An intelligent identification and classification method for ventilator-patient ventilation asynchrony

Abstract

The present invention discloses a method for intelligently identifying and classifying ventilator-patient ventilation asynchronous phenomena. This method addresses the problems of low automated detection and insufficient detection accuracy of existing clinical ventilator-machine asynchronous phenomena, and provides a set of high-precision and high-accuracy identification and classification methods. Based on feature extraction and intelligent identification of ventilator waveform data, the present invention deeply mines ventilator waveform data, has high feasibility and universality, and does not require invasive data collection methods such as esophageal electrode catheter technology on patients, which will not affect the ongoing patient treatment process. At the same time, it can intelligently identify the start time of inspiration and exhalation of patients during ventilation, which can uncover more effective information. Different human-machine asynchronous phenomena can be accurately defined based on clinical diagnostic standards, achieving high-precision identification and accurate classification of ventilator-patient ventilation asynchronous phenomena.

Description

Technical Field

[0001] The present invention relates to the field of biomedical engineering technology, and more particularly to an intelligent identification and classification method for ventilator-patient ventilation asynchronous phenomena. Background Art

[0002] Mechanical ventilation refers to a medical method that uses a ventilator to assist the patient's breathing, maintain the patient's oxygenation function, and buy time for the treatment of the primary disease. It plays a vital role in the treatment of respiratory diseases such as chronic obstructive pulmonary disease and acute respiratory syndrome, as well as in respiratory support for critically ill patients with respiratory failure.

[0003] However, during the patient's mechanical ventilation process, the asynchronous phenomenon between the ventilator and the patient will directly affect the patient's mechanical ventilation treatment process, making it impossible to ensure adequate respiratory support, thereby delaying weaning from the ventilator and even causing ventilator complications in the patient. The difficulty in identifying human-machine asynchronous phenomenon lies in identifying the patient's own respiratory effort. Clinically, identifying human-machine asynchronous phenomenon through ventilator waveforms requires sufficient clinical experience of internists. There are also methods at home and abroad that use esophageal electrodes to detect the patient's diaphragm electrical signals, perform signal processing and analysis, and then detect the patient's own respiratory effort and the starting point of each breath. However, this method is harmful to the patient himself, is not conducive to the treatment process of mechanical ventilation, and is not universally applicable.

[0004] In addition, there is also the use of artificial intelligence technologies such as machine learning for identification. A recognition model based on long short-term memory artificial neural network (LSTM) can be adopted to identify and classify human-computer asynchronous phenomena. However, this method has low sensitivity and accuracy in identifying the patient's respiratory effort.

[0005] Through testing and verification of calibrated clinical data, the recognition method of this project has a higher accuracy rate of human-computer asynchronous recognition.

[0006] Therefore, how to improve the accuracy of identifying human-machine asynchronous phenomena in a ventilator while avoiding harm to the patient is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides a method for intelligent identification and classification of ventilator-patient ventilation asynchronous phenomena. In order to address the problems of low degree of automated detection and insufficient detection accuracy of human-machine asynchronous phenomena in existing clinical ventilators, a set of high-precision and high-accuracy methods for identifying and classifying human-machine asynchronous phenomena during mechanical ventilation of patients by ventilators is provided. The present invention is based on feature extraction and intelligent identification of ventilator waveform data, deeply mines ventilator waveform data, has high feasibility and universality, and does not require invasive data collection methods such as esophageal electrode catheter technology for patients, and will not affect the ongoing patient treatment process. At the same time, it can intelligently identify the starting time of inspiration and exhalation of patients during ventilation, can discover more effective information, and accurately define different human-machine asynchronous phenomena based on clinical diagnostic standards, to achieve high-precision identification and accurate classification of ventilator-patient ventilation asynchronous phenomena.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A method for intelligently identifying and classifying ventilator-patient ventilation asynchrony, comprising the following steps:

[0010] Step 1: Acquire waveform data of the ventilator during mechanical ventilation, and calibrate the waveform data to calibrate the start time of inspiration and the start time of expiration of each cycle; acquire mechanical ventilation simulation data under different patient parameters based on a mechanical ventilation model of an actively breathing patient; and construct a database using the waveform data and the mechanical ventilation simulation data;

[0011] Step 2: resampling the database according to the set sampling frequency, and then performing standardization and batch processing in sequence; the resampling ensures that the sampling frequency of the respiratory data obtained in clinical and simulation data is unified to 1kHz;

[0012] Step 3: Based on the processed sampling data, a deep learning algorithm of the UNet neural network is used to establish a patient spontaneous breathing recognition model.

[0013] Step 4: Acquire waveform data and control parameters of the ventilator to be identified, and obtain the respiratory support trigger time and inspiration-expiration switching time of each cycle of the ventilator to be identified according to the control parameters;

[0014] Step 5: Input the waveform data to be identified into the patient's spontaneous breathing recognition model, and calculate the patient's spontaneous breathing start time stamp corresponding to the ventilator to be identified; the spontaneous breathing start time stamp includes the patient's physiological spontaneous inspiration start time and spontaneous exhalation start time; calculate the time difference between the spontaneous breathing start time and the ventilator triggering and switching time;

[0015] Step 6: Determine the time difference according to the preset human-computer asynchronous phenomenon classification standard to obtain an intelligent classification result.

[0016] Preferably, the waveform data includes pressure data, flow data and tidal volume data of the ventilator; the data resampling frequency depends on the sampling frequency of the ventilator itself or the sampling frequency of the waveform data acquisition device.

[0017] Preferably, the Z-Score standardization method is used for standardization processing, and the waveform data collected during the mechanical ventilation process of the same patient under the action of the ventilator are processed separately, that is, the pressure data, flow data and tidal volume data are processed separately; the control parameter is the feedback data of the ventilator control system;

[0018] The mean and standard deviation of pressure data, flow data, and tidal volume data were calculated respectively, and standardized according to the mean and standard deviation. The formula is:

[0019]

[0020] Among them, y is the standardized data; x is the data before standardization; mean is the mean; σ is the standard deviation.

[0021] Preferably, a Mini-Batch method is used for batch processing, with a batch length of 512 sampling points. In step 3, the data is input in batches when the UNet neural network is used to train the patient's spontaneous breathing recognition model. This method achieves an optimal balance between memory efficiency and memory capacity. Parallelization improves memory utilization, reduces the number of required training iterations, increases model recognition speed, and improves the accuracy of the gradient descent direction of the training model while reducing the amplitude of training vibrations.

[0022] Preferably, the patient spontaneous breathing recognition model includes a downsampling part, a splicing part and an upsampling part, and a UNet model structure is built based on a convolutional neural network to construct a patient spontaneous breathing recognition model;

[0023] Among them, the downsampling part extracts shallow feature information and compresses the data information through the convolution layer, using 4 layers of compression; the upsampling part extracts deep feature information and expands the data information through the deconvolution layer, using 4 layers of expansion;

[0024] The activation function of each convolutional layer is a linear rectification function, the convolution layer of the output layer is 1, and the activation function is a normalized exponential function. In the middle of the convolutional layer, a Dropout layer is added for regularization to prevent overfitting in the training process;

[0025] The splicing part fuses deep feature information and shallow feature information.

[0026] Preferably, the Adam optimizer is used to optimize the training process of the patient spontaneous breathing recognition model, and the Adam optimizer parameters are: α=0.001, β1=0.9, β2=0.999, ∈=1e-8.

[0027] Preferably, the loss function of the patient's spontaneous breathing recognition model adopts a cross entropy function, and the timestamp of the spontaneous breathing starting point calibrated by the expert is used as the true value of the patient's spontaneous breathing recognition model; and cross validation is adopted.

[0028] Preferably, the classification criteria for human-machine asynchronous phenomena include delayed inspiration, early switching, delayed switching and invalid triggering;

[0029] Delayed inspiration: The ventilator's respiratory support triggering time in step 5 is delayed by 250 milliseconds compared to the patient's physiological spontaneous inspiration start time;

[0030] Early switching: The ventilator inhalation-exhalation switching time in step 5 is 100 milliseconds earlier than the patient's physiological spontaneous exhalation start time;

[0031] Delayed switching: The ventilator inhalation-exhalation switching time in step 5 is delayed by 300 milliseconds compared to the patient's physiological spontaneous exhalation start time;

[0032] Invalid triggering: The waveform data to be identified is lower than the preset support level value; the waveform data is manifested as a drop in airway pressure and a small fluctuation in flow, but the ventilator cannot be triggered to make the pressure waveform reach the preset support level pressure value. The patient's breathing effort is unable to trigger the ventilator's pressure support, that is, the airway pressure or flow changes generated by the patient's breathing effort cannot reach the pressure or flow trigger sensitivity set by the ventilator. According to the preset ventilator sensitivity threshold, the patient's spontaneous breathing recognition model is used to identify whether the patient's breathing effort is unable to trigger the ventilator's pressure support.

[0033] Through the above technical solutions, it can be seen that compared with the existing technology, the present invention provides an intelligent identification and classification method for ventilator-patient ventilation asynchronous phenomena. In order to solve the problem of automatic identification and classification of human-machine asynchronous phenomena, an image feature extraction model based on the U-Net convolutional neural network in deep learning is proposed. The input data is the pressure, flow and tidal volume waveform data of the ventilator calibrated by clinical experts and the starting point of the patient's respiratory effort calibrated by the experts, that is, the start time of inspiration and the start time of expiration of each cycle. The model is trained with sufficient calibration data to effectively extract the trigger time and switching time of the ventilator, as well as the start time of the patient's inspiration and expiration phases, avoiding the previous invasive method of obtaining the start time of each stage of the patient's breathing, and realizing the automatic detection and identification of human-machine asynchronous phenomena. In addition, the classification standard of human-machine asynchronous phenomena is set based on expert clinical knowledge, and the identified asynchronous phenomena are automatically classified according to the classification standard, ultimately realizing the intelligent identification and classification of human-machine asynchronous phenomena between the ventilator and the patient during mechanical ventilation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0035] Figure 1 The accompanying drawing is a flow chart of the intelligent identification and classification method for ventilator-patient ventilation asynchrony provided by the present invention;

[0036] Figure 2 The accompanying figure is a schematic diagram of the UNet neural network structure of the patient spontaneous breathing recognition model provided by the present invention;

[0037] Figure 3 The accompanying drawing is a schematic diagram of the intelligent classification process of human-computer asynchronous phenomena provided by the present invention.

[0038] Figure 4 The accompanying drawing is a schematic diagram of a mechanical ventilation model for an active breathing patient provided by the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0040] The embodiment of the present invention discloses a method for intelligently identifying and classifying asynchrony between ventilator and patient ventilation. Figure 1 As shown in the figure, it is mainly divided into two parts. One part is to establish a patient spontaneous breathing recognition model based on the UNet neural network to realize the patient spontaneous breathing recognition; the other part is to classify and judge the human-computer asynchronous phenomenon according to the identified patient spontaneous breathing situation to realize the intelligent recognition of human-computer asynchronous phenomenon.

[0041] S1: Data collection and database retrieval

[0042] First, the waveform data of the ventilator during mechanical ventilation is obtained as clinical data, including pressure data, flow data and tidal volume data, and the waveform data is calibrated by experts to calibrate the start time of inspiration and the start time of expiration of each cycle; the sampling frequency of the data depends on the sampling frequency of the ventilator itself or the sampling frequency of the respiratory data acquisition device. The clinical data is stored in a data list or a database is established for data processing in the identification process. At the same time, based on the mechanical ventilation model of patients with active breathing, mechanical ventilation simulation data under different patient parameters is obtained, and the simulation data is stored in a data list or a database is established for training and testing of the identification model. The mechanical ventilation model of patients with active breathing based on pneumatic principles, such as Figure 4 As shown, it includes an air source 1, a control valve 2, a pressure sensor 3, a flow sensor 4, a throttle valve 5, an elastic chamber 6, a brake valve 7, a vacuum pump 8, an AD sampling module 9 and a host computer 10.

[0043] S2: Data Preprocessing

[0044] The collected ventilator waveform data has varying sampling frequencies, so the existing database needs to be resampled before inputting into the neural network. The processed sampling frequency is then unified to 100Hz. The average patient's respiratory cycle is 3 to 5 seconds, and resampling the data ensures a sampling point every 0.01 seconds, ensuring a certain degree of timeliness.

[0045] After resampling the data, data standardization (normalization) is performed. The present invention adopts the Z-Score standardization method to process the data of the same patient and the same patient parameters separately. The mean and standard deviation of the patient's pressure, flow, and tidal volume data are calculated respectively. The mean is subtracted from each sampling point and divided by the standard deviation to obtain the standardized data. The formula is:

[0046]

[0047] The normalized processed data is batched (Batches), with the batch size set to 512 samples. Later, during neural network training of the recognition model, data is input in batches. The correct batch size strikes an optimal balance between memory efficiency and capacity, improving memory utilization through parallelization, reducing the number of required training iterations, and increasing model recognition speed. It also increases the accuracy of the trained model's gradient descent and minimizes training fluctuations.

[0048] S3: Patient spontaneous breathing recognition model based on UNet neural network

[0049] The present invention uses a deep learning algorithm based on the UNet neural network to establish a recognition model to identify the time points of the patient's spontaneous breathing starting inspiration and breathing conversion, which are hereinafter referred to as the timestamps of the inspiration start point and the expiration start point (breathing start point timestamp). The deep learning algorithm based on the UNet neural network is usually used for edge segmentation of two-dimensional images. In the present invention, a new recognition model is established to perform feature recognition on one-dimensional respiratory data. The neural network structure of the model is as follows Figure 2 shown.

[0050] The model consists of three parts: downsampling, concatenation, and upsampling. The first half is the downsampling part, which mainly extracts shallow feature information; the second half is the upsampling part, which mainly extracts deep feature information; the middle part is the concatenation part, which fuses deep and shallow features to prevent feature information loss during the convolution process.

[0051] The downsampling part compresses the data through convolutional layers, while the upsampling part expands the data through deconvolutional layers. Four layers of compression and four layers of expansion are used for each part, respectively. The activation function of each convolutional layer is the Rectified Linear Unit (ReLU), while the activation function of the output layer is 1 and the activation function is the normalized exponential function (softmax). Dropout layers are added between the convolutional layers for regularization to prevent overfitting during training.

[0052] The recognition model uses the Adam optimizer to optimize the training process and accelerate convergence. Its design parameters are: α = 0.001, β1 = 0.9, β2 = 0.999, ∈ = 1e-8. The recognition model's loss function uses the cross-entropy function. The expert-calibrated timestamp of the patient's spontaneous breathing start point is used as the ground truth value (GroundTruth) of the recognition model.

[0053] Cross-validation was used to avoid overfitting. The data was divided into 15 groups, each containing respiratory data from the same patient. During training, 14 of these groups were used to train the neural network, and the training results were tested on the remaining group. This training process was repeated 15 times, with each group used as the test group, to obtain the final prediction results.

[0054] Identification result verification method

[0055] The detection results of the recognition model are the timestamps of the patient's inhalation start and exhalation start. If the time difference between the recognition model's detection timestamp and the expert calibration timestamp is within 0.5 seconds, the prediction result can be determined to be consistent with the expert calibration data and is considered a true value (TP). If a recognition result timestamp cannot be matched with any expert calibration timestamp, it is considered a false positive (FP). If an expert calibration timestamp cannot be matched with any recognition result timestamp, it is considered a false negative (FN). This method can be used to test the recognition results, obtain the test results, and then determine the accuracy of the recognition model for comparison with other recognition methods.

[0056] S4: Acquire waveform data and control parameters of the ventilator to be identified, and obtain the respiratory support trigger time and inspiration-expiration switching time of each cycle of the ventilator to be identified according to the control parameters.

[0057] S5: First, collect the waveform data and control parameters of the ventilator to obtain the trigger and switching time of each ventilator cycle. After data preprocessing, the waveform data is input into the patient's spontaneous breathing recognition model. Using the method introduced in the first part, the starting point timestamp of the patient's spontaneous breathing, as well as the start time of inspiration and the start time of expiration are identified. The time difference between the patient's spontaneous breathing start time and the ventilator trigger and switching time is automatically obtained;

[0058] S6: The time difference between the patient's spontaneous breathing start time and the ventilator triggering and switching time is substituted into the human-machine asynchronous classification standard for judgment to obtain the intelligent classification result.

[0059] Classification standard of human-computer asynchronous phenomenon:

[0060] Based on expert experience and clinical rules, the present invention divides the human-machine asynchronous phenomena that seriously affect the mechanical ventilation process into four categories: delayed inspiration, early switching, delayed switching and invalid triggering.

[0061] Delayed inspiration: defined as a delay of 250 milliseconds between the triggering time of the inspiratory support pressure provided by the ventilator and the initial time when the patient starts to inhale;

[0062] Early switching: defined as the time when the ventilator stops switching to support pressure is 100 milliseconds earlier than the initial time when the patient starts to exhale;

[0063] Delayed switching: defined as a delay of 300 milliseconds between the time the ventilator stops switching to stop support pressure and the time the patient starts exhaling;

[0064] Ineffective triggering: defined as the patient's respiratory effort being insufficient to trigger ventilator pressure support.

[0065] Intelligent classification technology of human-computer asynchronous phenomenon, the flow chart of this part is as follows Figure 3 shown.

[0066] Beneficial effects of the present invention:

[0067] A UNet neural network-based model for identifying patient spontaneous breathing was established, and a technology for intelligently classifying the types of human-machine asynchronous phenomena between the ventilator and the patient was proposed. Ultimately, a set of intelligent identification and classification technologies for human-machine asynchronous phenomena between the ventilator and the patient during mechanical ventilation was formed. Compared with existing technologies, this technology avoids invasive collection of the patient's spontaneous breathing effort, does not cause harm to the patient, and does not affect the normal operation and treatment of the ventilator. It also deeply mines the ventilator's waveform data and can automatically detect the characteristic timestamps of the patient's spontaneous breathing. The method proposed in this invention has high recognition accuracy, simple operation, high feasibility, and strong universality.

[0068] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0069] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligently identifying and classifying ventilator-patient ventilation asynchrony, characterized in that: The specific steps are as follows: Step 1: Acquire waveform data of the ventilator during mechanical ventilation, and calibrate the waveform data to calibrate the start time of inspiration and the start time of expiration of each cycle; Acquire mechanical ventilation simulation data under different patient parameters based on a mechanical ventilation model of an actively breathing patient; and construct a database using the waveform data and the mechanical ventilation simulation data; Step 2: resampling the database according to the set sampling frequency, and then performing standardization and batch processing on the sampled data in turn; Step 3: Based on the processed sampling data, a patient spontaneous breathing recognition model is established using the deep learning algorithm of the UNet neural network; the patient spontaneous breathing recognition model includes a downsampling part, a splicing part and an upsampling part. The UNet model structure is built based on the convolutional neural network to construct a patient spontaneous breathing recognition model; wherein, the downsampling part extracts shallow feature information, compresses the data information through the convolution layer, and adopts 4-layer compression; the upsampling part extracts deep feature information, expands the data information through the deconvolution layer, and adopts 4-layer expansion; the activation function of each convolution layer is a linear rectification function, the convolution layer of the output layer is 1, the activation function is a normalized exponential function, and a Dropout layer is added in the middle of the convolution layer for regularization; the splicing part fuses deep feature information and shallow feature information; Step 4: Acquire waveform data and control parameters of the ventilator to be identified, and obtain the respiratory support trigger time and inspiration-expiration switching time of each cycle of the ventilator to be identified according to the control parameters; Step 5: Input the waveform data to be identified into the patient's spontaneous breathing identification model, and calculate the patient's spontaneous breathing starting point timestamp corresponding to the ventilator to be identified; the spontaneous breathing starting point timestamp includes the patient's physiological spontaneous inspiration start time and spontaneous exhalation start time; calculate the time difference between the patient's spontaneous breathing starting point timestamp and the respiratory support trigger time and the inspiration-expiration switching time respectively; Step 6: Based on the preset human-machine asynchronous phenomenon classification standard, compare and distinguish the time difference and the waveform data to be identified to obtain the intelligent classification result.

2. The method for intelligently identifying and classifying ventilator-patient ventilation asynchrony according to claim 1, characterized in that: The waveform data includes the pressure data, flow data and tidal volume data of the ventilator; the resampling frequency of the data depends on the sampling frequency of the ventilator itself or the sampling frequency of the waveform data acquisition device.

3. The method for intelligently identifying and classifying ventilator-patient ventilation asynchrony according to claim 1, characterized in that: The Z-Score standardization method was used for standardization. The waveform data collected during mechanical ventilation of the same patient under ventilator action were processed separately. The standardization was performed according to the mean and standard deviation. The formula is: Among them, y is the standardized data; x is the data before standardization; mean is the mean; σ is the standard deviation.

4. The method for intelligently identifying and classifying ventilator-patient ventilation asynchrony according to claim 1, characterized in that: The Mini-Batch method is used for batch processing. In the batch processing, the batch length is set to 512 sampling points. When the UNet neural network is used to train the patient spontaneous breathing recognition model in step 3, data is input in batches.

5. The method for intelligent identification and classification of ventilator-patient ventilation asynchrony according to claim 1, characterized in that: The Adam optimizer is used to optimize the training process of the patient spontaneous breathing recognition model. The Adam optimizer parameters are: α = 0.001, β1 = 0.9, β2 = 0.999, ∈ = 1e -8 .

6. The method for intelligently identifying and classifying ventilator-patient ventilation asynchrony according to claim 1, characterized in that: The loss function of the patient's spontaneous breathing recognition model adopts a cross entropy function, and the timestamp of the spontaneous breathing starting point calibrated by the expert is used as the true value of the patient's spontaneous breathing recognition model; and cross validation is adopted.

7. The method for intelligently identifying and classifying ventilator-patient ventilation asynchrony according to claim 1, characterized in that: The classification criteria for human-machine asynchronous phenomena include delayed inspiration, early switching, delayed switching, and invalid triggering; Delayed inspiration: The ventilator's respiratory support triggering time in step 5 is delayed by 250 milliseconds compared to the patient's physiological spontaneous inspiration start time; Early switching: The ventilator inhalation-exhalation switching time in step 5 is 100 milliseconds earlier than the patient's physiological spontaneous exhalation start time; Delayed switching: The ventilator inhalation-exhalation switching time in step 5 is delayed by 300 milliseconds compared to the patient's physiological spontaneous exhalation start time; Invalid trigger: The waveform data to be identified is lower than the preset support level value.

Images