decision trees based on the learning rate to obtain the predicted value ŷ t corresponding to the sub-segment: ŷ t = ∑ k k=1 η· T k (X t ), where K is the total number of trees.
[0096] As an example, ŷ t can be directly used as an index such as "depth of muscle relaxation" or "muscle relaxation block rate".
[0097] Preferably, if it is necessary to discretize the prediction results (such as grading: no block, mild, moderate, severe), threshold segmentation or multi-class probability output (in the form of Softmax) can be added at the end of the muscle relaxation prediction model.
[0098] Numerical example: Suppose at time t, the dimension of the spatio-temporal feature V t is 128, the vital sign feature S t has 10 monitoring values, and the patient's static information P t has 5 dimensions, then the input feature vector X t is a total of 143 dimensions. XGBoost builds trees in sequence. For example, the first tree learns that the movement amplitude of certain frames and the BIS value can significantly distinguish between high and low muscle relaxation; the second tree then fine-tunes the residuals and pays attention to the correlation between the respiratory rate and the video opening and closing degree, etc.; the kth tree considers individual differences such as age and weight... Finally, we get ŷ t . If ŷ t ≈ 70, it means the patient is in a relatively deep muscle relaxation state; if ŷ t ≈ 20, the muscle relaxation degree is relatively shallow.
[0099] It should be noted that the training process of the muscle relaxation prediction model is the same as the above prediction process. The muscle relaxation prediction model learns the mapping relationship between the "spatio-temporal feature - vital sign feature - patient static information" three-in-one multi-dimensional input and the muscle relaxation value.
[0100] Based on the above, the present application also provides a muscle relaxation prediction system. As Figure 2 shown, the muscle relaxation prediction system includes a data acquisition device 210, a feature extraction module 220, and a prediction module 230.
[0101] The data acquisition device 210 is used to collect the patient's laryngeal video data and vital sign data in real time during the operation.
[0102] As an embodiment, as Figure 2 shown, the data acquisition device includes a visible laryngeal mask and a monitor.
[0103] The feature extraction module 220 is used to extract the laryngeal spatio-temporal feature and the vital sign feature from the laryngeal video data and the vital sign data respectively.
[0104] As an embodiment, as Figure 2 shown, the laryngeal spatio-temporal feature is extracted from the laryngeal video data through a video understanding model based on Transformer. The vital sign feature is extracted from the vital sign data through a machine learning model.
[0105] The prediction module 230 is used to input the laryngeal spatio-temporal feature, the vital sign feature, and the patient's static information as input data into the muscle relaxation prediction model to obtain the muscle relaxation prediction result.
[0106] As an embodiment, as Figure 2 shown, the muscle relaxation prediction model adopts the XGBoost algorithm.
[0107] Preferably, the feature extraction module 220 includes a video segmentation module and a spatio-temporal feature extraction module;
[0108] The video segmentation module is used to segment the laryngeal video data in the time domain into a plurality of sub-fragments with a preset duration;
[0109] The spatio-temporal feature extraction module is used to extract corresponding spatio-temporal features from each sub-fragment.
[0110] Preferably, the spatio-temporal features include laryngeal deformation features and changes in the tension and texture of laryngeal muscles.
[0111] Preferably, the spatio-temporal feature extraction module includes a capture module, a first feature acquisition module, and a second feature acquisition module;
[0112] The capture module is used to capture the inter-frame correlation within each sub-fragment;
[0113] The first feature acquisition module is used to obtain the deformation range of the glottis from full closure to opening and then back to closure based on the inter-frame correlation, so as to obtain the "peak-valley" feature;
[0114] The second feature acquisition module is used to calculate the local motion vectors of the upper and lower edges of the glottis based on the inter-frame correlation, and obtain the jitter or amplitude feature of the glottis.
[0115] Preferably, the spatio-temporal feature extraction module includes a preprocessing module, an identification module, and a third feature acquisition module;
[0116] The preprocessing module is used to extract edge features from the sub-fragments, and perform robustness processing on abnormal illumination and angle changes to obtain the processed sub-fragments;
[0117] The identification module is used to identify the "collapse" or "offset" parts in the processed sub-fragments;
[0118] The third feature acquisition module is used to obtain the changes in the tension and texture of laryngeal muscles based on the "collapse" or "offset" parts.
[0119] Preferably, the prediction module 230 includes an initialization module, a residual calculation module, a decision tree training module, a round output value calculation module, an iteration module, and a predicted value calculation module;
[0120] The initialization module is used to make the initial prediction output y (0) be the global mean and set the iteration increment.
[0121] The residual calculation module is used to calculate the residual between the output value of the previous round and the true muscle relaxation value in each round of training;
[0122] The decision tree training module is used to train a decision tree with the input data as the input and the residual as the target, and obtain the output value of the leaf nodes of this round of operation;
[0123] The round output value calculation module is used to calculate the output value of this round of training based on the output value of the leaf nodes and the learning rate;
[0124] The iteration module is used to add the decision tree of this round and the output value of this round of operation to the model for the next round of iterative operation;
[0125] The predicted value calculation module is used to sum all the decision trees based on the learning rate to obtain the predicted value.
[0126] Although some specific embodiments of the present application have been described in detail by way of examples, those skilled in the art should understand that the above examples are only for illustration purposes and not for limiting the scope of the present application. Those skilled in the art should understand that the above embodiments can be modified without departing from the scope and spirit of the present application. The scope of the present application is defined by the appended claims.
Claims
1. A muscle relaxation prediction method, characterized in that, Including: During the operation, the laryngeal video data and vital sign data of the patient are collected in real time; Laryngeal spatio-temporal features and vital sign features are respectively extracted from the laryngeal video data and the vital sign data; The laryngeal spatio-temporal features, the vital sign features and the patient's static information are used as input data to be input into a muscle relaxation prediction model to obtain a muscle relaxation prediction result.
2. The muscle relaxation prediction method according to claim 1, wherein Extracting laryngeal spatio-temporal features from the laryngeal video data includes: The laryngeal video data is segmented in the time domain into several sub-fragments with a preset duration; The corresponding spatio-temporal features are extracted from each sub-fragment.
3. The muscle relaxation prediction method according to claim 2, wherein, The spatio-temporal features include laryngeal deformation features and the features of the tension and texture changes of the laryngeal muscles.
4. The muscle relaxation prediction method according to claim 3, characterized in that, Extracting the laryngeal deformation features includes: Capturing the inter-frame correlation degree within each sub-fragment; According to the inter-frame correlation degree, obtaining the deformation range of the glottis from full closure to opening and then back to closure, so as to obtain the "peak-valley" feature; According to the inter-frame correlation degree, calculating the local motion vectors of the upper and lower edges of the glottis to obtain the jitter or amplitude feature of the glottis.
5. The muscle relaxation prediction method according to claim 1, characterized in that Using the superposition of multiple regression decision trees to fit the output value of the muscle relaxation prediction model, and using the muscle relaxation prediction model to obtain the muscle relaxation prediction result, including: Repeatedly execute the following steps until the number of operation rounds reaches a preset number of rounds or the error reaches a preset error: In each round of operation, calculating the residual between the output value of the previous round and the true muscle relaxation value; Using the input data as the input and the residual as the target to train a decision tree to obtain the output value of the leaf node of this round of operation; Calculating the output value of this round of operation based on the output value of the leaf node and the learning rate; Adding the decision tree of this round and the output value of this round of operation to the model for the next round of iterative operation; Based on the learning rate, summing all the decision trees to obtain the predicted value.
6. A muscle relaxation prediction system, characterized in that, Including a data acquisition device, a feature extraction module and a prediction module; The data acquisition device is used to collect the laryngeal video data and vital sign data of the patient in real time during the operation; The feature extraction module is used to extract laryngeal spatio-temporal features and vital sign features from the laryngeal video data and the vital sign data respectively; The prediction module is used to input the laryngeal spatio-temporal features, the vital sign features and the patient's static information as input data into a muscle relaxation prediction model to obtain a muscle relaxation prediction result.
7. The muscle relaxation prediction system according to claim 6, wherein The feature extraction module includes a video segmentation module and a spatio-temporal feature extraction module; The video segmentation module is used to segment the laryngeal video data in the time domain into several sub-fragments with a preset duration; The spatio-temporal feature extraction module is used to extract the corresponding spatio-temporal features from each sub-fragment.
8. The muscle relaxation prediction system according to claim 7, wherein The spatio-temporal features include laryngeal deformation features and the features of the tension and texture changes of the laryngeal muscles.
9. The muscle relaxation prediction system according to claim 8, characterized in that, The spatio-temporal feature extraction module includes a capturing module, a first feature obtaining module and a second feature obtaining module; The capturing module is used to capture the inter-frame correlation degree within each sub-fragment; The first feature obtaining module is used to obtain the deformation range of the glottis from full closure to opening and then back to closure according to the inter-frame correlation degree, so as to obtain the "peak-valley" feature; The second feature acquisition module is configured to calculate local motion vectors of the upper and lower edge positions of the glottis according to the inter-frame correlation degree, and acquire the jitter or amplitude feature of the glottis.
10. The muscle relaxation prediction system according to claim 6, wherein, The prediction module includes a residual calculation module, a decision tree training module, a round output value calculation module, an iteration module, and a predicted value calculation module; The residual calculation module is configured to calculate the residual between the output value of the previous round and the true muscle relaxation value in each round of training; The decision tree training module is configured to train a decision tree with the input data as the input and the residual as the target, and obtain the output value of the leaf node of this round of operation; The round output value calculation module is configured to calculate the output value of this round of training based on the output value of the leaf node and the learning rate; The iteration module is configured to add the decision tree of this round and the output value of this round of operation to the model for the next round of iterative operation; The predicted value calculation module is configured to sum all the decision trees based on the learning rate to obtain the predicted value.