Dynamic lung compliance monitoring method based on electrical impedance tomography

Abstract

The invention provides a dynamic lung compliance monitoring method based on electrical impedance tomography, and the method comprises the steps: obtaining an electrical impedance signal and pressure-volume curve data, recording a time sequence signal of lung tissue in a dynamic ventilation process through a synchronous collection device, and obtaining an original multi-dimensional data set; aiming at the second feature set, eliminating redundant parameters by applying a feature screening mechanism, and evaluating the importance of each feature in different breathing stages through an information gain algorithm to obtain a third feature set; constructing a dynamic compliance monitoring model through the weighted feature set, and fitting a lung compliance change trend by using a support vector regression algorithm to obtain a real-time prediction result; and according to the updated prediction result, evaluating the improvement degree of the calculation efficiency, and judging whether the real-time performance meets clinical requirements or not by recording the model operation time and the resource occupancy rate so as to obtain final monitoring output.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a dynamic lung compliance monitoring method based on electrical impedance tomography. Background Art

[0002] Problem background:

[0003] As a key area in respiratory physiology research and clinical diagnosis and treatment, lung compliance monitoring has irreplaceable value in revealing the state of lung function and guiding mechanical ventilation strategies. With the advancement of intensive care and respiratory support technology, the need for dynamic monitoring of changes in lung compliance has become increasingly prominent, especially in personalized ventilation therapy, where its accuracy directly affects the patient's prognosis. Electrical impedance tomography technology has become a research hotspot in this field due to its non-invasive and real-time characteristics. However, existing methods often expose obvious shortcomings when faced with complex respiratory patterns. Traditional monitoring methods mostly rely on single-dimensional signal analysis, such as being based only on pressure-volume curves or electrical impedance data, which makes it difficult to fully reflect the true compliance changes of lung tissue during dynamic ventilation. In addition, existing algorithms generally have problems with low efficiency and poor adaptability in feature extraction and processing, resulting in limited reliability of monitoring results under different ventilation modes.

[0004] The root of these limitations lies in several core challenges. First, the complementary features between the pressure-volume curve and the electrical impedance signal are difficult to effectively integrate, and the lack of a multi-dimensional cross-analysis mechanism leads to insufficient information utilization. Second, the importance of characteristic parameters is difficult to dynamically adjust in different respiratory stages, making the algorithm unable to adapt to the time-varying characteristics of lung tissue compliance. Finally, the accumulation of redundant features increases the computational burden and weakens the monitoring accuracy, which is particularly prominent in clinical scenarios with high real-time requirements. These unresolved technical factors have jointly led to the difficulty in balancing the accuracy and efficiency of dynamic lung compliance monitoring under complex ventilation conditions, forming a technical problem that urgently needs to be broken through.

[0005] Therefore, how to design an efficient cross-feature extraction algorithm to fuse the complementary information of the pressure-volume curve and the electrical impedance signal through multi-dimensional feature screening and adaptive weight adjustment, while eliminating redundant parameters to improve computational efficiency and monitoring accuracy, has become a key issue that needs to be urgently addressed in the dynamic lung compliance monitoring method based on electrical impedance tomography. Summary of the Invention

[0006] The present invention provides a dynamic lung compliance monitoring method based on electrical impedance tomography, which mainly includes:

[0007] Obtain electrical impedance signals and pressure-volume curve data, and use synchronous acquisition equipment to record the time series signals of lung tissue during dynamic ventilation to obtain the original multidimensional data set;

[0008] For the original multidimensional data set, signal preprocessing technology is used to remove noise interference and baseline drift, and the dynamic change characteristics within a continuous time period are extracted through the sliding window method to obtain the first feature set;

[0009] Based on the first feature set, a cross-analysis method is constructed to fuse the electrical impedance signal and the pressure-volume curve information. The complementary characteristics between the two types of signals are determined by calculating the correlation coefficient to obtain the second feature set.

[0010] For the second feature set, the feature screening mechanism is applied to eliminate redundant parameters, and the importance of each feature in different breathing stages is evaluated by the information gain algorithm to obtain the third feature set;

[0011] Obtain the feature importance distribution in the third feature set, dynamically update the weight coefficient using an adaptive weight adjustment method, and reflect the time-varying characteristics through the weighted average of the time series to obtain a weighted feature set;

[0012] By using weighted feature sets, a dynamic compliance monitoring model is constructed, and the support vector regression algorithm is used to fit the trend of lung compliance changes to obtain real-time prediction results.

[0013] If the deviation between the real-time prediction result and the actual measurement value exceeds the preset threshold, the correlation coefficient calculation formula in the cross-analysis method is adjusted, and the complementary feature extraction is optimized by introducing a nonlinear mapping function to obtain an updated prediction result;

[0014] Based on the updated prediction results, the degree of improvement in computing efficiency is evaluated, and by recording the model running time and resource occupancy rate, it is determined whether the real-time performance meets clinical needs to obtain the final monitoring output.

[0015] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0016] The present invention discloses a dynamic lung compliance monitoring method based on electrical impedance signals and pressure-volume curves. The method records multidimensional data during dynamic ventilation through synchronous acquisition equipment, and constructs a dynamic compliance monitoring model through signal preprocessing, feature extraction and cross-analysis. The present invention uses a support vector regression algorithm to fit the trend of lung compliance changes and achieves real-time prediction. When the deviation between the predicted result and the actual measured value is too large, the present invention can adaptively adjust the feature extraction method to optimize the predicted result. This method not only improves the accuracy and real-time performance of lung compliance monitoring, but can also be dynamically adjusted according to clinical needs, providing important technical support for the diagnosis and treatment of respiratory diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The flowchart of the dynamic lung compliance monitoring method based on electrical impedance tomography of the present invention is shown.

[0018] Figure 2 Schematic diagram of a dynamic lung compliance monitoring method based on electrical impedance tomography according to the present invention.

[0019] Figure 3 This is another schematic diagram of a dynamic lung compliance monitoring method based on electrical impedance tomography according to the present invention. DETAILED DESCRIPTION

[0020] The following will describe the technical solutions in the embodiments of the present invention in detail with reference to the accompanying drawings. The described embodiments are only a part of the embodiments of the present invention.

[0021] like Figure 1-3 In this embodiment, a dynamic lung compliance monitoring method based on electrical impedance tomography may specifically include:

[0022] S101. Acquire electrical impedance signals and pressure-volume curve data, and record time series signals of lung tissue during dynamic ventilation using a synchronous acquisition device to obtain an original multidimensional data set.

[0023] The time series signals of lung tissue during dynamic ventilation are acquired through synchronous acquisition equipment to obtain the original multidimensional data set. The electrical impedance signal and pressure-volume curve data are extracted from the original multidimensional data set to determine the signal feature vector. The principal component analysis method is used to perform dimensionality reduction processing on the signal feature vector to obtain a feature data set after dimensionality reduction. If there are outliers in the feature data set after dimensionality reduction, they are filtered out using a preset threshold to obtain a filtered feature data set. A time series model is constructed based on the filtered feature data set, and the support vector machine algorithm is used to determine the change trend during the ventilation process. The correlation between the lung tissue signal and the electrical impedance signal is analyzed by the change trend to obtain a correlation coefficient matrix. The dynamic mapping relationship between the pressure-volume curve data and the lung tissue signal is determined based on the correlation coefficient matrix to obtain the quantitative characteristics of the ventilation process.

[0024] Specifically, a 16-electrode electrical impedance tomography system was used to acquire lung tissue impedance signals during dynamic ventilation at a sampling rate of 100 frames / second. Simultaneously, airway pressure and tidal volume data were recorded synchronously at 500 Hz using an implantable pressure sensor and volume flowmeter. After removing high-frequency noise from the raw data using a fourth-order Butterworth low-pass filter (cutoff frequency 50 Hz), the impedance signals were time-series aligned with the pressure-volume curve using a dynamic time warping algorithm with a time window of 10 ms. A modified Gram angular field transform (ADAFT) was used to transform the 128×128 pixel two-dimensional impedance image sequence into Gram matrix features. Spatiotemporal features were then extracted using a three-dimensional convolutional neural network (kernel size 3×3×3, stride 2). The pressure-volume curve was resampled to 1000 data points using piecewise cubic Hermite interpolation, and a four-dimensional tensor (number of samples × time steps × feature dimension × number of channels) was constructed based on the impedance features. A long short-term memory network (256 hidden units) was used to map impedance changes to lung compliance. Training was performed using an adaptive moment estimation optimizer (initial learning rate 0.001, β1 = 0.9, β2 = 0.999). Model performance was evaluated through five-fold cross-validation, ultimately achieving a root mean square error (RMSE) of 0.38 cmH2O / mL on the test set. Outlier detection based on the Mahalanobis distance (threshold set at 3 standard deviations) was used for anomalous data points, and data repair was performed using a Kalman filter. All analytical procedures were implemented in Python using the TensorFlow 2.4 framework, with computational nodes configured on NVIDIA Tesla V100 graphics cards.

[0025] S102. For the original multidimensional data set, use signal preprocessing technology to remove noise interference and baseline drift, and use a sliding window method to extract dynamic change features within a continuous time period to obtain a first feature set.

[0026] Data is obtained from the original multidimensional dataset and filtered to remove noise and baseline drift, resulting in a second dataset. For the second dataset, the time series is segmented using a sliding window method to extract dynamic change features, resulting in a third dataset. If the dynamic features in the third dataset exceed a preset threshold, principal component analysis is used for dimensionality reduction, resulting in a fourth dataset. Based on the fourth dataset, the trend vector of the time series is calculated to determine the change trend. Based on the change trend, a clustering algorithm is used to group the fourth dataset to obtain grouping results. Based on the grouping results, the time series feature distribution of each group is obtained to determine feature stability. Based on feature stability, stable dynamic features are selected to obtain the final feature set.

[0027] Specifically, the original multidimensional data set was first preprocessed using wavelet transform. Using the db4 wavelet basis function, the signal was decomposed into a five-layer structure to remove high-frequency noise components while retaining the low-frequency useful signal. To address baseline drift, a median filter algorithm was used with a window length of 100 samples to smooth the signal, effectively eliminating baseline drift. Next, a sliding window method was used to extract dynamic features, with a window length of 256 samples and a step size of 64 samples. Within each window, time domain features (such as mean, variance, and peak-to-peak value) and frequency domain features (such as power spectral density and spectral centroid) were calculated, resulting in a total of 12 features. To enhance feature representation, the extracted features were normalized using the z-score method, ensuring that each feature had a mean of 0 and a standard deviation of 1. Finally, principal component analysis (PCA) was used to reduce the dimensionality of the feature set, retaining the principal components with a cumulative contribution of 95%. This yielded the first feature set for subsequent pattern recognition and classification tasks.

[0028] S103. Based on the first feature set, a cross-analysis method is constructed to fuse the electrical impedance signal and the pressure-volume curve information, and the complementary features between the two types of signals are determined by correlation coefficient calculation to obtain a second feature set.

[0029] The electrical impedance signal and pressure-volume data are obtained through the first feature set, and the two types of signal features are separated by a preset method. For the separated electrical impedance signal and pressure-volume data, a cross-analysis method is used to fuse the signal and curve information to obtain preliminary fusion features. Complementary features are extracted from the preliminary fusion features, and the correlation strength between the features is determined by calculating the correlation coefficient. If the correlation strength exceeds the preset threshold, the corresponding complementary features are retained to obtain a filtered feature subset. Based on the filtered feature subset, the principal component analysis algorithm is used to compress the feature dimensions and generate an optimized feature set. By comparing the optimized feature set with the first feature set, the completeness of the feature extraction is judged to obtain the second feature set. For the second feature set, the support vector machine algorithm is used to classify the features to determine the final signal analysis results.

[0030] Specifically, when constructing the cross-analysis method, time-domain features such as peak amplitude (e.g., 1.2V) and rise time (e.g., 0.5ms) were first extracted from the electrical impedance signal. The slope (e.g., 0.8mL / mmHg) and inflection point pressure (e.g., 15mmHg) were simultaneously extracted from the pressure-volume curve, forming a first feature set containing 20 dimensional features. The Pearson correlation coefficient was used to calculate the correlation between the two types of signals, with a threshold of 0.3. When the correlation coefficient between the frequency-domain energy of the electrical impedance signal (e.g., 0.7dB power at 3.5kHz) and the end-diastolic volume of the pressure-volume curve (e.g., 110mL) reached 0.12, it was determined to be a low-correlation feature and retained in the second feature set. Principal component analysis (PCA) was used to reduce feature dimensionality, and the top five principal components with a contribution greater than 85% were selected. The first principal component (42% of the variance) incorporated the impedance phase angle (-35°) and the pressure change rate (1.2 mmHg / s), while the second principal component (23% of the variance) included the real part of the impedance (50 Ω) and the bulk elastic modulus (0.6 mL-1). Finally, a random forest algorithm was used to rank feature importance, with a set of 100 trees. When the Gini coefficient decreased by more than 0.05, eight key features were selected as the second feature set, including the interaction term between the impedance amplitude difference (0.8 V) and the pressure-volume loop area (950 mmHg mL).

[0031] S104. For the second feature set, a feature screening mechanism is applied to eliminate redundant parameters, and the importance of each feature in different breathing stages is evaluated by an information gain algorithm to obtain a third feature set.

[0032] A screening mechanism is used to remove redundant parameters from the feature set to obtain a preliminary processing result. An information gain algorithm is used to evaluate the importance of each feature in the preliminary processing result and determine the feature importance ranking. Based on the feature importance ranking, stage difference data related to the respiratory stage is obtained. If the stage difference data exceeds a preset threshold, the feature set is adjusted through data processing to obtain an optimized feature combination. Based on the optimized feature combination, the weight distribution of each feature in different respiratory stages is determined. By comparing the weight distribution with the information gain result, the core features of the third feature set are determined. The core features are used to verify the data processing results to obtain the third feature set.

[0033] Specifically, in the second feature set, the importance of each feature in different breathing phases is first evaluated using an information gain algorithm. For example, during the inhalation phase, the information gain values ​​for feature A are 0.85, feature B is 0.72, and feature C is 0.65. During the exhalation phase, the information gain values ​​for feature A are 0.78, feature B is 0.68, and feature C is 0.60. These values ​​clearly indicate that feature A is most important during the inhalation phase, while feature C is relatively less important during the exhalation phase. Next, a feature filtering mechanism is applied to eliminate redundant parameters. For example, if the information gain threshold is set to 0.70, features B and C will be eliminated during the inhalation phase because their information gain values ​​are below the threshold. Feature C will also be eliminated during the exhalation phase. Ultimately, the third feature set is obtained, in which only feature A is retained during the inhalation phase, and both features A and B are retained during the exhalation phase. This process ensures that each feature in the third feature set has a high information gain across different breathing phases, thereby improving the accuracy and efficiency of the model.

[0034] S105 , obtaining the feature importance distribution in the third feature set, dynamically updating the weight coefficients using an adaptive weight adjustment method, reflecting the time-varying characteristics through a weighted average of the time series, and obtaining a weighted feature set.

[0035] Obtain the third feature set and calculate the feature importance of each feature to obtain its distribution. Use the random forest algorithm to extract key features from the distribution and determine the initial weight coefficients. Dynamically update the weight coefficients using an adaptive adjustment method to obtain adjusted weight coefficients. Obtain time series data and calculate a weighted average of the adjusted weight coefficients to reflect time-varying characteristics. If the weighted average exceeds a preset threshold, optimize the weighted features using the gradient boosting algorithm to obtain an optimized feature set. Based on the optimized feature set, update the feature importance distribution to generate a weighted feature set. By comparing the generated weighted feature set with the initial feature set, determine the changing trend of the time-varying characteristics and obtain the final weighted feature set.

[0036] Specifically, when obtaining the feature importance distribution of the third feature set, the random forest algorithm is first used to calculate the Gini importance of each feature. For example, the importance of features A, B, and C are 0.35, 0.28, and 0.17, respectively, while the sum of the remaining features is 0.20. Based on this distribution, an adaptive weight adjustment method is designed: the initial weight coefficient is set to the normalized value of the feature importance (for example, A: 0.35 / 1.0 = 0.35), and a time decay factor λ = 0.9 is introduced. The weight is dynamically updated according to the formula w_t = λ·w_{t-1} + (1-λ)·I_t, where I_t is the feature importance at the current time step.

[0037] For example, if the importance of feature A suddenly increases to 0.45 at time t, its weight is adjusted from 0.35 to 0.9×0.35+0.1×0.45=0.36. Next, the time-varying characteristics are calculated by weighted averaging of the time series: taking the sliding window T=5 as an example, the weighted average of the weight sequence of feature A [0.35, 0.34, 0.36, 0.37, 0.38] is calculated to obtain

[0038] 0.35 × 0.1 + 0.34 × 0.15 + 0.36 × 0.2 + 0.37 × 0.25 + 0.38 × 0.3 = 0.368. Finally, the weighted feature value of each feature is multiplied by the original feature value to generate a weighted feature set. For example, if the original value of feature A is x = 10, the weighted feature value is 10 × 0.368 = 3.68. This process uses the Python sklearn library to implement random forest training and the numpy library to perform matrix operations, ensuring full automation.

[0039] S106. A dynamic compliance monitoring model is constructed by using a weighted feature set, and a support vector regression algorithm is used to fit the change trend of lung compliance to obtain real-time prediction results.

[0040] The feature set data is extracted through weighted features to construct a dynamic compliance monitoring model. The support vector regression algorithm is used to train the feature set to obtain the initial fitting function of lung compliance. Real-time monitoring data is obtained, and the trend of lung compliance changes is calculated using the initial fitting function. The dynamic compliance state is judged based on the change trend, and the real-time prediction value is determined. If the real-time prediction value exceeds the preset threshold, the fitting parameters are adjusted through the regression algorithm to obtain an updated prediction result. The compliance fluctuation is analyzed for the updated prediction result, and the mapping relationship between the fluctuation feature and the weighted feature is obtained. The feature set weight is optimized through the mapping relationship to obtain the adjusted dynamic compliance monitoring model.

[0041] Specifically, when constructing a dynamic compliance monitoring model, the patient's respiratory mechanics parameters, such as airway pressure, tidal volume, and flow, are first collected. The sampling frequency is set to 100 Hz to ensure data accuracy. The raw data is preprocessed using a sliding window method with a window size of 200 milliseconds and a step size of 50 milliseconds. A Butterworth low-pass filter is used to eliminate high-frequency noise, with a cutoff frequency of 5 Hz. During the feature extraction phase, 12-dimensional features, including peak pressure, positive end-expiratory pressure, and tidal volume change rate, are selected. Maximum-minimum normalization is used to map the feature values ​​to the [0, 1] interval. The weighted feature set is calculated using a random forest algorithm. For example, the weight for airway pressure is set to 0.35, and the weight for tidal volume is set to 0.28. The weights for the remaining features are automatically assigned based on the Gini impurity metric. The support vector regression model uses a radial basis kernel function with a kernel parameter γ set to 0.1 and a penalty coefficient C of 10. Model parameters are updated every 30 seconds using incremental training. During real-time prediction, the feature vector within the current 5 seconds is input and the predicted lung compliance value is output. For example, when the input features are [0.72, 0.65, 0.81], the model outputs a dynamic compliance value of 45 ml / cmH2O. Model performance is evaluated using a holdout method, with 70% of the data used for training and 30% for testing, and the root mean square error is controlled within 3.2 ml / cmH2O. The prediction results are smoothed using a Kalman filter, with the state transition matrix coefficient set to 0.9 and the observation noise covariance adjusted to 1.5. The final output is a compliance trend curve updated every 100 milliseconds. If the monitored compliance value falls below 30 ml / cmH2O for 10 consecutive times, an early warning mechanism is triggered and the abnormal timestamp and associated parameters are automatically recorded.

[0042] S107. If the deviation between the real-time prediction result and the actual measurement value exceeds a preset threshold, the correlation coefficient calculation formula in the cross analysis method is adjusted, and the complementary feature extraction is optimized by introducing a nonlinear mapping function to obtain an updated prediction result.

[0043] If the deviation between the real-time prediction and the actual measurement exceeds a deviation threshold, the deviation value is extracted from the real-time prediction and the measurement. By comparing the deviation value with a preset threshold, it is determined whether the calculation formula should be adjusted, generating an adjustment trigger signal. Based on the adjustment trigger signal, the correlation coefficient is extracted from the cross-analysis. The mapping relationship between the correlation coefficient and the calculation formula is analyzed to determine the adjustment direction and obtain the formula adjustment parameters. The calculation formula is updated using the formula adjustment parameters, and the calculation formula is transformed using a nonlinear mapping function to obtain an optimized calculation formula. Complementary features are obtained from the optimized calculation formula and processed using a feature extraction algorithm to obtain an enhanced feature set. The real-time prediction is updated using the enhanced feature set, and regression analysis of the enhanced feature set using a support vector machine algorithm is performed to obtain a preliminary updated prediction result. If the deviation between the preliminary updated prediction result and the measurement value still exceeds the deviation threshold, the parameters of the nonlinear mapping function are iteratively adjusted to obtain the final prediction result. Based on the deviation between the final prediction result and the actual measurement, the prediction stability is assessed by recording the deviation trend and obtaining stability assessment data.

[0044] Specifically, during the real-time prediction process, when the deviation between the prediction result and the actual measurement value exceeds a preset threshold (for example, the deviation exceeds 5%), the system will automatically adjust the correlation coefficient calculation formula in the cross-analysis method. Specifically, the traditional Pearson correlation coefficient calculation may not be able to capture nonlinear relationships, so a nonlinear mapping function (such as the Sigmoid function) is introduced to optimize the extraction of complementary features. Assuming that the original correlation coefficient is 0.7, after mapping through the Sigmoid function, the correlation coefficient may be adjusted to 0.85, thereby better reflecting the nonlinear association between features. Next, the system will recalculate the feature weights based on the updated correlation coefficients, for example, adjusting the weight of feature A from 0.3 to 0.4, and the weight of feature B from 0.5 to 0.45. Through this optimization, the system can more accurately extract complementary features and generate updated prediction results.

[0045] For example, if the original predicted value is 100 and the actual measured value is 105, with a deviation of 5%, after optimization, the updated predicted value may be adjusted to 103.5, with the deviation reduced to 1.5%. This process is achieved through an automated algorithm without human intervention, ensuring the accuracy and real-time nature of the prediction results.

[0046] S108. Based on the updated prediction results, evaluate the degree of improvement in computing efficiency, determine whether the real-time performance meets clinical needs by recording the model running time and resource occupancy rate, and obtain the final monitoring output.

[0047] Through the preset threshold, the running time and resource occupancy data are obtained from the model operation to obtain the time record and occupancy judgment results. The time record is used to analyze the running time change trend, determine whether the computing efficiency meets the preset standard, and obtain the efficiency evaluation result. According to the efficiency evaluation result, the improvement degree is determined to obtain the quantitative output of the computing efficiency. If the real-time performance is lower than the clinical demand threshold, the model operation parameters are adjusted to obtain new running time data and obtain the optimized time record. Through the optimized time record, it is judged whether the real-time performance meets the clinical demand and the real-time confirmation result is obtained. The real-time confirmation result is used to generate the final monitoring output data and obtain the matching judgment of the clinical demand. The key indicators are extracted from the matching judgment to determine the completeness of the monitoring output and obtain the final monitoring output.

[0048] Specifically, during the prediction update process, the model runtime was reduced from 120 seconds to 45 seconds by introducing parallel computing technology, improving computational efficiency by 62.5%. The GPU-based TensorFlow framework was used in the implementation, and computing resource utilization was further optimized by adjusting the data batch size from 32 to 128. Resource utilization was also reduced from 85% to 65%, ensuring system stability under high loads. To assess whether the model's real-time performance met clinical requirements, we measured the model's response time under different hardware configurations. The results showed that on a server equipped with an NVIDIA Tesla V100, a single inference time remained stable at under 0.8 seconds, fully meeting clinical real-time monitoring requirements. The final monitoring output, using an integrated multimodal data fusion algorithm, comprehensively analyzes physiological parameters such as electrocardiogram (ECG), blood pressure, and blood oxygen saturation to generate a dynamic patient health score. The score ranges from 0 to 100, with scores above 90 indicating good health and scores below 60 requiring immediate intervention. By introducing a deep learning-based anomaly detection model, the system can identify abnormal signals within 0.5 seconds and automatically trigger an early warning mechanism to ensure the timeliness and accuracy of clinical decision-making.

[0049] The description of the above embodiments is only used to help understand the technical solutions and core ideas of this application; ordinary technicians in this field should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A dynamic lung compliance monitoring method based on electrical impedance tomography, characterized in that: The method comprises: Obtain electrical impedance signals and pressure-volume curve data, and use synchronous acquisition equipment to record the time series signals of lung tissue during dynamic ventilation to obtain the original multidimensional data set; For the original multidimensional data set, signal preprocessing technology is used to remove noise interference and baseline drift, and the dynamic change characteristics within a continuous time period are extracted through the sliding window method to obtain the first feature set; Based on the first feature set, a cross-analysis method is constructed to fuse the electrical impedance signal and the pressure-volume curve information. The complementary characteristics between the two types of signals are determined by calculating the correlation coefficient to obtain the second feature set. For the second feature set, the feature screening mechanism is applied to eliminate redundant parameters, and the importance of each feature in different breathing stages is evaluated by the information gain algorithm to obtain the third feature set; Obtain the feature importance distribution in the third feature set, dynamically update the weight coefficient using an adaptive weight adjustment method, and reflect the time-varying characteristics through the weighted average of the time series to obtain a weighted feature set; By using weighted feature sets, a dynamic compliance monitoring model is constructed, and the support vector regression algorithm is used to fit the trend of lung compliance changes to obtain real-time prediction results. If the deviation between the real-time prediction result and the actual measurement value exceeds the preset threshold, the correlation coefficient calculation formula in the cross-analysis method is adjusted, and the complementary feature extraction is optimized by introducing a nonlinear mapping function to obtain an updated prediction result; Based on the updated prediction results, the degree of improvement in computing efficiency is evaluated, and by recording the model running time and resource occupancy rate, it is determined whether the real-time performance meets clinical needs to obtain the final monitoring output.

2. The method according to claim 1, characterized in that The electrical impedance signal and pressure-volume curve data are obtained, and the time series signals of the lung tissue during dynamic ventilation are recorded by synchronous acquisition equipment to obtain an original multidimensional data set, including: The time series signals of lung tissue during dynamic ventilation are acquired through synchronous acquisition equipment to obtain the original multidimensional data set; Extracting electrical impedance signals and pressure-volume curve data from the original multidimensional data set and determining the signal feature vector; The principal component analysis method is used to reduce the dimension of the signal feature vector to obtain the feature data set after dimension reduction; If there are outliers in the feature data set after dimensionality reduction, they are filtered out using a preset threshold to obtain a filtered feature data set; A time series model was constructed based on the filtered feature data set, and the support vector machine algorithm was used to determine the changing trend during ventilation. The correlation between lung tissue signals and electrical impedance signals was analyzed by changing trends to obtain the correlation coefficient matrix; The dynamic mapping relationship between the pressure-volume curve data and the lung tissue signal is determined according to the correlation coefficient matrix to obtain the quantitative characteristics of the ventilation process.

3. The method according to claim 1, characterized in that The original multidimensional data set is subjected to signal preprocessing technology to remove noise interference and baseline drift, and the dynamic change features within a continuous time period are extracted by a sliding window method to obtain a first feature set, including: Acquire data from the original multidimensional data set, remove noise interference and baseline drift using a filtering method, and obtain a second data set; For the second data set, the time series is segmented by the sliding window method, and the dynamic change features are extracted to obtain the third data set; If the dynamic features in the third data set exceed a preset threshold, principal component analysis is used to reduce the dimension to obtain a fourth data set; Calculate the trend vector of the time series based on the fourth data set to determine the change trend; According to the change trend, the clustering algorithm is used to group the fourth data set to obtain the grouping results; Based on the grouping results, obtain the time series feature distribution of each group and determine the feature stability; According to the feature stability, stable dynamic features are screened out to obtain the final feature set.

4. The method according to claim 1, wherein According to the first feature set, a cross analysis method is constructed to fuse the electrical impedance signal and the pressure-volume curve information, and the complementary characteristics between the two types of signals are determined by correlation coefficient calculation to obtain the second feature set, including: The electrical impedance signal and pressure-volume data are obtained through the first feature set, and the two types of signal features are separated using a preset method; For the separated electrical impedance signal and pressure-volume data, a cross-analysis method is used to fuse the signal and curve information to obtain preliminary fusion features; Extract complementary features from the preliminary fusion features and determine the strength of the correlation between features by calculating the correlation coefficient; If the correlation strength exceeds the preset threshold, the corresponding complementary features are retained to obtain the filtered feature subset; Based on the filtered feature subset, the principal component analysis algorithm is used to compress the feature dimensions and generate an optimized feature set; By comparing the optimized feature set with the first feature set, the completeness of feature extraction is judged and the second feature set is obtained; For the second feature set, the support vector machine algorithm is used to classify the features and determine the final signal analysis results.

5. The method according to claim 1, wherein For the second feature set, a feature screening mechanism is applied to eliminate redundant parameters, and the importance of each feature in different respiratory stages is evaluated by an information gain algorithm to obtain a third feature set, including: The redundant parameters are removed from the feature set through the screening mechanism to obtain the preliminary processing results; The information gain algorithm is used to evaluate the importance of each feature in the preliminary processing results and determine the feature importance ranking; Based on the feature importance ranking, phase difference data related to the respiratory phase are obtained; If the stage difference data exceeds the preset threshold, the feature set is adjusted through data processing to obtain the optimized feature combination; According to the optimized feature combination, the weight distribution of each feature in different breathing stages is determined; By comparing the weight distribution and information gain results, the core features of the third feature set are determined; The core features are used to verify the data processing results and obtain the third feature set.

6. The method according to claim 1, characterized in that The method of obtaining the feature importance distribution in the third feature set, dynamically updating the weight coefficient using an adaptive weight adjustment method, and reflecting the time-varying characteristics through the weighted average of the time series to obtain a weighted feature set includes: Obtain the third feature set and obtain the distribution by calculating the feature importance of each feature; The random forest algorithm is used to extract key features from the distribution and determine the initial weight coefficients; The weight coefficient is dynamically updated through an adaptive adjustment method to obtain an adjusted weight coefficient; Obtain time series data and calculate the weighted average value based on the adjusted weight coefficient to reflect the time-varying characteristics; If the weighted average value exceeds the preset threshold, the weighted features are optimized using the gradient boosting algorithm to obtain the optimized feature set; According to the optimized feature set, the feature importance distribution is updated to generate a weighted feature set; By comparing the generated weighted feature set with the initial feature set, the changing trend of the time-varying characteristics is judged and the final weighted feature set is obtained.

7. The method according to claim 1, characterized in that The dynamic compliance monitoring model is constructed by weighting the feature set, and the support vector regression algorithm is used to fit the change trend of lung compliance to obtain real-time prediction results, including: Extract feature set data through weighted features and build a dynamic compliance monitoring model; The support vector regression algorithm is used to train the feature set to obtain the initial fitting function of lung compliance; Acquire real-time monitoring data and calculate the change trend of lung compliance through the initial fitting function; Determine the dynamic compliance status based on the changing trend and determine the real-time prediction value; If the real-time prediction value exceeds the preset threshold, the fitting parameters are adjusted through the regression algorithm to obtain an updated prediction result; Analyze the compliance fluctuations based on the updated prediction results and obtain the mapping relationship between the fluctuation characteristics and the weighted characteristics; The feature set weights are optimized through mapping relationships to obtain the adjusted dynamic compliance monitoring model.

8. The method according to claim 1, characterized in that If the deviation between the real-time prediction result and the actual measurement value exceeds a preset threshold, the correlation coefficient calculation formula in the cross analysis method is adjusted, and the complementary feature extraction is optimized by introducing a nonlinear mapping function to obtain an updated prediction result, including: If the deviation between the real-time prediction and the actual measurement exceeds the deviation threshold, the deviation value is obtained from the real-time prediction and the measurement value, and the calculation formula is determined by comparing the deviation value with the preset threshold to obtain an adjustment trigger signal; According to the adjustment trigger signal, the correlation coefficient is extracted from the cross analysis, and the adjustment direction is determined by analyzing the mapping relationship between the correlation coefficient and the calculation formula to obtain the formula adjustment parameter; The calculation formula is updated by adjusting the parameters of the formula, and the calculation formula is transformed by using a nonlinear mapping function to obtain an optimized calculation formula; Obtain complementary features from the optimized calculation formula, process the complementary features through the feature extraction algorithm, and obtain an enhanced feature set; By updating the real-time prediction through the enhanced feature set, the support vector machine algorithm is used to perform regression analysis on the enhanced feature set to obtain the preliminary updated prediction results; If the deviation between the initial updated prediction result and the measured value still exceeds the deviation threshold, the parameters of the nonlinear mapping function are iteratively adjusted to obtain the final prediction result; According to the deviation value between the final prediction result and the actual measurement, the prediction stability is judged by recording the deviation change trend and the stability evaluation data is obtained.

9. The method according to claim 1, characterized in that Based on the updated prediction results, the improvement in computing efficiency is evaluated, and by recording the model running time and resource occupancy rate, whether the real-time performance meets clinical needs is determined, and the final monitoring output is obtained, including: Through the preset threshold, the running time and resource occupancy data are obtained from the model operation to obtain the time record and occupancy judgment results; Use time records to analyze the running time change trend, determine whether the calculation efficiency meets the preset standard, and obtain the efficiency evaluation result; According to the efficiency evaluation results, the improvement degree is determined and the quantitative output of the calculation efficiency is obtained; If the real-time performance is lower than the clinical requirement threshold, the model operation parameters are adjusted to obtain new operation time data and obtain optimized time records; Through optimized time records, it is determined whether the real-time performance meets clinical needs and the real-time confirmation results are obtained; Use real-time confirmation results to generate final monitoring output data and obtain matching judgments for clinical needs; Extract key indicators from the matching judgment, determine the completeness of the monitoring output, and obtain the final monitoring output.

Images