Respirator man-machine asynchronous event intelligent identification method based on self-developed sensor

Abstract

The invention discloses a breathing machine man-machine asynchronous event intelligent identification method based on a self-developed sensor, and the method comprises the steps: S1, integrating multi-channel physiological signals collected by a self-developed body surface sensor and expert labeling data, and constructing a standardized data set with a uniform structure through signal purification, time alignment, window segmentation and category balance processing; and S2, a ResNet-Attention BiLSTM fusion framework is adopted, respiratory signal time sequence features are extracted, and abnormal key segments are focused. According to the method, integration, cleaning, time synchronization and category equalization of heterogeneous physiological signals are realized through a standardized data preprocessing process, core pain points of missing value, abnormal value interference and category imbalance in traditional data processing are solved, and a high-robustness and high-quality data set is provided for model training.

Description

Technical Field

[0001] This invention belongs to the field of respiratory monitoring technology, specifically referring to an intelligent identification method for ventilator-machine asynchrony events based on a self-developed sensor. Background Technology

[0002] Currently, ventilators are the core equipment for respiratory support in critically ill patients, and the timely identification of patient-ventilator asynchrony (PVA) events directly affects treatment outcomes and patient safety. In clinical practice, patient-ventilator asynchrony can lead to risks such as airway damage and decreased ventilation efficiency. However, traditional identification methods heavily rely on medical staff manually interpreting physiological signals, resulting in low efficiency, high subjectivity, and a high rate of missed detections. Existing technologies also face multiple bottlenecks: First, multi-source sensor data exhibits strong heterogeneity, inconsistent timestamps, and significant noise interference, making it difficult to form standardized training data; second, class imbalance is prominent (normal respiratory samples far outnumber asynchrony events), causing models to favor the majority class and have weak ability to identify rare events; third, conventional time-series models struggle to simultaneously capture the forward and backward dynamic dependencies and deep abstract features of respiratory signals, and lack the ability to focus on key diagnostic segments; finally, performance evaluation metrics are singular and cannot fully meet the core clinical needs for "comprehensive detection and accurate interpretation." Therefore, it is urgent to build an integrated framework that goes from data standardization to intelligent identification and then to comprehensive verification, so as to realize the automated and high-precision identification of human-machine asynchronous events and provide reliable support for clinical decision-making. Summary of the Invention

[0003] The purpose of this invention is to overcome the technical shortcomings of existing technologies, such as low automation and accuracy of human-machine asynchrony events, and the lack of comprehensive matching and clinical application of performance evaluation indicators. This invention provides an intelligent identification method for human-machine asynchrony events in ventilators based on self-developed sensors.

[0004] To achieve the above objectives, the present invention employs the following technical solution: an intelligent recognition method for human-machine asynchrony events in ventilators based on self-developed sensors, comprising the following steps:

[0005] S1. Integrate multi-channel physiological signals collected by self-developed body surface sensors with expert-annotated data, and construct a standardized dataset with a unified structure through signal purification, time alignment, window segmentation and category balancing.

[0006] S2. Employ the ResNet-Attention BiLSTM fusion architecture to extract temporal features of respiratory signals and focus on abnormal key segments;

[0007] S3. Based on the weighted loss function and multi-round iterative optimization strategy, the model with the best generalization ability is dynamically selected by combining the macro F1 score;

[0008] S4. The reliability of the model in the identification of human-machine asynchronous events is quantified by using multi-dimensional indicators such as sensitivity, specificity, ROC-AUC and confusion matrix.

[0009] Furthermore, the signal purification in step S1 includes the following: First, linear interpolation is used to fill the gaps inside the signal, and zero values ​​are used to fill the missing parts at both ends; Second, extreme outliers are identified and pruned based on the interquartile range statistical method to reduce noise interference.

[0010] The category balancing process described in step S1 includes: first, performing stratified downsampling on normal breathing samples to preserve their key feature distribution; second, merging the stratified downsampled normal samples with all human-machine asynchronous event samples, and then constructing a training dataset with balanced category distribution after global random shuffling.

[0011] The “window segmentation” mentioned in step S1 refers to taking the four synchronous physiological signals collected by the self-developed body surface sensor, using the unified timestamp as the index, and extracting a segment every 1 minute from the start time of the signal to form a continuous and non-overlapping window sequence.

[0012] The "time alignment" mentioned in step S1 includes unifying the time format of multi-source data and matching the timestamps of signals and tags; wherein, unifying the time format of multi-source data refers to converting the four physiological signals collected by the self-developed body surface sensor and the asynchronous events labeled by experts into a unified standard time format; and matching the timestamps of signals and tags refers to accurately associating the multi-channel physiological signals with the event tags labeled by experts using the unified timestamp as an index.

[0013] The "linear interpolation filling of signal gaps" refers to calculating missing values ​​uniformly over time intervals based on valid data points before and after the missing segment. This includes the following steps: First, by detecting data continuity, identify time intervals in the signal where values ​​are empty or marked as "invalid." Second, obtain the last valid data point A before the missing segment and the first valid data point B after the missing segment. Finally, based on the time length of the missing segment, allocate the numerical changes from the last valid data point A to the first valid data point B in a "linear proportion" to fill the missing positions.

[0014] The "missing portions at both ends" refers to the missing data at the beginning or end of the sensor acquisition. This type of missing data at the beginning and end means that only one side of the missing segment has valid data, and the missing value cannot be calculated by interpolation. The "filling with zero values" specifically means filling all the missing segments at the beginning or end of the signal with 0.

[0015] The "identification and trimming of extreme outliers based on interquartile range statistical method to reduce noise interference" refers to defining outlier boundaries based on interquartile range and trimming extreme values ​​that exceed the boundaries. This includes the following steps: First, calculate the quartiles of the signal. For complete signal data from a certain sensor, sort the data by value and calculate the lower quartile Q1 and the upper quartile Q3. Second, calculate the outlier boundaries. The outlier determination range is (Q1 - 1.5 × IQR) to (Q3 + 1.5 × IQR), where IQR = Q3 - Q1. Finally, trim extreme values. Replace the minimum extreme value in the signal that is less than (Q1 - 1.5 × IQR) with (Q1 - 1.5 × IQR); replace the maximum extreme value that is greater than (Q3 + 1.5 × IQR) with (Q3 + 1.5 × IQR).

[0016] The ResNet-Attention BiLSTM fusion architecture described in step S2 includes at least: a Bi-LSTM bidirectional layer for processing breathing signals to capture forward and backward contextual dependencies; a residual connection layer for passing the original input information through cross-layer shortcuts to alleviate gradient vanishing; and an attention mechanism layer for dynamically assigning high weights to key temporal segments that foreshadow asynchronous events.

[0017] The "strategy" mentioned in step S3 includes: dividing the standardized dataset into a training set and a validation set proportionally, setting 200 rounds of iterative training; and using a weighted cross-entropy loss function to increase the learning weight of rare asynchronous events.

[0018] The "macro F1 score" mentioned in step S3 is a combination of the macro average and the F1 score. The F1 score refers to the harmonic mean of the precision and recall of a single class, calculated by combining the precision and recall of each class. The formula is: F1 = 2 × (precision × recall) / (precision + recall), where precision is the proportion of samples that are actually of a certain class as predicted by the model, and recall is the proportion of samples that are actually of a certain class as correctly predicted by the model. The macro average refers to the arithmetic mean of the F1 scores of all classes, calculated separately for each class in the dataset.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] (1) This invention achieves the integration, cleaning, time synchronization and class balance of heterogeneous physiological signals through a standardized data preprocessing process, which solves the core pain points of missing values, outlier interference and class imbalance in traditional data processing, and provides a highly robust and high-quality dataset for model training.

[0021] (2) The ResNet-Attention BiLSTM model in this invention integrates the contextual understanding ability of bidirectional LSTM, the deep feature preservation ability of residual connections, and the key information focusing ability of attention mechanism. Compared with traditional models, it can more accurately capture the feature patterns of PVA events, especially improving the recognition recall rate of rare events. Combined with the weighted loss function and macro average F1 score screening strategy, the generalization ability of the model is further enhanced, adapting to different clinical scenarios.

[0022] (3) The present invention constructs a multi-dimensional evaluation system covering comprehensive classification performance, misjudgment analysis, key clinical indicators and overall diagnostic capabilities, which not only ensures that PVA events are "not missed" (high sensitivity) but also controls that normal breathing is "not misdiagnosed" (high specificity), effectively reducing the clinical false alarm rate and providing reliable decision-making basis for medical staff.

[0023] (4) This invention realizes full-process automation from data processing and model training to performance evaluation. It can complete the intelligent identification of PVA events without human intervention, greatly reduce the workload of medical staff, improve the efficiency of clinical monitoring, and provide timely support for real-time respiratory therapy adjustment.

[0024] (5) The present invention adopts a modular design, which can be expanded and adjusted according to different sensor types and PVA event categories. The preprocessing process and evaluation system can also be adapted to more physiological signal monitoring scenarios, and have broad clinical promotion value. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example

[0028] like Figure 1 As shown, the intelligent recognition method for ventilator-machine asynchrony events based on self-developed sensors described in this embodiment includes four main steps, S1 to S4, as follows:

[0029] S1. Integrate multi-channel physiological signals collected by self-developed body surface sensors with expert-annotated data, and construct a standardized dataset with a unified structure through signal purification, time alignment, window segmentation and category balancing.

[0030] This step is the data foundation and core prerequisite for the entire intelligent recognition of human-ventilator asynchrony events. It is mainly used to solve the defects of the original data, adapt to the model training requirements, and ensure the reliability of clinical recognition. The signal purification includes the following: First, linear interpolation is used to fill the gaps inside the signal, and the missing parts at both ends are filled with zero values; Second, extreme outliers are identified and pruned based on the interquartile range statistical method to reduce noise interference.

[0031] The core purpose of the aforementioned signal purification is to repair the missing data in the original sensor data, suppress noise interference, ensure the integrity and stability of physiological signals, and provide high-quality data for subsequent model training. The "linear interpolation filling of signal gaps" refers to uniformly calculating missing values ​​based on valid data points before and after the missing segment, at time intervals. This includes the following steps: First, through data continuity detection, identify time intervals in the signal where values ​​are empty or marked as "invalid"; second, obtain the last valid data point A before the missing segment and the first valid data point B after the missing segment; finally, according to the time length of the missing segment, distribute the numerical changes from the last valid data point A to the first valid data point B in a "linear proportion" to fill the missing positions.

[0032] The "missing portions at both ends" refers to the missing data at the beginning or end of the sensor acquisition. This type of missing data at the beginning and end means that only one side of the missing segment has valid data, and the missing value cannot be calculated by interpolation. The "filling with zero values" specifically means filling all the missing segments at the beginning or end of the signal with 0.

[0033] The phrase "identifying and pruning extreme outliers based on interquartile range statistical methods to reduce noise interference" refers to defining outlier boundaries based on interquartile ranges and pruning extreme values ​​that exceed these boundaries. This includes the following steps: First, calculate the quartiles of the signal. For complete signal data from a specific sensor, after sorting by numerical value, calculate the lower quartile Q1 and the upper quartile Q3. The lower quartile Q1 is the 25th quartile, meaning 25% of the data are less than this value; the upper quartile Q3 is the 75th quartile, meaning 75% of the data are less than this value.

[0034] Secondly, the outlier boundary is calculated. The outlier determination range is (Q1-1.5×IQR) to (Q3+1.5×IQR), where IQR=Q3-Q1 (i.e., the fluctuation range of the middle 50% of the data).

[0035] Finally, extreme values ​​are pruned. Extreme minimum values ​​less than (Q1 - 1.5 × IQR) are uniformly replaced with (Q1 - 1.5 × IQR); extreme maximum values ​​greater than (Q3 + 1.5 × IQR) are uniformly replaced with (Q3 + 1.5 × IQR). For example, if a signal has Q1 = 1.2 and Q3 = 2.8, then IQR = 1.6, and the outlier boundaries are (1.2 - 1.5 × 1.6) = -1.2 and (2.8 + 1.5 × 1.6) = 5.2, meaning the outlier boundaries are -1.2 to 5.2. If the signal contains -2.0 (minimum) or 6.0 (maximum), then these values ​​are pruned to -1.2 and 5.2 respectively.

[0036] It should be noted that the above steps are performed through "pruning" rather than "deletion." Pruning eliminates the interference of extreme noise on the signal, preventing the model from learning false features, while maintaining the signal's duration and integrity. Furthermore, defining boundaries using statistical methods ensures the processing is objective and reproducible, improving data stability and model robustness.

[0037] The category balancing process includes two aspects: First, stratified downsampling is performed on normal breathing samples to preserve their key feature distribution; second, the stratified downsampled normal samples are merged with all human-machine asynchronous event samples, and after global random shuffling, a training dataset with balanced category distribution is constructed.

[0038] The "stratified downsampling of normal respiratory samples to preserve their key feature distribution" is an optimization strategy designed to address the class imbalance problem in clinical data where there are far more normal samples than abnormal samples. Its core is to ensure that the key physiological features contained in normal samples are not lost or their distribution is not shifted while reducing the number of normal samples, so as to avoid the model misjudging abnormal events due to the lack of features of normal samples.

[0039] The "stratification" here is not simply divided into a single category of "normal." Instead, all normal breathing samples are first divided into multiple sub-layers (groups) based on key physiological characteristics that characterize the diversity of normal breathing patterns, ensuring that each sub-layer represents a typical normal breathing state. These key physiological characteristics are extracted from multi-channel signals collected by self-developed sensors and mainly include respiratory rate-related features, respiratory waveform features, and temporal pattern features. Among them, the respiratory rate-related features include, but are not limited to, the range of respiratory rates per minute (RR); the respiratory waveform features include, but are not limited to, the peak / trough range of airway pressure waveforms and the stability of respiratory flow (such as the normal fluctuation range of tidal volume Vt); and the temporal pattern features include, but are not limited to, the regularity of the respiratory cycle (no sudden fluctuations, no rhythmic disturbances).

[0040] The term "downsampling" refers to the random cropping of sub-layer samples according to a predetermined ratio without destroying the features within the layer. After stratification, downsampling does not involve randomly deleting all normal samples, but rather randomly selecting samples within each sub-layer according to a preset ratio (such as a target balance ratio) to ensure that the number of samples in each sub-layer is reduced by the same ratio, rather than deleting samples from only one sub-layer. The specific steps are as follows: First, determine the target balance ratio: based on the total number of abnormal samples, calculate the total number of normal samples to be retained (e.g., if there are 1000 abnormal samples, and the target normal:abnormal = 1:1, then 1000 normal samples need to be retained); Second, allocate the sample size of each sub-layer: according to the proportion of each sub-layer in the original normal samples, allocate the number of samples to be retained in each sub-layer (e.g., if there are 10000 original normal samples, with the "standard frequency group" accounting for 60%, the "faster group" accounting for 25%, and the "slower group" accounting for 15%, then 600, 250, and 150 samples need to be retained respectively); Finally, random sampling within the group: in each sub-layer, select the samples to be retained by random sampling and delete redundant samples that exceed the limit.

[0041] The aforementioned "window segmentation" refers to the process of extracting a continuous and non-overlapping window sequence from four synchronous physiological signals collected by a self-developed surface sensor, indexed by a unified timestamp, starting from the signal's initial moment and accumulating a segment every minute. The core of this window segmentation is to transform continuous multi-channel physiological signals into independent sample units with a unified structure and associative diagnostic labels.

[0042] The original data segmented here consists of four synchronous physiological signals collected by a self-developed body surface sensor, such as respiratory flow, airway pressure, chest movement displacement, and abdominal movement signals. These signals have already achieved preliminary synchronization in the time dimension during acquisition (i.e., the instantaneous values ​​of the four signals at the same time point). During segmentation, the four signals need to be treated as a whole to ensure that the single window contains the complete multi-channel features within a certain 1 minute, and to avoid signal misalignment between channels.

[0043] This "window segmentation" is a prerequisite for the aforementioned "time alignment." The "time alignment" includes unifying the time format of multi-source data and matching the timestamps of signals with labels. Unifying the time format of multi-source data refers to converting the four physiological signals collected by the self-developed body surface sensor and the asynchronous events annotated by experts into a unified standard time format. Matching the timestamps of signals with labels refers to accurately associating multi-channel physiological signals with expert-annotated event labels using the unified timestamp as an index.

[0044] S2. Employing a ResNet-Attention BiLSTM fusion architecture, the temporal features of respiratory signals are extracted and key abnormal segments are focused. This step integrates bidirectional temporal capture, deep feature abstraction, and key segment focusing capabilities through the ResNet-Attention BiLSTM fusion architecture, achieving accurate extraction of respiratory signal temporal features and efficient localization of abnormal information.

[0045] The ResNet-Attention BiLSTM fusion architecture has at least a Bi-LSTM bidirectional layer for processing breathing signals to capture forward and backward contextual dependencies; a residual connection layer for passing original input information through cross-layer shortcuts to alleviate gradient vanishing; and an attention mechanism layer for dynamically assigning high weights to key temporal segments that foreshadow asynchronous events.

[0046] S3. Based on a weighted loss function and a multi-round iterative optimization strategy, the model with the best generalization ability is dynamically selected using macro F1 scores. This step, through a combination of "weighted loss function, multi-round iterative optimization, and dynamic selection using macro F1 scores," designs a training optimization system to address the three core needs of training on imbalanced data, ensuring sufficient model learning, and guaranteeing generalization ability. Ultimately, this serves to construct the optimal model for clinical use.

[0047] The purpose of the weighted loss function is to assign high weights to "asynchronous events" (minority class) and low weights to "normal breathing" (majority class), so that when the model makes a prediction error, it will impose a higher loss penalty on "missed / misjudged asynchronous events", forcing the model to prioritize learning minority class features during training and avoid recognition bias caused by class dominance.

[0048] The multi-round iterative optimization strategy includes two aspects: dividing the standardized dataset into training and validation sets proportionally, setting 200 rounds of iterative training, and using a weighted cross-entropy loss function to enhance the learning weights of rare asynchronous events. Through a cycle of "prediction per round → loss calculation → error backpropagation → parameter update," the model gradually converges from "initially learning normal patterns" to "accurately identifying rare anomalies," avoiding "underfitting" (the model not fully learning the features, resulting in low accuracy in identifying new data) caused by insufficient iterations.

[0049] The "macro F1 score" is a combination of the macro average and the F1 score. Its purpose is to use the "macro F1 score" (treating all classes equally, integrating precision and recall) as the core evaluation metric. After each training round, the macro F1 score is calculated using an independent validation set. Only models with the highest macro F1 score are saved, ensuring that the final selected model not only performs well on the training set but also stably identifies all classes (including rare anomalies) on "unseen validation data," thus avoiding overfitting during evaluation and selection. Simultaneously, the "class fairness" metric ensures the model's ability to identify minority classes, guaranteeing that the selected model has the generalization ability to "work stably on new data."

[0050] The F1 score refers to the harmonic mean of precision and recall for each class, calculated by combining the precision and recall of each class. The formula is: F1 = 2 × (precision × recall) / (precision + recall). Precision is the proportion of samples that are actually of a certain class as predicted by the model; recall is the proportion of samples that are actually of a certain class as correctly predicted by the model. The macro-average refers to the arithmetic mean of the F1 scores of all classes, calculated separately for each class in the dataset.

[0051] S4. The reliability of the model in the identification of human-machine asynchronous events is quantified by using multi-dimensional indicators such as sensitivity, specificity, ROC-AUC and confusion matrix.

[0052] This step is the core method of the model performance evaluation stage in this embodiment. The confusion matrix is ​​a visualization error analysis tool that displays the model's prediction results for four types of events, namely "human-machine asynchrony events" and "normal breathing," in matrix form. It directly presents the source of the model's misjudgment and provides direction for subsequent optimization.

[0053] Sensitivity is a core metric for focusing on asynchronous events. It quantifies the proportion of all real asynchronous events correctly identified by the model, directly impacting the clinical need to avoid missing key abnormalities. The calculation is: Sensitivity = True Positive (TP) / [True Positive (TP) + False Negative (FN)]. A higher sensitivity means a lower probability of the model missing asynchronous events. For example, a sensitivity of 95% means that out of 100 real asynchronous events, 95 are correctly detected, with only 5 missed.

[0054] Specificity, a complementary indicator focusing on "normal breathing," quantifies the proportion of truly normal breaths correctly identified by the model, directly impacting the clinical need to avoid unnecessary interventions. It is calculated as: Specificity = True Negative (TN) / [True Negative (TN) + False Positive (FP)]. The higher the specificity, the lower the probability of the model misclassifying normal breathing as abnormal. For example, a specificity of 92% means that out of 100 truly normal breaths, 92 are correctly identified as normal, with only 8 false positives.

[0055] The ROC-AUC is used to comprehensively evaluate the model's overall ability to distinguish asynchronous events from normal breathing, reflecting generalization reliability. The ROC curve (Receiver Operating Characteristic curve) and AUC value (Area Under the Curve) are the gold standard for measuring the model's overall discriminative performance. By traversing all possible classification thresholds, they demonstrate the dynamic relationship between "sensitivity" and "specificity," while AUC quantifies the area under the curve (ranging from 0 to 1), comprehensively reflecting the model's discriminative ability.

[0056] As described above, the present invention can be well implemented.

Claims

1. A method for intelligent recognition of human-machine asynchrony events in ventilators based on self-developed sensors, characterized in that, Includes the following steps: S1. Integrate multi-channel physiological signals collected by self-developed body surface sensors with expert-annotated data, and construct a standardized dataset with a unified structure through signal purification, time alignment, window segmentation and category balancing. S2. Employ the ResNet-Attention BiLSTM fusion architecture to extract temporal features of respiratory signals and focus on abnormal key segments; S3. Based on the weighted loss function and multi-round iterative optimization strategy, the model with the best generalization ability is dynamically selected by combining the macro F1 score; S4. The reliability of the model in the identification of human-machine asynchronous events is quantified by using multi-dimensional indicators such as sensitivity, specificity, ROC-AUC and confusion matrix.

2. The intelligent recognition method for ventilator human-machine asynchrony events based on self-developed sensors according to claim 1, characterized in that, The signal purification described in step S1 includes the following: First, linear interpolation is used to fill the gaps inside the signal, and zero values ​​are used to fill the missing parts at both ends. Second, extreme outliers are identified and pruned based on the interquartile range statistical method to reduce noise interference.

3. The intelligent recognition method for ventilator-machine asynchrony events based on self-developed sensors according to claim 1, characterized in that, The category balancing process described in step S1 includes: first, performing stratified downsampling on normal breathing samples to preserve their key feature distribution; second, merging the stratified downsampled normal samples with all human-machine asynchronous event samples, and then constructing a training dataset with balanced category distribution after global random shuffling.

4. The intelligent identification method for ventilator-machine asynchrony events based on self-developed sensors according to claim 1, characterized in that, The "window segmentation" mentioned in step S1 refers to taking the four synchronous physiological signals collected by the self-developed body surface sensor, using the unified timestamp as the index, and extracting a segment every 1 minute from the start time of the signal to form a continuous and non-overlapping window sequence.

5. The intelligent identification method for ventilator-machine asynchrony events based on self-developed sensors according to claim 1, characterized in that, The "time alignment" mentioned in step S1 includes unifying the time format of multi-source data and matching the timestamps of signals and tags; wherein, unifying the time format of multi-source data refers to converting the four physiological signals collected by the self-developed body surface sensor and the asynchronous events labeled by experts into a unified standard time format; and matching the timestamps of signals and tags refers to accurately associating the multi-channel physiological signals with the event tags labeled by experts using the unified timestamp as an index.

6. The intelligent identification method for ventilator-machine asynchrony events based on self-developed sensors according to claim 2, characterized in that, The "linear interpolation filling of signal gaps" refers to calculating missing values ​​uniformly over time intervals based on valid data points before and after the missing segment. This includes the following steps: First, by detecting data continuity, identify time intervals in the signal where values ​​are empty or marked as "invalid." Second, obtain the last valid data point A before the missing segment and the first valid data point B after the missing segment. Finally, based on the time length of the missing segment, allocate the numerical changes from the last valid data point A to the first valid data point B in a "linear proportion" to fill the missing positions.

7. The intelligent recognition method for ventilator human-machine asynchrony events based on self-developed sensors according to claim 6, characterized in that, The "missing portions at both ends" refers to the missing data at the beginning or end of the sensor acquisition. This type of missing data at the beginning and end means that only one side of the missing segment has valid data, and the missing value cannot be calculated by interpolation. The "filling with zero values" specifically means filling all the missing segments at the beginning or end of the signal with 0.

8. The intelligent recognition method for ventilator-machine asynchrony events based on self-developed sensors according to claim 7, characterized in that, The "identification and trimming of extreme outliers based on interquartile range statistical method to reduce noise interference" refers to defining outlier boundaries based on interquartile range and trimming extreme values ​​that exceed the boundaries. This includes the following steps: First, calculate the quartiles of the signal. For complete signal data from a certain sensor, sort the data by value and calculate the lower quartile Q1 and the upper quartile Q3. Second, calculate the outlier boundaries. The range for determining outliers is (Q1 - 1.5 × IQR) to (Q3 + 1.5 × IQR), where IQR = Q3 - Q1. Finally, trim extreme values. Replace the minimum extreme value in the signal that is less than (Q1 - 1.5 × IQR) with (Q1 - 1.5 × IQR); replace the maximum extreme value that is greater than (Q3 + 1.5 × IQR) with (Q3 + 1.5 × IQR).

9. A method for intelligent identification of human-machine asynchrony events in ventilators based on self-developed sensors according to any one of claims 1 to 8, characterized in that, The ResNet-Attention BiLSTM fusion architecture described in step S2 includes at least: a Bi-LSTM bidirectional layer for processing breathing signals to capture forward and backward contextual dependencies; a residual connection layer for passing the original input information through cross-layer shortcuts to alleviate gradient vanishing; and an attention mechanism layer for dynamically assigning high weights to key temporal segments that foreshadow asynchronous events.

10. The intelligent identification method for ventilator-machine asynchrony events based on self-developed sensors according to claim 9, characterized in that, The "strategy" mentioned in step S3 includes: dividing the standardized dataset into a training set and a validation set proportionally, setting 200 rounds of iterative training; and using a weighted cross-entropy loss function to increase the learning weight of rare asynchronous events. The "macro F1 score" mentioned in step S3 is a combination of the macro average and the F1 score. The F1 score refers to the harmonic mean of the precision and recall of a single class, calculated by combining the precision and recall of the two classes. The formula is: F1 = 2 × (precision × recall) / (precision + recall), where precision is the proportion of samples that are actually of a certain class as predicted by the model, and recall is the proportion of samples that are actually of a certain class as correctly predicted by the model. The macro average refers to the arithmetic mean of the F1 scores of all classes, calculated separately for each class in the dataset.

Images