A method for monitoring and early warning of emergency transport ventilators and its application

Abstract

This invention discloses a method and application for monitoring and early warning of ventilators during emergency transport, comprising: S1: collecting ambulance bump data and airway pressure time-series data and preprocessing them; S2: extracting bump interference features and airway pressure features respectively; calculating airway pressure features after interference suppression; calculating pure airway pressure features; S3: calculating bump time-series features, and then completing the classification through a bump intensity differentiation mechanism to obtain bump patterns; calculating cross-modal fusion features to generate initial calibration parameters and final calibration parameters; S4: calculating calibrated airway pressure time-series data; and then combining the pure airway pressure features to calculate ventilation trend abnormality markers and the proportion of numerical abnormality, and generating graded early warning levels; S5: graded early warning and visualization display. This invention can solve the problem of poor monitoring and early warning accuracy in traditional ventilator monitoring and early warning methods due to insufficient consideration of vehicle bumps during transport.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of respirator technology, in particular to an emergency transport respirator monitoring and early warning method and application. BACKGROUND

[0002] The emergency transport respirator early warning technology is a key support for real-time monitoring of ventilation safety and avoiding risks such as insufficient or excessive ventilation in advance during the transport of critical patients. The core lies in accurately capturing the abnormality of key parameters such as airway pressure and the change of patient's respiratory state to provide timely early warning feedback for medical personnel, and the effectiveness of parameter calibration directly determines the reliability of the early warning result.

[0003] Traditional early warning methods are mostly based on preset fixed thresholds or simple signal filtering algorithms, and the early warning is realized by judging whether the parameters exceed the set range, but the parameter data relied on is easily disturbed by the jolting of the ambulance and fluctuates, and the traditional calibration method has a lagging response and lacks self-adaptive ability to different jolting intensities, resulting in distorted parameter data directly affecting the accuracy of the early warning threshold judgment, and further causing false negatives and false positives.

[0004] The existing related schemes that apply deep learning to medical device early warning mostly focus on signal feature extraction and abnormality identification in static scenes, do not fully consider the relationship between the dynamic time sequence characteristics of jolting and the parameter deviation in emergency transport, and do not design a synchronous calibration mechanism for jolting interference, resulting in a lagging nature and insufficient self-adaptation of parameter calibration, which also becomes a difficulty in restricting the accuracy of early warning, and it is difficult to meet the demand for real-time and reliability of early warning in the emergency transport scene. SUMMARY

[0005] Therefore, the present application aims to provide an emergency transport respirator monitoring and early warning method and application to solve the problem of poor monitoring and early warning accuracy in traditional respirator monitoring and early warning methods due to insufficient consideration of vehicle jolting during transport.

[0006] An emergency transport respirator monitoring and early warning method, comprising: S1: collecting ambulance jolting state data and airway pressure time series data and preprocessing to obtain preprocessed jolting state time series data and preprocessed airway pressure time series data; S2: constructing a double-path convolutional neural network according to the preprocessed jolting state time series data and the preprocessed airway pressure time series data, respectively extracting jolting interference features and airway pressure features; then calculating the airway pressure features after interference suppression through a jolting-pressure double-path feature decoupling gating mechanism; and finally calculating the pure airway pressure features through a pressure feature enhancement residual mechanism; The turbulence-pressure dual-path feature decoupling gating mechanism processes the spliced ​​turbulence interference features and airway pressure features through a multilayer perceptron, calculates the weight coefficient of the turbulence interference features using the Sigmoid function, and then uses the turbulence physical constraint matrix and the pressure physiological constraint matrix to constrain the turbulence interference features and airway pressure features respectively, to obtain the airway pressure features after interference suppression. S3: Based on the preprocessed turbulence time series data, calculate the turbulence time series characteristics, and then complete the classification through the turbulence intensity differentiation mechanism to obtain the turbulence mode; combined with the pure airway pressure characteristics, calculate the cross-modal fusion characteristics to generate the initial calibration parameters and the final calibration parameters. S4: Correct the preprocessed airway pressure data according to the final calibration parameters to obtain the calibrated airway pressure time series data; then combine the pure airway pressure characteristics to calculate the ventilation trend abnormality markers, the percentage of numerical abnormality, and generate a graded early warning level. S5: Provide tiered early warnings and visual displays based on the tiered warning levels.

[0007] Furthermore, step S1 also includes: S11: The ambulance bump state data is collected through MEMS gyroscope and MEMS accelerometer. The data type is multi-dimensional time series data, including angular velocity, angular acceleration and linear acceleration. The data is processed by three standard deviation method and low-pass filtering method for noise reduction, and then the data is normalized to obtain the preprocessed bump state time series data. S12: Collect time-series airway pressure data via ventilator. The data type is one-dimensional time-series data, including peak inspiratory pressure, positive end-expiratory pressure, mean airway pressure, and airway fluctuation pressure. Perform sliding window smoothing and noise reduction, linear interpolation to complete missing values, and then perform data normalization to obtain preprocessed time-series airway pressure data.

[0008] Furthermore, in step S2, the pressure feature enhancement residual mechanism processes the airway pressure features after interference suppression using convolutional residual blocks to obtain the pressure features after residual processing. The pressure features after residual processing are enhanced by combining feature enhancement coefficients, and residual interference is suppressed by using the Sigmoid function and residual interference suppression coefficient. Finally, the pure airway pressure features are obtained by Hadamard product and global average pooling operation.

[0009] Furthermore, step S2 also includes: S21: Based on the preprocessed turbulence state time-series data and the preprocessed airway pressure time-series data, a dual-path convolutional neural network is constructed to extract turbulence interference features and airway pressure features, respectively. The calculation method is as follows: ; ; in, Characterized by turbulence and interference. It is a convolutional layer. To add element by element, For batch normalization, This is the preprocessed time series data of the turbulence state. Characteristics of airway pressure. This is the preprocessed airway pressure time series data. For Hadama accumulation, For ReLU functions; S22: Based on the characteristics of turbulence interference and airway pressure, the airway pressure characteristics after interference suppression are calculated through a turbulence-pressure dual-path feature decoupling gating mechanism. The calculation method is as follows: ; ; ; in, The weighting coefficients for turbulence interference characteristics. For the Sigmoid function, It is a multilayer perceptron. For splicing operations, The characteristics of the turbulence and interference after stripping. The bump constraint coefficient is... For the Laplace operator, For the physical constraint matrix of the turbulence, To obtain the maximum value, The physiological constraint matrix of stress. The airway pressure characteristics after interference suppression; S23: Based on the airway pressure characteristics after interference suppression, a residual enhancement mechanism is used to strengthen effective characteristics and suppress residual interference, calculating the pure airway pressure characteristics. The calculation method is as follows: ; ; in, The pressure characteristics after residual processing. For convolution residual blocks, For feature enhancement coefficients, This represents the characteristics of a pure airway pressure. The residual interference suppression coefficient is... This is global average pooling.

[0010] It should be further explained that continuous or sudden bumps may occur during ambulance transport. These bumps can cause the pre-treatment airway pressure time series data and the pre-treatment bump state time series data to overlap and become confused, making it impossible to accurately distinguish the patient's true airway pressure signal from the interference components caused by the bumps. This confusion will directly affect subsequent ventilation monitoring: the bump interference may be misjudged as ventilation abnormality, leading to ineffective treatment by medical staff; or the true airway pressure abnormality may be masked by the bump interference and fail to be detected in time, ultimately threatening the patient's ventilation safety during transport. To address the aforementioned issues, this invention constructs a turbulence-pressure dual-path feature decoupling gating mechanism. Through multi-stage collaborative operation, it achieves precise separation of coupled features: First, by constructing a dual-path convolutional neural network, independent feature extraction is performed on the preprocessed turbulence state time-series data and the preprocessed airway pressure time-series data, resulting in separate turbulence interference features and airway pressure features—avoiding the problem of the two types of features being mixed together in single-path processing, and providing clear basic data for subsequent decoupling; Next, the two types of features are integrated through a concatenation operation and input into a multilayer perceptron. The turbulence interference feature weight coefficient is calculated using the Sigmoid function and dynamically adjusted according to changes in real-time turbulence intensity and airway pressure signals. For example, when the turbulence is strong, the weight of the interference feature is increased. The weighting is reduced when the turbulence is weak to ensure the decoupling strategy is adaptable to different transport scenarios. Subsequently, the turbulence physical constraint matrix is ​​used in conjunction with the Laplace operator to process the stripped turbulence interference features, preventing the separated interference features from deviating from the actual turbulence situation and becoming distorted. At the same time, the airway pressure features are first normalized to their maximum value, and then processed in conjunction with the pressure physiological constraint matrix to ensure that the retained airway pressure features conform to the normal range of human airway pressure and do not exceed the physiologically reasonable value. Throughout the process, dual-path extraction ensures feature independence, dynamic weighting ensures decoupling adaptability, and constraint matrix ensures data authenticity. This achieves effective separation of turbulence interference and airway pressure features while avoiding the loss of effective features or interference residue, ultimately obtaining the airway pressure features after interference suppression. Existing technologies for handling airway pressure data interference from turbulence often employ single-path filtering (such as low-pass filtering or wavelet filtering) or fixed threshold separation to uniformly filter mixed airway pressure signals containing interference, making it difficult to distinguish between effective components and interference in the signal. Alternatively, a fixed threshold is set, and all data exceeding the threshold is treated as interference and removed, leading to the accidental deletion of effective airway pressure features with high amplitude, or the threshold being too low to completely remove turbulence interference, resulting in residual interference. In contrast, the turbulence-pressure dual-path feature decoupling gating mechanism of this invention has significant advantages over existing technologies: on the one hand, it achieves independent extraction of the two types of features through a dual-path convolutional neural network, reducing confusion at the processing end; on the other hand, the dynamically generated turbulence interference feature weight coefficients can adapt to different turbulence intensities, and combined with a dedicated turbulence physical constraint matrix and pressure physiological constraint matrix, it achieves decoupling of dynamic adjustment and scene constraints, avoiding both over-processing and loss of effective features, and under-processing and residual interference. The decoupling effect and scene adaptability are superior to existing technologies. In terms of integration and fusion, this invention overcomes the difficulty of isolated feature extraction and constraint optimization in existing technologies. Existing technologies typically separate feature extraction and interference suppression into two independent steps. For example, a single network is used to extract mixed airway pressure features containing interference, and then the mixed features are constrained separately. This approach leads to poor connection between the two steps, easily resulting in excessive interference after feature extraction, difficulty in subsequent constraint correction, or constraints that deviate from the actual situation of the features in the early stages. In contrast, this invention deeply integrates three core steps: dual-path feature extraction, dynamic weight calculation, and domain constraints (bump physical constraints and pressure physiological constraints). The features extracted from the dual-path are directly used as input for weight calculation, and the weight calculation results directly guide the application of the constraint matrix. The three steps form a closed-loop linkage, rather than a simple superposition of technologies. Each step serves decoupling, breaking the limitations of the decentralized processing of each step in traditional technologies. This achieves more efficient and accurate separation of coupled features, providing reliable support for obtaining pure airway pressure features through pressure feature enhancement residual mechanisms.

[0011] Furthermore, step S3 also includes: The turbulence intensity differentiation mechanism is as follows: based on the preprocessed turbulence state time series data, turbulence time series features are calculated through a long short-term memory network, and pattern matching weights are calculated based on dynamic time warping similarity; peak values ​​of turbulence time series features are extracted based on a sliding window peak detection algorithm, and then the ratio is calculated with the mean of turbulence intensity in the preprocessed turbulence state time series data, multiplied by the pattern matching weight, and rounded to obtain the turbulence pattern.

[0012] Furthermore, step S3 also includes: S31: Based on the preprocessed turbulence time series data, calculate the turbulence time series characteristics, and then complete the classification through the turbulence intensity differentiation mechanism to obtain the turbulence mode. The calculation method is as follows: ; ; ; ; in, The characteristics of the turbulent timing are as follows: For Long Short-Term Memory (LSTM) networks, This is the preprocessed time series data of the turbulence state. For dynamic time-warped similarity, The average cluster center for each turbulence pattern, To take the absolute value, Weights for pattern matching For the Sigmoid function, Tumble mode, This is a rounding operation. This is a peak detection algorithm based on a sliding window. This represents the mean value of the turbulence intensity in the preprocessed turbulence state time series data; S32: The pure airway pressure characteristics and turbulence timing characteristics are fused using an attention mechanism to obtain cross-modal fusion features. The calculation method is as follows: ; ; ; in, For cross-modal splicing features, For splicing operations, This represents the characteristics of a pure airway pressure. For attention weights, For the Softmax function, It is a multilayer perceptron. For cross-modal fusion features, For Hadama accumulation, For element-wise addition, For layer normalization; S33: Based on the turbulence mode and cross-modal fusion features, initial calibration parameters are generated through a dynamic calibration module. The calculation method is as follows: ; ; ; in, This is the weight vector for the turbulence mode. For embedding layer, For feature-parameter mapping features, For Transformer encoder operations, For calibration parameter bias terms, These are the initial calibration parameters; S34: Perform constraint optimization on the initial calibration parameters to obtain the final calibration parameters, calculated as follows: ; ; in, For calibration parameters after safety constraints, To obtain the minimum value, To obtain the maximum value, This is the lower limit of the calibration parameter. To calibrate the upper limit of the parameter, This is a physiological adaptation constraint compensation term. This is a matrix representing the physiological adaptation range of airway pressure. These are the final calibration parameters.

[0013] It should be further explained that during ambulance transport, the intensity of bumps changes dynamically and in various patterns—there are both continuous mild vibrations and sudden severe bumps. The degree of interference with ventilator airway pressure data varies significantly depending on the intensity and pattern of the bumps. If the intensity of the bumps cannot be accurately classified, the subsequent calibration parameters will not be suitable for the real-time bump scenario: if a high-intensity calibration strategy is used during mild bumps, it will lead to over-correction of airway pressure; if the calibration strategy is insufficient during severe bumps, it will not be able to effectively counteract the interference. Ultimately, both will affect the stability of ventilator ventilation parameters and threaten the ventilation safety of patients during transport. In terms of technical logic, firstly, this invention employs a Long Short-Term Memory (LSTM) network to extract turbulence temporal features from preprocessed turbulence state time-series data. Because turbulence data exhibits significant temporal dependencies (such as the intensity correlation between consecutive turbulence events), the LTM network effectively captures these temporal relationships, avoiding the problem of traditional feature extraction methods focusing only on data from a single moment while ignoring the overall turbulence trend. This provides comprehensive temporal feature support for subsequent grading. Secondly, the matching degree between real-time turbulence temporal features and the average cluster centers of each turbulence pattern is calculated using dynamic time warping similarity. This is then combined with a Sigmoid function to generate pattern matching weights, thereby distinguishing the current turbulence pattern (such as continuous vibration, sudden turbulence, etc.). The LSTM temporal feature extraction ensures the completeness of the features, while the dynamic time warping similarity and pattern matching weights ensure the adaptability of the patterns. This allows the generated turbulence patterns to accurately reflect the real-time turbulence intensity and match different turbulence characteristics, providing a precise grade basis for the subsequent dynamic generation of calibration parameters. The LSTM temporal feature extraction ensures the completeness of the features, while the dynamic time warping similarity and pattern matching weights ensure pattern adaptability. This ensures that the generated turbulence patterns accurately reflect the real-time turbulence intensity and match different turbulence characteristics, providing a precise grade basis for the subsequent dynamic generation of calibration parameters. Existing technologies for handling bump intensity often employ either single-indicator quantification (e.g., based solely on peak acceleration) or fixed-threshold grading. The former only reflects the bump intensity at a specific moment, failing to capture the temporal trends and pattern differences. For example, at the same peak value, the impact of continuous bumps differs drastically from that of sudden bumps, which a single peak indicator cannot distinguish. The latter, by setting fixed thresholds, categorizes bumps into a few levels (e.g., light, moderate, and heavy). However, these thresholds cannot be dynamically adjusted according to the actual transport scenario. For instance, in the different bumpy environments of urban and rural roads, fixed thresholds can lead to grading deviations. This leads to insufficient adaptability of subsequent calibration parameters. In contrast, the turbulence pattern generation scheme of this invention has significant advantages: on the one hand, by extracting temporal features through LSTM and combining dynamic time warping similarity to distinguish turbulence patterns, it breaks through the limitations of existing technologies that only focus on a single indicator or fixed pattern, making the classification results more consistent with the temporal characteristics and pattern differences of actual turbulence; on the other hand, by realizing intensity quantification through the ratio of peak value to mean value and combining it with pattern matching weight adjustment, the classification results can not only accurately reflect the intensity level, but also adapt to the interference requirements of different patterns. The classification accuracy and scene adaptability are better than existing technologies.

[0014] It should also be noted that both the turbulence pattern and the patient's airway pressure status are dynamic during transport. Fluctuations in turbulence intensity will cause the degree of interference with airway pressure data to change in real time. Furthermore, the patient's airway pressure (such as peak inspiratory pressure and positive end-expiratory pressure) will vary due to individual differences in the patient's condition. If the generated calibration parameters cannot simultaneously adapt to real-time turbulence interference and the patient's physiological characteristics, it is easy for the calibration parameters to fail to meet the patient's physiological needs for airway pressure, resulting in over-correction or under-correction, which will ultimately affect ventilation and patient safety. To address the aforementioned issues, this invention designs final calibration parameters. First, in the cross-modal feature fusion stage, the pure airway pressure features and turbulence timing features are concatenated. Then, attention weights are generated using a multilayer perceptron and a Softmax function. Finally, layer normalization is combined to obtain the cross-modal fused features. The core value of this process lies in dynamically allocating the weight ratios of the two types of features: when turbulence interference is strong, the attention weights will increase the contribution of the turbulence timing features, ensuring that subsequent calibration parameters can effectively counteract the impact of turbulence; when the patient's airway pressure shows minor abnormalities, the weights will focus on the pure airway pressure features, allowing the calibration parameters to take into account abnormality correction and avoiding calibration bias caused by a single feature input; at the same time, layer normalization ensures the numerical stability of the fused features, providing high-quality input for subsequent parameter generation. Secondly, in the initial calibration parameter generation stage, the turbulence level is transformed into a turbulence mode weight vector through an embedding layer. At the same time, the Transformer encoder is used to capture the global correlation of cross-modal fusion features (such as the overall adaptation relationship between turbulence mode and pressure anomaly), and the Long Short-Term Memory network is used to capture the temporal dependence of features (such as the temporal coherence of calibration parameters under continuous turbulence). The outputs of the two are superimposed to form a feature-parameter mapping feature, which is then subjected to a Hadamard product operation with the turbulence mode weight vector and the calibration parameter bias term is added to obtain the initial calibration parameters. This stage, through the collaboration of the Transformer encoder and the Long Short-Term Memory network, not only ensures the adaptability of the calibration parameters to the global scene, but also avoids parameter fluctuations in the temporal dimension. Meanwhile, the turbulence mode weight vector ensures that the parameters match the current turbulence mode. Finally, in the constraint optimization stage, on the one hand, by taking the minimum and maximum values, the initial calibration parameters are limited to the preset lower and upper limits of the calibration parameters, forming equipment constraints to ensure that the parameters do not exceed the operating range of the ventilator hardware; on the other hand, after normalizing the pure airway pressure characteristics using the Sigmoid function, a Hadamard product is performed with the airway pressure physiological adaptation range matrix to obtain the physiological adaptation constraint compensation term. This allows the calibration parameters to be adjusted according to the patient's actual airway pressure characteristics. For example, for patients with low positive end-expiratory pressure, the compensation term will appropriately increase the calibration parameters to meet physiological needs; finally, the parameters after equipment constraints are superimposed with the physiological adaptation compensation term to obtain the final calibration parameters. Existing technologies for generating ventilator calibration parameters often employ fixed parameter presets or single-factor driven methods (such as those based solely on turbulence intensity). Fixed parameter preset schemes set calibration parameters to fixed values, failing to adjust according to changes in turbulence intensity or patient physiological differences, resulting in poor correction performance in complex transport scenarios. Single-factor driven schemes generate parameters based solely on turbulence intensity, lacking consideration for the physiological characteristics of patient airway pressure, easily leading to calibration parameters exceeding physiologically reasonable ranges or ignoring individual patient differences, resulting in insufficient correction. In contrast, the final calibration parameters of this invention, through cross-modal feature fusion, simultaneously incorporate pure airway pressure characteristics and turbulence temporal characteristics, overcoming the limitations of single-factor driven technologies in existing technologies. This allows parameter generation to balance interference cancellation and physiological adaptation. Furthermore, the use of a dynamic calibration module that combines Transformer and LSTM captures global scenario correlations while ensuring temporal stability, avoiding parameter fluctuations caused by existing technologies focusing only on local data. This aligns with patient physiological needs and solves the safety or adaptability issues caused by single constraints in existing technologies.

[0015] Furthermore, step S4 also includes: S41: Correct the preprocessed airway pressure data according to the final calibration parameters to obtain the calibrated airway pressure time series data. The calculation method is as follows: ; in, This is the calibrated airway pressure time series data. This is the preprocessed airway pressure time series data. The physiological constraint matrix for stress; S42: Based on the characteristics of pure airway pressure and calibrated airway pressure data, determine abnormal respiratory states and obtain abnormal ventilation trend markers and the percentage of numerical abnormalities. The calculation method is as follows: ; ; in, Mark abnormal ventilation trends. To select the category with the highest probability, The percentage of numerical anomalies. This is the normal reference value for airway pressure; S43: Based on the abnormal ventilation trend markers and the proportion of abnormal numerical values, a graded early warning level is generated. The calculation method is as follows: in, It is classified into different warning levels.

[0016] It should be further explained that the transport process needs to quickly and accurately identify ventilation abnormalities. It is necessary to avoid false alarms (such as judging instantaneous data fluctuations as abnormalities, interfering with the emergency operation of medical staff) and also to prevent missed alarms (such as ignoring trend abnormalities, delaying the timing of ventilation intervention). Bumps and disturbances can easily lead to distortion of airway pressure data, and a single judgment dimension is difficult to take into account both the accuracy and timeliness of the warning. This invention designs a graded early warning system. First, the preprocessed airway pressure data is corrected using the final calibration parameters to obtain calibrated airway pressure time-series data, ensuring that the pressure data used for judgment truly reflects the patient's ventilation status and avoiding judgment bias caused by distortion of the original data. Next, ventilation trend abnormality markers are generated based on the characteristics of pure airway pressure, and the percentage of numerical abnormality is calculated based on the calibrated airway pressure data, forming a trend-numerical dual judgment dimension. Finally, a graded early warning level is generated based on the combination of the dual-dimensional results: Level 0 (no warning) is when the trend is normal and the numerical deviation is less than 10%; Level 1 (mild warning) is when the trend is normal but the numerical deviation is 10%-30% or the trend is abnormal but the numerical deviation is less than 10%; and Level 2 (severe warning) is when the trend is abnormal and the numerical deviation is ≥30%. The complementary dual dimensions ensure that the early warning does not miss trend risks and is not interfered with by instantaneous numerical fluctuations. Existing technologies for ventilation anomaly warnings often rely on a single dimension, or simply on whether the pressure value exceeds the standard, which is prone to false alarms due to instantaneous fluctuations; or they only focus on whether the ventilation trend is disordered, making it difficult to quantify the degree of anomaly and resulting in a coarse warning classification. This invention combines two dimensions, using trend markers to capture overall risk and using numerical proportions to quantify the degree of anomaly. Compared with existing technologies, this significantly reduces the probability of false alarms and missed alarms, and the warning classification is more refined, which can match the intervention needs of different levels of urgency.

[0017] Furthermore, step S5 also includes: S51: Add timestamps and patient identifiers to the graded early warning signal to obtain a graded early warning signal with associated information; S52: Encapsulate the hierarchical early warning signal with associated information using a data transmission protocol, convert it into a standard data format compatible with the emergency transport early warning terminal and the ventilator control module, and obtain standardized early warning execution data; S53: Standardized early warning execution data are transmitted to the ventilator parameter control module to realize the visualization and push of graded early warning information.

[0018] Furthermore, the above-mentioned emergency transport ventilator monitoring and early warning methods are applied to emergency transport ventilator monitoring and early warning, including: Step 1: Collect ambulance bump data and airway pressure time series data and preprocess them to obtain preprocessed bump data and preprocessed airway pressure time series data. Step 2: Based on the preprocessed turbulence state time series data and the preprocessed airway pressure time series data, a dual-path convolutional neural network is constructed to extract turbulence interference features and airway pressure features respectively; then, through the turbulence-pressure dual-path feature decoupling gating mechanism, the airway pressure features after interference suppression are calculated; finally, through the pressure feature enhancement residual mechanism, the pure airway pressure features are calculated. Step 3: Based on the preprocessed turbulence time series data, calculate the turbulence time series characteristics, and then complete the classification through the turbulence intensity differentiation mechanism to obtain the turbulence mode; combined with the pure airway pressure characteristics, calculate the cross-modal fusion characteristics to generate the initial calibration parameters and the final calibration parameters. Step 4: Correct the preprocessed airway pressure data according to the final calibration parameters to obtain the calibrated airway pressure time series data; then combine the pure airway pressure characteristics to calculate the ventilation trend abnormality markers, the percentage of numerical abnormality, and generate a graded early warning level. Step 5: Based on the graded warning level, conduct graded warnings and visualize the results.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention addresses the problem of poor monitoring and early warning accuracy caused by insufficient consideration of vehicle bumps during transport. It obtains pure airway pressure characteristics through a bump-pressure dual-path feature decoupling gating mechanism and a pressure feature enhancement residual mechanism. It generates suitable final calibration parameters in combination with the bump mode, and then uses these parameters to correct the airway pressure data. It combines the pure airway pressure characteristics to judge the ventilation trend and numerical anomalies and generates graded early warnings. This effectively counteracts bump interference, solves the problem of insufficient self-adaptation in traditional calibration, and significantly reduces the rate of missed and false alarms in early warnings, thus meeting the core requirements of real-time and reliable early warnings in emergency transport scenarios.

[0020] (2) To address the issues of confusion between airway pressure and turbulence data caused by ambulance turbulence and the easy misjudgment or omission of ventilation monitoring, a turbulence-pressure dual-path feature decoupling gating mechanism is constructed. Through a dual-path convolutional neural network, features of the preprocessed turbulence state time series data and airway pressure time series data are extracted independently to avoid feature confusion in single-path processing. Then, through splicing operation and multilayer perceptron combined with the Sigmoid function, the turbulence interference feature weight coefficient is dynamically generated to adapt to different turbulence intensity scenarios. At the same time, the turbulence physical constraint matrix and the pressure physiological constraint matrix are used to ensure that the stripped turbulence interference features are real and the retained airway pressure features are within the physiological range, thus achieving accurate separation of the two types of features.

[0021] (3) In view of the problem that traditional turbulence classification methods are difficult to capture turbulence pattern changes due to the dynamic changes, diverse patterns and large differences in interference of ambulance turbulence, this invention uses a long short-term memory network to extract turbulence time-series data after preprocessing to capture the temporal dependence of data and avoid the limitation of traditional methods that only focus on a single moment and ignore the overall trend. Then, the real-time features are calculated by dynamic time warping similarity and the matching degree of the average cluster center of each turbulence pattern is calculated. Combined with Sigmoid to generate pattern matching weights to distinguish turbulence patterns. Finally, the intensity peak is extracted by sliding window peak detection, and the ratio of the peak intensity to the mean turbulence intensity is multiplied by the weight and rounded to obtain the turbulence pattern that accurately reflects the real-time turbulence intensity and pattern, providing a reliable basis for subsequent dynamic generation of adaptation calibration parameters.

[0022] (4) To address the problem of improper correction caused by the dynamic changes in patient airway pressure during transport and the difficulty in adapting traditional calibration parameters, this invention designs a final calibration parameter. First, the pure airway pressure characteristics and turbulence time sequence characteristics are concatenated, and attention weights are generated by multilayer perceptron and Softmax. Combined with layer normalization to obtain cross-modal fusion characteristics, the weights are dynamically allocated to adapt to turbulence interference and pressure abnormalities. Then, the turbulence level is converted into a weight vector, and the Transformer encoder is used to capture global correlations and the Long Short-Term Memory Network is used to capture temporal dependencies, generating initial calibration parameters that are adapted to the global context and are temporally stable. Finally, the final calibration parameters are obtained by optimizing through equipment constraints and physiological adaptation constraints, which can simultaneously adapt to real-time turbulence and patient physiological characteristics, avoid over-correction or under-correction, and ensure ventilation effect.

[0023] (5) The present invention designs a graded early warning level. First, the preprocessed airway pressure data is corrected with the final calibration parameters to obtain calibration data that can truly reflect the ventilation status and avoid the judgment bias of the original data distortion. Then, the ventilation trend abnormality mark is generated by combining the pure airway pressure characteristics. Simultaneously, the proportion of numerical abnormality is calculated based on the calibration data, and a trend-numerical dual judgment dimension is constructed. It does not miss trend risks and is not disturbed by instantaneous fluctuations, thus meeting the emergency needs for early warning accuracy and timeliness. Attached Figure Description

[0024] Figure 1 A flowchart illustrating a method for monitoring and early warning of ventilators used in emergency transport, provided by the present invention; Figure 2 This is a system interface diagram for monitoring emergency transport ventilators provided by the present invention. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0026] Example 1: A method for monitoring and early warning of ventilators used in emergency transport, such as... Figure 1 As shown, it includes the following steps: S1: Collect ambulance bump data and airway pressure time-series data and preprocess them to obtain preprocessed bump data and airway pressure time-series data, including: S11: The ambulance bump state data is collected through MEMS gyroscope and MEMS accelerometer. The data type is multi-dimensional time series data, including angular velocity, angular acceleration and linear acceleration. The data is processed by three standard deviation method and low-pass filtering method for noise reduction, and then the data is normalized to obtain the preprocessed bump state time series data. S12: Collect time-series airway pressure data via ventilator. The data type is one-dimensional time-series data, including peak inspiratory pressure, positive end-expiratory pressure, mean airway pressure, and airway fluctuation pressure. Perform sliding window smoothing and noise reduction, linear interpolation to complete missing values, and then perform data normalization to obtain preprocessed time-series airway pressure data.

[0027] S2: Based on the preprocessed turbulence state time-series data and the preprocessed airway pressure time-series data, a dual-path convolutional neural network is constructed to extract turbulence interference features and airway pressure features, respectively; then, through a turbulence-pressure dual-path feature decoupling gating mechanism, the airway pressure features after interference suppression are calculated; finally, through a pressure feature enhancement residual mechanism, the pure airway pressure features are calculated, including: S21: Based on the preprocessed turbulence state time-series data and the preprocessed airway pressure time-series data, a dual-path convolutional neural network is constructed to extract turbulence interference features and airway pressure features, respectively. The calculation method is as follows: ; ; in, Characterized by turbulence and interference. It is a convolutional layer. To add element by element, For batch normalization, This is the preprocessed time series data of the turbulence state. Characteristics of airway pressure. This is the preprocessed airway pressure time series data. For Hadama accumulation, For ReLU functions; S22: Based on the characteristics of turbulence interference and airway pressure, the airway pressure characteristics after interference suppression are calculated through a turbulence-pressure dual-path feature decoupling gating mechanism. The calculation method is as follows: ; ; ; in, The weighting coefficients for turbulence interference characteristics. For the Sigmoid function, It is a multilayer perceptron. For splicing operations, The characteristics of the turbulence and interference after stripping. The bump constraint coefficient is... For the Laplace operator, For the physical constraint matrix of the turbulence, To obtain the maximum value, The physiological constraint matrix of stress. The airway pressure characteristics after interference suppression; S23: Based on the airway pressure characteristics after interference suppression, a residual enhancement mechanism is used to strengthen effective characteristics and suppress residual interference, calculating the pure airway pressure characteristics. The calculation method is as follows: ; ; in, The pressure characteristics after residual processing. For convolution residual blocks, For feature enhancement coefficients, This represents the characteristics of a pure airway pressure. The residual interference suppression coefficient is... This is global average pooling.

[0028] S3: Based on the preprocessed turbulence time-series data, calculate the turbulence time-series characteristics, and then complete the classification through a turbulence intensity differentiation mechanism to obtain the turbulence mode; combined with the pure airway pressure characteristics, calculate the cross-modal fusion characteristics to generate initial calibration parameters and final calibration parameters, including: S31: Based on the preprocessed turbulence time series data, calculate the turbulence time series characteristics, and then complete the classification through the turbulence intensity differentiation mechanism to obtain the turbulence mode. The calculation method is as follows: ; ; ; ; in, The characteristics of the turbulent timing are as follows: For Long Short-Term Memory (LSTM) networks, This is the preprocessed time series data of the turbulence state. For dynamic time-warped similarity, The average cluster center for each turbulence pattern, To take the absolute value, Weights for pattern matching For the Sigmoid function, Tumble mode, This is a rounding operation. This is a peak detection algorithm based on a sliding window. This represents the mean value of the turbulence intensity in the preprocessed turbulence state time series data; S32: The pure airway pressure characteristics and turbulence timing characteristics are fused using an attention mechanism to obtain cross-modal fusion features. The calculation method is as follows: ; ; ; in, For cross-modal splicing features, For splicing operations, This represents the characteristics of a pure airway pressure. For attention weights, For the Softmax function, It is a multilayer perceptron. For cross-modal fusion features, For Hadama accumulation, For element-wise addition, For layer normalization; S33: Based on the turbulence mode and cross-modal fusion features, initial calibration parameters are generated through a dynamic calibration module. The calculation method is as follows: ; ; ; in, This is the weight vector for the turbulence mode. For embedding layer, For feature-parameter mapping features, For Transformer encoder operations, For calibration parameter bias terms, These are the initial calibration parameters; S34: Perform constraint optimization on the initial calibration parameters to obtain the final calibration parameters, calculated as follows: ; ; ; in, For calibration parameters after safety constraints, To obtain the minimum value, To obtain the maximum value, This is the lower limit of the calibration parameter. To calibrate the upper limit of the parameter, This is a physiological adaptation constraint compensation term. This is a matrix representing the physiological adaptation range of airway pressure. These are the final calibration parameters.

[0029] S4: Correct the preprocessed airway pressure data according to the final calibration parameters to obtain calibrated airway pressure time-series data; then, combine the pure airway pressure characteristics to calculate ventilation trend anomaly markers, the proportion of numerical anomalies, and generate graded warning levels, including: S41: Correct the preprocessed airway pressure data according to the final calibration parameters to obtain the calibrated airway pressure time series data. The calculation method is as follows: ; in, This is the calibrated airway pressure time series data. This is the preprocessed airway pressure time series data. The physiological constraint matrix for stress; S42: Based on the characteristics of pure airway pressure and calibrated airway pressure data, determine abnormal respiratory states and obtain abnormal ventilation trend markers and the percentage of numerical abnormalities. The calculation method is as follows: ; ; in, Mark abnormal ventilation trends. To select the category with the highest probability, The percentage of numerical anomalies. This is the normal reference value for airway pressure; S43: Based on the abnormal ventilation trend markers and the proportion of abnormal numerical values, a graded early warning level is generated. The calculation method is as follows: ; in, It is classified into different warning levels.

[0030] For example, the normal reference value for airway pressure is set to 20 for peak inspiratory pressure. Positive end-expiratory pressure 8 Focusing on the peak inspiratory pressure, the effectiveness of the graded early warning system is verified through three typical scenarios, and the shortcomings of traditional methods are compared: Scenario 1: Normal ventilation under mild turbulence; like Figure 2 As shown, the ambulance was traveling on a city asphalt road, encountering a few bumps, with a bump intensity of approximately 1.2. The peak inhalation pressure was between 19.5 and 21.5 due to slight vibrations. Intermittent fluctuations; final calibration parameters: -0.3 to +0.2 After correction, the calibrated airway pressure data stabilized between 19.2 and 21.3. Based on the characteristics of pure airway pressure, the respiratory cycle is stable without trend fluctuations, and abnormal ventilation trends are marked as normal. The percentage of abnormal values ​​is calculated, and the results are between 1.0% and 6.5% (all less than 10%), ultimately outputting a level 0 warning (no warning). Traditional methods do not correct data, and the value is 21.5 at some times. Approaching the preset threshold 22 This can easily trigger a "suspected abnormality" prompt, interfering with medical staff's judgment of normal ventilation status; Scenario 2: Instantaneous numerical fluctuations under moderate turbulence; The ambulance traveled on rural dirt roads, a continuous stretch of gravel, with a bumpiness level of 2.5. The peak inspiratory pressure is affected between 18.0 and 23.5. Fluctuations occurred intermittently, with a single instantaneous peak value reaching 23.5. The final calibration parameters are -0.5 to +0.4. After correction, the calibrated airway pressure data is 17.5-23.1. Based on the characteristics of pure airway pressure, although there are transient fluctuations in the respiratory cycle, there is no overall sustained upward or downward trend, and the abnormal ventilation trend is marked as normal; the percentage of abnormal values ​​is calculated, with a maximum of 15.5% (within the 10%~30% range), ultimately outputting a Level 1 warning (mild warning), reminding medical staff to pay attention to transient fluctuations without requiring emergency intervention; traditional methods do not calibrate data and directly use 23.5 Substituting the values, the percentage of abnormal values ​​was 17.5%, exceeding the single threshold of "reporting abnormality at 15%", which led to an incorrect output of "moderate warning". This caused medical staff to interrupt intravenous infusion and other operations to make ineffective ventilation adjustments. Scenario 3: Trends of insufficient ventilation masked by severe turbulence; The ambulance traveled along winding mountain roads, experiencing a bumpiness level of 4.0. The peak inhalation pressure was between 16.0 and 22.0 due to severe turbulence. Intermittent oscillations masked the patient's insufficient ventilation due to increased airway secretions; the final calibration parameters were -0.8 to +0.6. After correction, the calibrated airway pressure data showed a gradual downward trend, decreasing from 19.8 within 5 minutes. It dropped to 15.2 Within the next 30 seconds, the percentage of abnormal values ​​rose to 31% (exceeding 30%). Based on the characteristics of clean airway pressure, the peak inspiratory pressure continued to decline, and the abnormal ventilation trend was marked as abnormal, ultimately outputting a Level 2 warning (severe warning), triggering an audible and visual alarm to prompt medical staff to immediately clear airway secretions; traditional methods rely on uncalibrated data, 16.0-22.0 Still in "15-25" Within the safety threshold, and lacking a trend judgment dimension, it is impossible to identify the downward trend masked by data fluctuations, leading to the risk of underreporting insufficient ventilation.

[0031] Specifically, for scenarios involving sudden, severe turbulence when ambulances traverse sections of road with continuous high elevation changes and potholes, this invention also provides a calibration calculation method with instantaneous interference compensation to replace step S41. The calculation method is as follows: ; in, This is the compensation coefficient for instantaneous turbulence interference, dynamically adjusted according to the intensity of the instantaneous turbulence, with a value ranging from -1.2 to +0.8. .

[0032] S5: Based on the graded warning level, provide graded warnings and visualizations, including: S51: Add timestamps and patient identifiers to the graded early warning signal to obtain a graded early warning signal with associated information; S52: Encapsulate the hierarchical early warning signal with associated information using a data transmission protocol, convert it into a standard data format compatible with the emergency transport early warning terminal and the ventilator control module, and obtain standardized early warning execution data; S53: Standardized early warning execution data are transmitted to the ventilator parameter control module to realize the visualization and push of graded early warning information.

[0033] Example 2: Applying the above-mentioned emergency transport ventilator monitoring and early warning method to emergency transport ventilator monitoring and early warning, including: Step 1: Collect ambulance bump data and airway pressure time series data and preprocess them to obtain preprocessed bump data and preprocessed airway pressure time series data. Step 2: Based on the preprocessed turbulence state time series data and the preprocessed airway pressure time series data, a dual-path convolutional neural network is constructed to extract turbulence interference features and airway pressure features respectively; then, through the turbulence-pressure dual-path feature decoupling gating mechanism, the airway pressure features after interference suppression are calculated; finally, through the pressure feature enhancement residual mechanism, the pure airway pressure features are calculated. Step 3: Based on the preprocessed turbulence time series data, calculate the turbulence time series characteristics, and then complete the classification through the turbulence intensity differentiation mechanism to obtain the turbulence mode; combined with the pure airway pressure characteristics, calculate the cross-modal fusion characteristics to generate the initial calibration parameters and the final calibration parameters. Step 4: Correct the preprocessed airway pressure data according to the final calibration parameters to obtain the calibrated airway pressure time series data; then combine the pure airway pressure characteristics to calculate the ventilation trend abnormality markers, the percentage of numerical abnormality, and generate a graded early warning level. Step 5: Based on the graded warning level, conduct graded warnings and visualize the results.

[0034] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0035] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0036] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for monitoring and early warning of emergency transport ventilators, characterized in that, Includes the following steps: S1: Collect ambulance bumpy state data and airway pressure time series data and preprocess them to obtain preprocessed bumpy state time series data and preprocessed airway pressure time series data. S2: Based on the preprocessed turbulence state time series data and the preprocessed airway pressure time series data, a dual-path convolutional neural network is constructed to extract turbulence interference features and airway pressure features respectively; then, through the turbulence-pressure dual-path feature decoupling gating mechanism, the airway pressure features after interference suppression are calculated; finally, through the pressure feature enhancement residual mechanism, the pure airway pressure features are calculated. The turbulence-pressure dual-path feature decoupling gating mechanism processes the spliced ​​turbulence interference features and airway pressure features through a multilayer perceptron, calculates the weight coefficient of the turbulence interference features using the Sigmoid function, and then uses the turbulence physical constraint matrix and the pressure physiological constraint matrix to constrain the turbulence interference features and airway pressure features respectively, to obtain the airway pressure features after interference suppression. S3: Based on the preprocessed turbulence time series data, calculate the turbulence time series characteristics, and then complete the classification through the turbulence intensity differentiation mechanism to obtain the turbulence mode; combined with the pure airway pressure characteristics, calculate the cross-modal fusion characteristics to generate the initial calibration parameters and the final calibration parameters. S4: Correct the preprocessed airway pressure data according to the final calibration parameters to obtain the calibrated airway pressure time series data; then combine the pure airway pressure characteristics to calculate the ventilation trend abnormality markers, the percentage of numerical abnormality, and generate a graded early warning level. S5: Provide tiered early warnings and visual displays based on the tiered warning levels.

2. The emergency transport ventilator monitoring and early warning method according to claim 1, characterized in that, Step S1 includes: S11: The ambulance bump state data is collected through MEMS gyroscope and MEMS accelerometer. The data type is multi-dimensional time series data, including angular velocity, angular acceleration and linear acceleration. The data is processed by three standard deviation method and low-pass filtering method for noise reduction, and then the data is normalized to obtain the preprocessed bump state time series data. S12: Collect time-series airway pressure data via ventilator. The data type is one-dimensional time-series data, including peak inspiratory pressure, positive end-expiratory pressure, mean airway pressure, and airway fluctuation pressure. Perform sliding window smoothing and noise reduction, linear interpolation to complete missing values, and then perform data normalization to obtain preprocessed time-series airway pressure data.

3. The emergency transport ventilator monitoring and early warning method according to claim 1, characterized in that, In step S2, the pressure feature enhancement residual mechanism processes the airway pressure features after interference suppression using a convolutional residual block to obtain the residual-processed pressure features. The residual-processed pressure features are then enhanced by combining feature enhancement coefficients, and residual interference is suppressed by using the Sigmoid function and residual interference suppression coefficients. Finally, the pure airway pressure features are obtained through Hadamard product and global average pooling operations.

4. The emergency transport ventilator monitoring and early warning method according to claim 3, characterized in that, Step S2 includes: S21: Based on the preprocessed turbulence state time-series data and the preprocessed airway pressure time-series data, a dual-path convolutional neural network is constructed to extract turbulence interference features and airway pressure features, respectively. The calculation method is as follows: ； ； in, Characterized by turbulence and interference. It is a convolutional layer. To add element by element, For batch normalization, This is the preprocessed time series data of the turbulence state. Characteristics of airway pressure. This is the preprocessed airway pressure time series data. For Hadama accumulation, For ReLU functions; S22: Based on the characteristics of turbulence interference and airway pressure, the airway pressure characteristics after interference suppression are calculated through a turbulence-pressure dual-path feature decoupling gating mechanism. The calculation method is as follows: ； ； ； in, The weighting coefficients for turbulence interference characteristics. For the Sigmoid function, It is a multilayer perceptron. For splicing operations, The characteristics of the turbulence and interference after stripping. The bump constraint coefficient is... For the Laplace operator, For the physical constraint matrix of the turbulence, To obtain the maximum value, The physiological constraint matrix of stress. The airway pressure characteristics after interference suppression; S23: Based on the airway pressure characteristics after interference suppression, a residual enhancement mechanism is used to strengthen effective characteristics and suppress residual interference, calculating the pure airway pressure characteristics. The calculation method is as follows: ； ； in, The pressure characteristics after residual processing. For convolution residual blocks, For feature enhancement coefficients, This is a characteristic of pure airway pressure. The residual interference suppression coefficient is... This is global average pooling.

5. The emergency transport ventilator monitoring and early warning method according to claim 3, characterized in that, Step S3 includes: The turbulence intensity differentiation mechanism is as follows: based on the preprocessed turbulence state time series data, turbulence time series features are calculated through a long short-term memory network, and pattern matching weights are calculated based on dynamic time warping similarity; peak values ​​of turbulence time series features are extracted based on a sliding window peak detection algorithm, and then the ratio is calculated with the mean of turbulence intensity in the preprocessed turbulence state time series data, multiplied by the pattern matching weight, and rounded to obtain the turbulence pattern.

6. The emergency transport ventilator monitoring and early warning method according to claim 5, characterized in that, Step S3 includes: S31: Based on the preprocessed turbulence time series data, calculate the turbulence time series characteristics, and then complete the classification through the turbulence intensity differentiation mechanism to obtain the turbulence mode. The calculation method is as follows: ； ； ； ； in, The characteristics of the turbulent timing are as follows: For Long Short-Term Memory (LSTM) networks, This is the preprocessed time series data of the turbulence state. For dynamic time-warped similarity, The average cluster center for each turbulence pattern, To take the absolute value, Weights for pattern matching For the Sigmoid function, Tumble mode, This is a rounding operation. This is a peak detection algorithm based on a sliding window. This represents the mean value of the turbulence intensity in the preprocessed turbulence time series data. S32: The pure airway pressure characteristics and turbulence timing characteristics are fused using an attention mechanism to obtain cross-modal fusion features. The calculation method is as follows: ； ； ； in, For cross-modal splicing features, For splicing operations, This is a characteristic of pure airway pressure. For attention weights, For the Softmax function, It is a multilayer perceptron. For cross-modal fusion features, For Hadama accumulation, For element-wise addition, For layer normalization; S33: Based on the turbulence mode and cross-modal fusion features, the initial calibration parameters are generated through the dynamic calibration module. The calculation method is as follows: ； ； ； in, This is the turbulence mode weight vector. For embedding layer, For feature-parameter mapping features, For Transformer encoder operations, For calibration parameter bias terms, These are the initial calibration parameters; S34: Perform constraint optimization on the initial calibration parameters to obtain the final calibration parameters, calculated as follows: ； ； ； in, For calibration parameters after safety constraints, To obtain the minimum value, To obtain the maximum value, This is the lower limit of the calibration parameter. To calibrate the upper limit of the parameter, This is a physiological adaptation constraint compensation term. This is a matrix representing the physiological adaptation range of airway pressure. These are the final calibration parameters.

7. The emergency transport ventilator monitoring and early warning method according to claim 6, characterized in that, Step S4 includes: S41: Correct the preprocessed airway pressure data according to the final calibration parameters to obtain the calibrated airway pressure time series data. The calculation method is as follows: ； in, For calibrated airway pressure time series data, This is the preprocessed airway pressure time series data. The physiological constraint matrix for stress; S42: Based on the characteristics of pure airway pressure and calibrated airway pressure data, determine abnormal respiratory states and obtain abnormal ventilation trend markers and the percentage of numerical abnormalities. The calculation method is as follows: ； ； in, Mark abnormal ventilation trends. To select the category with the highest probability, The percentage of numerical anomalies. This is the normal reference value for airway pressure. This represents the pressure characteristics of a clean airway; S43: Based on the abnormal ventilation trend markers and the proportion of abnormal numerical values, a graded early warning level is generated. The calculation method is as follows: ； in, It is classified into different warning levels.

8. The emergency transport ventilator monitoring and early warning method according to claim 7, characterized in that, Step S5 includes: S51: Add timestamps and patient identifiers to the graded early warning signal to obtain a graded early warning signal with associated information; S52: Encapsulate the hierarchical early warning signal with associated information using a data transmission protocol, convert it into a standard data format compatible with the emergency transport early warning terminal and the ventilator control module, and obtain standardized early warning execution data; S53: Standardized early warning execution data are transmitted to the ventilator parameter control module to realize the visualization and push of graded early warning information.

9. The application of the method according to any one of claims 1-8 in the monitoring and early warning of ventilators used in assisted emergency transport, characterized in that, include: Step 1: Collect ambulance bump data and airway pressure time series data and preprocess them to obtain preprocessed bump data and preprocessed airway pressure time series data. Step 2: Based on the preprocessed turbulence state time series data and the preprocessed airway pressure time series data, a dual-path convolutional neural network is constructed to extract turbulence interference features and airway pressure features respectively; then, through the turbulence-pressure dual-path feature decoupling gating mechanism, the airway pressure features after interference suppression are calculated; finally, through the pressure feature enhancement residual mechanism, the pure airway pressure features are calculated. Step 3: Based on the preprocessed turbulence time series data, calculate the turbulence time series characteristics, and then complete the classification through the turbulence intensity differentiation mechanism to obtain the turbulence mode; combined with the pure airway pressure characteristics, calculate the cross-modal fusion characteristics to generate the initial calibration parameters and the final calibration parameters. Step 4: Correct the preprocessed airway pressure data according to the final calibration parameters to obtain the calibrated airway pressure time series data; then combine the pure airway pressure characteristics to calculate the ventilation trend abnormality markers, the percentage of numerical abnormality, and generate a graded early warning level. Step 5: Based on the graded warning levels, conduct graded warnings and provide visual displays; The application, by executing steps 1 to 5, utilizes the generated graded early warning levels to provide assistance for monitoring and early warning of ventilators used in emergency transport.

Images