A method for recognizing the installation state of a high-frequency jet ventilator laryngeal mask and a storage medium

Abstract

The application discloses a kind of high-frequency jet breathing machine laryngeal mask installation state identification method and storage medium, method includes: in initial calibration period, reference benchmark model of pressure waveform and frequency distribution are established;Real-time extraction peak pressure variation rate, energy spectrum offset and acoustic spectrum feature constitute dynamic feature vector;Dynamic feature vector is compared with benchmark model, and the determination result of sealing failure or partial obstruction is output.The application utilizes the multi-dimension feature fusion of pressure waveform characteristic analysis and jet acoustic monitoring, adapts to the open loop characteristics of high-frequency jet ventilation, without additional implantation in vivo sensor, non-contact laryngeal mask installation state identification can be realized, and interference signal caused by surgical operation can be dynamically eliminated.

Description

Technical Field

[0001] This invention relates to the field of ventilators, and more particularly to a method and storage medium for identifying the installation status of a laryngeal mask in a high-frequency jet ventilator. Background Technology

[0002] High-frequency ventilators need to provide precise ventilation at frequencies of several hertz or even tens of hertz in clinical scenarios such as neonatal respiratory distress and acute respiratory failure. High-frequency jet ventilators need to provide precise ventilation support at frequencies of several hertz to tens of hertz in clinical scenarios such as neonatal respiratory distress syndrome, acute respiratory failure, and laryngeal mask airway support. As a non-invasive airway management tool, the installation status of the laryngeal mask airway (e.g., sealing, whether it is partially blocked) directly affects the ventilation effect and patient safety.

[0003] Currently, clinical monitoring of laryngeal mask airway (LMA) installation status mainly relies on experience-based judgment, rough observation of airway pressure waveforms, or the addition of specialized sensors (such as esophageal pressure and end-tidal CO2 monitoring). This approach has the following shortcomings: (1) It lacks real-time, quantitative, and non-invasive automatic identification methods, making continuous monitoring during surgery difficult; (2) The unique open-loop characteristics and high-speed jet noise of high-frequency jet ventilation make traditional pressure or flow monitoring susceptible to interference, making feature extraction difficult; (3) Surgical procedures (such as suction, electrocautery, and changes in body position) introduce transient pressure fluctuations and acoustic artifacts, often leading to false alarms or missed detections. Existing technologies have not effectively solved the problem of online identification of LMA seal failure and partial obstruction in high-frequency jet ventilation scenarios. Therefore, an intelligent identification method that adapts to the dynamic characteristics of high-frequency ventilation and is resistant to surgical interference is urgently needed. Summary of the Invention

[0004] The purpose of this invention is to provide a method for identifying the installation status of a laryngeal mask in a high-frequency jet ventilator, which solves the technical problem of abnormal installation status monitoring pointed out in the prior art.

[0005] This invention provides a method for identifying the installation status of a laryngeal mask airway on a high-frequency jet ventilator, comprising the following steps:

[0006] S10. During the initial calibration period after the start of high-frequency jet ventilation, the pressure timing signal output by the ventilator airway pressure sensor and the acoustic timing signal output by the jet acoustic sensor are collected. The collected pressure timing signals are subjected to baseline drift correction and noise filtering to establish a pressure waveform reference model and a frequency distribution reference model. The reference models together constitute the initial reference reference.

[0007] S20. During normal operation of high-frequency jet ventilation, extract the peak pressure variation rate feature, energy spectrum shift feature and acoustic spectrum feature at the current moment in real time, and fuse the three types of features to form a dynamic feature vector V(t).

[0008] S30. Perform a multi-dimensional comparison operation between the dynamic feature vector V(t) and the corresponding benchmark model in the initial reference benchmark, calculate the feature deviation comprehensive evaluation index δ(t), and output the sealing failure judgment result or partial blockage judgment result based on the relationship between the feature deviation comprehensive evaluation index δ(t) and the preset judgment threshold.

[0009] S40. During the operation of steps S20 and S30, transient interference signals caused by surgical operations are detected in real time. The transient interference signals are marked and eliminated by the interference signal identification model to ensure the effectiveness of the dynamic feature vector V(t). The elimination results are fed back to the comparison operation in step S30 to correct the feature deviation comprehensive evaluation index δ(t).

[0010] Preferably, as one possible implementation, the dynamic feature extraction stage in step S20 further includes the following sub-steps:

[0011] S21. Perform cycle-by-cycle peak detection on the real-time acquired airway pressure time-series signal P(t), extract the peak pressure sequence {Pmax,1,Pmax,2,…,Pmax,N} over N consecutive ventilation cycles, calculate and normalize the peak pressure difference between adjacent cycles, and obtain the peak pressure variation rate characteristic F. P ;

[0012] S22. Perform a Fast Fourier Transform on the real-time pressure time-series signal P(t) within a sliding time window [t-Tw,t] to obtain the current energy power spectral density distribution E(f,t). Calculate the offset between the centroid frequency of the power spectrum of the current energy power spectral density distribution E(f,t) and the corresponding reference frequency in the initial reference, and obtain the energy spectrum offset characteristic quantity. ;

[0013] S23. Acoustic spectral feature extraction and fusion: real-time extraction and fusion of acoustic time-series signals from jet acoustic sensors. Extracting Mel frequency cepstral coefficient sequence As a characteristic quantity of acoustic spectrum The peak pressure variation rate characteristic quantity Energy spectral shift characteristic quantity and acoustic spectral characteristics The dynamic feature vector is constructed by concatenating the features along their respective dimensions. .

[0014] Preferably, as one possible implementation, the feature comparison and anomaly determination stage in step S30 further includes the following sub-steps:

[0015] S31. Based on the adaptive kernel density estimation algorithm, the pressure waveform benchmark model and frequency distribution benchmark model in the waveform feature analysis scenario are modeled with probability density distribution, and the dynamic feature vector is... Projecting the vector onto the feature space defined by the probability density distribution model, the dynamic feature vector is calculated. Deviation values ​​of each component in the pressure waveform dimension, energy spectrum dimension, and acoustic spectrum dimension. and The deviation values ​​of each component together constitute the characteristic deviation component set;

[0016] S32. Calculate the comprehensive evaluation index of the feature deviation. Based on the set of feature deviation components, calculate the comprehensive evaluation index of the feature deviation according to the preset multidimensional fusion rules. ;

[0017] S33. Then, the classification result is output, and the comprehensive evaluation index is based on the aforementioned feature deviation. Between the preset seal failure value and preset partial blocking threshold The size relationship between them is determined by a three-level decision logic: when... Output the seal failure judgment result at the time; when Output partial blocking determination results; when The system outputs a result indicating that the installation status is normal.

[0018] Preferably, as one possible implementation, the interference signal dynamic elimination stage in step S40 further includes the following sub-steps:

[0019] S41. Identify surgical operation interference signals and model transient interference signal patterns in waveform feature analysis scenarios based on hidden Markov model algorithms, including real-time pressure time-series signals. With real-time acoustic timing signals The Hidden Markov Model is input synchronously to calculate the posterior probability that the signal at the current moment belongs to a surgical operation interference state. When the posterior probability Exceeding the preset interference identification threshold At that time, the data frame at the current moment is marked as an interference frame;

[0020] S42. Perform a removal operation on all data segments marked as interference frames, and use cubic spline interpolation to smoothly fill the data gaps formed after removal, so as to restore the complete effective timing signal.

[0021] S43. Set the time positions of the interference frames The characteristic deviation comprehensive evaluation index is passed to step S30. In the calculation stage, for the above The deviation calculation results within the time period are shielded to ensure that interference signals do not participate in the final judgment logic calculation. The shielding results are synchronously updated to the initial reference benchmark to achieve rolling benchmark self-calibration.

[0022] Preferably, as one possible implementation, the adaptive kernel density estimation algorithm in step S31, when modeling the waveform feature analysis scenario, further includes the following sub-steps:

[0023] S311. On the pressure waveform sample set acquired during the initial calibration period, the optimal kernel bandwidth for each feature dimension is calculated using the cross-validation maximum likelihood estimation method. For the first For each feature layer, the optimal kernel bandwidth is determined by the following formula:

[0024] ;

[0025] in To leave one cross-validation kernel density estimate, This represents the total number of samples during the calibration period.

[0026] S312. In view of the characteristic that the high-frequency jet ventilation pressure waveform may exhibit a multi-peak distribution under different ventilation conditions, a mixture Gaussian kernel function is used instead of a single Gaussian kernel function to perform multimodal modeling of the probability density distribution of each feature layer. The number of components of the mixture Gaussian kernel function is determined by the Bayesian information criterion.

[0027] S313. During normal operation, every preset update cycle. The latest effective feature vector samples are incrementally incorporated into the kernel density estimation model to achieve dynamic rolling updates of the baseline model. The incremental updates are only performed on data frames that are determined to be in normal installation status, ensuring that abnormal data does not contaminate the baseline model.

[0028] Preferably, as one possible implementation; when the Hidden Markov Model models the surgical operation interference signal in step S41, the following sub-steps are further included:

[0029] S411. Definition of Hidden Markov Model State Space: The hidden state space of the Hidden Markov Model is defined as follows: These correspond to normal ventilation, mild surgical interference, and severe surgical interference, respectively. The observation space consists of pressure time-series signals. Short-time statistical eigenvectors and acoustic time-series signals The short-time spectral eigenvectors are jointly constructed;

[0030] S412. The Baum-Welch expectation-maximization algorithm is used to initialize the transition probability matrix of the hidden Markov model during the preoperative preparation phase. and initial state probability vector Offline pre-training is performed, and the recursive online Baum-Welch algorithm is used to fine-tune the model parameters in real time during the operation, so that the model can continuously adapt to the interference characteristics of the current surgical scenario.

[0031] S413. In the scenario of high-frequency jet ventilation acoustic spectrum characteristic analysis, respectively at short time scales Mesoscale and long-term scale The Hidden Markov Model is run independently to decode and obtain the posterior probability of interference at three time scales. The geometric mean of the three values ​​is taken to obtain the final fused posterior probability. The fusion posterior probability is used as the criterion for marking the interference frame in step S41.

[0032] Preferably, as one possible implementation, it also includes sending a warning signal to the host computer when the output seal failure judgment result or the output partial blockage judgment result is output.

[0033] A storage medium includes a memory for executing a program for identifying the installation status of a high-frequency jet ventilator laryngeal mask.

[0034] Compared with the prior art, the embodiments of the present invention have at least the following technical advantages:

[0035] Analysis of the above-mentioned method for identifying the installation status of a high-frequency jet ventilator laryngeal mask provided by the present invention shows that the method includes: during the initial calibration period after the high-frequency jet ventilation is started, collecting the pressure timing signal output by the ventilator airway pressure sensor and the acoustic timing signal output by the jet acoustic sensor, performing baseline drift correction and noise filtering on the collected pressure timing signal, and establishing a pressure waveform reference model and a frequency distribution reference model, wherein the reference models together constitute the initial reference reference.

[0036] During normal operation of high-frequency jet ventilation, the peak pressure variation rate, energy spectrum shift, and acoustic spectrum features at the current moment are extracted in real time, and the three types of features are fused to form a dynamic feature vector V(t). The dynamic feature vector V(t) is compared with the corresponding benchmark model in the initial reference benchmark in a multi-dimensional comparison operation to calculate the feature deviation comprehensive evaluation index δ(t). Based on the relationship between the feature deviation comprehensive evaluation index δ(t) and the preset judgment threshold, the seal failure judgment result or partial blockage judgment result is output. Transient interference signals caused by surgical operations are detected in real time. The transient interference signals are marked and eliminated by the interference signal identification model to ensure the effectiveness of the dynamic feature vector V(t), and the elimination result is fed back to the comparison operation in the aforementioned steps to correct the feature deviation comprehensive evaluation index δ(t).

[0037] The significant advantages of this invention are: (1) It integrates the characteristics of airway pressure waveform and jet acoustic spectrum, eliminating the need for additional implanted sensors and enabling non-contact, continuous online monitoring; (2) It designs multi-dimensional features such as peak pressure variation rate, energy spectrum shift and acoustic Mel-Cepstral coefficient for the open loop characteristics of high-frequency jet ventilation, which can effectively distinguish between seal failure and partial obstruction; (3) It introduces hidden Markov model and cubic spline interpolation to achieve accurate identification, elimination and data repair of transient surgical interference, while ensuring the stability and accuracy of long-term monitoring through rolling benchmark self-calibration; (4) The three-level judgment logic and adaptive kernel density estimation can adapt to individual differences and intraoperative changes, reduce false alarm rate, and improve clinical safety and ease of operation. Attached Figure Description

[0038] Figure 1 The main flowchart of a method for identifying the installation status of a high-frequency jet ventilator laryngeal mask provided in Embodiment 1;

[0039] Figure 2 The following is a flowchart illustrating the specific operation of a method for identifying the installation status of a laryngeal mask on a high-frequency jet ventilator, as provided in Example 1. Detailed Implementation

[0040] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0041] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0042] Example 1

[0043] See Figure 1 Embodiment 1 of the present invention provides a method for identifying the installation status of a laryngeal mask in a high-frequency jet ventilator, comprising the following steps:

[0044] S10. During the initial calibration period after the start of high-frequency jet ventilation, the pressure timing signal output by the ventilator airway pressure sensor and the acoustic timing signal output by the jet acoustic sensor are collected. The collected pressure timing signals are subjected to baseline drift correction and noise filtering to establish a pressure waveform reference model and a frequency distribution reference model. The reference models together constitute the initial reference reference.

[0045] S20. During normal operation of high-frequency jet ventilation, extract the peak pressure variation rate feature, energy spectrum shift feature and acoustic spectrum feature at the current moment in real time, and fuse the three types of features to form a dynamic feature vector V(t).

[0046] S30. Perform a multi-dimensional comparison operation between the dynamic feature vector V(t) and the corresponding benchmark model in the initial reference benchmark, calculate the feature deviation comprehensive evaluation index δ(t), and output the sealing failure judgment result or partial blockage judgment result based on the relationship between the feature deviation comprehensive evaluation index δ(t) and the preset judgment threshold.

[0047] S40. During the operation of steps S20 and S30, transient interference signals caused by surgical operations are detected in real time. The transient interference signals are marked and eliminated by the interference signal identification model to ensure the effectiveness of the dynamic feature vector V(t). The elimination results are fed back to the comparison operation in step S30 to correct the feature deviation comprehensive evaluation index δ(t).

[0048] Analysis of the above scheme shows that Embodiment 1 of the present invention provides a method for identifying the installation status of the laryngeal mask of a high-frequency jet ventilator. During the initial calibration period after the start of high-frequency jet ventilation (the initial calibration period is the time period of 30 to 60 ventilation cycles continuously collected after the start of high-frequency jet ventilation, or the period of 10 to 30 seconds and the end after the kernel density estimation model convergence criterion confirms stability), airway pressure time-series signals and jet acoustic time-series signals are collected. After baseline drift correction and noise filtering, a pressure waveform reference model and a frequency distribution reference model are established, which together constitute the initial reference.

[0049] Then, during normal operation, peak pressure variation rate features, energy spectrum shift features, and acoustic spectrum features are extracted in real time and fused to form a dynamic feature vector. The dynamic feature vector is compared with the initial reference benchmark in multiple dimensions to calculate the feature deviation comprehensive evaluation index. Based on the comparison result of this index with the preset judgment threshold, the seal failure judgment result or partial obstruction judgment result is output. At the same time, transient interference signals caused by surgical operations are detected in real time, and they are marked and eliminated through the interference signal identification model to ensure the effectiveness of the dynamic feature vector. The elimination result is fed back into the comparison calculation to correct the feature deviation comprehensive evaluation index. Analysis shows that this invention utilizes the fusion of multiple dimensions of pressure waveform feature analysis and jet acoustic monitoring to adapt to the open loop characteristics of high-frequency jet ventilation, and can achieve non-contact laryngeal mask installation status identification without the need for additional implanted sensors.

[0050] See Figure 2 The dynamic feature extraction stage in step S20 further includes the following sub-steps:

[0051] S21. Perform cycle-by-cycle peak detection on the real-time acquired airway pressure time-series signal P(t), extract the peak pressure sequence {Pmax,1,Pmax,2,…,Pmax,N} over N consecutive ventilation cycles, calculate and normalize the peak pressure difference between adjacent cycles, and obtain the peak pressure variation rate characteristic F. P ;

[0052] The calculation process of the peak pressure variation rate characteristic FP is illustrated with an example:

[0053] Assume continuous The peak-pressure sequence is as follows Calculate the adjacent difference (in The range normalization or Z-score normalization is used, and the calculation formula is: where N is the preset number of sliding window cycles, which ranges from 5 to 20 ventilation cycles, and preferably N=10;

[0054] ;in This refers to the peak pressure average. This value represents the degree of fluctuation under unit average pressure.

[0055] S22. Perform a Fast Fourier Transform on the real-time pressure time-series signal P(t) within a sliding time window [t−Tw,t] (where Tw is the preset sliding time window width, ranging from 0.5 to 5 seconds, preferably Tw=2 seconds) to obtain the current energy power spectral density distribution E(f,t). Calculate the offset between the centroid frequency of the power spectrum of the current energy power spectral density distribution E(f,t) and the corresponding reference frequency in the initial reference reference, and obtain the energy spectrum offset characteristic quantity. ;

[0056] Energy spectral shift characteristic The calculation process includes: first, calculating the power spectral density as... Center of gravity frequency The calculation formula is: Let the centroid reference established during the initial calibration period be... Then the offset feature (using relative offset ratio) is: ;

[0057] S23. Acoustic spectral feature extraction and fusion: real-time extraction and fusion of acoustic time-series signals from jet acoustic sensors. Extracting Mel frequency cepstral coefficient sequence As a characteristic quantity of acoustic spectrum (Where the typical MFCC coefficients are of order 12 or 13, i.e., M is 12 or 13, preferably M=13); the peak pressure variation rate characteristic quantity Energy spectral shift characteristic quantity and acoustic spectral characteristics The dynamic feature vector is constructed by concatenating the features along their respective dimensions. .

[0058] The above process involves the analysis of acoustic signals. Perform pre-emphasis, frame segmentation with Hamming windowing, and short-time Fourier transform to obtain the energy spectrum; pass the energy spectrum through a Mel-triangle filter bank; take the logarithm of the filtered output energy and then perform a discrete cosine transform (DCT), taking the first... Order coefficient composition .

[0059] In step S21, the peak pressure variation rate is calculated by periodic peak detection and adjacent difference normalization. Since high-frequency jet ventilation outputs the jet at a fixed frequency, the periodic variation of the peak pressure directly reflects the change in airway resistance or sealing conditions. When the seal fails, leakage causes the peak pressure to drop and the fluctuation to increase. When there is partial blockage, the peak pressure rises and the variation intensifies. Therefore, this characteristic quantity can sensitively capture the impact of abnormal laryngeal mask installation status on the steady state of the pressure waveform.

[0060] Step S22 above utilizes the physical characteristic that the frequency band position of the energy concentration of the high-frequency jet ventilation energy spectrum is closely related to the airway impedance by performing a fast Fourier transform within a sliding time window and calculating the power spectrum centroid frequency shift. That is, when the laryngeal mask seal fails or is partially blocked, the propagation path and reflection conditions of the jet in the airway change, causing the energy spectrum centroid to shift to low or high frequencies. This shift, as a feature quantity, can reflect the abnormal state from the frequency domain dimension. Step S23 above uses Mel-scale cepstral coefficients as acoustic spectral features, utilizing the nonlinear frequency resolution characteristics of human hearing perception by a Mel-scale filter bank to effectively extract acoustic resonance features in the jet noise related to the laryngeal mask airway geometry and sealing state. The low correlation between MFCC coefficients ensures the compactness of the feature expression. By splicing the above three types of feature quantities dimensionally to form a dynamic feature vector, three-dimensional information fusion of the time domain (peak pressure variation rate), frequency domain (energy spectrum shift), and acoustic domain (MFCC) is achieved. This allows subsequent judgment to comprehensively utilize abnormal signals from multiple physical dimensions, significantly improving the ability to distinguish between seal failure and partial blockage.

[0061] Preferably, as one possible implementation, the feature comparison and anomaly determination stage in step S30 further includes the following sub-steps:

[0062] S31. Based on the adaptive kernel density estimation algorithm, the pressure waveform benchmark model and frequency distribution benchmark model in the waveform feature analysis scenario are modeled with probability density distribution, and the dynamic feature vector is... Projecting the vector onto the feature space defined by the probability density distribution model, the dynamic feature vector is calculated. Deviation values ​​of each component in the pressure waveform dimension, energy spectrum dimension, and acoustic spectrum dimension. and The deviation values ​​of each component together constitute the characteristic deviation component set;

[0063] S32. Calculate the comprehensive evaluation index of the feature deviation. Based on the set of feature deviation components, calculate the comprehensive evaluation index of the feature deviation according to the preset multidimensional fusion rules. ;

[0064] S33. Then, the classification result is output, and the comprehensive evaluation index is based on the aforementioned feature deviation. Between the preset seal failure value and preset partial blocking threshold The size relationship between them is determined by a three-level decision logic: when... Output the seal failure judgment result at the time; when Output partial blocking determination results; when The system outputs a result indicating that the installation status is normal. > > 0;

[0065] In step S32, the component deviation values ​​of the three dimensions are integrated into a single evaluation index through a preset multidimensional fusion rule. The technical effect is to compress high-dimensional heterogeneous information into a scalar index with clear physical meaning, facilitating subsequent threshold comparison. Simultaneously, multidimensional fusion can utilize the complementary nature of different features' sensitivity to noise, reducing the excessive influence of accidental fluctuations in any dimension on the final judgment result. Step S33 employs a dual-threshold, three-level judgment logic (sealing failure threshold). With partial blocking threshold This method has significant advantages over single-threshold judgment: sealing failure usually leads to pressure reduction and changes in leaky acoustic characteristics, which are more serious than partial blockage. Therefore, placing the two types of abnormalities with different severity in different threshold ranges for judgment avoids clinical confusion caused by alarming all abnormalities at the same time.

[0066] Preferably, as one possible implementation, the interference signal dynamic elimination stage in step S40 further includes the following sub-steps:

[0067] S41. Identify surgical operation interference signals and model transient interference signal patterns in waveform feature analysis scenarios based on hidden Markov model algorithms, including real-time pressure time-series signals. With real-time acoustic timing signals The Hidden Markov Model is input synchronously to calculate the posterior probability that the signal at the current moment belongs to a surgical operation interference state. When the posterior probability Exceeding the preset interference identification threshold At that time, the data frame at the current moment is marked as an interference frame;

[0068] S42. Perform a removal operation on all data segments marked as interference frames, and use cubic spline interpolation to smoothly fill the data gaps formed after removal, so as to restore the complete effective timing signal.

[0069] S43. Set the time positions of the interference frames The characteristic deviation comprehensive evaluation index is passed to step S30. In the calculation stage, for the above The deviation calculation results within the time period are shielded to ensure that interference signals do not participate in the final judgment logic calculation. The shielding results are synchronously updated to the initial reference benchmark to achieve rolling benchmark self-calibration.

[0070] In step S41, a Hidden Markov Model (HMM) is used to model the transient interference signal pattern. This leverages the inherent advantages of HMM in modeling time-series stochastic processes. Interference signals caused by surgical procedures (such as suction, electrocautery, and changes in body position) exhibit state persistence and transition patterns over time. HMM can capture this temporal dependency through the hidden state sequence. Simultaneous input of pressure and acoustic signals allows the model to use the interference features of two channels for joint discrimination, which significantly improves the accuracy and robustness of interference identification compared to single-channel discrimination.

[0071] Step S42 involves removing data segments marked as interference frames and then filling the data gaps using cubic spline interpolation. Compared to simple linear interpolation or direct discarding, cubic spline interpolation ensures the continuity of the first and second derivatives of the data at the nodes after filling, smoothly connecting the effective signals before and after the interference without introducing additional high-frequency artifacts. This maximizes the restoration of the ventilation characteristics of the original signal after removing the interference and avoids blind spots in judgment caused by missing data. Step S43 involves shielding the interference frame location set by passing it to the evaluation index calculation stage and combining it with rolling benchmark self-calibration. This has a dual technical effect: on the one hand, it ensures that abnormal deviation values ​​during the interference period do not participate in the judgment logic, thereby eliminating the root cause of false alarms; on the other hand, it feeds back the shielded effective data to the benchmark model for incremental updates, enabling the benchmark model to track the slow drift of the ventilation status caused by changes in patient position and anesthesia depth during surgery. This achieves automatic adaptive adjustment of the benchmark in long-term monitoring and avoids cumulative judgment deviations caused by mismatch between the initial benchmark and the current state.

[0072] Preferably, as one possible implementation scheme; the characteristic deviation comprehensive evaluation index It is calculated using the following vectorized formula:

[0073] ;

[0074] in: For feature layer index, These correspond to the pressure waveform feature layer, energy spectrum feature layer, and acoustic spectrum feature layer, respectively. ; For the first The first feature layer Each feature component index For the first On the feature layer, the th Each feature component in Real-time deviation value at any given moment; and The first in the initial reference datum On the feature layer, the th The mean and standard deviation of each characteristic component; For acoustic evaluation dimensions index, , corresponding to different orders of the Mel frequency cepstral coefficients; For the first The first dimension of acoustic evaluation Index of time-frequency evaluation values, For the first On the first acoustic evaluation dimension Each time-frequency evaluation value is in Actual measurement of time; The first corresponding to the initial reference datum On the first acoustic evaluation dimension One time-frequency reference value; The scaling factor between acoustic and pressure characteristics is adaptively determined by statistical analysis of the initial calibration data. .

[0075] It should be noted that, For the first On the feature layer, the th Each feature component in The real-time deviation value at a given moment is (real-time value - real-time value at the previous moment).

[0076] Preferably, as one possible implementation, the adaptive kernel density estimation algorithm in step S31, when modeling the waveform feature analysis scenario, further includes the following sub-steps:

[0077] S311. On the pressure waveform sample set acquired during the initial calibration period, the optimal kernel bandwidth for each feature dimension is calculated using the cross-validation maximum likelihood estimation method. For the first For each feature layer, the optimal kernel bandwidth is determined by the following formula:

[0078] ;

[0079] in To leave one cross-validation kernel density estimate, This represents the total number of samples during the calibration period.

[0080] S312. In view of the characteristic that the high-frequency jet ventilation pressure waveform may exhibit a multi-peak distribution under different ventilation conditions, a mixture Gaussian kernel function is used instead of a single Gaussian kernel function to perform multimodal modeling of the probability density distribution of each feature layer. The number of components of the mixture Gaussian kernel function is determined by the Bayesian information criterion.

[0081] S313. During normal operation, every preset update cycle. The latest effective feature vector samples are incrementally incorporated into the kernel density estimation model to achieve dynamic rolling updates of the baseline model. The incremental updates are only performed on data frames that are determined to be in normal installation status, ensuring that abnormal data does not contaminate the baseline model.

[0082] Preferably, as one possible implementation; when the Hidden Markov Model models the surgical operation interference signal in step S41, the following sub-steps are further included:

[0083] S411. Definition of Hidden Markov Model State Space: The hidden state space of the Hidden Markov Model is defined as follows: These correspond to normal ventilation, mild surgical interference, and severe surgical interference, respectively. The observation space consists of pressure time-series signals. Short-time statistical eigenvectors and acoustic time-series signals The short-time spectral eigenvectors are jointly constructed;

[0084] S412. The Baum-Welch expectation-maximization algorithm is used to initialize the transition probability matrix of the hidden Markov model during the preoperative preparation phase. and initial state probability vector Offline pre-training is performed, and the recursive online Baum-Welch algorithm is used to fine-tune the model parameters in real time during the operation, so that the model can continuously adapt to the interference characteristics of the current surgical scenario.

[0085] S413. In the scenario of high-frequency jet ventilation acoustic spectrum characteristic analysis, respectively at short time scales Mesoscale and long-term scale The Hidden Markov Model is run independently to decode and obtain the posterior probability of interference at three time scales. The geometric mean of the three values ​​is taken to obtain the final fused posterior probability. The fusion posterior probability is used as the criterion for marking the interference frame in step S41.

[0086] It should be noted that step S411 above defines a state space containing three hidden states. The technical effect of distinguishing between mild and severe interference states is that different surgical operations (such as slight touching of tubes vs. large changes in body position vs. intraoperative suction) have significant differences in the amplitude and duration of interference to pressure and acoustic signals. By dividing the interference states into multiple levels, HMM can more finely characterize the intensity levels of interference signals and provide richer discrimination information for subsequent interference frame marking.

[0087] Step S412 adopts a strategy that combines preoperative offline pre-training with intraoperative online fine-tuning. Offline pre-training uses the relatively stable signal environment during the preoperative preparation stage to establish the initial estimation of HMM parameters, providing a good starting point for model convergence. Intraoperative recursive online Baum-Welch algorithm can continuously fine-tune the model parameters using newly arrived data during the operation, enabling the model to adaptively track the changes in the statistical characteristics of interference signals in the current surgical scenario (such as different surgeons' operating habits and differences in interference types at different surgical stages). The combination of the two ensures both the stability of model initialization and the flexibility of intraoperative adaptation.

[0088] However, step S413 independently runs HMM decoding and uses geometric mean fusion at three time scales: short, medium, and long. That is, surgical operation interference has diversity at different time scales—transient spike interference is obvious at the short time scale, while persistent interference is easier to identify at the medium and long time scales. Single time scale analysis is prone to missing some types of interference. Three-scale fusion uses geometric mean to comprehensively weigh the interference detection results at each scale, ensuring that interference can be effectively identified regardless of its duration. At the same time, the geometric mean is not sensitive to extreme values, which avoids the excessive influence of occasional false detections at a certain scale on the fusion results, thus improving the reliability of interference labeling.

[0089] Preferably, as one possible implementation, it also includes sending a warning signal to the host computer when the output seal failure judgment result or the output partial blockage judgment result is output.

[0090] It should be noted that the above steps, by actively sending an early warning signal to the host computer when the output seal failure or partial obstruction judgment result is obtained, realize a complete closed loop from "monitoring and identification" to "risk warning" - the monitoring system not only completes the automatic judgment of the installation status, but also actively pushes the judgment result to the host device such as the monitor, anesthesia workstation or central monitoring system in the form of an early warning, so that clinical medical staff can obtain information on abnormal laryngeal mask installation status in a timely manner without having to continuously pay attention to the monitoring interface, thus shortening the time delay from the occurrence of abnormality to clinical intervention;

[0091] Especially in scenarios where anesthesiologists manage multiple monitoring indicators simultaneously during general anesthesia surgery, this proactive early warning mechanism effectively reduces the probability of missed risk reports due to distraction, thereby improving the level of patient safety.

[0092] Example 2

[0093] This invention provides a computer-readable storage medium including a memory for executing a program including the method for identifying the installation status of a high-frequency jet ventilator laryngeal mask according to Embodiment 1.

[0094] In summary, the method for identifying the installation status of a high-frequency jet ventilator laryngeal mask provided in this embodiment of the invention integrates airway pressure waveform characteristics and jet acoustic spectrum characteristics, eliminating the need for additional implanted sensors and enabling non-contact, continuous online monitoring. This method, tailored to the open-loop characteristics of high-frequency jet ventilation, incorporates multi-dimensional features such as peak pressure variation rate, energy spectrum shift, and acoustic Mel-Cepstral coefficients, effectively distinguishing between seal failure and partial obstruction. Furthermore, the method introduces a hidden Markov model and cubic spline interpolation to accurately identify, eliminate, and repair transient surgical interference, while maintaining stability and accuracy over long-term monitoring through rolling benchmark self-calibration.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; those skilled in the art can modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying the installation status of a laryngeal mask airway on a high-frequency jet ventilator, characterized in that, Includes the following steps: S10. During the initial calibration period after the start of high-frequency jet ventilation, the pressure timing signal output by the ventilator airway pressure sensor and the acoustic timing signal output by the jet acoustic sensor are collected. The collected pressure timing signals are subjected to baseline drift correction and noise filtering to establish a pressure waveform reference model and a frequency distribution reference model. The reference models together constitute the initial reference reference. S20. During normal operation of high-frequency jet ventilation, extract the peak pressure variation rate feature, energy spectrum shift feature and acoustic spectrum feature at the current moment in real time, and fuse the three types of features to form a dynamic feature vector V(t). S30. Perform a multi-dimensional comparison operation between the dynamic feature vector V(t) and the corresponding benchmark model in the initial reference benchmark, calculate the feature deviation comprehensive evaluation index δ(t), and output the sealing failure judgment result or partial blockage judgment result based on the relationship between the feature deviation comprehensive evaluation index δ(t) and the preset judgment threshold. S40. During the operation of steps S20 and S30, transient interference signals caused by surgical operations are detected in real time. The transient interference signals are marked and eliminated by the interference signal identification model to ensure the effectiveness of the dynamic feature vector V(t). The elimination results are fed back to the comparison operation in step S30 to correct the feature deviation comprehensive evaluation index δ(t).

2. The method according to claim 1, characterized in that, The dynamic feature extraction stage in step S20 further includes the following sub-steps: S21. Perform cycle-by-cycle peak detection on the real-time acquired airway pressure time-series signal P(t), extract the peak pressure sequence {Pmax,1,Pmax,2,…,Pmax,N} over N consecutive ventilation cycles, calculate and normalize the peak pressure difference between adjacent cycles, and obtain the peak pressure variation rate characteristic F. P ; S22. Perform a Fast Fourier Transform on the real-time pressure time-series signal P(t) within a sliding time window [t−Tw,t] to obtain the current energy power spectral density distribution E(f,t). Calculate the offset between the centroid frequency of the power spectrum of the current energy power spectral density distribution E(f,t) and the corresponding reference frequency in the initial reference, and obtain the energy spectrum offset characteristic quantity. ; S23. Acoustic spectral feature extraction and fusion: real-time extraction and fusion of acoustic time-series signals from jet acoustic sensors. Extracting Mel frequency cepstral coefficient sequence As a characteristic quantity of acoustic spectrum The peak pressure variation rate characteristic quantity Energy spectral shift characteristic quantity and acoustic spectral characteristics The dynamic feature vector is constructed by concatenating the features along their respective dimensions. .

3. The method according to claim 1, characterized in that, The feature comparison and anomaly determination stage in step S30 further includes the following sub-steps: S31. Based on the adaptive kernel density estimation algorithm, the pressure waveform benchmark model and frequency distribution benchmark model in the waveform feature analysis scenario are modeled with probability density distribution, and the dynamic feature vector is... Projecting the vector onto the feature space defined by the probability density distribution model, the dynamic feature vector is calculated. Deviation values ​​of each component in the pressure waveform dimension, energy spectrum dimension, and acoustic spectrum dimension. and The deviation values ​​of each component together constitute the characteristic deviation component set; S32. Calculate the comprehensive evaluation index of the feature deviation. Based on the set of feature deviation components, calculate the comprehensive evaluation index of the feature deviation according to the preset multidimensional fusion rules. ; S33. Then, the classification result is output, and the comprehensive evaluation index is based on the aforementioned feature deviation. Between the preset seal failure value and preset partial blocking threshold The size relationship between them is determined by a three-level decision logic: when... Output the seal failure judgment result at the time; when Output partial blocking determination results; when The system outputs a result indicating that the installation status is normal.

4. The method according to claim 1, characterized in that, The dynamic interference signal elimination stage in step S40 further includes the following sub-steps: S41. Identify surgical operation interference signals and model transient interference signal patterns in waveform feature analysis scenarios based on hidden Markov model algorithms, including real-time pressure time-series signals. With real-time acoustic timing signals The Hidden Markov Model is input synchronously to calculate the posterior probability that the signal at the current moment belongs to a surgical operation interference state. When the posterior probability Exceeding the preset interference identification threshold At that time, the data frame at the current moment is marked as an interference frame; S42. Perform a removal operation on all data segments marked as interference frames, and use cubic spline interpolation to smoothly fill the data gaps formed after removal, so as to restore the complete effective timing signal. S43. Set the time positions of the interference frames. The characteristic deviation comprehensive evaluation index is passed to step S30. In the calculation stage, for the above The deviation calculation results within the time period are shielded to ensure that interference signals do not participate in the final judgment logic calculation. The shielding results are synchronously updated to the initial reference benchmark to achieve rolling benchmark self-calibration.

5. The method according to claim 1, characterized in that, When the adaptive kernel density estimation algorithm models the waveform feature analysis scenario in step S31, it further includes the following sub-steps: S311. On the pressure waveform sample set acquired during the initial calibration period, the optimal kernel bandwidth for each feature dimension is calculated using the cross-validation maximum likelihood estimation method. For the first For each feature layer, the optimal kernel bandwidth is determined by the following formula: ； in To leave one cross-validation kernel density estimate, This represents the total number of samples during the calibration period. S312. In view of the characteristic that the high-frequency jet ventilation pressure waveform may exhibit a multi-peak distribution under different ventilation conditions, a mixture Gaussian kernel function is used instead of a single Gaussian kernel function to perform multimodal modeling of the probability density distribution of each feature layer. The number of components of the mixture Gaussian kernel function is determined by the Bayesian information criterion. S313. During normal operation, every preset update cycle. The latest effective feature vector samples are incrementally incorporated into the kernel density estimation model to achieve dynamic rolling updates of the baseline model. The incremental updates are only performed on data frames that are determined to be in normal installation status, ensuring that abnormal state data does not contaminate the baseline model.

6. The method according to claim 4, characterized in that, When the Hidden Markov Model models the surgical operation interference signal in step S41, the following sub-steps are further included: S411. Definition of Hidden Markov Model State Space: The hidden state space of the Hidden Markov Model is defined as follows: These correspond to normal ventilation, mild surgical interference, and severe surgical interference, respectively. The observation space consists of pressure time-series signals. Short-time statistical eigenvectors and acoustic time-series signals The short-time spectral eigenvectors are jointly constructed; S412. The Baum-Welch expectation-maximization algorithm is used to initialize the transition probability matrix of the hidden Markov model during the preoperative preparation phase. and initial state probability vector Offline pre-training is performed, and the recursive online Baum-Welch algorithm is used to fine-tune the model parameters in real time during the operation, so that the model can continuously adapt to the interference characteristics of the current surgical scenario. S413. In the scenario of high-frequency jet ventilation acoustic spectrum characteristic analysis, respectively at short time scales Mesoscale and long-term scale The Hidden Markov Model is run independently to decode and obtain the posterior probability of interference at three time scales. The final posterior probability is obtained by taking the geometric mean of the three values: The fusion posterior probability is used as the criterion for marking the interference frame in step S41.

7. The method according to claim 1, characterized in that, It also includes sending a warning signal to the host computer when the output seal failure judgment result or the output partial blockage judgment result is output.

8. A storage medium, characterized in that, Includes a memory for executing a program for the method of identifying the installation status of the laryngeal mask of a high-frequency jet ventilator according to any one of claims 1-7.

Images