A single-stage identification method and system for mechanical ventilation patient-machine asynchrony

Abstract

The present invention discloses a single-stage identification method and system for mechanical ventilation human-machine asynchrony, comprising: obtaining pressure time series, flow rate time series and volume time series data of a ventilator; after feature embedding and feature extraction of the pressure time series, flow rate time series and volume time series data, obtaining a feature map characterizing the pressure time series, flow rate time series and volume time series; after the feature map passes through a classifier, an output tensor is obtained, and based on the output tensor, the location where human-machine asynchrony occurs and the type of human-machine asynchrony are obtained. In the single-stage identification method for mechanical ventilation human-machine asynchrony of the present invention, positioning and classification can be integrated into one feature map; compared with a two-stage network design, the single stage can achieve a faster recognition speed. At the same time, the classification and positioning are integrated into one feature map, achieving certain feature constraints and having higher positioning and classification accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of ventilator data analysis, and in particular to a single-stage identification method and system for mechanical ventilation human-machine asynchrony. Background Art

[0002] Ventilators are crucial life-support devices in the intensive care unit (ICU), providing respiratory support, resolving hypoxia, facilitating recovery from underlying illnesses, and ultimately saving lives. However, the interaction between ventilators and patients carries inevitable risks, particularly frequent patient-ventilator asynchrony (i.e., patient-ventilator asynchrony). Literature reports suggest that patient-ventilator asynchrony can lead to various adverse effects, including discomfort, air starvation, increased work of breathing, muscle damage, decreased sleep quality, increased need for sedatives or muscle relaxants, ventilator-associated lung injury, prolonged mechanical ventilation, increased difficulty weaning from the ventilator, and even increased mortality.

[0003] Most current ventilators provide positive pressure ventilation, using a variety of sensors to detect respiratory status indicators such as the patient's respiratory flow rate and pressure, and using the device's built-in algorithms to control respiratory rhythms such as respiratory triggering, switching, and pausing. Ventilators generally have a control mode (mainly program-controlled ventilation rhythm) and an autonomous mode (mainly patient-controlled ventilation rhythm). In clinical practice, volume control and pressure support ventilation modes are mostly used. Ventilators from various manufacturers in countries around the world are equipped with a variety of parameters. Professional respiratory therapists are required to set the ventilator parameters and provide different mechanical ventilation modes to patients based on the patient's condition. More importantly, by monitoring and analyzing the various waveform data on the ventilator screen, it is possible to determine whether there are abnormalities such as human-machine asynchrony between the patient and the ventilator, and to adjust the parameters in a timely manner to avoid irreversible harm to the patient. Human-machine asynchrony mainly includes invalid triggering and repeated triggering, which can occur in various ventilator modes. Unfortunately, there is a shortage of respiratory therapists in countries around the world, and daily mechanical ventilation monitoring is mostly undertaken by ICU critical care nurses, resulting in a high rate of missed diagnoses and misdiagnoses. Data from multiple clinical studies from 2011 to 2020 confirmed that ICU doctors and nurses, including respiratory therapists, have a large problem of missed diagnoses and misdiagnoses when visually identifying waveform data and human-machine asynchrony.

[0004] Ventilators manufactured worldwide feature intelligent technical alarms, such as critical value alarms and circuit disconnects, but lack the ability to identify patient-ventilator asynchrony. Currently, some studies are attempting to use rule-based methods to mine waveform features such as gradients, curvatures, and thresholds, translating expert experience into automated methods for identifying patient-ventilator asynchrony. Other studies are attempting to identify the relationship between a small number of patient groups and specific types of patient-ventilator asynchrony, using automated methods such as machine learning to classify specific types of asynchrony. These methods simply abstract the identification of patient-ventilator asynchrony into a classification problem, with little or no consideration of localization, which inherently presents significant limitations. Summary of the Invention

[0005] To address the above issues, the present invention proposes a one-stage detection method and system for mechanical ventilation human-machine asynchrony (OSD). Based on three time-varying waveforms of pressure, flow rate, and volume, the system utilizes machine learning, especially deep learning technology, as a feature extractor, and common machine learning classifiers such as fully connected networks as classifiers. The output tensor of the classifier can simultaneously characterize the location and classification of human-machine asynchrony, thereby achieving the goal of real-time identification of various human-machine asynchronies.

[0006] In some embodiments, the following technical solutions are adopted:

[0007] A single-stage method for identifying patient-ventilator asynchrony during mechanical ventilation, comprising:

[0008] Obtain the pressure time series, flow rate time series and volume time series data of the ventilator;

[0009] The pressure time series, flow rate time series and volume time series data are converted into original time series representations through feature embedding to obtain multi-time series embedding of pressure, flow rate and volume;

[0010] After feature extraction, the multi-time series embedding obtains a feature map representing the pressure time series, flow rate time series and volume time series;

[0011] The feature map is passed through a classifier to obtain an output tensor, and the location where human-computer asynchrony occurs and the type of human-computer asynchrony are obtained based on the output tensor.

[0012] The output tensor includes a positioning tensor and a classification tensor. The first element of each row in the output tensor is the positioning tensor, and a positioning tensor of 1 represents the location where human-computer asynchrony occurs. The remaining elements of each row are classification tensors, and the category corresponding to a classification tensor of 1 is the type of human-computer asynchrony.

[0013] The objective function of the classifier is specifically:

[0014]

[0015] in,

[0016]

[0017]

[0018]

[0019] Where, Represent classification items, positioning items, and auxiliary items respectively; α, β, γ represent the weights of positioning items and auxiliary items respectively; C represents the category of mechanical ventilation human-machine asynchrony, N represents the number of batches, K represents the total number of positioning positions, y∈{0,1} is a onehot encoded tensor, is the network output tensor with the same dimension as y; δ∈[0,∞) is the control parameter of the auxiliary term; w c is the weight of each category in the classification item; It is the output tensor of the cth class in the nth batch of classification items and the output tensor of a certain i class in the nth batch; y n，i They represent the tensors composed of the output tensor of the i-th position of the n-th batch during prediction and the i-th position of the true label of the n-th batch; y n They represent the tensors consisting of the nth batch of output tensors and the nth batch of true labels during prediction.

[0020] In other embodiments, the following technical solutions are adopted:

[0021] A single-stage recognition system for mechanical ventilation patient-ventilator asynchrony, comprising:

[0022] A data acquisition module is used to obtain the pressure time series, flow rate time series and volume time series data of the ventilator;

[0023] A feature embedding module is used to convert the pressure time series, flow rate time series and volume time series data into original time series representations through feature embedding to obtain multi-time series embedding of pressure, flow rate and volume;

[0024] A feature extraction module is used to extract features from the multiple time series embeddings to obtain feature maps representing pressure time series, flow rate time series and volume time series;

[0025] The feature classification module is used to obtain an output tensor after passing the feature map through a classifier, and obtain the location and type of human-computer asynchrony based on the output tensor.

[0026] In other embodiments, the following technical solutions are adopted:

[0027] A terminal device includes a processor and a memory, wherein the processor is used to implement various instructions; the memory is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor for the above-mentioned single-stage identification method for mechanical ventilation human-machine asynchrony.

[0028] In other embodiments, the following technical solutions are adopted:

[0029] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device to implement the above-mentioned single-stage identification method for mechanical ventilation human-machine asynchrony.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] (1) The single-stage recognition method for mechanical ventilation human-machine asynchronous detection of the present invention can integrate positioning and classification into one feature map. Compared with the two-stage network design, the single-stage method can achieve faster recognition speed. At the same time, the classification and positioning are integrated into one feature map, which realizes certain feature constraints and has higher positioning and classification accuracy.

[0032] (2) The present invention selects three waveform data, namely pressure, flow rate and volume, as input data for asynchronous positioning and classification of mechanical ventilation humans and machines. Compared with the conventional method that only uses pressure and flow rate as input data, the present invention adds volume data and uses three-dimensional data to express the human-machine interaction state, which has a better effect and can obtain more accurate positioning and classification results.

[0033] (3) The classifier of the present invention integrates positioning and classification into one output tensor and designs its corresponding objective function, which can simplify the entire process of human-machine asynchronous target recognition, obtain better accuracy and faster speed, and meet the clinical needs for real-time and accurate recognition.

[0034] (4) The present invention provides clinicians, nurses, and respiratory therapists with the ability to quickly identify the time and location of various human-machine asynchrony, and to make targeted parameter adjustments in a timely manner to correct human-machine asynchrony, thereby improving the quality of mechanical ventilation, reducing complications, and promoting patient recovery. It has wide applicability.

[0035] Other features and advantages of additional aspects of the present invention will be given in part in the following description and in part will become obvious from the following description or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of a single-stage identification method for mechanical ventilation patient-machine asynchrony in an embodiment of the present invention;

[0037] Figure 2 Schematic diagram of a specific example of a single-stage identification method for mechanical ventilation patient-machine asynchrony in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0039] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0040] Example 1

[0041] In one or more embodiments, a single-stage identification method for mechanical ventilation patient-machine asynchrony is disclosed, combining Figure 1 , specifically including the following process:

[0042] S101: Acquire pressure time series, flow rate time series, and volume time series data of the ventilator;

[0043] In this embodiment, a ventilator data acquisition module is used to collect data, for example: a pressure sensor is used to collect the pressure data waveform of the ventilator, a flow rate sensor is used to collect the flow rate data waveform of the ventilator, the waveform collected by the flow rate sensor is integrated, or other calculations are performed to obtain a volume data waveform.

[0044] Because data export protocols vary among ventilators from different manufacturers, the data acquisition module must adapt to these protocols and send them. The data acquisition module collects pressure, flow rate, and volume waveforms based on the protocol. These waveforms are then sent in stream or batch mode via transport layer protocols such as TCP, UDP, and QUIC.

[0045] After obtaining the pressure, flow rate, and volume waveform data, the data is preprocessed.

[0046] A ventilator is a real-time device that processes large amounts of data. Data delays and loss may occur during data transmission. Data preprocessing involves calibrating the received data and marking missing data with blanks. Calibrated and missing data requires sampling, with the sampling rate adjustable based on network load and system capacity.

[0047] After calibration, missing blank marks and sampling, the data will enter a sliding window with adjustable step size and window length. The sliding window slides in time and outputs pressure time series, flow rate time series and volume time series data with a length of the window length respectively.

[0048] In addition, data enhancement techniques such as random addition of Gaussian noise, data inversion, and random data cropping (mostly used in the training phase of machine learning) are not limited in this embodiment.

[0049] S102: The obtained pressure time series, flow rate time series, and volume time series data are converted into original time series representations through feature embedding, highlighting features such as cycle and trend, and obtaining multi-time series embedding of pressure, flow rate, and volume;

[0050] In this embodiment, the feature embedding component primarily serves to perform high-dimensional characterization of the three input time series data (i.e., multivariate time series). Generally speaking, the input data is separable in a low-dimensional space in a high-dimensional space. This embodiment can utilize feature embedding techniques such as Ts2vec, Time2vec, time-frequency conversion, or time delay embedding based on Takens theory to enhance the characterization capabilities of the input data and improve the performance of subsequent feature characterization classification locators.

[0051] This embodiment does not limit a specific feature embedding technology. Feature embedding before feature extraction can be similar to certain structures, technologies, and components in feature extraction. This embodiment does not limit the specific internal implementation and technical characteristics of the feature embedding part, but limits its input and output data structure. The input of the feature embedding part is a tensor of window length × 3 (three time series), and the output is a tensor of embedding dimension × feature dimension, which serves as the feature map for the next step of feature extraction.

[0052] S103: After feature extraction, the multi-time series embedding obtains a feature map representing the pressure time series, flow rate time series, and volume time series;

[0053] In this embodiment, the feature extractor is mainly a further expression of the embedded features, which can be implemented by a multi-layer perceptron, a one-dimensional convolutional neural network, a one-dimensional recurrent convolutional neural network, a wavelet-based neural network, an autoencoder or a Transformer or a hybrid technology thereof. The feature extractor mainly uses various technologies in neural networks or machine learning theory to characterize tensors, and uses a deep and wide network to express input data in order to accurately identify the target of interest in the classifier. This embodiment does not limit the specific internal implementation and technical characteristics of the feature extractor, but limits its input and output data structure. The input of the feature extractor is a tensor of embedding dimension × feature dimension, and the output is a tensor of characterization dimension × feature dimension, which serves as a feature map characterizing pressure time series, flow rate time series and volume time series.

[0054] There are many feature extractors in the field of machine learning, such as convolutional neural networks, recurrent neural networks, gradient boosting trees, etc., which are not limited in this embodiment.

[0055] S104: The feature map is passed through the classifier to obtain an output tensor, and the location where human-computer asynchrony occurs and the type of human-computer asynchrony are obtained based on the output tensor.

[0056] In this embodiment, the input of the classifier is the feature map of the extraction dimension × feature dimension extracted by the feature extractor, and the output is the window length × (category (one-hot encoding) + 1 (representing positioning)).

[0057] When the output tensor of the classifier is a two-dimensional tensor (matrix), the first dimension (i.e., row) is the window length, including the positioning tensor and the classification tensor; the first element of each row is the positioning tensor, and a positioning tensor of 1 represents the location where human-computer asynchrony occurs, and there can be multiple elements representing multiple locations; the remaining elements of each row are classification tensors, and the category corresponding to a classification tensor of 1 is the type of human-computer asynchrony.

[0058] The classifier's output tensor is relatively sparse overall, and can be considered a multi-target, multi-label problem. Backpropagation is implemented using an entropy-based loss and a distance metric to achieve successive approximations. The classification tensor involved in the backpropagation algorithm is controlled by the classification part of the objective function, while the localization tensor involved in the backpropagation algorithm is controlled by the localization part of the objective function.

[0059] This embodiment designs a multi-task classifier objective function that includes a classification term, a localization term, and an auxiliary term. The classification term uses cross-entropy to calculate the classification tensor and the true label, while the localization term uses cross-entropy combined with the maximum entropy of multiple labels. The classification and localization terms are each given adjustable weights, and an auxiliary term, smoothed mean absolute error, is added for enhancement.

[0060] Assume that the multivariate time series after preprocessing is p(t), f(t), v(t) represents pressure, flow rate, and volume respectively. Assume that the sliding window length is T, the step size is S, and the multivariate time series is p(T+nS), f(T+nS), v(T+nS), The feature embedding part is assumed to be θ emb , the feature extraction module assumes θ fea , then the feature map can be calculated as

[0061] For the real labeled data y, the multi-task objective function can be expressed as:

[0062] The objective function of the classifier is specifically:

[0063]

[0064] in,

[0065]

[0066]

[0067]

[0068] Where, Represent classification items, positioning items and auxiliary items respectively; α, β, γ represent the weights of classification items, positioning items and auxiliary items respectively; C represents the mechanical ventilation human-machine asynchrony category, N represents the number of batches, y∈{0,1} is a onehot encoded tensor, is the network output tensor with the same dimension as y; δ∈[0,∞) is the control parameter of the auxiliary term.

[0069] L cls Middle w c It is the weight of each category in the classification item. The default weight of all categories is 1; It is the output tensor of the cth category in the nth batch of classification items and the output tensor of a certain i category (a total of C categories) in the nth batch;

[0070] L det middle y n，i They represent the tensors composed of the output tensor of the i-th position of the n-th batch during prediction and the i-th position of the true label of the n-th batch;

[0071] L assit middle y n They represent the tensors consisting of the nth batch of output tensors and the nth batch of true labels during prediction.

[0072] S105: Post-process the output tensor output by the classifier;

[0073] The post-processing process is:

[0074] For the positioning tensor, use functions such as topK to obtain the positions with the highest confidence in the positioning tensor;

[0075] For the classification tensor, for the obtained positions with the highest confidence, use functions such as topK to obtain the category with the highest confidence in the classification tensor.

[0076] The location and type of human-machine asynchrony after post-processing are sent to the human-machine asynchrony alarm distribution module to be forwarded to the corresponding control unit.

[0077] There are many types of post-processing logic, but the general principle is to combine positioning and classification, and only consider classification when positioning is true.

[0078] It can be understood that the process from data collection to alarm distribution in this embodiment supports batch processing, and can utilize the acceleration capabilities of parallel computing devices such as GPU and TPU, which will not be repeated here.

[0079] It should also be noted that based on Figure 1The data processing architecture shown uses offline data collected from a ventilator and manually labeled as a sample data set. The sample data set is used to train the entire processing flow, and the back-propagation algorithm is used to optimize the various model parameters.

[0080] As a specific example, combining Figure 2 The ventilator data acquisition module is composed of hardware such as a single-chip microcomputer, an I / O conversion module, a power supply circuit, and a network communication module. The software adapts to the data export protocol of the ventilator, implements a server that communicates with the ventilator, and uses the transport layer protocol to send data to the preprocessing module.

[0081] The data preprocessing module includes data calibration, data enhancement, data resampling, and data standardization. It performs preprocessing such as calibration, resampling, and standardization on the collected pressure, flow rate, and volume waveforms. At the same time, data enhancement technology can be used to increase the data volume.

[0082] The collected pressure, flow rate, and volume waveform data are respectively passed through a sliding window with a length of 512 and a step size of 1. These waveforms are sent in a stream or batch manner through transport layer protocols such as TCP, UDP, and QUIC, generating a 512×3 tensor each time.

[0083] The data embedding part uses Time2vec technology to convert the 512×3 tensor into a 512×16 embedding tensor.

[0084] The feature extraction part includes a front-end multi-level structure and a multi-head attention module. The embedded tensor passes through a three-layer feature pyramid, and the feature tensor of each layer is sent to the multi-head attention at the back. The multi-head attention is then connected to a multi-layer recurrent convolutional network (such as the gated recurrent unit (GRU)) to extract features, and there are skip connections between layers.

[0085] The classifier consists of two fully connected layers, which convert the extracted 512×16 feature tensor into a 512×8 output tensor. The output tensor is as follows: Figure 2 As shown, it consists of a positioning tensor and a classification tensor (seven categories in total, six categories of human-machine asynchrony plus a normal category).

[0086] The post-processing part is sent to the alarm forwarding module after processing such as threshold judgment, argmax / argmin, topk, etc.

[0087] The alarm forwarding module receives the corresponding results, processes them as needed, and forwards them to the subsequent system.

[0088] First, a portion of ventilator data was collected and annotated by medical professionals, including experts and respiratory specialists. The data was categorized into six categories of patient-machine synchronization asynchrony and one category of normal asynchrony. The patient-machine synchronization asynchrony was annotated point by point within the 512-byte time series data. The labeled data was used to train the entire recognition model constructed above, optimizing the model parameters using the objective function and backpropagation algorithm described in this example.

[0089] Example 2

[0090] In one or more embodiments, a single-stage identification system for mechanical ventilation patient-ventilator asynchrony is disclosed, comprising:

[0091] A data acquisition module is used to obtain the pressure time series, flow rate time series and volume time series data of the ventilator;

[0092] A feature embedding module is used to convert the pressure time series, flow rate time series and volume time series data into original time series representations through feature embedding to obtain multi-time series embedding of pressure, flow rate and volume;

[0093] A feature extraction module is used to extract features from the multiple time series embeddings to obtain feature maps representing pressure time series, flow rate time series and volume time series;

[0094] The feature classification module is used to obtain an output tensor after passing the feature map through a classifier, and obtain the location and type of human-computer asynchrony based on the output tensor.

[0095] The specific implementation of each of the above modules is the same as that disclosed in Example 1 and will not be described in detail here.

[0096] Example 3

[0097] In one or more embodiments, a terminal device is disclosed, including a server, the server including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the single-stage identification method for mechanical ventilation patient-ventilator asynchrony in Example 1 is implemented. For the sake of brevity, this description is omitted here.

[0098] It should be understood that in this embodiment, the processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), off-the-shelf field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0099] The memory may include a read-only memory and a random access memory, and provides instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0100] During implementation, each step of the above method may be completed by an integrated logic circuit of hardware in a processor or by instructions in the form of software.

[0101] Example 4

[0102] In one or more embodiments, a computer-readable storage medium is disclosed, storing a plurality of instructions suitable for loading and executing the single-stage identification method for mechanical ventilation human-machine asynchrony described in Example 1 by a processor of a terminal device.

[0103] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A single-stage identification method for mechanical ventilation patient-machine asynchrony, characterized by: include: Obtain the pressure time series, flow rate time series and volume time series data of the ventilator; The pressure time series, flow rate time series and volume time series data are converted into original time series representations through feature embedding to obtain multi-time series embedding of pressure, flow rate and volume; After feature extraction, the multi-time series embedding obtains a feature map representing the pressure time series, flow rate time series and volume time series; The feature map is passed through a classifier to obtain an output tensor, and the location and type of human-computer asynchrony are obtained based on the output tensor; The output tensor includes a positioning tensor and a classification tensor; wherein the first element of each row in the output tensor is the positioning tensor, and a positioning tensor with a value of 1 represents the location where human-machine asynchrony occurs; the remaining elements in each row are classification tensors, and the category corresponding to a classification tensor with a value of 1 is the type of human-machine asynchrony; The objective function of the classifier is specifically: in, Where, 、 、 Represent classification items, positioning items and auxiliary items respectively; Represent the weights of classification items, positioning items and auxiliary items respectively; Represents the category of mechanical ventilation patient-ventilator asynchrony, Represents the batch number, Represents the total number of positioning positions, is a onehot encoded tensor, is the network output tensor, with the same dimension as same; is the control parameter of the auxiliary item; is the weight of each category in the classification item; It is the nth batch of prediction in the classification item Output tensor of the class; , They represent the tensors composed of the output tensor of the i-th position of the n-th batch during prediction in the positioning item and the i-th position of the n-th batch of the true label; , They represent the tensors consisting of the nth batch of output tensors and the nth batch of true labels during prediction in the auxiliary items.

2. A single-stage identification method for mechanical ventilation patient-machine asynchrony according to claim 1, characterized in that: Also includes: Post-process the output tensor of the classifier. The post-processing process is: For the positioning tensor, use the topK function to obtain the positions with the highest confidence in the positioning tensor; For the classification tensor, for the several positions with the highest confidence obtained, use the topK function to obtain the category with the highest confidence in the classification tensor.

3. A single-stage identification method for mechanical ventilation patient-machine asynchrony according to claim 2, characterized in that: Also includes: The location and type of human-machine asynchrony after post-processing are sent to the human-machine asynchrony alarm distribution module to be forwarded to the corresponding control unit.

4. A single-stage identification method for mechanical ventilation patient-machine asynchrony according to claim 1, characterized in that: Obtain the pressure time series, flow rate time series, and volume time series data of the ventilator. The specific process is as follows: Obtain the pressure, flow rate and volume data of the ventilator and pre-process the acquired data; The preprocessed data enters a sliding window with adjustable step size and window length. The sliding window slides in time and outputs pressure time series, flow rate time series and volume time series data with a length equal to the window length.

5. A single-stage identification method for mechanical ventilation patient-machine asynchrony according to claim 1, characterized in that: Assume that the sliding window length is , the step size is , the pressure time series, flow rate time series and volume time series are , The feature embedding part assumes that ,The feature extraction part assumes that , then the feature map is calculated as: 。 6. A single-stage recognition system for mechanical ventilation patient-machine asynchrony, characterized by: include: A data acquisition module is used to obtain the pressure time series, flow rate time series and volume time series data of the ventilator; A feature embedding module is used to convert the pressure time series, flow rate time series and volume time series data into original time series representations through feature embedding to obtain multi-time series embedding of pressure, flow rate and volume; A feature extraction module is used to extract features from the multiple time series embeddings to obtain feature maps representing pressure time series, flow rate time series and volume time series; A feature classification module, configured to obtain an output tensor after passing the feature map through a classifier, and obtain the location and type of human-computer asynchrony based on the output tensor; The output tensor includes a positioning tensor and a classification tensor; wherein the first element of each row in the output tensor is the positioning tensor, and a positioning tensor with a value of 1 represents the location where human-machine asynchrony occurs; the remaining elements in each row are classification tensors, and the category corresponding to a classification tensor with a value of 1 is the type of human-machine asynchrony; The objective function of the classifier is specifically: in, Where, 、 、 Represent classification items, positioning items and auxiliary items respectively; Represent the weights of classification items, positioning items and auxiliary items respectively; Represents the category of mechanical ventilation patient-ventilator asynchrony, Represents the batch number, Represents the total number of positioning positions, is a onehot encoded tensor, is the network output tensor, with the same dimension as same; is the control parameter of the auxiliary item; is the weight of each category in the classification item; It is the nth batch of prediction in the classification item Output tensor of the class; , They represent the tensors composed of the output tensor of the i-th position of the n-th batch during prediction in the positioning item and the i-th position of the n-th batch of the true label; , They represent the tensors consisting of the nth batch of output tensors and the nth batch of true labels during prediction in the auxiliary items.

7. A terminal device comprising a processor and a memory, wherein the processor is used to implement various instructions; the memory is used to store multiple instructions, characterized in that: The instructions are suitable for being loaded by a processor and executing the single-stage identification method for mechanical ventilation human-machine asynchrony according to any one of claims 1-5.

8. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing the single-stage identification method for mechanical ventilation human-machine asynchrony according to any one of claims 1-5.

Images