Difficult airway prediction method and prediction device
The difficult airway prediction method based on multimodal data and dynamic analysis solves the problem of prediction accuracy under the influence of body position changes in existing technologies, and achieves more efficient difficult airway prediction and intubation tool selection.
Patent Information
- Application Number
- CN202510660943.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-22
- Publication Date
- 2025-10-17
AI Technical Summary
The existing difficult airway prediction methods rely on single modality data, lack real-time dynamic analysis capabilities, and cannot adapt to changes in body position, resulting in poor prediction accuracy.
Multimodal data (facial images, ultrasound images, voiceprint data and respiratory data) are combined with dynamic analysis to monitor body position changes in real time, generate body position status labels and input them into the trained difficult airway prediction model to generate the final prediction results.
It improves the accuracy of difficult airway prediction, can adapt to changes in body position, provides more accurate intubation tool selection, and reduces the risks of hypoxemia, brain damage, etc.
Smart Images
Figure CN120809029A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of clinical medical treatment, and particularly relates to a difficult airway prediction method and a prediction device. BACKGROUND
[0002] Difficult airway refers to a clinical situation in which a mask ventilation or tracheal intubation is difficult for an anesthesiologist or an emergency physician with professional training. Difficult airway is a major challenge in anesthesia, emergency and critical care medicine, and improper handling may lead to hypoxemia, brain injury and even death. Therefore, difficult airway prediction is of great significance in clinical medical treatment.
[0003] The difficult airway prediction method in the related art relies on single modal data and lacks real-time dynamic analysis capability, and cannot adapt to the influence of body position change on difficult airway, so the difficult airway prediction accuracy is poor. SUMMARY
[0004] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the present application provides a difficult airway prediction method and a prediction device, which rely on multi-modal data and have dynamic analysis capability, can adapt to the influence of body position change on difficult airway, and improve the difficult airway prediction accuracy.
[0005] In a first aspect, the present application provides a difficult airway prediction method, comprising:
[0006] Collecting real-time physiological data of a patient, wherein the real-time physiological data at least includes a facial image, an ultrasound image, voiceprint data and breathing data;
[0007] Obtaining current body position state information of the patient, and generating a first body position state label according to the current body position state information;
[0008] Labeling the first body position state label on the real-time physiological data;
[0009] Inputting the real-time physiological data labeled with the first body position state label into a trained difficult airway prediction model to generate a first prediction result;
[0010] In the process of difficult airway prediction, if it is detected that the current body position state information of the patient is updated, a second body position state label is generated according to the updated current body position state information, the second body position state label is labeled on the real-time physiological data, the real-time physiological data labeled with the second body position state label is inputted into the trained difficult airway prediction model to generate a second prediction result, and the second prediction result is determined as a final prediction result;
[0011] If no patient current body position state information update is detected during the difficult airway prediction process, the first prediction result is determined as a final prediction result.
[0012] In some embodiments, labeling the first body position state label on the real-time physiological data comprises:
[0013] Feature extraction is performed on the facial image, ultrasound image, voiceprint data, and respiratory data to obtain first feature data, second feature data, third feature data, and fourth feature data.
[0014] The first feature data, second feature data, third feature data, and fourth feature data are spliced to obtain spliced feature data.
[0015] The first body position state label is associated with the spliced feature data to obtain the real-time physiological data labeled with the first extracorporeal label.
[0016] In some embodiments, before the feature extraction is performed on the facial image, ultrasound image, voiceprint data, and respiratory data, the difficult airway prediction method further comprises:
[0017] The facial image, ultrasound image, voiceprint data, and respiratory data are sequentially subjected to denoising processing and sliding time window cutting processing to obtain the facial image, ultrasound image, voiceprint data, and respiratory data located in the same sliding time window.
[0018] In some embodiments, if the patient current body position state information update is detected, before a second body position state label is generated according to the updated current body position state information, the difficult airway prediction method further comprises:
[0019] It is determined that the current body position state information is not information generated according to a body position instruction input by a user.
[0020] In some embodiments, the current body position state information is one of supine position information, slope position information, lateral position information, and prone position information.
[0021] In some embodiments, detecting whether the patient current body position state information is updated comprises:
[0022] A first basic feature of the current body position state information at a first time is determined.
[0023] A second basic feature of the current body position state information at a second time is determined, wherein the second time is later than the first time.
[0024] A difference value between the first basic feature and the second basic feature is calculated.
[0025] determining that a patient current body position state information update is detected if the difference value exceeds a difference value threshold;
[0026] determining that a patient current body position state information update is not detected if the difference value does not exceed the difference value threshold.
[0027] In some embodiments, the difference threshold between the basic features corresponding to the supine position information, the slope position information, the lateral position information and the prone position information are all different from each other.
[0028] In some embodiments, the difficult airway prediction method further comprises:
[0029] obtaining a result determination instruction input by a user, wherein the result determination instruction represents whether the final prediction result is correct;
[0030] labeling a corresponding result label on the real-time physiological data labeled with the body position state label according to the result determination instruction;
[0031] adding the real-time physiological data labeled with the result label to a training sample set as a training sample to obtain an updated training sample set;
[0032] training the difficult airway prediction model using the updated training sample set to obtain an updated difficult airway prediction model.
[0033] In some embodiments, the difficult airway prediction method further comprises:
[0034] if the result determination instruction represents that the final prediction result is correct, determining that the result label is a correct result label, and adding the real-time physiological data labeled with the result label to the training sample set as a positive training sample;
[0035] if the result determination instruction represents that the final prediction result is incorrect, determining that the result label is an incorrect result label, and adding the real-time physiological data labeled with the result label to the training sample set as a negative training sample.
[0036] In a second aspect, the present application provides a difficult airway prediction device, comprising:
[0037] a data acquisition module for acquiring real-time physiological data of a patient, wherein the real-time physiological data at least includes a facial image, an ultrasound image, a voiceprint data and a breathing data;
[0038] an information acquisition module for acquiring current body position state information of a patient, and generating a first body position state label according to the current body position state information;
[0039] a label marking module, configured to mark the first body position state label on the real-time physiological data;
[0040] a result generation module, configured to input the real-time physiological data marked with the first body position state label into the trained difficult airway prediction model to generate a first prediction result;
[0041] a result determination module, configured to, in the process of predicting the difficult airway, if it is detected that the current body position state information of the patient is updated, generate a second body position state label according to the updated current body position state information, mark the second body position state label on the real-time physiological data, input the real-time physiological data marked with the second body position state label into the trained difficult airway prediction model to generate a second prediction result, and determine the second prediction result as a final prediction result; if it is not detected that the current body position state information of the patient is updated, determine the first prediction result as the final prediction result.
[0042] The difficult airway prediction method of the present application comprises collecting real-time physiological data of a patient, the real-time physiological data being multi-modal data at least including a facial image, an ultrasound image, a voiceprint data and a breathing data of the patient. Current body position state information of the patient is obtained, and a first body position state label is generated according to the current body position state information. The first body position state label is marked on the real-time physiological data. The real-time physiological data marked with the first body position state label is input into a trained difficult airway prediction model to generate a first prediction result.
[0043] In the process of predicting the difficult airway, if it is detected that the current body position state information of the patient is updated, a second body position state label is generated according to the updated current body position state information, the second body position state label is marked on the real-time physiological data, the real-time physiological data marked with the second body position state label is input into the trained difficult airway prediction model to generate a second prediction result, and the second prediction result is determined as a final prediction result. If it is not detected that the current body position state information of the patient is updated, the first prediction result is determined as the final prediction result. That is, in the process of predicting the difficult airway, not only the multi-modal data is relied on, but also the change of the body position of the patient is monitored in real time, the influence of the change of the body position on the difficult airway is considered, so that the dynamic analysis capability is possessed, and the prediction accuracy of the difficult airway is improved.
[0044] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS
[0045] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, including the appended drawings.
[0046] Figure 1 is a flowchart of a method for predicting a difficult airway provided by an embodiment of the present application;
[0047] Figure 2 is a schematic diagram of the acquisition principle of a body position state information acquisition device provided by an embodiment of the present application;
[0048] Figure 3 is a flowchart of a method for marking a first body position state label on real-time physiological data provided by an embodiment of the present application;
[0049] Figure 4 is a flowchart of a method for detecting whether the current body position state information of a patient is updated provided by an embodiment of the present application;
[0050] Figure 5 is a schematic diagram of the overall structure of a flexible film pressure sensor provided by an embodiment of the present application;
[0051] Figure 6 is a flowchart of a method for updating a difficult airway prediction model according to a prediction result provided by an embodiment of the present application;
[0052] Figure 7 is a schematic diagram of the principle of updating a difficult airway prediction model according to a prediction result provided by an embodiment of the present application;
[0053] Figure 8 is a schematic diagram of the structure of a difficult airway prediction device provided by an embodiment of the present application. DETAILED DESCRIPTION
[0054] The terms "first", "second", "third", "fourth" and the like in the description and in the claims of the present application, if any, are used for distinguishing between similar objects talking about the embodiments and do not necessarily have to appear in the description or claims of the present application in this particular order. It is to be understood that the data used in these embodiments can be interchanged, where appropriate, so that the embodiments described herein can be carried out in other sequences than the ones illustrated or described herein. Furthermore, the terms "comprising", "having", "including" and "containing" and any variations thereof in the present description and in the claims of the present application are intended to cover the instance in which the stated feature is present but not to exclude the presence of other features. The techniques and solutions described herein will be explained with reference to the accompanying drawings and figures that show some embodiments of the application. Obviously, the embodiments described herein are only some of the possible embodiments and do not limit the scope of the application in any way.
[0055] Difficult airway refers to a clinical situation in which a mask ventilation or tracheal intubation is difficult for an anesthesiologist or an emergency physician with professional training. Difficult airway is a major challenge in anesthesia, emergency and critical medicine, and improper handling may lead to hypoxemia, brain injury and even death. Therefore, difficult airway prediction is of great significance in clinical treatment.
[0056] The difficult airway prediction method in the related art relies on single modal data and lacks real-time dynamic analysis capability, and cannot adapt to the influence of body position change on difficult airway, so the difficult airway prediction accuracy is poor.
[0057] For example, the related art relies on the clinical experience and naked-eye observation of an anesthesiologist, adopts subjective indexes such as Mallampati (airway classification) score and Cormack-Lehane (laryngoscope exposure) classification, and has low sensitivity and specificity (for example, the prediction sensitivity of Mallampati for difficult tracheal intubation is only 75%, and there is still a 25% risk of missed diagnosis, so the accuracy is low).
[0058] Alternatively, anatomic parameters (such as hyoid bone position and mandibular spacing) are measured by X-ray, CT (Computed Tomography) and other imaging methods, and dynamic physiological signals (such as respiratory waveform and voiceprint characteristics) are not integrated.
[0059] In addition, the related art only provides static image or simple parameter display, and lacks dynamic analysis capability, and cannot adapt to the influence of intraoperative body position change on difficult airway. The above problems all lead to poor difficult airway prediction accuracy.
[0060] In view of the above problems, the embodiments of the present application provide a difficult airway prediction method and a prediction device, which rely on multi-modal data and have dynamic analysis capability, can adapt to the influence of body position change on difficult airway, and improve the difficult airway prediction accuracy.
[0061] Figure 1 is a flowchart of the difficult airway prediction method provided by the embodiments of the present application. As shown in Figure 1 The difficult airway prediction method provided by the embodiments of the present application comprises:
[0062] Step S101, collecting real-time physiological data of a patient, wherein the real-time physiological data at least includes a face image, an ultrasound image, voiceprint data and respiratory data.
[0063] The facial image can be collected by a three-dimensional facial scanner. The three-dimensional facial scanner is integrated with a high-resolution camera and an infrared sensor, and can capture the facial anatomical features of the patient in the front, side and supine positions (such as mandibular retrusion and tongue hypertrophy). The three-dimensional facial scanner can be connected to the data processing module through a USB 3.0 interface. The ultrasound image can be collected by a portable ultrasound probe. The portable ultrasound probe uses a miniature linear array sensor to measure key anatomical parameters such as hyoid bone position and atlantooccipital distance. The portable ultrasound probe can transmit the collected data to the data processing module through wireless transmission.
[0064] The voiceprint data can be collected by a voiceprint detector. The voiceprint detector is built-in with a vibration analyzer and an elastic fixing ring, which can be attached to the neck skin to collect tracheal vibration signals. The voiceprint detector can be connected to the data processing module through a shielded wire. The respiratory data can be collected by a respiratory waveform sensor. The respiratory waveform sensor measures the chest movement by impedance method, and synchronously records the respiratory frequency and rhythm changes.
[0065] In step S102, the current body position state information of the patient is obtained, and a first body position state label is generated according to the current body position state information.
[0066] The current body position state information of the patient can be supine position information, slope position information, lateral position information and prone position information, that is, the current body position of the patient can be one of supine position, slope position, lateral position and prone position.
[0067] The body position state information of the patient can be collected by a camera to collect the current whole body image of the patient. Then the collected current whole body image of the patient is compared with the preset image corresponding to each body position, and the current body position of the patient is determined according to the comparison result.
[0068] Specifically, key comparison parts can be set, including head comparison part, shoulder blade comparison part, lumbar comparison part, knee comparison part and foot comparison part. In the preset image of supine position, the head comparison part can expose all facial organs, the slope position can expose most of the facial organs, the lateral position can expose a small part of the facial organs, and the prone position can hide all the facial organs. It can be seen that in the ideal state, only by comparing the image of the head comparison part, the current body position state information of the patient can be determined.
[0069] In order to further accurately determine the current body position information of the patient, other comparison parts can also be compared, for example, for the knee comparison part, the supine position and the slope position can expose two knees, the lateral position can expose only one knee, and the prone position can hide both knees.
[0070] In some embodiments, a body position state information acquisition device can also be provided, which is placed on a mattress for use. Figure 2 The body position state information acquisition device provided by the embodiments of the present application has the acquisition principle diagram shown in FIG. 1. The body position acquisition device includes a plurality of flexible film pressure sensors a uniformly distributed. Each flexible film pressure sensor a at a point can detect the pressure data of the corresponding point. Specifically, the body position acquisition device can be arranged at the scapular part. According to the pressure data detected by the flexible film pressure sensors, the current body position information of the patient can be obtained.
[0071] For example, when the patient is in a supine position, the entire body position acquisition device at the scapular part can detect uniform pressure data. When the patient is in an inclined position, although the entire body position acquisition device at the scapular part can detect pressure data, the detected pressure data is obviously smaller than that when the patient is in a supine position. This is because when the patient is in an inclined position, the gravity of the patient is decomposed into two forces, while when the patient is in a supine position, the detected pressure data is equal to the gravity of the patient.
[0072] When the patient is in a lateral position, only the middle segment of the body position acquisition device at the scapular part can detect pressure data. When the patient is in a prone position, the pressure distribution detected by the body position acquisition device at the scapular part is different from that when the patient is in a supine position. For example, when prone, only the shoulders at both ends exert pressure on the body position acquisition device, so only the flexible film pressure sensors at both ends of the body position acquisition device detect pressure data.
[0073] In addition, when the patient is in different body position states, the airway may be affected, so the present application needs to dynamically predict the difficult airway based on the change of the body position of the patient.
[0074] When the patient is in a supine position, the head can be in a neutral position or use a sniffing position. In this body position, the tongue base falling back can affect the smooth airway. In addition, obese patients in this position can have a higher risk because abdominal fat can compress the diaphragm, affecting ventilation. When the patient is in an inclined position, the head is raised by 25-30 degrees, the effect of gravity can reduce the pressure of abdominal contents on the diaphragm, increase the functional residual capacity, improve oxygenation, and the position of the tongue and mandible can be more conducive to the view of the glottis, making it easier to expose the laryngoscope.
[0075] When the patient is in a lateral position, the position of the airway anatomy can be changed, making it easier to expel secretions and reducing the risk of aspiration, but increasing the difficulty of intubation. When the patient is in a prone position, the difficulty of intubation is the highest, that is, the risk of difficult airway is the highest.
[0076] When the patient is in different postures, the current posture state information of the patient can be acquired through the above manner, and then the first posture state label is generated according to the current posture state information.
[0077] In step S103, the first posture state label is marked on the real-time physiological data.
[0078] Figure 3 is a flowchart of marking the first posture state label on the real-time physiological data according to an embodiment of the present application. Marking the first posture state label on the real-time physiological data specifically includes:
[0079] In step S301, feature extraction is performed on the face image, the ultrasound image, the voiceprint data and the respiratory data to obtain first feature data, second feature data, third feature data and fourth feature data.
[0080] Specifically, feature extraction is performed on the face image to obtain the first feature data, which can include:
[0081] A plurality of feature points of the face image are detected, and based on the plurality of feature points (eyes, nose tip, corners of the mouth, etc.), the face in the face image is normalized to obtain a normalized face and a plurality of normalized feature points on the normalized face corresponding to the plurality of feature points in the face image. The face posture in the face image is estimated. Based on the feature region of the front face and the face posture, the feature region in the normalized face is located, and finally the first feature data is extracted from the feature region in the normalized face.
[0082] The above method of feature extraction on the face image is only an example, and other methods can also be used for feature extraction on the face image, which is not limited in the present application.
[0083] Feature extraction is performed on the ultrasound image to obtain the second feature data, which can include:
[0084] The ultrasound image is preprocessed: a segmentation model is trained through a deep learning method to remove the frame of the ultrasound image, leaving only the ultrasound image located in the middle position. Each ultrasound image has information such as composition, echo, strong echo, shape, edge, etc. Then the ultrasound image is changed in size, and all data are adjusted to the same size image, which is convenient for the next training and testing. Finally, data enhancement is performed, and the data enhancement method includes mirror image, rotation, folding, normalization, etc. The training, hyperparameter adjustment and testing are performed in a 5-fold cross-validation manner. The preprocessed ultrasound image is used as the input of the convolutional neural network model.
[0085] A convolutional neural network model suitable for ultrasound imaging was constructed. This model was used to extract features from ultrasound images and output classification results. Based on the ultrasound image's composition, echo, hyperechoic, morphology, and edge labels, five classifiers were constructed within the convolutional neural network model for composition, echo, hyperechoic, morphology, and edge, suitable for ultrasound image feature extraction.
[0086] Feature extraction: Convolution processing is performed on the ultrasound image to fully extract ultrasound image features. A pooling layer is used to extract deep features and reduce the image size. Multiple layers of convolution and pooling are performed to further extract features from the input ultrasound image. A fully connected layer linearly combines the extracted features to generate the output. The output layer uses a normalized exponential function to output the classification results, resulting in the second feature data.
[0087] The above method for extracting features from ultrasound images is only an example. Other methods may be used to extract features from ultrasound images, and this application does not limit this.
[0088] Extracting features from the voiceprint data to obtain third feature data may include:
[0089] Voiceprint data can be a speech segment. First, the speech segment is preprocessed. Specifically, a first-order high-pass filter can be used to enhance the high-frequency portion and compensate for the high-frequency attenuation of the vocal system. The preprocessed data is then framed and windowed. Specifically, the signal can be divided into short time frames (20-30ms, such as 25ms), with a frame shift of typically 10ms. A Hamming window is applied to each frame to reduce spectral leakage. Each frame is then converted into an amplitude spectrum or power spectrum in the frequency domain, the linear frequency is converted to a Mel scale, and the energy is extracted through a bank of 40 triangular filters. The logarithm of the filter bank energy is taken, and a discrete cosine transform (DCT) is applied to obtain the cepstral coefficients. The first 12-13 coefficients and energy are retained to form a 13-dimensional static feature. The first-order (delta) and second-order (delta-delta) differences are calculated, expanding the feature to 39 dimensions to obtain the third feature data corresponding to the voiceprint data.
[0090] The above method for extracting features from voiceprint data is only an example. Other methods may be used to extract features from voiceprint data, and this application does not limit this.
[0091] Extracting the respiratory data features to obtain fourth feature data may include:
[0092] The respiratory data is usually a respiratory waveform signal. The respiratory waveform signal usually contains noise (such as motion artifacts, baseline drift, electromagnetic interference, etc.), and needs to be preprocessed first. The preprocessing can include filtering, segmentation, normalization and other processing. Then the same and morphological features are extracted from the time series of the preprocessed waveform. The frequency components of the respiratory signal are analyzed by Fourier transform (FFT) or power spectral density (PSD). The dynamic characteristics of the signal are analyzed by wavelet transform (Wavelet Transform) or short-time Fourier transform (STFT). The time domain, frequency domain and nonlinear features are combined into a multi-dimensional feature vector to obtain the fourth feature data.
[0093] The above method of extracting features from respiratory data is only an example. Other methods can be used to extract features from respiratory data, and the application is not limited.
[0094] In step S302, the first feature data, the second feature data, the third feature data and the fourth feature data are spliced to obtain spliced feature data.
[0095] The first feature data, the second feature data, the third feature data and the fourth feature data can be in the form of a vector or a matrix. The four feature vectors can be directly spliced, for example, simply concatenated, or can be fused by weighting, attention mechanism or neural network automatic learning feature interaction. For example, the first feature data is a 10-dimensional feature vector, and the second feature vector is a 5-dimensional feature vector. The first feature data and the second feature data are simply concatenated to obtain a 15-dimensional feature vector.
[0096] In step S303, the first body position state label is associated with the spliced feature data to obtain the real-time physiological data marked with the first body position label.
[0097] The spliced feature data obtained after step S303 can be table data. The first body position state label can be associated with the spliced feature data by directly adding the first body position state label to the last column of the table data. For example, the spliced feature data is [feature1, feature2, feature3, feature4], and after adding the first body position state label, the spliced feature data marked with the label is obtained [feature1, feature2, feature3, feature4, label1]. Wherein, feature1, feature2, feature3, feature4 are the first feature data, the second feature data, the third feature data and the fourth feature data, and label1 is the first body position state label.
[0098] Step S104, inputting the real-time physiological data marked with the first body position state label into the trained difficult airway prediction model to generate a first prediction result.
[0099] In the embodiments of the present application, the first feature data can be data generated by identifying Mallampati (airway classification) classification features (such as soft palate visibility) through a CNN model, so the first feature data represents the probability of anatomical abnormalities. The second feature data can be data generated by extracting laryngoscope exposure difficulty related parameters (such as glottic opening area) through a U-Net (image segmentation network) segmentation algorithm, so the second feature data represents the laryngoscope exposure difficulty. The third feature data can be data generated by detecting abnormal harmonics through an LSTM (Long Short-Term Memory) network after the voiceprint signal is transformed through MFCC (Mel Frequency Cepstral Coefficient), so the third feature data represents the intubation resistance. The fourth feature data can be data generated by calculating the coefficient of variation (CV) to evaluate the influence of sedation depth on airway muscles, so the fourth feature data represents the influence of sedation depth on airway muscles.
[0100] The difficult airway prediction model of the embodiments of the present application can be a decision tree algorithm. The decision tree algorithm is a method for approximating discrete function values. It is a typical classification method, which first processes data, generates readable rules and decision trees using inductive algorithms, and then analyzes new data using decisions. Essentially, the decision tree is a process of classifying data through a series of rules.
[0101] The embodiments of the present application can be based on a basic decision tree, and then train the basic decision tree using a training set to obtain a trained difficult airway prediction model. Among them, a large amount of data (including face images, ultrasound images, voiceprint data, respiratory data, and corresponding difficult airway prediction results) in the database can be divided into a training set and a test set. The training set is used to train the decision tree, and then the test set is used to test the trained decision tree to evaluate the accuracy of the decision tree, and finally obtain the trained difficult airway prediction model.
[0102] The embodiments of the present application can also use gradient boosting algorithms (such as CatBoost, XGBoost) as prediction models to predict difficult airways, and the embodiments of the present application do not limit the type of difficult airway prediction model used.
[0103] In the difficult airway prediction model, different weights can be set for the first feature data, the second feature data, the third feature data, and the fourth feature data, that is, different weights are set for the facial image, the ultrasound image, the voiceprint data, and the respiratory data. For example, the weights of the facial image, the ultrasound image, the voiceprint data, and the respiratory data are set to 30%, 40%, 20%, and 10% respectively. If the data of a certain data is unreliable, the weight of the data in the difficult airway prediction model can be reduced. For example, if the facial image is blurred, the weight of the facial image can be reduced to 10%, and the weights of other data are increased.
[0104] The real-time physiological data marked with the first body position state label is input into the trained difficult airway prediction model to generate a first prediction result. The generated first prediction result can be a difficult airway comprehensive risk index (0-100%) representing the possibility of difficult airway. For example, the first prediction result is 5%, indicating a low possibility of difficult airway, and the first prediction result is 80%, indicating a high possibility of difficult airway.
[0105] In step S105, during the difficult airway prediction process, if the patient's current body position state information is detected to be updated, a second body position state label is generated according to the updated current body position state information, the second body position state label is marked on the real-time physiological data, the real-time physiological data marked with the second body position state label is input into the trained difficult airway prediction model to generate a second prediction result, and the second prediction result is determined as the final prediction result.
[0106] During the difficult airway prediction process, the current body position state information of the patient can be monitored in real time. If the patient's current body position state information is detected to be updated, that is, the patient's current body position state information is detected to be changed, for example, from a supine position to an inclined position. A second body position state label can be generated according to the updated current body position information. Then mark the second body position state label on the real-time physiological data. The real-time physiological data with the re-labeled body position state label is re-input into the trained difficult airway prediction model to generate a second prediction result, and the second prediction result is determined as the final prediction result.
[0107] In step S106, during the difficult airway prediction process, if the patient's current body position state information is not detected to be updated, the first prediction result is determined as the final prediction result. If the patient's body position state is not detected to be changed, the first prediction result can be directly determined as the final prediction result.
[0108] After determining the final prediction result, a corresponding intubation tool can be selected according to the final prediction result. In some embodiments, different intubation tools can be set according to different prediction results. For example, 0-20% in the difficult airway comprehensive risk index (0-100%) can correspond to intubation tool A, 20%-40% can correspond to intubation tool B, 40%-60% can correspond to intubation tool C, 60%-80% can correspond to intubation tool D, and 80%-100% can correspond to intubation tool E. The difference between intubation tools A-E can be the difference in pipe thickness or pipe material.
[0109] For example, the initial body position state information of the patient is supine position, the first prediction result is 45%, but it is monitored that the body position state of the patient changes from supine position to prone position, and the second prediction result is 80%, at this time intubation tool C cannot be selected, and intubation tool E needs to be selected. It can be seen that the method of the present application can not only improve the prediction accuracy of difficult airway, but also accurately select the intubation tool according to the prediction result of difficult airway.
[0110] The difficult airway prediction method of the present application comprises collecting real-time physiological data of a patient, the real-time physiological data being multi-modal data, at least comprising facial image, ultrasound image, voiceprint data and breathing data of the patient. The current body position state information of the patient is obtained, and a first body position state label is generated according to the current body position state information. The first body position state label is marked on the real-time physiological data. The real-time physiological data marked with the first body position state label is input into the trained difficult airway prediction model to generate a first prediction result.
[0111] In the process of predicting difficult airway, if it is detected that the current body position state information of the patient is updated, a second body position state label is generated according to the updated current body position state information, the second body position state label is marked on the real-time physiological data, the real-time physiological data marked with the second body position state label is input into the trained difficult airway prediction model to generate a second prediction result, and the second prediction result is determined as the final prediction result. If it is not detected that the current body position state information of the patient is updated, the first prediction result is determined as the final prediction result. That is, in the process of predicting difficult airway, not only multi-modal data is relied on, but also the change of the body position of the patient is monitored in real time, the influence of the change of the body position on difficult airway is considered, so as to have dynamic analysis ability and improve the prediction accuracy of difficult airway.
[0112] In some embodiments, before feature extraction is performed on the facial image, ultrasound image, voiceprint data and breathing data, the difficult airway prediction method further comprises:
[0113] The face image, the ultrasonic image, the voiceprint data and the respiratory data are sequentially subjected to denoising processing and sliding time window cutting processing to obtain the face image, the ultrasonic image, the voiceprint data and the respiratory data located in the same sliding time window.
[0114] Since the physiological data of the patient can change over time, that is, the face image, the ultrasonic image, the voiceprint data and the respiratory data have a certain correlation on the time axis. Therefore, the face image, the ultrasonic image, the voiceprint data and the respiratory data can be cut by using the same sliding time window to obtain the face image, the ultrasonic image, the voiceprint data and the respiratory data located in the same sliding time window. The face image, the ultrasonic image, the voiceprint data and the respiratory data with strong correlation are applied to the difficult airway prediction process, and the difficult airway prediction accuracy is further improved.
[0115] In some embodiments, if it is detected that the patient's current body position state information is updated, the method for predicting a difficult airway further comprises the following steps before generating a second body position state label according to the updated current body position state information:
[0116] It is determined that the current body position state information is not information generated according to a body position instruction input by a user.
[0117] That is, the embodiment of the present application can also generate the current body position state information according to the body position instruction input by the user. If the current body position state information is information generated according to the body position instruction input by the user, even if it is detected that the body position state information is updated, the prediction result is not regenerated, but the first prediction result is determined as the final prediction result. If the current body position state information is not information generated according to the body position instruction input by the user, and it is detected that the body position state information is updated, the prediction result needs to be regenerated, that is, the generation process of the second prediction result is performed, and the second prediction result is determined as the final prediction result. In this way, when the current body position state information needs to be manually input, the difficult airway prediction result can be obtained more quickly.
[0118] Figure 4 is a method flow diagram for detecting whether the patient's current body position state information is updated. In some embodiments, detecting whether the patient's current body position state information is updated comprises the following steps:
[0119] Step S401, determining a first basic feature of the current body position state information at a first time;
[0120] Step S402, determining a second basic feature of the current body position state information at a second time, wherein the second time is later than the first time;
[0121] Step S403, calculating the difference between the first basic feature and the second basic feature;
[0122] Step S404: if the difference value exceeds the difference value threshold, it is determined that the patient's current body position status information is updated;
[0123] Step S405: If the difference value does not exceed the difference value threshold, it is determined that no update of the patient's current body position status information is detected.
[0124] like Figure 5 The illustrated body position information acquisition device includes segments A, B, and C, each of which can measure three pressure data points. When the scapular body position information acquisition device is used to detect the current body position information, the first basic feature of the current body position information at a first time is [X1, X2, X3], representing the pressure data measured in segments A, B, and C, respectively. The second basic feature of the current body position information at a second time is [Y1, Y2, Y3], similarly representing the pressure data measured in segments A, B, and C.
[0125] The difference between the first basic feature [X1, X2, X3] and the second basic feature [Y1, Y2, Y3] is then calculated. Specifically, the difference is [Y1-X1, Y2-X2, Y3-X3]. The three pressure data have difference thresholds M1, M2, and M3, respectively. The difference between [Y1-X1, Y2-X2, and Y3-X3] is compared with the corresponding difference thresholds. As long as the difference between any one of the pressure data exceeds the difference threshold, it can be determined that the current body position status information has been updated.
[0126] For example, when the patient is in the supine position, the first basic feature is [10, 9, 10], and the difference threshold is 0.5, 0.5, 0.5. If the second basic feature is [9.8, 9, 10.1], the difference between the second basic feature and the first basic feature is [0.2, 0, 0.1], both of which are less than the difference threshold. Therefore, it is determined that the current body position state information has not been updated. In this case, it may be that the patient moved slightly, causing the pressure value to change, and the patient did not change the body position. If the second basic feature is [5, 4, 5], the difference between the second basic feature and the first basic feature is [5, 5, 5], both of which are greater than the difference threshold. Therefore, it is determined that the current body position state information has been updated. In this case, it may be that the patient changed from the supine position to the sloped position.
[0127] In some embodiments, the difference thresholds between any two of the basic features corresponding to the supine position information, the slope position information, the lateral position information, and the prone position information are all different.
[0128] That is, there is a difference threshold Z1 between the supine position and the slope position, a difference threshold Z2 between the supine position and the lateral position, a difference threshold Z3 between the supine position and the prone position, a difference threshold Z4 between the slope position and the lateral position, a difference threshold Z5 between the slope position and the prone position, and a difference threshold Z6 between the lateral position and the prone position. With different difference thresholds between different body positions, not only whether the body position information changes can be determined, but also the updated body position state information can be determined, so that the difficult airway prediction efficiency can be improved.
[0129] For example, if the initial state of the patient is in the supine position, if it is detected that the body position state information is updated, and the difference value between the two body positions exceeds the difference threshold Z1, it can be determined that the updated body position is the slope position. If it is detected that the body position state information is updated, and the difference value between the two body positions exceeds the difference threshold Z2, it can be determined that the updated body position is the lateral position. If it is detected that the body position state information is updated, and the difference value between the two body positions exceeds the difference threshold Z3, it can be determined that the updated body position is the prone position. Correspondingly, in other initial body position states, the updated body position state information can be determined by using these difference thresholds.
[0130] It should be noted that each difference threshold is data generated according to a large number of test results, so that the updated body position state information can be accurately determined according to the initial body position state information, the difference value and each difference threshold, and the effect of automatically and dynamically predicting the difficult airway is realized.
[0131] Figure 6 Fig. 6 is a method flow diagram of updating a difficult airway prediction model according to a prediction result according to an embodiment of the present application. Figure 7 Fig. 7 is a principle diagram of updating a difficult airway prediction model according to a prediction result according to an embodiment of the present application. In some embodiments, the method of updating a difficult airway prediction model according to a prediction result comprises the following steps:
[0132] In step S601, a result determination instruction input by a user is acquired, wherein the result determination instruction represents whether the final prediction result is correct; the user can input an instruction of whether the result is correct on an operation panel according to the obtained prediction result and the actual situation. If the final prediction result is the same as the actual situation, the user can input a correct result instruction. If the final prediction result is different from the actual situation, the user can input an incorrect result instruction.
[0133] In step S602, a corresponding result label is marked on the real-time physiological data marked with a body position state label according to the result determination instruction.
[0134] Step S603, adding the real-time physiological data marked with the result label to the training sample set as a training sample, to obtain an updated training sample set;
[0135] Step S604, training the difficult airway prediction model by using the updated training sample set, to obtain an updated difficult airway prediction model.
[0136] Specifically, as shown in the figure, Figure 7 if the result determination instruction represents that the final prediction result is correct, determining the result label as a correct result label, and adding the real-time physiological data marked with the result label to the training sample set as a positive training sample.
[0137] if the result determination instruction represents that the final prediction result is incorrect, determining the result label as an incorrect result label, and adding the real-time physiological data marked with the result label to the training sample set as a negative training sample.
[0138] In this way, according to the real-time feedback of the user, the final prediction result is taken as a training set to continue optimizing the trained difficult airway prediction model, so as to further improve the accuracy of the difficult airway prediction result.
[0139] In a second aspect, the present application provides a difficult airway prediction device, Figure 8 is a structural schematic diagram of a difficult airway prediction device provided by an embodiment of the present application, and the difficult airway prediction device provided by the embodiment of the present application comprises:
[0140] A data acquisition module 1001 is configured to acquire real-time physiological data of a patient, wherein the real-time physiological data at least comprises a facial image, an ultrasonic image, voiceprint data and breathing data.
[0141] An information acquisition module 1002 is configured to acquire current body position state information of the patient, and generate a first body position state label according to the current body position state information
[0142] A label marking module 1003 is configured to mark the first body position state label on the real-time physiological data.
[0143] A result generation module 1004 is configured to input the real-time physiological data marked with the first body position state label into a trained difficult airway prediction model, to generate a first prediction result.
[0144] The result determining module 1005 is configured to, in the process of predicting the difficult airway, if the current body position state information update of the patient is detected, generate a second body position state label according to the updated current body position state information, mark the real-time physiological data with the second body position state label, input the real-time physiological data marked with the second body position state label into the trained difficult airway prediction model, generate a second prediction result, and determine the second prediction result as the final prediction result; if the current body position state information update of the patient is not detected, determine the first prediction result as the final prediction result.
[0145] It should be noted that if the current body position state information update of the patient is detected, the whole process needs to return to the result generating module 1004, generate a second prediction result by using the result generating module 1004, and finally determine the second prediction result as the final prediction result by the result determining module 1005.
[0146] Those skilled in the art should understand that the embodiments of the present application can be provided as a method or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or a combination of software and hardware.
[0147] Each of the embodiments in the present application is described in a progressive manner, and the same or similar parts of each of the embodiments can be referred to each other. Each of the embodiments mainly explains the difference from other embodiments. In particular, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the related parts can be referred to the part of the method embodiments.
[0148] It should be further noted that the terms "comprising", "containing" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or apparatus. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or apparatus including the element.
[0149] The above only describes the embodiments of the present application and is not intended to limit the present application. The present application can have various changes and modifications for those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the scope of claims of the present application.
Claims
1. A difficult airway prediction method, characterized in that: include: Collecting real-time physiological data of the patient, wherein the real-time physiological data includes at least facial images, ultrasound images, voiceprint data and respiratory data; Acquiring the patient's current body position state information, and generating a first body position state tag according to the current body position state information; Marking the first body position state label on the real-time physiological data; Inputting the real-time physiological data marked with the first body position state label into a trained difficult airway prediction model to generate a first prediction result; During the difficult airway prediction process, if it is detected that the patient's current body position state information is updated, a second body position state label is generated according to the updated current body position state information, the second body position state label is marked on the real-time physiological data, the real-time physiological data marked with the second body position state label is input into the trained difficult airway prediction model, a second prediction result is generated, and the second prediction result is determined as the final prediction result; During the difficult airway prediction process, if no update of the patient's current body position status information is detected, the first prediction result is determined as the final prediction result.
2. The difficult airway prediction method according to claim 1, characterized in that: Marking the first body position state label on the real-time physiological data includes: Performing feature extraction on the facial image, ultrasonic image, voiceprint data, and respiratory data to obtain first feature data, second feature data, third feature data, and fourth feature data; Performing feature splicing on the first feature data, the second feature data, the third feature data, and the fourth feature data to obtain spliced feature data; The first body position state tag is associated with the splicing feature data to obtain the real-time physiological data marked with the first external tag.
3. The difficult airway prediction method according to claim 1, characterized in that: Before extracting features from the facial image, ultrasound image, voiceprint data, and respiratory data, the difficult airway prediction method further includes: The facial image, ultrasound image, voiceprint data and respiratory data are subjected to denoising and sliding time window cutting processing in sequence to obtain the facial image, ultrasound image, voiceprint data and respiratory data located in the same sliding time window.
4. The difficult airway prediction method according to claim 1, characterized in that: If it is detected that the patient's current body position state information is updated, before generating the second body position state label according to the updated current body position state information, the difficult airway prediction method further includes: It is determined that the current body posture state information is not information generated according to a body posture instruction input by a user.
5. The difficult airway prediction method according to claim 1, characterized in that: The current body position state information is one of supine position information, slope position information, lateral position information and prone position information.
6. The difficult airway prediction method according to claim 5, characterized in that: Detecting whether the patient's current posture status information is updated includes: determining a first basic feature of the current body position state information at a first time; determining a second basic feature of the current body position state information at a second time, wherein the second time is later than the first time; calculating a difference between the first basic feature and the second basic feature; If the difference value exceeds the difference value threshold, it is determined that an update of the patient's current body position state information is detected; If the difference value does not exceed the difference value threshold, it is determined that no update of the patient's current body position status information is detected.
7. The difficult airway prediction method according to claim 6, characterized in that: The difference thresholds between any two of the basic features corresponding to the supine position information, the slope position information, the lateral position information, and the prone position information are all different.
8. The difficult airway prediction method according to claim 1, characterized in that: The difficult airway prediction method further includes: Obtaining a result determination instruction input by a user, wherein the result determination instruction indicates whether the final prediction result is correct; Determine an instruction according to the result, and mark a corresponding result tag on the real-time physiological data marked with the body position state tag; adding the real-time physiological data marked with the result label as a training sample to a training sample set to obtain an updated training sample set; The difficult airway prediction model is trained using the updated training sample set to obtain the updated difficult airway prediction model.
9. The difficult airway prediction method according to claim 8, characterized in that: The difficult airway prediction method further includes: If the result determination instruction indicates that the final prediction result is correct, determining that the result label is a correct result label, and adding the real-time physiological data marked with the result label as a positive training sample to the training sample set; If the result determination instruction indicates that the final prediction result is wrong, the result label is determined to be an incorrect result label, and the real-time physiological data marked with the result label is added to the training sample set as a negative training sample.
10. A difficult airway prediction device, characterized in that: include: A data acquisition module, configured to collect real-time physiological data of the patient, wherein the real-time physiological data includes at least facial images, ultrasound images, voiceprint data, and respiratory data; An information acquisition module, configured to acquire the patient's current body position status information and generate a first body position status tag according to the current body position status information; a label marking module, configured to mark the first body position state label on the real-time physiological data; A result generating module, configured to input the real-time physiological data marked with the first body position state label into a trained difficult airway prediction model to generate a first prediction result; A result determination module is used to, during the difficult airway prediction process, generate a second body position state label based on the updated current body position state information if an update is detected for the patient's current body position state information, mark the second body position state label on the real-time physiological data, input the real-time physiological data marked with the second body position state label into a trained difficult airway prediction model, generate a second prediction result, and determine the second prediction result as the final prediction result; if no update is detected for the patient's current body position state information, determine the first prediction result as the final prediction result.