CN115068760A - Intelligent identification and classification method for asynchronous phenomenon of breathing machine-patient - Google Patents

Abstract

The invention discloses an intelligent identification and classification method for asynchronous phenomena of breathing machine-patient ventilation, and provides a set of high-precision and high-accuracy identification and classification method aiming at the problems of low automatic detection degree and insufficient detection accuracy of the man-machine asynchronous phenomena of the breathing machine in the prior clinic. The invention deeply excavates the waveform data of the breathing machine based on the characteristic extraction and intelligent identification of the waveform data of the breathing machine, has higher feasibility and universality, does not need to carry out invasive data acquisition modes such as esophageal electrode catheter technology and the like on a patient, and does not influence the ongoing treatment process of the patient. Meanwhile, the starting time of inspiration and expiration of a patient in the ventilation process can be intelligently identified, more effective information can be found out, different man-machine asynchronous phenomena are accurately defined based on clinical diagnosis standards, and high-precision identification and accurate classification of the breathing machine-patient ventilation asynchronous phenomena are realized.

Description

Intelligent identification and classification method for asynchronous phenomenon of breathing machine-patient

Technical Field

The invention relates to the technical field of biomedical engineering, in particular to an intelligent identification and classification method for asynchronous phenomena of breathing machine-patient ventilation.

Background

Mechanical ventilation is a medical means for assisting the respiration of a patient by using a ventilator and maintaining the oxygenation function of the patient to strive for time for the treatment of primary diseases, and plays an important role in the treatment process of respiratory diseases such as new coronary pneumonia, atypical pneumonia, chronic obstructive pulmonary disease, acute respiratory syndrome and the like and the respiratory support of respiratory failure of critically ill patients.

However, during the mechanical ventilation of the patient, the asynchronous phenomenon between the ventilator and the patient directly affects the mechanical ventilation treatment process of the patient, and cannot ensure sufficient respiratory support, thereby delaying the taking off and even causing ventilator complications of the patient. The difficulty of the identification of the human-computer asynchronous phenomenon lies in the identification of the respiratory effort degree of a patient, the clinical identification of the human-computer asynchronous phenomenon through a ventilator waveform requires sufficient clinical experience of a physician, and methods for detecting the respiratory effort degree of the patient and the starting point of each section of respiration after signal processing and analysis are carried out by detecting the diaphragm electric signal of the patient through an esophageal electrode are also available at home and abroad. However, this method is not universal, and is not harmful to the patient himself, and is not favorable to the treatment process of mechanical ventilation.

In addition, the artificial intelligence technology such as machine learning is utilized for identification, and an identification model based on a long-short term memory artificial neural network (LSTM) can be adopted for identifying and classifying the human-computer asynchronous phenomenon, however, the method has lower sensitivity and accuracy for identifying the respiratory effort degree of the patient.

Through detection and verification of calibrated clinical data, the identification method of the project has higher human-computer asynchronous identification accuracy.

Therefore, how to improve the accuracy of the recognition of the asynchronous phenomenon of the breathing machine on the basis of avoiding the damage to the patient is a problem that needs to be solved by those skilled in the art.

Disclosure of Invention

In view of this, the invention provides an intelligent identification and classification method for asynchronous phenomena of ventilator-patient ventilation, which aims at the problems of low automatic detection degree and insufficient detection accuracy of the human-machine asynchronous phenomena of the existing clinical ventilators, and provides a set of high-precision and high-accuracy identification and classification method for the human-machine asynchronous phenomena existing in the mechanical ventilation process of patients. The invention deeply excavates the waveform data of the breathing machine based on the characteristic extraction and intelligent identification of the waveform data of the breathing machine, has higher feasibility and universality, does not need to carry out invasive data acquisition modes such as esophageal electrode catheter technology and the like on a patient, and does not influence the ongoing treatment process of the patient. Meanwhile, the starting time of inspiration and expiration of a patient in the ventilation process can be intelligently identified, more effective information can be found out, different man-machine asynchronous phenomena are accurately defined based on clinical diagnosis standards, and high-precision identification and accurate classification of the breathing machine-patient ventilation asynchronous phenomena are realized.

In order to achieve the purpose, the invention adopts the following technical scheme:

an intelligent identification and classification method for asynchronous phenomenon of breathing of a breathing machine-patient comprises the following specific steps:

step 1: acquiring waveform data of a breathing machine in a mechanical ventilation process, calibrating the waveform data, and calibrating inspiration starting time and expiration starting time of each period; acquiring mechanical ventilation simulation data under different patient parameters based on a mechanical ventilation model of an active respiration patient; constructing a database using the waveform data and the mechanical ventilation simulation data;

step 2: resampling the database according to a set sampling frequency, and then sequentially carrying out standardization and batch processing; resampling to ensure that the sampling frequency of the clinical and simulated respiratory data is unified to 1 kHz;

and 3, step 3: establishing a patient autonomous respiration recognition model by adopting a deep learning algorithm of a UNet neural network according to the processed sampling data,

and 4, step 4: acquiring waveform data to be identified and control parameters of a breathing machine to be identified, and acquiring breathing support trigger time and inhalation-exhalation switching time of each period of the breathing machine to be identified according to the control parameters;

and 5: inputting waveform data to be identified into the patient spontaneous respiration identification model, and calculating to obtain a spontaneous respiration starting point timestamp of the patient corresponding to the breathing machine to be identified; the spontaneous breathing onset timestamp comprises a patient physiological spontaneous inspiration onset time and a spontaneous expiration onset time; calculating the time difference between the spontaneous respiration starting time and the trigger and switching time of the respirator;

step 6: and judging the time difference according to a preset classification standard of the man-machine asynchronous phenomenon to obtain an intelligent classification result.

Preferably, the waveform data includes ventilator pressure data, flow data, and tidal volume data; 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.

Preferably, a Z-Score standardization method is adopted for standardization processing, and waveform data in a mechanical ventilation process which is acquired by the same patient under the action of a breathing machine in a single time are respectively processed, namely pressure data, flow data and tidal volume data are respectively processed; the control parameter is feedback data of a breathing machine control system;

respectively calculating the mean value and the standard deviation of the pressure data, the flow data and the tidal volume data, and carrying out standardization processing according to the mean value and the standard deviation, wherein the formula is as follows:

wherein y is the normalized data; x is pre-normalization data; mean is the mean value; σ is the standard deviation.

Preferably, a Mini-Batch method is adopted for Batch processing, the length of a Batch is set to be 512 sampling points in the Batch processing, and data is input by taking the Batch as a unit when a UNet neural network is adopted to train the patient spontaneous respiration recognition model in step 3. The optimal balance between the memory efficiency and the memory capacity can be searched, the memory utilization rate is improved through parallelization, the required training iteration times are reduced, the recognition speed of the model is improved, the accuracy of the gradient descending direction of the training model can be increased, and the amplitude of training vibration is reduced.

Preferably, the patient spontaneous respiration identification model comprises a down-sampling part, a splicing part and an up-sampling part, and a UNet model structure is built based on a convolutional neural network to build the patient spontaneous respiration identification model;

wherein, the down-sampling part extracts shallow layer characteristic information, compresses data information through the convolution layer, and adopts 4-layer compression; the up-sampling part extracts deep characteristic information, and data information is expanded through a deconvolution layer, wherein 4 layers of expansion are adopted;

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 each convolution layer for regularization to prevent overfitting in the training process;

the splicing section fuses the deep layer feature information and the shallow layer feature information.

Preferably, an Adam optimizer is adopted to optimize the training process of the patient spontaneous respiration recognition model, and the parameters of the Adam optimizer are as follows: α ═ 0.001, β ₁ ＝0.9，β ₂ ＝0.999，∈＝1e ^-8 。

Preferably, the loss function of the patient spontaneous respiration recognition model adopts a cross entropy function, and a spontaneous respiration starting point timestamp calibrated by an expert is used as a true value of the patient spontaneous respiration recognition model; and cross validation was used.

Preferably, the criteria for classifying the human-machine asynchronous phenomenon include delayed inspiration, early switching, delayed switching and ineffective triggering;

and (3) delaying inspiration: delaying the ventilator's breathing support trigger time from the patient's physiological spontaneous inspiration start time by 250 milliseconds in step 5;

early handover: the switching time of the breathing machine inhalation-exhalation in the step 5 is 100 milliseconds ahead of the starting time of the physiological spontaneous exhalation of the patient;

and (3) delayed switching: delaying the ventilator inhale-exhale switching time in step 5 by 300 milliseconds from the patient physiological spontaneous exhale starting time;

invalid triggering: if the waveform data to be identified is lower than a preset support level value; the waveform data shows that the airway pressure sinks, the flow fluctuates slightly, but the breathing machine cannot be triggered to enable the pressure waveform to reach a pressure value of a preset support level, and the pressure support of the breathing machine cannot be triggered for the respiratory effort degree of the patient, namely the airway pressure or flow change generated by the respiratory effort of the patient cannot reach the pressure or flow trigger sensitivity set by the breathing machine, and whether the respiratory effort degree of the patient cannot trigger the pressure support of the breathing machine or not is identified through the patient spontaneous respiration identification model according to a preset sensitivity threshold of the breathing machine.

Through the technical scheme, compared with the prior art, the invention discloses and provides an intelligent identification and classification method for asynchronous phenomena of breathing machine-patient ventilation, which aims at the automatic identification and classification problems of man-machine asynchronous phenomena, provides an image feature extraction model based on a U-Net convolutional neural network in deep learning, inputs data which are pressure, flow and tidal volume waveform data of a breathing machine calibrated by a clinical expert and a respiratory effort starting point of a patient calibrated by the expert, namely the inspiration starting time and the expiration starting time in each period, the model is trained by adopting enough calibration data, the triggering time and the switching time of the breathing machine are effectively extracted, and the starting time of the inspiration phase and the expiration phase of the patient, thereby avoiding the original invasive method for obtaining the starting time of each phase of the respiration of the patient and realizing the automatic detection and identification of the man-machine asynchronous phenomenon. In addition, a man-machine asynchronous phenomenon classification standard is set based on expert clinical knowledge, the identified asynchronous phenomenon is automatically classified according to the classification standard, and finally intelligent identification and classification of the man-machine asynchronous phenomenon existing between the breathing machine and a patient in the mechanical ventilation process are achieved.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the embodiments or the prior art descriptions will be briefly described below, it is obvious that the drawings in the following description are only embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the provided drawings without creative efforts.

FIG. 1 is a schematic flow chart of a method for intelligently identifying and classifying ventilator-patient asynchronous ventilation phenomena according to the present invention;

FIG. 2 is a schematic diagram of a UNet neural network structure of a patient spontaneous respiration recognition model provided by the invention;

FIG. 3 is a schematic diagram of an intelligent classification process of the human-machine asynchronous phenomenon provided by the invention.

Fig. 4 is a schematic diagram of a mechanical ventilation model of an active breathing patient according to the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The embodiment of the invention discloses a breathing machine-patient ventilation asynchronous phenomenon intelligent identification and classification method, the flow is shown in figure 1, and the method is mainly divided into two parts, wherein one part is used for establishing a patient spontaneous respiration identification model based on a UNet neural network to realize patient spontaneous respiration identification; and the other part is to classify and judge the human-machine asynchronous phenomenon according to the identified spontaneous respiration condition of the patient so as to realize intelligent identification of the human-machine asynchronous phenomenon.

S1: data acquisition and database retrieval

Firstly, acquiring waveform data of a breathing machine in a mechanical ventilation process as clinical data, wherein the waveform data comprises pressure data, flow data and tidal volume data, carrying out expert calibration on the waveform data, and calibrating inspiration starting time and expiration starting time of each period; 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. Storing the clinical data into a data list or establishing a database for data processing in the identification process. Meanwhile, based on the mechanical ventilation model of the active respiration patient, mechanical ventilation simulation data under different patient parameters are obtained, and the simulation data are stored in a data list or a database is established for training and checking of the recognition model. A mechanical ventilation model of an active respiration patient based on the pneumatic principle, as shown in fig. 4, includes an air source 1, a control valve 2, a pressure sensor 3, a flow sensor 4, a throttle valve 5, an elastic cavity 6, a brake valve 7, a vacuum pump 8, an AD sampling module 9, and an upper computer 10.

S2: data pre-processing

The acquired waveform data of the breathing machine have different sampling frequencies, so that the existing database needs to be resampled before being input into the neural network, and the sampling frequency after processing is unified to 100 Hz. The breathing cycle of a patient is averagely 3-5 seconds, and after data is resampled, a sampling point is guaranteed to be one every 0.01 second, so that the method has certain timeliness.

Data is resampled and then normalized. The invention adopts a Z-Score standardization method to respectively process the data of the same patient and the parameter of the same patient. And respectively calculating the mean value and the standard deviation of the pressure, flow and tidal volume data of the patient, subtracting the mean value from each sampling point, and dividing the result by the standard deviation to obtain the standardized data. The formula is as follows:

the normalized processed data is subjected to batch processing (Batches), and the length (Batchsize) of the batch is set to 512 sample points. When performing neural network training of a subsequent recognition model, data is input in units of Batches (Batches). The correct batch length (Batchsize) can find the optimal balance between the memory efficiency and the memory capacity, the memory utilization rate is improved through parallelization, the required training iteration times are reduced, the recognition speed of the model is improved, the accuracy of the gradient descending direction of the training model is increased, and the amplitude of the training vibration is reduced.

S3: patient autonomous respiration recognition model based on UNet neural network

The invention adopts a deep learning algorithm based on a UNet neural network to establish an identification model, and identifies the time point of the switching of the inspiration and expiration of the spontaneous respiration of a patient, which is called as the time stamp of the inspiration starting point and the expiration starting point (the time stamp of the respiration starting point). A deep learning algorithm based on a UNet neural network is usually used for edge segmentation of a two-dimensional image, and a novel identification model is established in the invention to perform feature identification on one-dimensional respiratory data. The neural network structure of the model is shown in fig. 2.

The model is divided into three parts, a down sampling part, a splicing part and an up sampling part. The first half part is a down-sampling part which is mainly used for extracting shallow layer characteristic information; the second half part is an upper sampling part and mainly used for extracting deep characteristic information; the middle part is a splicing part, and deep layer features and shallow layer features are fused, so that feature information loss in the convolution process is prevented.

The down-sampling part compresses the data information by a convolution layer, the up-sampling part expands the data information by a deconvolution layer, and the two parts adopt 4-layer compression and 4-layer expansion respectively. The activation function for each convolutional layer is a linear rectification function (ReLU), the convolutional layer for the output layer is 1, and the activation function is a normalized exponential function (softmax). And in the middle of the convolutional layer, a Dropout layer is added for regularization, so that overfitting in the training process is prevented.

The recognition model adopts an Adam optimizer to optimize the training process, accelerates convergence, and has the design parameters as follows: alpha is 0.001, beta ₁ ＝0.9，β ₂ ＝0.999，∈＝1e ^-8 . The loss function of the recognition model uses a cross-entropy function (cross-entropy). And adopting the timestamp of the spontaneous respiration starting point of the patient, which is calibrated by an expert, as a true value (GroudTruth) of the identification model.

Cross validation was used to avoid overfitting. The data was divided into 15 sets, each set containing respiratory data of the same patient. In the training process, 14 groups of data are used for training the neural network, and the training result is checked in the remaining 1 group of data. The training process is repeated for 15 times, each group of data is circularly used as a test group, and a prediction result is finally obtained.

Recognition result checking method

And if the time difference between the detection time stamp of the identification model and the expert calibration time stamp is within 0.5s, the prediction result is considered to be consistent with the expert calibration data and is regarded as a true positive value (TP). If a recognition result timestamp cannot match any expert calibration timestamp, it is considered as a false positive value (FP). If an expert calibration timestamp cannot match any of the recognition result timestamps, it is considered a false negative value (FN). By the method, the recognition result can be checked to obtain the check result, so that the accuracy of the recognition model is obtained and compared with other recognition methods.

S4: acquiring waveform data to be identified and control parameters of the breathing machine to be identified, and acquiring the breathing support trigger time and the inhalation-exhalation switching time of each period of the breathing machine to be identified according to the control parameters.

S5: firstly, waveform data and control parameters of a breathing machine are collected, and triggering and switching time of each period of the breathing machine is obtained. And preprocessing the waveform data, inputting the preprocessed waveform data into a spontaneous respiration recognition model of the patient, and recognizing a starting point time stamp, an inspiration starting time and an expiration starting time of spontaneous respiration of the patient by adopting the method described in the first part. Automatically acquiring the time difference between the spontaneous respiration starting time of the patient and the trigger and switching time of a respirator;

s6: and substituting the time difference between the spontaneous respiration starting time of the patient and the triggering and switching time of the respirator into a human-computer asynchronous classification standard for judgment to obtain an intelligent classification result.

Classification standard of man-machine asynchronous phenomenon:

based on expert experience and clinical rules, the invention divides the man-machine asynchronous phenomenon which seriously affects the mechanical ventilation process into four types of delayed inspiration, early switching, delayed switching and ineffective triggering.

And (3) delaying inspiration: the trigger time defined as the inspiratory support pressure provided by the ventilator is delayed by 250 milliseconds from the initial time the patient begins inspiration;

early handover: the switching time defined as the ventilator stopping support pressure is 100 milliseconds ahead of the initial time at which the patient begins to exhale;

and (3) delayed switching: the switching time defined as the ventilator stopping support pressure is 300 milliseconds later than the initial time the patient begins to exhale;

invalid triggering: defined as the inability of the patient's respiratory effort to trigger the ventilator's pressure support.

The flow chart of the method of the part of the technology for intelligently classifying the human-machine asynchronous phenomenon is shown in figure 3.

The invention has the beneficial effects that:

a patient autonomous respiration recognition model based on a UNet neural network is established, a technology for intelligently classifying the man-machine asynchronous phenomenon between a respirator and a patient is provided, and finally a set of intelligent recognition and classification technology for the man-machine asynchronous phenomenon between the respirator and the patient in the mechanical ventilation process is formed. Compared with the prior art, the technology avoids an invasive acquisition mode for the spontaneous breathing effort degree of the patient, does not cause damage to the patient, and does not influence the normal work and treatment of the breathing machine; the waveform data of the breathing machine is deeply mined, and the characteristic time stamp of spontaneous breathing of the patient can be automatically detected. The method provided by the invention has the advantages of high identification accuracy, simple operation, high feasibility and strong universality.

The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. The device disclosed by the embodiment corresponds to the method disclosed by the embodiment, so that the description is simple, and the relevant points can be referred to the method part for description.

The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (8)

1. An intelligent identification and classification method for asynchronous phenomenon of breathing of a breathing machine-patient is characterized by comprising the following specific steps:

step 1: acquiring waveform data of a breathing machine in a mechanical ventilation process, calibrating the waveform data, and calibrating inspiration starting time and expiration starting time of each period; acquiring mechanical ventilation simulation data under different patient parameters based on a mechanical ventilation model of an active respiration patient; constructing a database using the waveform data and the mechanical ventilation simulation data;

step 2: resampling the database according to a set sampling frequency, and then sequentially carrying out standardization and batch processing on the sampling data;

and step 3: establishing a patient autonomous respiration recognition model by adopting a deep learning algorithm of a UNet neural network according to the processed sampling data;

and 4, step 4: acquiring waveform data to be identified and control parameters of a breathing machine to be identified, and acquiring breathing support trigger time and inhalation-exhalation switching time of each period of the breathing machine to be identified according to the control parameters;

and 5: inputting waveform data to be identified into the patient spontaneous respiration identification model, and calculating to obtain a spontaneous respiration starting point timestamp of the patient corresponding to the breathing machine to be identified; the spontaneous breathing onset timestamp comprises a patient physiological spontaneous inspiration onset time and a spontaneous expiration onset time; calculating the time difference between the time stamp of the spontaneous respiration starting point of the patient and the respiration support trigger time and the inhalation-exhalation switching time respectively;

step 6: and comparing and judging the time difference and the waveform data to be identified according to a preset human-computer asynchronous phenomenon classification standard to obtain an intelligent classification result.

2. The method for intelligently identifying and classifying ventilator-patient ventilation asynchrony phenomena as claimed in claim 1, wherein the waveform data comprises ventilator pressure data, flow data and tidal volume data; 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 intelligent identification and classification method for asynchronous ventilator-patient ventilation phenomena according to claim 1, wherein a Z-Score standardization method is adopted to perform standardization processing, and waveform data in a mechanical ventilation process acquired by the same patient under the action of a ventilator in a single time are respectively processed; and (4) carrying out standardization treatment according to the mean value and the standard deviation, wherein the formula is as follows:

wherein y is the normalized data; x is pre-normalization data; mean is the mean value; σ is the standard deviation.

4. The method for intelligently identifying and classifying the asynchronous phenomenon of ventilator-patient ventilation according to claim 1, wherein a Mini-Batch method is adopted to perform Batch processing, the length of a Batch is set to be 512 sampling points in the Batch processing, and data is input in a Batch unit when a UNet neural network is adopted to train the patient spontaneous respiration identification model in step 3.

5. The intelligent identification and classification method for the asynchronous phenomenon of ventilator-patient ventilation according to claim 1 is characterized in that the patient spontaneous respiration identification model comprises a down-sampling part, a splicing part and an up-sampling part, a UNet model structure is built based on a convolutional neural network, and the patient spontaneous respiration identification model is built;

wherein, the down-sampling part extracts shallow layer characteristic information, compresses data information through the convolution layer, and adopts 4-layer compression; the up-sampling part extracts deep characteristic information, and data information is expanded through a deconvolution layer, wherein 4 layers of expansion are adopted;

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 each convolution layer for regularization;

the splicing section fuses the deep layer feature information and the shallow layer feature information.

6. The method for intelligently identifying and classifying the asynchronous phenomenon of ventilator-patient ventilation according to claim 1, wherein an Adam optimizer is adopted to optimize the training process of the patient spontaneous respiration identification model, and the parameters of the Adam optimizer are as follows: alpha is 0.001, beta ₁ ＝0.9，β ₂ ＝0.999，∈＝1e ^-8 。

7. The intelligent recognition and classification method for asynchronous phenomenon of ventilator-patient ventilation according to claim 1, wherein the loss function of the patient spontaneous respiration recognition model adopts a cross entropy function, and takes a self respiration starting point timestamp calibrated by an expert as the true value of the patient spontaneous respiration recognition model; and cross validation was used.

8. The method for intelligently identifying and classifying ventilator-patient ventilation asynchrony as claimed in claim 1, wherein the classification criteria for human-machine asynchrony comprise delayed inspiration, early switch, delayed switch and ineffective trigger;

and (3) delaying inspiration: delaying the ventilator's breathing support trigger time from the patient's physiological spontaneous inspiration start time by 250 milliseconds in step 5;

early handover: the switching time of the breathing machine inhalation-exhalation in the step 5 is 100 milliseconds ahead of the starting time of the physiological spontaneous exhalation of the patient;

and (3) delayed switching: delaying the ventilator inhale-exhale switching time in step 5 by 300 milliseconds from the patient physiological spontaneous exhale starting time;

invalid triggering: and judging that the waveform data to be identified is lower than a preset support level value.

Images