Pressure adjusting and monitoring method and system for intubation of nasogastric tube

Abstract

The invention relates to the technical field of nasogastric tube intubation, and discloses a pressure adjusting and monitoring method and system for nasogastric tube intubation, and the pressure adjusting and monitoring method for nasogastric tube intubation comprises the steps: obtaining real-time pressure monitoring data in the nasogastric tube intubation process, and generating a pressure mode feature vector; mapping the pressure mode feature vector to a low-dimensional Hash space by using a region sensitive Hash algorithm to generate a Hash code; similar case retrieval is carried out in the multi-level hash index structure based on hash coding, and a candidate similar pressure mode set is output; analyzing the retrieved similar pressure mode and the corresponding historical control strategy, and generating a control strategy recommendation of the current intubation scene; according to the method, the technical problem of searching similar pressure modes in real time in a large-scale historical case library is solved, millisecond-level pressure mode recognition and control strategy recommendation are realized, and the decision-making efficiency and safety in the nasogastric tube intubation process are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of nasogastric tube intubation, more particularly, it relates to a pressure regulation and monitoring method and system for nasogastric tube intubation. BACKGROUND

[0002] In the clinical operation of nasogastric tube intubation, it is necessary to monitor and control the pressure change in the intubation process in real time to ensure the safety and success rate of intubation. A large amount of pressure monitoring data is generated during intubation, which contains rich pressure change patterns. Each pattern often corresponds to a specific physiological state and a corresponding control strategy. In actual application, when encountering a certain pressure pattern, if similar pressure patterns and their corresponding successful control strategies can be quickly retrieved from the historical case library, the safety and efficiency of intubation operation will be greatly improved.

[0003] However, the existing pressure pattern recognition and retrieval system has serious efficiency problems. The traditional sequential comparison method needs to calculate the similarity of the current pressure pattern with each case in the historical case library one by one. When the size of the historical case library reaches tens of thousands or even hundreds of thousands of records, the time complexity of this linear search method is , where is the number of historical cases. In the clinical scene that requires real-time response, such retrieval delay is unacceptable.

[0004] In addition, pressure pattern data usually has high-dimensional features, and traditional similarity calculation methods have high computational complexity in high-dimensional space, further exacerbating the retrieval efficiency problem. This causes the system to be unable to find similar historical pressure patterns within milliseconds, so it cannot provide effective control strategy suggestions to medical staff in time, affecting the safety and success rate of nasogastric tube intubation. SUMMARY

[0005] The present application provides a pressure regulation and monitoring method and system for nasogastric tube intubation, which solves the technical problem of low retrieval efficiency of traditional sequential comparison method in related technology.

[0006] The present application provides a pressure regulation and monitoring method and system for nasogastric tube intubation, which solves the technical problem of low retrieval efficiency of traditional sequential comparison method in related technology. Obtain real-time pressure monitoring data of the nasogastric tube intubation process, and generate a pressure pattern feature vector; Map the pressure pattern feature vector to a low-dimensional hash space using a region-sensitive hash algorithm, and generate a hash code; Based on the hash code, perform similar case retrieval in a pre-established multi-level hash index structure, and output a candidate similar pressure pattern set; Analyze the retrieved candidate similar pressure patterns and their corresponding historical control strategies to generate a control strategy recommendation for the current intubation scenario, and adjust the pressure based on the control strategy recommendation; The region-sensitive hashing algorithm includes a hash function based on random projection and a hash function based on a random hyperplane. The hash function is selected according to the similarity measurement method of the stress pattern feature vector.

[0007] Furthermore, the pressure mode feature vector includes a pressure mode time-domain feature vector, a pressure mode frequency-domain feature vector, and a correlation feature vector, and is generated by the following steps: The sliding window algorithm is used to calculate the time-domain feature vector of the pressure pattern from the real-time pressure monitoring data, including the pressure mean, variance, peak value and valley value within each window; The Fast Fourier Transform algorithm is used to extract features from real-time pressure monitoring data to obtain the frequency domain feature vector of the pressure mode, including the main frequency component and the spectral energy distribution. The corresponding autocorrelation function and cross-correlation function are calculated based on real-time pressure monitoring data, and a correlation feature vector is generated based on the autocorrelation function and cross-correlation function.

[0008] Furthermore, the hash function based on random projection Defined as: ; in, For feature vectors, It is a random vector sampled from a standard normal distribution. From Random numbers sampled from a uniform distribution It is the width parameter of the hash bucket. This indicates that a rounding down operation is being performed.

[0009] Furthermore, the hash function based on the random hyperplane Defined as: ; in, For feature vectors, It is a random vector sampled from a standard normal distribution. Represents a sign function, when Returns 1 when Returns -1 when Returns 0.

[0010] Furthermore, the pre-established multi-level hash index structure includes Each hash table uses a different combination of hash functions. The value ranges from 2 to 10 and is dynamically adjusted according to the size of the historical case library and the requirements for retrieval efficiency. The multi-level hash index structure is constructed based on the historical pressure pattern feature vectors and their corresponding hash codes in the historical case library.

[0011] Furthermore, the generation of the candidate similar pressure pattern set also includes the following steps: For each similar case in the candidate similar stress pattern set, calculate its similarity score with the stress pattern corresponding to the retrieved hash code; Sort the similarity scores and output the top-ranked similarity scores. The stress patterns of similar cases were used as the final search results, among which The value range is from 5 to 20.

[0012] Furthermore, the similarity score is calculated using one of the following similarity functions: Euclidean distance function : ; Cosine similarity function : ; Mahalanobis distance function : ; in, The retrieved hash code corresponds to the pressure pattern feature vector. For similar cases, the stress pattern feature vectors are... The covariance matrix of the eigenvectors. express The i-th component, express The i-th component, The dimension of the feature vector. This indicates the transpose operation.

[0013] Furthermore, the generation of the recommended control strategy includes the following steps: Extract the control strategy parameters used in each similar case, including intubation speed, angle adjustment, and pressure threshold; The parameters of each control strategy are normalized, and the normalized similarity scores of similar cases are used as weights to perform weighted fusion of the normalized control strategy parameters to generate recommended control strategy values ​​suitable for the current intubation scenario. The recommended values ​​of the control strategy are restored to their original dimensions through inverse transformation, and then truncated and adjusted in combination with the effective range of parameters in clinical practice.

[0014] Further, the generation of the control strategy recommendation further comprises the following steps: The correlation between similar cases is analyzed by using a collaborative filtering algorithm, a control strategy combination with a success rate exceeding a preset success rate threshold under a similar pressure mode is identified, and a control strategy recommendation based on case correlation is generated; The rare pressure mode case is input into the generative adversarial network, the generator in the generative adversarial network synthesizes virtual case data, and the virtual case data is integrated into the historical case library; Real-time feedback data after pressure regulation based on the control strategy recommendation is obtained, the multi-level hash index structure is updated based on the real-time feedback data, including adding new pressure mode corresponding cases and their control effects to the historical case library, and dynamically adjusting the hash function parameters used by the hash table in the multi-level hash index structure.

[0015] The present application provides a pressure regulation and monitoring system for nasogastric tube intubation, for performing the aforementioned pressure regulation and monitoring method for nasogastric tube intubation, comprising: A data acquisition module for acquiring real-time pressure monitoring data during nasogastric tube intubation; A feature generation module for generating a pressure pattern feature vector based on real-time pressure monitoring data; A hash code generation module for mapping the pressure pattern feature vector to a low-dimensional hash space using a region-sensitive hash algorithm to generate a hash code; wherein the region-sensitive hash algorithm includes a hash function based on random projection and a hash function based on random hyperplane, and the hash function is selected according to the similarity measurement method of the pressure pattern feature vector; A similar case retrieval module for retrieving similar cases based on the hash code in a pre-established multi-level hash index structure, and outputting a candidate similar pressure pattern set; A control strategy generation and regulation module for analyzing the retrieved candidate similar pressure patterns and their corresponding historical control strategies, generating a control strategy recommendation for the current intubation scenario, and adjusting the pressure based on the control strategy recommendation; A human-computer interaction interface for displaying the retrieval results and control strategy recommendation to medical personnel and receiving operation feedback. The present application has the following advantages: 1) The present application provides a pressure regulation and monitoring method and system for nasogastric tube intubation, which extracts features from real-time pressure monitoring data in time domain, frequency domain and correlation, improving the comprehensiveness and accuracy of the acquired data; 2) The application provides a pressure regulation and monitoring method and system for nasogastric tube intubation, which maps a pressure pattern feature vector to a low-dimensional hash space by using a region-sensitive hash algorithm, so as to reduce the storage space and calculation complexity of data; and the region-sensitive hash algorithm comprises a hash function based on random projection and a hash function based on random hyperplane, and the hash function is selected according to the similarity measurement mode of the pressure pattern feature vector, so that similar pressure pattern feature vectors have higher similarity in the hash space, thereby facilitating subsequent similar case retrieval; 3) The application provides a pressure regulation and monitoring method and system for nasogastric tube intubation, which carries out similar case retrieval in a pre-constructed multi-level hash index structure based on hash coding, and outputs a candidate similar pressure pattern set, wherein the multi-level hash index structure comprises a plurality of independent hash tables, each table uses a different hash function combination, the accuracy and efficiency of similar case retrieval are improved, and the number of independent hash tables can be dynamically adjusted according to actual conditions, so that the system can flexibly adapt to different sizes of historical case bases and retrieval efficiency requirements, and the adaptability and practicality of the system are enhanced; 4) The application provides a pressure regulation and monitoring method and system for nasogastric tube intubation, which comprehensively considers the parameters of similar cases, the correlation between cases and rare cases, generates control strategy recommendations in multiple ways, so that the recommended control strategy is more suitable for the current intubation scene; and the application also updates the multi-level hash index structure and the historical case base based on real-time feedback data, so that the system can continuously learn and adapt to new scenes, and realize continuous optimization and dynamic evolution of the system. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of the pressure regulation and monitoring method for nasogastric tube intubation of the application. DETAILED DESCRIPTION

[0017] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can be changed in function and arrangement without departing from the scope of the content of the present specification. Various processes or components can be omitted, replaced, or added according to needs of various examples. In addition, features described in some examples can also be combined in other examples.

[0018] Example 1: The present embodiment provides a pressure regulation and monitoring method for nasogastric tube intubation, as shown in Figure 1 The method comprises the following steps: Step 100: Obtain real-time pressure monitoring data of the nasogastric tube intubation process, and generate a pressure pattern feature vector.

[0019] The step collects real-time pressure monitoring data in the cannulation process through a pressure sensor, including pressure value sequence, pressure change rate, pressure peak value, pressure waveform characteristics and other parameters. Before feature extraction, the original pressure data (i.e. real-time pressure monitoring data) is preprocessed, including denoising, outlier detection and processing, data smoothing and standardization processing, to eliminate data deviation caused by different measurement environments and equipment differences. The standardization processing adopts Z-score standardization method to convert each pressure measurement value into a standard normal distribution with a mean of 0 and a standard deviation of 1, ensuring that pressure data of different dimensions and ranges can be effectively compared and calculated. After preprocessing, the data is feature extracted to generate a feature vector describing the current pressure pattern. The dimension of the feature vector is determined according to the actual application requirements, usually containing time domain features, frequency domain features and statistical features and other multiple dimension information.

[0020] It should be noted that the generation of the pressure pattern feature vector can use multiple feature extraction methods.

[0021] For example, step 101: using a sliding window method to analyze the pressure data sequence of real-time pressure monitoring data, calculating the mean, variance, peak and valley of pressure in each window, and outputting a time domain feature vector of pressure pattern; Step 102: input the pressure data sequence into the fast Fourier transform algorithm to extract the main frequency component and frequency spectrum energy distribution, and output the frequency domain feature vector of pressure pattern; Step 103: calculate the autocorrelation function and cross-correlation function based on the pressure data sequence, identify the periodic pattern of pressure change, and output the correlation feature vector.

[0022] Further, the sliding window method in step 101 receives input as time series : ; Wherein , , are the 1st, 2nd, time series unit, and the output is the feature vector : ; Wherein , , correspond to the 1st, 2nd, mth feature of , each feature corresponds to a statistical quantity such as mean, variance, peak and valley.

[0023] Further, the fast Fourier transform algorithm in step 102 receives input as time series , and the output is the frequency domain feature vector : ; where 、 、 denote the amplitude of the 1st, 2nd, kth corresponding frequency component, respectively.

[0024] Further, the autocorrelation function and cross-correlation function analysis in step 103 receives the input as a time series ; The output is a correlation feature vector: ; where 、 、 denote the correlation coefficient under the 1st, 2nd, kth time delay, respectively.

[0025] Step 200: Map the stress pattern feature vector to a low-dimensional hash space using a locality sensitive hashing algorithm to generate a hash code.

[0026] In this step, the locality sensitive hashing (LSH) algorithm is used to map the high-dimensional stress pattern feature vector to a low-dimensional hash space. To ensure the effectiveness of the hash mapping, first, the feature vector generated in step 100 is normalized to scale the feature values of different dimensions to a unified numerical range, eliminating the influence of dimensional differences on hash calculation. For time domain, frequency domain and correlation features, the min-max normalization method is used to map them to the [0, 1] interval, ensuring that each type of feature has the same weight in hash calculation. The LSH algorithm uses a special hash function family to make similar feature vectors have a higher probability of being mapped to the same hash bucket, and dissimilar feature vectors have a lower probability of being mapped to the same hash bucket. For an input dimensional feature vector , use independent hash functions to map it to generate a hash code of bits : ; 、 、 denote the 1st, 2nd, kth bit of the hash code.

[0027] It should be noted that the selection of the locality sensitive hashing function depends on the similarity measurement method of the feature vector.

[0028] Step 201: When using Euclidean distance to measure similarity, use a hash function based on random projection ​: ; where is a random vector sampled from the standard normal distribution, is a random number sampled from the uniform distribution, is the width parameter of the hash bucket, denotes the floor function that takes the integer part of ; Step 202: When the cosine similarity measure is adopted, the hash function based on random hyperplane is used : ; where is a random vector sampled from the standard normal distribution, denotes the sign function that returns 1 when , -1 when , and 0 when .

[0029] Further, the input of the region-sensitive hash algorithm in this embodiment is the pressure pattern feature vector : ; This is a high-dimensional vector that contains time-domain features, frequency-domain features, and correlation features; The output of the algorithm is a low-dimensional hash code : ; where , , denote the 1st, 2nd, zk-bit of the low-dimensional hash code , which is a binary value of 0 or 1, is much smaller than the original dimension of the feature vector.

[0030] Further, the random projection hash function in step 201 receives the input of the feature vector , its corresponding random vector , and random offset , and outputs an integer hash value; the random projection hash function calculates the projection of the feature vector on the random vector, then quantizes it, and discretizes the continuous feature space into a finite number of hash buckets. To ensure the dimensional consistency of the calculation, the random vector is constructed with the same dimension as the normalized feature vector , and each component is sampled from the standard normal distribution, so that the dot product becomes a dimensionless quantity; the random offset ​And bucket width parameter Also designed as dimensionless parameter, ensure the entire hash function dimensional consistency. In the specific implementation, the random vector The same dimension as the feature vector, the element obeys the standard normal distribution; random offset From Uniform sampling interval; bucket width parameter According to the data distribution characteristics, the value is 4.0 in this embodiment, which is determined by experimental test, and can be adjusted between 1.0 and 10.0. Smaller Value will make the hash more sensitive (resulting in more hash buckets), while larger Value makes the hash more robust (resulting in fewer hash buckets); when the feature space dimension is high (such as more than 100 dimensions), the projection dimension decreasing strategy is adopted, and each hash function only uses a random subset of the feature vector to calculate the projection, so as to reduce the calculation complexity, and the subset size is usually set to 10%-30% of the total dimension.

[0031] Further, the random hyperplane hash function in step 202 receives the input of the feature vector And its corresponding random vector , the output is a binary hash value (1 or-1); the random hyperplane hash function realizes the bisection of the feature space by judging which side of the random hyperplane the feature vector is located, so as to maintain the cosine similarity relationship between vectors. In practical application, the generation of random vector Adopt pseudo-random number generator combined with fixed seed value (such as seed value set to 42) to ensure the reproducibility of the hash function, which has the same dimension as the feature vector , each component is independently sampled from the standard normal distribution ; In order to improve the stability of the hash, the input feature vector First, the L2 normalization processing is carried out, that is , the normalized feature vector Is obtained, which eliminates the interference caused by the difference in vector length; in order to deal with the boundary condition that the dot product result is exactly 0, the rule is set to return 1 when , which ensures the complete definition of the hash function.

[0032] Step 300: Based on the hash coding, the similar case retrieval is carried out in the multi-level hash index structure, and the candidate similar pressure pattern set is output.

[0033] This step uses the generated hash coding to carry out fast retrieval in the pre-established multi-level hash index structure. The multi-level hash index structure contains one independent hash table, each hash table uses different hash function combination. For the retrieved stress pattern feature vector, locate to the corresponding hash bucket in each hash table, and combine all the historical cases in all hash buckets into a candidate set. This multi-level index structure reduces the time complexity of retrieval from to , where is the average number of cases in each hash bucket, is the total number of historical cases, usually .

[0034] In the embodiments of the present application, in order to further improve the retrieval accuracy, step 301 is further included on the basis of step 300: For each historical case in the candidate similar stress pattern set, calculate its exact similarity score with the query stress pattern, sort according to the similarity score, and output the top stress patterns with the highest similarity as the final retrieval result. In the present embodiment, the value range is 5 to 20, which is set according to the complexity of the specific clinical scene and the decision-making time requirement. In the emergency scene, a smaller value (such as 5-10) is taken to quickly give recommendations, and in the regular operation, a larger value (such as 15-20) is taken to improve the diversity and comprehensiveness of the recommendations. The similarity score can be calculated by the function , which returns the similarity value between two feature vectors according to the selected metric method (such as Euclidean distance, cosine similarity, etc.).

[0035] Further, the multi-level hash index structure is a data structure composed of one independent hash table, each hash table uses different hash function combination, which improves the recall rate of retrieval. In the present embodiment, the value range is 2 to 10, which is dynamically adjusted according to the size of the historical case library and the retrieval efficiency requirement. When the case library size is large (such as more than 100,000 records), the value is close to the upper limit to improve the recall rate of retrieval. The input received by the multi-level hash index structure is the hash code of the stress pattern; the output is the candidate similar case set : ; wherein , , are the 1st, 2nd, and mth historical stress pattern cases, respectively.

[0036] Further, the exact similarity calculation in step 301 receives the input of the retrieved stress pattern feature vector and the feature vector of each case in the candidate case set , the output is a similarity score ; To ensure the dimensional consistency between different feature dimensions, the feature vector is twice normalized before calculating the similarity to eliminate any possible dimensional differences.

[0037] According to the application scenario, the similarity calculation can use the Euclidean distance function : ; Pressure pattern feature vector and each case feature vector in the candidate case set , represents the i-th component of , represents the i-th component of . Cosine similarity function

[0038] : ; Mahalanobis distance function : ; These measurement methods can accurately reflect the similarity between pressure patterns, where represents the covariance matrix of the features. Among them, the cosine similarity function naturally eliminates the dimensional influence by dividing the dot product of the vectors by the product of the vector lengths; the Mahalanobis distance function appropriately compensates for the dimensional differences and correlations of each feature dimension by introducing the inverse matrix of the covariance matrix as the weight, represents the transpose operation.

[0039] Step 400: Analyze the retrieved similar pressure patterns and their corresponding historical control strategies to generate a control strategy recommendation for the current intubation scene.

[0040] ​This step analyzes similar historical cases and extracts the control strategy parameters used in each case, including intubation speed, angle adjustment, and pressure threshold. Since these control strategy parameters have different units and dimensions (e.g., intubation speed in mm / s, angle in degrees, and pressure threshold in kPa), each parameter needs to be normalized before fusion to eliminate the impact of dimensional differences on the weighted average. These normalized control strategy parameters are then weighted and fused according to similarity weights to generate recommended control strategy values ​​suitable for the current intubation scenario. Finally, the fused normalized parameter values ​​are restored to their original dimensions through inverse transformation and truncated based on the effective range of parameters in clinical practice to ensure that the recommended control parameters are within a safe and effective range. The output control strategy includes specific operational parameters and implementation suggestions, presented in an intuitive and easy-to-understand manner for medical staff to reference.

[0041] In this embodiment of the application, in order to improve the accuracy of the control strategy recommendation, step 401 is further included in addition to step 400: By using collaborative filtering algorithms to analyze the correlation between similar cases, we can identify control strategy combinations that have a success rate exceeding a preset success rate threshold (e.g., 80%) under similar stress patterns and generate strategy recommendations based on case correlation. Step 402: Input rare stress pattern cases into the generative adversarial network, and synthesize similar but varied virtual case data through the generator network to expand the case library of rare scenarios and improve the recommendation ability in scarce scenarios. Step 403: Update the hash index structure based on real-time feedback data, add new stress mode cases and their control effects to the historical case library, dynamically adjust the hash function parameters, and continuously optimize retrieval accuracy and recommendation quality.

[0042] Step 403 specifically employs a time decay weighting model, assigning different weights to historical cases at different time points. The weight calculation formula is as follows:

[0043] in Indicates the current time. Timestamps representing historical cases The time decay coefficient ranges from [0.001, 0.1], and is set to 0.01 in this embodiment. The hash function parameters aim to maximize retrieval recall, minimize normalized retrieval time, and maximize retrieval precision. The dynamic adjustment of the hash function parameters employs a time-based multi-objective optimization method, with the objective function being... : ; in It is the number of hash tables. is the hash code length, is the hash bucket width parameter, represents the retrieval recall rate, represents the normalized retrieval time, represents the time relevance score, which measures the retrieval accuracy of the most recent cases, , and are the coefficients for balancing the respective objectives, satisfying In the present embodiment, they are respectively set to 0.4, 0.3 and 0.3.

[0044] Further, the input of the weighting fusion strategy in step 400 is the set of retrieved similar cases : ; and the corresponding similarity scores : ; and the control strategy parameters P: ; The output is the fused control strategy recommendation ; , , represent the 1st, 2nd, similar cases, respectively; , , represent the similarity scores of the 1st, 2nd, similar cases, respectively, , , represent the control strategy parameters of the 1st, 2nd, respectively. The weighting fusion strategy ensures that the control strategy of a more similar case obtains a higher influence weight through similarity weighted average.

[0045] Specifically, for each control strategy parameter , first, the minimum-maximum normalization processing is performed to obtain the normalized control strategy parameter :

[0046] wherein and are the minimum and maximum values of the parameter in all candidate cases, respectively; Then, the normalized similarity score : is used. ; As weights, the weighted average formula is used : ; The recommended value of the parameter is calculated; Finally, the recommended value of the normalized space is converted back to the original parameter space to obtain the recommended value of the normalized parameter : ; According to the effective range of the parameter, threshold limiting is performed to ensure that the final recommended value is within the safe and effective range.

[0047] Further, the collaborative filtering algorithm in step 401 is a recommendation algorithm based on historical data analysis of user preferences or item similarity. In the present embodiment, the input received by the collaborative filtering algorithm is the pressure mode-control strategy matrix , where each row represents a pressure mode and each column represents a control strategy parameter, and the matrix element represents the success degree of using the control strategy under the pressure mode. Since the different control strategy parameters (such as speed, angle, pressure threshold, etc.) in the matrix have different dimensions and score ranges, before applying the collaborative filtering algorithm, the matrix is first column-normalized to convert the success degree scores of each control strategy parameter into a unified 0-1 range, eliminating the influence of dimensional differences on similarity calculation. The column normalization process specifically adopts the Z-score standardization method, that is, for each column of the matrix , the mean and standard deviation are calculated, and then each element of the column is transformed: ; Then the standardized value is mapped to the interval through the Sigmoid function, as the Sigmoid function is a common technical means in the art and will not be described here. The output of the collaborative filtering algorithm is the optimal control strategy combination for the current pressure mode, which is a recommended value in the normalized space, and needs to be mapped back to the original parameter space through inverse transformation and converted into specific clinical operation guidance, such as inserting at a speed of 2.5 mm / s, maintaining an angle of 15 degrees, pausing and waiting for 2 seconds when the pressure exceeds 1.8 kPa, and other instructions that medical personnel can directly execute. The collaborative filtering algorithm uses a similarity function to calculate the similarity between the pressure mode and , which in the present embodiment actually adopts cosine similarity, that is: ; where is the total number of control policy parameters; predict the current stress pattern by weighted average function score of the control policy , which is expressed in formula as: ; where is the neighbor set most similar to the current stress pattern. In this embodiment, the size of the neighbor set ranges from 10 to 50, which is dynamically adjusted according to the data distribution characteristics and algorithm performance: and represent the element in the i-th row and s-th column and the element in the j-th row and s-th column of the stress pattern-control policy matrix ; When the data is intensive and the similarity distribution is uniform, a smaller value (such as 10-20) is taken to improve the recommendation accuracy; When the data is sparse or the similarity distribution is uneven, a larger value (such as 30-50) is taken to increase the robustness of the recommendation.

[0048] The value range of the similarity function is [-1, 1], where 1 represents complete similarity, 0 represents no association, and -1 represents complete opposition.

[0049] Further, the generative adversarial network in step 402 is composed of a generator and a discriminator. The input of the generator network is a random noise vector and a rare stress pattern feature , and the output is a synthesized stress pattern feature ; the input of the discriminator network is a stress pattern feature (real or synthesized), and the output is the probability that the feature is real data. The training goal of the generative adversarial network is to enable the generator network to generate synthesized data that the discriminator network cannot distinguish between true and false. In actual implementation, the random noise vector is sampled using a standard normal distribution with a dimension of 100, i.e. ; The rare stress pattern feature is reduced to 50 dimensions through principal component analysis, retaining more than 90% of the variance information to ensure that the main feature structure is captured while reducing computational complexity; The training of the generative adversarial network uses a loss function WGAN-GP based on the Wasserstein distance, where the gradient penalty coefficient , and the number of iterations of the discriminator The learning rate of the generator and the discriminator is set to 0.0002, and the training stability is ensured by batch normalization and gradient clipping (threshold value is 0.01).

[0050] Further, the generator network is a feedforward neural network, and the layer structure thereof comprises: an input layer receiving random noise vectors and conditional information of rare pressure pattern features, wherein the rare pressure pattern features are normalized before input; a plurality of fully connected hidden layers, each followed by batch normalization and a ReLU activation function, for feature conversion and nonlinear mapping; an output layer generating a synthetic feature vector of the same dimension as the real pressure pattern features, using a Tanh activation function to scale the output value to the range of [-1, 1].

[0051] The generated synthetic feature vector needs to be converted back to the original feature space through inverse normalization processing, and then converted into a clinically interpretable pressure pattern representation through a feature decoder, including typical pressure waveform parameters, key point time series, and pressure change pattern descriptions, to facilitate intuitive understanding and use by medical personnel. The training of the generator network uses an adversarial loss function : ; wherein denotes the mathematical expectation, denotes the generator function, denotes the discriminator function, and the goal is to maximize the probability that the discriminator judges the generated sample as a real sample. In the actual training process, in order to ensure the continuity and physical reasonableness of the generated pressure pattern features in the time dimension, a time smoothness constraint term : ; is added to the generator loss function, wherein denotes the value of the synthetic feature at time point , is a weight coefficient of the smoothness constraint, and the value range is [0.01, 0.1], which is set to 0.05 in the embodiment. The complete generator loss function is : ; By optimizing this loss function, the generated pressure pattern is ensured to maintain authenticity while meeting the time continuity constraint.

[0052] Further, the discriminator network is a feedforward neural network, and the layer structure thereof comprises: an input layer receiving pressure pattern feature vectors; A plurality of fully connected hidden layers, each followed by a Leaky ReLU activation function, to extract discriminative representations of features; An output layer using a Sigmoid activation function to output a probability value between 0 and 1, representing the likelihood of the input data being real data. The training of the discriminator network uses a binary cross-entropy loss function : ; The goal is to maximize the log probability of real samples and maximize the log inverse probability of synthetic samples. To ensure the sensitivity of the discriminator to the temporal characteristics, a time correlation constraint is added to the discriminator loss function : ; wherein and represent the values of the real and synthetic features at time point , and are the feature values at the previous time point, is the weight coefficient of the time correlation constraint, which is in the range of [0.1, 0.5], and is set to 0.3 in the present embodiment. The complete discriminator loss function is: ; By optimizing this loss function, the discriminator can identify the physical reasonableness of the stress pattern in the time dimension.

[0053] It can be understood that the data preprocessing methods known to those skilled in the art include data cleaning, data conversion, and data reduction, wherein the data conversion includes type conversion and normalization and standardization, although the dimensions and types of the data are ignored in the description of the foregoing embodiments, the data preprocessing is a technical knowledge known to those skilled in the art and a prerequisite for data processing, and therefore the known data preprocessing steps are not independently described in the foregoing content.

[0054] The embodiments of the present application are described above, but the embodiments are not limited to the specific embodiments described above, and the specific embodiments described above are merely illustrative and not limiting, and those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection of the embodiments.

Claims

1. A method of pressure regulation and monitoring for intubation of a nasogastric tube, characterized by, The method comprises the following steps: obtaining real-time pressure monitoring data of a nasogastric tube intubation process to generate a pressure pattern feature vector; mapping the pressure pattern feature vector to a low-dimensional hash space using a region-sensitive hash algorithm to generate a hash code; performing similar case retrieval in a pre-established multi-level hash index structure based on the hash code to output a set of candidate similar pressure patterns; analyzing the retrieved candidate similar pressure patterns and their corresponding historical control strategies to generate a control strategy recommendation for the current intubation scenario, and adjusting the pressure based on the control strategy recommendation; wherein the region-sensitive hash algorithm comprises a hash function based on random projection and a hash function based on random hyperplane, and the hash function is selected according to the similarity measurement method of the pressure pattern feature vector.

2. The method of claim 1, wherein, The pressure pattern feature vector comprises a pressure pattern time domain feature vector, a pressure pattern frequency domain feature vector and a correlation feature vector, and is generated by the following steps: using a sliding window algorithm to calculate the real-time pressure monitoring data to obtain the pressure pattern time domain feature vector, including the pressure mean, variance, peak and valley in each window; using a fast Fourier transform algorithm to extract features from the real-time pressure monitoring data to obtain the pressure pattern frequency domain feature vector, including the main frequency component and the frequency spectrum energy distribution; calculating the corresponding autocorrelation function and cross-correlation function based on the real-time pressure monitoring data, and generating the correlation feature vector based on the autocorrelation function and the cross-correlation function.

3. The method of claim 1, wherein, The random projection based hash function is defined as: ； wherein, is a feature vector, is a random vector sampled from a standard normal distribution, is a random number sampled from a uniform distribution, is a width parameter of the hash bucket, denotes a floor operation.

4. The method of claim 1, wherein, The hash function based on random hyperplanes is defined as: ； wherein, is a feature vector, is a random vector sampled from a standard normal distribution, denotes a sign function that returns 1 when returns -1 when returns 0 when .

5. The method of claim 1, wherein, The pre-established multi-level hash index structure comprises independent hash tables, each hash table using a different hash function combination, wherein the value range of k is 2 to 10, which is dynamically adjusted according to the size of the historical case library and the retrieval efficiency requirement, and the multi-level hash index structure is constructed based on the historical stress pattern feature vectors in the historical case library and the corresponding hash codes.

6. The method of claim 1, wherein, The generation of the set of candidate similar pressure patterns further comprises the following steps: for each similar case in the set of candidate similar pressure patterns, calculate the similarity score of the similar case with the retrieved hash code corresponding pressure pattern; ranking the similarity scores and outputting as final search results the stress patterns of the top cases with the highest similarity scores, wherein the value of k ranges from 5 to 20.

7. The method of claim 6, wherein, The similarity score is calculated using one of the following similarity functions: Euclidean distance function : ； Cosine similarity function : ； Mahalanobis distance function : ； wherein, is the hash code corresponding to the pressure pattern feature vector of the retrieved case, is the pressure pattern feature vector of the similar case, denotes the covariance matrix of the feature vector, denotes the i-th component of denotes the i-th component of denotes the dimension of the feature vector, denotes the transpose operation.

8. The method of claim 6, wherein, The generation of the control strategy recommendation comprises the following steps: extracting the control strategy parameters used in each similar case, including intubation speed, angle adjustment, pressure threshold; normalizing each control strategy parameter, and weighting and fusing the normalized control strategy parameters using the normalized similarity score of the similar case as the weight to generate a control strategy recommendation value suitable for the current intubation scenario; restoring the control strategy recommendation value to the original dimension through inverse transformation, and truncating and adjusting it in combination with the effective range of parameters in clinical practice.

9. The method of claim 5, wherein, The generation of the control strategy recommendation further comprises the following steps: using a collaborative filtering algorithm to analyze the association between similar cases, identifying control strategy combinations with a success rate exceeding a preset success rate threshold under similar pressure patterns, and generating a control strategy recommendation based on case association; inputting rare pressure pattern cases into a generative adversarial network, synthesizing virtual case data through a generator in the generative adversarial network, and integrating the virtual case data into a historical case library; obtaining real-time feedback data after adjusting the pressure based on the control strategy recommendation, updating the multi-level hash index structure based on the real-time feedback data, including adding new pressure pattern corresponding cases and their control effects to the historical case library, and dynamically adjusting the hash function parameters used by the hash tables in the multi-level hash index structure.

10. A pressure regulation and monitoring system for nasogastric tube intubation for performing the pressure regulation and monitoring method for nasogastric tube intubation according to any one of claims 1 to 9, characterized in that, The method comprises the following steps: a data acquisition module for obtaining real-time pressure monitoring data of a nasogastric tube intubation process; The feature generation module is configured to generate a pressure pattern feature vector based on real-time pressure monitoring data; The hash code generation module is configured to map the pressure pattern feature vector to a low-dimensional hash space by using a region-sensitive hash algorithm to generate a hash code; wherein the region-sensitive hash algorithm includes a hash function based on random projection and a hash function based on random hyperplane, and the hash function is selected according to a similarity measurement manner of the pressure pattern feature vector; The similar case retrieval module is configured to perform similar case retrieval in a pre-established multi-level hash index structure based on the hash code, and output a candidate similar pressure pattern set; The control strategy generation and adjustment module is configured to analyze the retrieved candidate similar pressure pattern and the corresponding historical control strategy, generate a control strategy recommendation for the current cannulation scene, and the medical staff performs pressure adjustment based on the control strategy recommendation; The human-computer interaction interface is configured to display the retrieval result and the control strategy recommendation to the medical staff, and receive operation feedback.

Images