A lung compliance prediction model training method, application method and system

Abstract

This invention relates to the field of lung compliance prediction technology, specifically to a method, application method, and system for training a lung compliance prediction model. The training method includes: acquiring sample data; training a pre-defined prediction model using multi-parameter curve data and corresponding lung compliance values ​​from the samples to obtain a lung compliance prediction model; during training, the prediction model outputs predicted lung compliance values ​​based on the parameter curve data, and updates the trainable parameters in the model through an error backpropagation mechanism based on the error between the predicted lung compliance value and the corresponding lung compliance value. The trainable parameters include convolutional layer parameters in the feature extraction module and fully connected layer parameters in the regression prediction module. This invention can utilize data from lung function testing devices such as spirometers (e.g., parameter curves) for non-invasive, non-surgical prediction of lung compliance.

Description

Technical Field

[0001] This invention relates to the field of lung compliance prediction technology, specifically to a lung compliance prediction model training method, application method, and system. Background Technology

[0002] Lung compliance assessment has applications throughout the entire process of respiratory support therapy. For example, in the intensive care unit, it can help doctors identify high-risk patients, such as those with acute respiratory distress syndrome, before implementing invasive mechanical ventilation, thereby developing individualized lung protection strategies. For instance, patent application CN120299663A proposes a method and system for predicting weaning behavior from medical ventilators. This method uses a neural network model for intelligent prediction, providing automated weaning decision support based on multiple key parameters such as respiratory rate, tidal volume, oxygenation index, and lung compliance index.

[0003] However, all current analyses related to lung compliance rely on pressure measurements. Pressure measurements often require invasive procedures (such as requiring endotracheal intubation or using an esophageal balloon), which can easily cause patient discomfort or even refusal to cooperate, making the examination difficult to complete.

[0004] For example, patent application CN121221097A proposes a closed-loop assessment system for respiratory muscle strength and lung compliance based on EIT, including: a multi-frequency EIT injection / sampling unit; a pressure measurement and synchronization unit; an FPGA clock synchronization triggering unit; an EIT preprocessing unit; a compression sensing reconstruction unit; a ΔZ–Pressure fitting unit; a feature extraction unit; a control decision unit; and an adaptive NMES modulation and trend recording unit.

[0005] For example, patent application CN121102662A discloses an adaptive regulation and control system for mechanical ventilation in acute respiratory distress syndrome. This system includes a data acquisition module, a signal preprocessing module, a physiological parameter calculation module, a prediction module, and a decision and control logic module. The prediction module uses a long short-term memory neural network model to predict the dynamic trend of lung compliance changes, and the decision and control logic module generates ventilation parameter adjustment instructions based on the prediction results and clinical safety rules.

[0006] In summary, traditional methods for analyzing lung compliance often rely on invasive procedures, which presents several challenges in practical application: 1) Invasive procedures place extremely high demands on the skills of medical staff; 2) Patients often have a psychological fear of invasive procedures, which can lead to non-cooperation, resulting in excessively long examination cycles or even making the procedures difficult to complete. Summary of the Invention

[0007] The purpose of this invention is to provide a method, application method and system for training a lung compliance prediction model, which partially solves or alleviates the above-mentioned shortcomings in the prior art and can improve the accuracy of lung compliance prediction results.

[0008] To solve the aforementioned technical problems, the present invention specifically adopts the following technical solution: A first aspect of the present invention is to provide a method for training a lung compliance prediction model, comprising the steps of: S100, acquire sample data, the sample data including: parameter curves measured during the subject's breathing process, and corresponding lung compliance values, the parameter curves including: (1) The parameter curves used to reflect the respiratory snapshot include: inspiratory flow rate-volume curve and expiratory flow rate-volume curve; (2) The parameter curves used to reflect the respiratory accumulation include: inspiratory time-volume curve and expiratory time-volume curve. S101, the preset prediction model is trained using the sample data to obtain a prediction model for lung compliance; wherein, the prediction model includes: a feature extraction module, a global feature aggregation module, and a regression prediction module; wherein, S101 includes the following steps: S1011, the feature extraction module extracts feature information from the parameter curve, the feature information including: (1) the rate of change and / or magnitude of change of at least one curve segment; (2) the smoothness or fluctuation of at least one curve segment; the parameter curve can be divided into at least two curve segments; S1012, the feature information is integrated by the global feature aggregation module to form a feature vector; wherein, one feature vector is used to characterize at least one breathing test, and the breathing test includes at least one breathing cycle; S1013, The regression prediction module learns the mapping relationship between the feature vector and the lung compliance prediction value; S1014, determine whether the prediction model has converged based on the model prediction error. If so, end the training. The model prediction error is the difference between the predicted lung compliance value and the corresponding lung compliance value.

[0009] In some embodiments, S1011 includes: The parameter curves are aligned according to a uniform scale, and a multi-channel input is constructed so that different parameter curves correspond to each other at the same physiological stage, thereby forming a multi-channel respiratory data sequence. The feature extraction module performs feature extraction on the respiratory data sequence to extract feature information; And / or, the feature extraction module includes: at least one one-dimensional convolutional layer; And / or, the feature extraction module employs a one-dimensional convolutional neural network, and the regression prediction module employs a fully connected neural network.

[0010] In some embodiments, S1011 further includes: The parameter curves are preprocessed; the preprocessing methods include resampling or interpolation.

[0011] In some embodiments, the parameter curve data is acquired through a pulmonary function testing device, which includes a spirometer, a plethysmometer, or a sensor device for measuring respiratory flow and volume.

[0012] A second aspect of the present invention is to provide a method for applying a lung compliance prediction model, the method comprising: Obtain the measured parameter curve data of the object under test; After preprocessing the measured parameter curve data, it is input into the lung compliance prediction model, and the lung compliance prediction model outputs the corresponding lung compliance value.

[0013] In some embodiments, the parametric curve is labeled with weight information, which includes: curve weight and / or segment weight, wherein the curve weight is the overall influence weight of a parametric curve, and the segment weight is the local influence weight of a local region of the parametric curve; correspondingly, preprocessing the measured parametric curve includes the following steps: A noise score is generated for the parameter curve using a noise assessment method. If the noise score is greater than the preset first score, the parameter curve is marked as a discard curve; When the noise score is greater than a preset second score and less than or equal to the first score, a curve weight reduction suggestion is generated for the parameter curve, and the curve weight reduction suggestion is used to reduce the curve weight.

[0014] In some embodiments, when generating the curve weighting suggestion, the method further includes the step of: To obtain the physiological state of the subject under test; The lung function level of the test subject is predicted based on the physiological state, wherein the higher the lung function level, the stronger the lung function integrity of the test subject, or the better the lung health status of the test subject. When the lung function level is greater than the set first function level, the curve weighting suggestion is accepted. The segment weighting suggestion is used to generate segment weights for at least one curve segment, or to increase the segment weights of at least one curve segment.

[0015] In some embodiments, including: When the noise score is greater than the preset third score and less than or equal to the second score, a weighted suggestion for the curve generation segment is made.

[0016] In some embodiments, the steps further include: Recommended settings for ventilator parameters are generated based on the lung compliance values.

[0017] A third aspect of the present invention is to provide a lung compliance prediction model training system, comprising: The sample data acquisition module is used to acquire sample data, which includes: parameter curves measured during the subject's breathing process, and corresponding lung compliance values. The parameter curves include: (1) The parameter curves used to reflect the respiratory snapshot include: inspiratory flow rate-volume curve and expiratory flow rate-volume curve; (2) The parameter curves used to reflect the respiratory accumulation include: inspiratory time-volume curve and expiratory time-volume curve. A prediction model training module is used to train a preset prediction model using the sample data to obtain a prediction model for lung compliance; wherein the prediction model includes: a feature extraction module, a global feature aggregation module, and a regression prediction module; wherein the prediction model training module includes: The feature information extraction unit is used to extract feature information from the parameter curve through the feature extraction module. The feature information includes: (1) the rate of change and / or the magnitude of change of at least one curve segment; (2) the smoothness or fluctuation of at least one curve segment; the parameter curve can be divided into at least two curve segments. The feature vector integration unit is used to integrate the feature information through the global feature aggregation module to form a feature vector; wherein, one of the feature vectors is used to characterize at least one breathing test, and the breathing test includes at least one breathing cycle; A mapping relationship learning unit is used to learn the mapping relationship between the feature vector and the lung compliance prediction value through the regression prediction module; The prediction error evaluation unit is used to determine whether the prediction model has converged based on the model prediction error. If it has, the training ends. The model prediction error is the difference between the predicted lung compliance value and the corresponding lung compliance value.

[0018] Beneficial technical effects: This invention proposes a pressureless (or non-invasive) method for measuring lung compliance, which involves appropriately screening and preprocessing respiratory data (parameter curves) to indirectly provide feedback on lung compliance using multidimensional respiratory parameters, thereby avoiding direct measurement of lung pressure or airway pressure.

[0019] Specifically, this invention sets key features from both the time-series and curve data dimensions. Specifically, this means: 1) In the time-series dimension, this invention uses a flow-volume curve to reflect respiration from a snapshot perspective, and a time-volume curve to reflect respiration from a cumulative perspective. 2) In the curve data dimension, this invention focuses on extracting the curve's trend (equivalent to reflecting the speed or magnitude of change, such as rate of change or amplitude of change) and volatility (or disorder, such as degree of volatility or smoothness) to provide feedback on the curve's detailed features.

[0020] Therefore, by selecting features in the time-series dimension and the curve data dimension, the present invention can achieve reliable prediction of lung compliance based solely on respiratory data (i.e., it can get rid of the dependence on pressure measurement data, such as esophageal pressure).

[0021] Alternatively, this invention extracts the feature information of the parametric curve and learns the mapping relationship between the feature vector and lung compliance, that is, it uses the parametric curve to predict lung compliance, thereby avoiding the physiological damage (such as the use of invasive instruments such as esophageal balloons and endotracheal intubation) or psychological burden (such as fear, pain or discomfort) to patients caused by invasive operations.

[0022] This invention also proposes a preprocessing mechanism for parameter curves, which can perform noise reduction on parameter curves to eliminate noise caused by abnormal conditions such as sudden changes, abnormal peaks, or long plateaus that do not conform to physiological meaning, thereby avoiding interference from abnormal noise on prediction and causing distortion of prediction results.

[0023] Specifically, this invention proposes a restrictive parametric curve denoising process. In the scenario of non-invasive prediction of lung compliance, by adopting a more conservative denoising strategy for the parametric curve (such as discarding high-noise curves and reducing the weight of medium-noise curves), the basic accuracy of the prediction results can be ensured through denoising, while the data processing pressure of the model can be reduced by appropriately reducing the data. This achieves a better balance between prediction accuracy and data processing pressure, thereby quickly outputting a more accurate lung compliance prediction result, and thus supporting doctors to make better subsequent decisions.

[0024] Furthermore, this invention employs a local weighting mechanism for curve segments, which is a further refinement of the overall restrictive noise reduction strategy. Specifically, it applies weighted adjustments to the segments of interest within the parametric curve, while retaining the original weights of the remaining segments.

[0025] In other words, when noise is low, the local weighting mechanism does not need to reacquire the complete parameter curve, but only performs weight adjustments on local segments, significantly reducing the computational cost and memory usage of data preprocessing. This mechanism helps avoid high computational overhead, enabling the model to maintain a fast response even in continuous monitoring scenarios. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. The elements or parts in the drawings are not necessarily drawn to scale. Obviously, the drawings described below are some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0027] Figure 1 This is a flowchart illustrating a lung compliance prediction model training method provided by the present invention. Figure 2 This is a schematic diagram of the structure of a lung compliance prediction model training system provided by the present invention; Figure 3 A schematic block diagram of the structure of a computer device provided by the present invention; Figure 4 A schematic diagram illustrating the process from data acquisition to clinical deployment of the lung compliance prediction model provided by this invention; Figure 5 A schematic flowchart of the lung compliance prediction method based on parametric curves provided by the present invention; Figure 6 This is a diagram illustrating the complete technical architecture of the lung compliance prediction model provided by this invention. Figure 7 Example diagram of parameter curves provided for this invention; Figure 8 Example diagram for verifying the prediction performance of the model provided by this invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0029] In this document, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, "module," "part," or "unit" may be used interchangeably.

[0030] In this document, the terms "upper," "lower," "inner," "outer," "front," "rear," "one end," and "the other end," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the present invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0031] In this document, unless otherwise explicitly specified and limited, the terms "installed," "equipped with," "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection, a direct connection, or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0032] In this document, "and / or" includes any and all combinations of one or more of the listed related items.

[0033] In this article, "multiple" means two or more, that is, it includes two, three, four, five, etc.

[0034] As used in this specification, the term "about" typically means + / -5% of the value, more typically + / -4% of the value, more typically + / -3% of the value, more typically + / -2% of the value, even more typically + / -1% of the value, and even more typically + / -0.5% of the value.

[0035] In this specification, certain embodiments may be disclosed in a range-bound format. It should be understood that this "range-bound" description is merely for convenience and brevity and should not be construed as a rigid limitation on the disclosed range. Therefore, the description of a range should be considered as having specifically disclosed all possible subranges and the individual numerical values ​​within those ranges. For example, a description of the range 1-6 should be considered as having specifically disclosed subranges such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6, etc., and the individual numbers within those ranges, such as 1, 2, 3, 4, 5, and 6. This rule applies regardless of the breadth of the range.

[0036] Definition of noun: Lung compliance refers to the change in lung volume caused by a unit change in pressure, reflecting the elasticity and expansion capacity of lung tissue under external force.

[0037] Example 1: Please see Figure 4 , Figure 6 This illustrates an example of how a lung compliance prediction model progresses from clinical need to model deployment and clinical application.

[0038] This invention relates to the fields of medical signal processing and artificial intelligence, specifically to a deep learning-based method and system for estimating lung compliance. The method uses parametric curve data (or, in some embodiments, parametric curves) collected by a spirometer as input (see [link to relevant documentation]). Figure 7 By constructing an end-to-end deep learning model, the system achieves automatic estimation of lung compliance in subjects. The spirometry data includes inspiratory flow rate-volume curves (i.e.,...). Figure 7 The Insp Flow curve), expiratory flow rate-volume curve (i.e. Figure 7 The Exp Flow curve and the inspiratory time-volume curve (i.e., the inspiratory time-volume curve) are shown in the figure. Figure 7 The Insp-Vol curve) and the expiratory time-volume curve (i.e. Figure 7 The method first performs uniform length processing and normalization preprocessing on different types of curve data (Exp Vol curves), and then combines them into a multi-channel one-dimensional signal input model according to a preset format. The deep learning model includes a one-dimensional convolutional neural network feature extraction module, used to automatically extract time-series features related to lung compliance from the multi-channel parametric curves, and a regression prediction module, used to output the corresponding lung compliance assessment based on the extracted features. By training the model on a dataset containing real spirometry measurement data and corresponding lung compliance annotations, the method of this invention can directly learn respiratory mechanics features from the original spirometry curves without manually designed features or ventilator testing, achieving rapid, objective, and consistent assessment of the subject's lung compliance. This invention can be widely applied to lung mechanics parameter monitoring in clinical respiratory function assessment, intensive care, and respiratory support therapy, and has the advantages of high automation, strong adaptability, and easy integration.

[0039] The parameter curves (or breathing curves) are preprocessed and then uniformly resampled into multi-channel one-dimensional time-series data of consistent length. The prediction model is a one-dimensional convolutional neural network model, including a feature extraction module, a global feature aggregation module, and a regression prediction module. During training, the prediction model outputs predicted lung compliance values ​​based on the breathing curve data, and updates the trainable parameters in the model through an error backpropagation mechanism based on the error between the predicted lung compliance values ​​and the corresponding lung compliance values. The trainable parameters include the convolutional layer parameters in the feature extraction module and the fully connected layer parameters in the regression prediction module.

[0040] Please see Figure 5In some embodiments, the lung compliance prediction method proposed in this invention may include the following steps: 1. Acquire spirometer parameter curve data, wherein the spirometer parameter curve data includes at least the inspiratory flow rate-volume curve, the expiratory flow rate-volume curve, the inspiratory time-volume curve, and the expiratory time-volume curve; 2. The spirometer parameter curve data is preprocessed, including length unification and numerical normalization of different types of parameter curves to obtain standardized multi-channel one-dimensional time-series data; 3. Construct an end-to-end deep learning model, wherein the deep learning model includes: A one-dimensional convolutional neural network feature extraction module is used to extract features from the multi-channel one-dimensional time-series data to obtain a high-dimensional feature representation related to lung compliance. The regression prediction module is used to output the corresponding lung compliance estimate based on the high-dimensional feature representation; 4. The deep learning model is trained using a dataset containing real spirometry data and corresponding lung compliance annotations, and the model parameters are optimized by minimizing the error between the predicted values ​​and the labeled values; 5. Input the spirometry parameter curve data of the test subject into the trained deep learning model, and output the lung compliance estimation result of the test subject.

[0041] Traditional lung compliance assessments typically require invasive or time- and technically costly methods for data acquisition. For example, forced oscillation techniques or oscillatory lung function tests can be used to measure lung impedance data, or imaging data such as CT scans can be acquired to perform lung compliance analysis. However, these interventional impedance and CT scans often cause some degree of harm to patients or impose significant physical, time, and financial burdens.

[0042] Furthermore, this type of invasive testing requires a high degree of patient cooperation. However, especially for children, impedance testing may cause discomfort, leading to potential non-cooperation and increasing the difficulty of diagnosis and treatment for doctors.

[0043] In this regard, this application overcomes the limitations of destructive / high-cost measurement techniques and provides technical support for non-destructive, low-cost, and rapid lung compliance testing. Specifically, this application directly uses simple respiratory data curves as raw data, and through comprehensive analysis of the multi-dimensional characteristics of the respiratory data curves, it can predict the patient's lung compliance.

[0044] For example, the present invention provides a method for training a lung compliance prediction model, comprising the steps of: Acquire sample data, which includes: the object's parameter curves and the corresponding lung compliance values. The parameter curve data includes: inspiratory flow rate-volume curves, expiratory flow rate-volume curves, inspiratory time-volume curves, and expiratory time-volume curves. The preset prediction model is trained using the parameter curve data and the corresponding lung compliance values ​​to obtain a lung compliance prediction model. The prediction model includes a feature extraction module, a global feature aggregation module, and a regression prediction module. During training, the prediction model outputs predicted lung compliance values ​​based on the parameter curve data, and updates the trainable parameters in the model using an error backpropagation mechanism based on the error between the predicted lung compliance values ​​and the corresponding lung compliance values. The trainable parameters include the convolutional layer parameters in the feature extraction module and the fully connected layer parameters in the regression prediction module. The training also includes the following steps: Feature information is extracted from the parameter curve using the feature extraction module; The feature information is integrated by a global feature aggregation module to form a feature vector; wherein, a feature vector is used to characterize a breathing test, which contains at least one breathing cycle; The regression prediction module outputs the mapping relationship of lung compliance prediction values ​​based on the feature vector, and updates the model parameters based on the difference between the lung compliance prediction values ​​and the corresponding lung compliance values.

[0045] In other words, please see Figure 1 The present invention proposes a method for training a lung compliance prediction model, comprising the following steps: S100, acquire sample data, the sample data including: parameter curves measured during the subject's breathing process, and corresponding lung compliance values, the parameter curves including: (1) The parameter curves used to reflect the respiratory snapshot include: inspiratory flow rate-volume curve and expiratory flow rate-volume curve; (2) The parameter curves used to reflect the respiratory accumulation include: inspiratory time-volume curve and expiratory time-volume curve. S101, the preset prediction model is trained using the sample data to obtain a prediction model for lung compliance; wherein, the prediction model includes: a feature extraction module, a global feature aggregation module, and a regression prediction module; wherein, S101 includes the following steps: S1011, the feature extraction module extracts feature information from the parameter curve, the feature information including: (1) the rate of change and / or magnitude of change of at least one curve segment; (2) the smoothness or fluctuation of at least one curve segment; the parameter curve can be divided into at least two curve segments; S1012, the feature information is integrated by the global feature aggregation module to form a feature vector; wherein, one feature vector is used to characterize at least one breathing test, and the breathing test includes at least one breathing cycle; S1013, The regression prediction module learns the mapping relationship between the feature vector and the lung compliance prediction value; S1014, determine whether the prediction model has converged based on the model prediction error. If so, end the training. The model prediction error is the difference between the predicted lung compliance value and the corresponding lung compliance value.

[0046] In some embodiments, the prediction model can continue to be trained even after it has converged.

[0047] For example, sample data can be selected from a pre-defined sample pool to train the prediction model. For instance, a larger sample size, such as 50%, can be used in the early training process, and the sample size can be reduced to 10% after the prediction model converges.

[0048] Alternatively, in some embodiments, prior to S1014, it is permissible to continuously update or adjust the training samples during model training, as long as the final convergence condition (such as meeting the requirements of the loss function) is met. For example, in some embodiments, training can be stopped and the final prediction model can be given when the prediction error after model training is less than a preset value. Or, in some embodiments, even if the prediction error after model training is less than the preset value, the prediction model can still be retrained and optimized if there are untrained samples (such as updated supplementary samples).

[0049] In some embodiments, the rate of change of a curve segment can refer to how fast or slow the airflow or volume changes during a breathing process of a set unit length (such as a set volume range or a set time), such as the slope of the curve segment.

[0050] In some embodiments, the variation range of a curve segment may refer to the difference in airflow or volume changes during a breathing process of a set unit length (such as a set volume range or a set time).

[0051] In some embodiments, a curve segment may correspond to a rate of change and / or a magnitude of change.

[0052] In some embodiments, the smoothness of a curve segment can refer to whether the change of the curve segment is stable and continuous, and can be represented by the mean or integral of the absolute value of the second derivative of the curve segment; the smaller the second derivative, the smoother the curve.

[0053] In some embodiments, the degree of fluctuation of a curve segment can refer to the magnitude of the jitter or undulation that characterizes the deviation of the curve segment from a smooth trend, and can be represented by the variance / standard deviation of the first derivative of the curve segment.

[0054] In some embodiments, the smoothness or variability of a curve segment can also be referred to as a phased change characteristic. That is, a curve segment can correspond to at least two smoothness or variability levels.

[0055] For example, a respiratory snapshot refers to observing changes in flow rate synchronously with respect to volume, similar to taking a momentary photograph of respiratory data to directly reflect the ventilation status at a certain volume, such as an inspiratory flow rate-volume curve or an expiratory flow rate-volume curve. Respiratory accumulation, on the other hand, refers to recording the cumulative changes in volume over time with respect to time, reflecting the temporal cumulative effect of the entire respiratory process, such as an inspiratory time-volume curve or an expiratory time-volume curve.

[0056] This invention divides parametric curves into respiratory snapshot curves and respiratory accumulation curves for more targeted processing of respiratory data. Respiratory snapshots capture instantaneous mechanical response characteristics during respiration, while respiratory accumulation reflects the overall cumulative change in lung and airway volume over time. By merging and processing both snapshot and accumulation types of respiratory data through multiple channels, respiratory mechanical characteristics can be comprehensively characterized at both the instantaneous morphology and temporal accumulation levels.

[0057] It is worth noting that this invention sets key features from both the time-series and curve data dimensions. Specifically, this means: 1) In the time-series dimension, this invention uses a flow-volume curve to reflect respiration from a snapshot perspective, and a time-volume curve to reflect respiration from a cumulative perspective. 2) In the curve data dimension, this invention focuses on extracting the curve's trend (equivalent to reflecting the speed or magnitude of change, such as rate of change or amplitude of change) and volatility (such as the degree of volatility or smoothness) to provide feedback on the curve's detailed features.

[0058] Therefore, by selecting features in the time-series dimension and the curve data dimension, the present invention can achieve reliable prediction of lung compliance based solely on respiratory data (i.e., it can get rid of the dependence on pressure measurement data, such as esophageal pressure).

[0059] In some embodiments, if the model prediction error is less than or equal to a preset error threshold, the prediction model can be considered to have converged. The preset error threshold can be set by a technician. If the model prediction error is greater than the preset error threshold, training can continue until the model prediction error is less than or equal to the preset error threshold.

[0060] In this embodiment, the present invention proposes a pressureless (or non-invasive) lung compliance measurement method, which involves appropriately filtering and classifying respiratory data (parameter curves) (such as respiratory snapshots and respiratory accumulation) to indirectly provide feedback on lung compliance using multidimensional respiratory parameters, thereby avoiding direct measurement of lung pressure or airway pressure.

[0061] Specifically, this invention extracts the feature information of the parametric curve and learns the mapping relationship between the feature vector and lung compliance, that is, it uses the parametric curve to predict lung compliance. This can avoid the physiological damage (such as the use of invasive instruments such as esophageal balloons and endotracheal intubation) or psychological burden (such as fear, pain or discomfort) to patients caused by invasive procedures.

[0062] Furthermore, this hierarchical processing method for multidimensional respiratory data (distinguishing between respiratory snapshot curves and respiratory accumulation curves) enables the model to fully explore the deep information related to lung compliance in the respiratory data, thereby achieving more stable and accurate non-destructive and non-invasive lung compliance prediction.

[0063] In some embodiments, S1011 includes: Align the inspiratory flow rate-volume, expiratory flow rate-volume, inspiratory time-volume, and expiratory time-volume curves according to a uniform scale (e.g., align the time axis to the range of 0-200 data points). Figure 7 As shown in the figure, a multi-channel input is constructed so that different curves correspond to each other at the same physiological stage, in order to form a multi-channel respiratory data sequence; The feature extraction module performs feature extraction on the respiratory data sequence to extract feature information; And / or, the feature extraction module includes: at least one one-dimensional convolutional layer; And / or, the feature extraction module employs a one-dimensional convolutional neural network, and the regression prediction module employs a fully connected neural network.

[0064] The correspondence between different curves at the same physiological stage refers to the correspondence between a set of curve data at a certain volume and that volume. For example, a set of inspiratory flow rate, expiratory flow rate, inspiratory time, and expiratory time corresponding to a volume of 5 liters.

[0065] In some embodiments, constructing a multi-channel input means inputting the inspiratory flow rate-volume curve, expiratory flow rate-volume curve, inspiratory time-volume curve, and expiratory time-volume curve into the model together, so that different curves form a corresponding relationship at the same physiological stage, or in other words, so that the same breathing moment corresponds to multiple physiological indicators, that is, forming a multi-channel respiratory data sequence.

[0066] In some embodiments, S1011 further includes: The parameter curves are preprocessed; the preprocessing methods include resampling or interpolation.

[0067] In some embodiments, resampling can refer to sampling with a fixed number of sampling points. Interpolation can refer to estimating intermediate values ​​to fill in curve gaps. Preprocessing steps involving resampling and interpolation can standardize data length, fill in missing points, and facilitate subsequent multi-channel input and feature extraction.

[0068] Alternatively, resampling can refer to transforming the original parameter curve according to a predetermined sampling strategy (such as uniform sampling length or uniform sampling interval); interpolation can refer to estimating intermediate data points based on existing sampling points to achieve curve continuity or alignment. Through the above preprocessing, sampling differences between different curves can be reduced, the consistency of multi-channel data fusion can be improved, and the stability of subsequent feature extraction and model training can be enhanced.

[0069] In some embodiments, the parameter curve data is acquired through a pulmonary function testing device, which includes a spirometer, a plethysmometer, or a sensor device for measuring respiratory flow and volume.

[0070] In some embodiments, the feature information is used to characterize the changes in the shape of the parametric curve, the rate of change, the stage differences, and the response relationship between channels; the feature information includes at least one of the following: the overall steepness and flatness of the curve, the distribution characteristics of the rate of volume change during inhalation and exhalation, the response differences between different curve channels within the same volume range, local inflection points of the curve, plateau segments, and trends of change.

[0071] In some embodiments, the feature extraction module includes at least one one-dimensional convolutional layer.

[0072] In some embodiments, S1011 includes: The inspiratory flow rate-volume curve, expiratory flow rate-volume curve, inspiratory time-volume curve, and expiratory time-volume curve are combined in a predetermined order to form a multi-channel respiratory data sequence. The feature extraction module performs feature extraction on the respiratory data sequence to extract feature information.

[0073] In some embodiments, S1011 further includes: The parameter curves are preprocessed; the preprocessing methods include resampling or interpolation.

[0074] In some embodiments, prior to S100, the following step is also included: The respiratory volume index is obtained through the parameter curve data, and the respiratory volume index is expiratory flow rate, inspiratory flow rate, expiratory volume or inspiratory volume. Calculate the rate of change of the respiratory volume index within a set period; If the rate of change is lower than a set threshold, the corresponding parameter curve data is marked as abnormal data; if the degree of abnormality of the abnormal data is greater than a preset degree of abnormality, the weight of the corresponding abnormal data in the model training is reduced, or the corresponding abnormal data is removed.

[0075] In some embodiments, the present invention proposes a hierarchical anomaly handling strategy, specifically: If the abnormality only occurs in a local segment of a certain respiratory cycle (e.g., a short-term loss of signal in the post-inspiratory phase). Furthermore, the length of the abnormal segment is lower than a preset percentage threshold (e.g., lower than 5% to 10% of the cycle length). Instead of discarding the entire cycle, local interpolation repair or labeling and weighting can be used. If the following situations occur, you can choose to discard the entire cycle: the flow signal is continuously lost, the respiratory cycle cannot be completely identified, or the data length is insufficient to constitute a complete cycle (a threshold can be set, such as containing at least 30 data points).

[0076] In some embodiments, the parameter curves of test subjects such as COPD patients, asthma patients, restrictive lung disease patients, the elderly, and children may exhibit characteristics such as: significantly prolonged expiratory phase, decreased peak flow rate, generally flat curve, and abnormally slow rate of volume change. In this case, a specific implementation can be achieved by setting cycle weights, that is, assigning a confidence weight w ∈ [0,1] to each respiratory cycle. For example: w=1 for a normal cycle; w=0.8 for mild local abnormalities; and w=0.5 for the presence of noise interference.

[0077] For example, patients with COPD typically exhibit prolonged expiratory phase and slow decrease in airflow. Correspondingly, this cycle can be retained, but given appropriate weights (by setting a threshold) based on its overall statistical distribution to prevent the model from treating this pattern as noise.

[0078] In some embodiments, the parameter curve data is acquired through a pulmonary function testing device, which includes, but is not limited to, a spirometer, a plethysmometer, or a sensor device for measuring respiratory flow and volume.

[0079] In some embodiments, the feature extraction module employs a one-dimensional convolutional neural network, and the regression prediction module employs a fully connected neural network.

[0080] For example, the global feature aggregation module can be one of the following: average pooling structure, max pooling structure, or attention mechanism.

[0081] In some embodiments, the model's front end employs a one-dimensional convolutional structure as a feature extraction module. This module does not rely on manually defined feature formulas but instead slides across the curve sequence via convolution operations to automatically learn the local curve morphology. This module primarily extracts the following types of information: the overall steepness and smoothness of the curve, the distribution characteristics of the rate of volume change during inhalation and exhalation, the response differences between different curve channels within the same volume range, local inflection points, plateau segments, and trends of change in the curve. Through the stacking of multiple convolutional structures, the model can progressively transition from local variation features to global breathing pattern features, resulting in a final feature representation that simultaneously includes instantaneous dynamic information and overall morphological information.

[0082] After curve feature extraction, the model integrates features from the entire breathing process through a global feature aggregation module. Features extracted from different time points and curve channels are compressed into a fixed-length feature vector, thus forming a holistic representation of a single spirometry test. This feature vector is then input into a regression prediction module to output the corresponding lung compliance estimate. The role of this regression module is not to perform complex calculations, but rather to establish a stable mapping relationship, enabling the model to provide a numerical output consistent with actual lung compliance based on the learned parametric curve features.

[0083] During the model training phase, a dataset containing real spirometry curve data and corresponding lung compliance annotations was used. The core objective of the training process was to gradually adjust the internal parameters of the model by repeatedly exposing it to curve data from different individuals and different breathing patterns, thereby improving the accuracy and stability of the lung compliance estimation results (see [link to training program]). Figure 8 In this model, TRUE represents the true compliance score; PREDICT represents the predicted compliance without the weighting mechanism; Weighted-1 represents the prediction result after introducing a noisy data reduction mechanism in the simulation experiment, showing that abnormal biased samples are significantly corrected; and Weighted-2 represents the result after further simulation experiments with the introduction of a data weighting optimization strategy (including adaptation processing for special samples), showing that the predicted values ​​further converge to the true values. This training process can be understood as a parameter calibration process based on sample learning. Its technical effect is reflected in the model's ability to adapt to the differences in lung function among different subjects without manually adjusting the feature extraction rules for different populations.

[0084] The model can be used as a signal processing and parameter estimation module embedded in the data processing unit of medical devices. Its target is the real physical signals (flow rate, volume, time) collected by the spirometer, and the output is a numerical estimation parameter of lung compliance that can be directly used to assist clinical decision-making.

[0085] This invention can be integrated into ventilator or medical monitoring systems, working in conjunction with sensor acquisition, data preprocessing, and display modules. It enables real-time or near-real-time estimation of lung compliance without altering the spirometer's operation; it automatically adapts to the different lung mechanics characteristics of patients, reducing the need for manual parameter adjustments; and it is suitable for continuous monitoring and trend analysis scenarios, rather than one-time measurements. The measurement is low-cost, non-invasive, and suitable for large-scale, rapid testing.

[0086] In other words, this invention employs a data-driven end-to-end estimation method. Starting from a different technical approach, it establishes a mapping relationship between curve characteristics and lung compliance by learning the overall shape and dynamic changes of the spirometer curve. This method is highly adaptable to real-world data with significant individual differences, complex curve shapes, or high levels of noise.

[0087] It should be noted that a single cycle may be affected by random respiratory fluctuations, short-term uneven exertion, coughing, and equipment noise. Preferably, when measuring multiple cycles, it is preferable to collect data from multiple cycles (in some embodiments, this can be set to collect data from no less than 3 complete respiratory cycles) to reflect the overall lung mechanics.

[0088] In some embodiments, interval measurements are applicable for: efficacy comparisons, and measurements before and after drug intervention. In some embodiments, continuous measurement refers to continuity in the subject's physiological state, including no change in posture, no drug intervention, no strenuous activity, and no long intervals (e.g., more than one hour). In actual measurements, intervals of a few minutes between each respiratory test can also be considered as continuous physiological state measurements. Generally, continuous measurement is more appropriate because a stable physiological state allows for data comparison and facilitates periodic consistency analysis.

[0089] In some embodiments, different collection standards can be set for different patients. For example, 3 to 5 cycles are sufficient for healthy individuals, while for COPD patients, whose expiratory time is prolonged, a longer collection time of 5 to 8 cycles is recommended. Alternatively, for pediatric patients with unstable breathing, it is recommended to extend the collection time and allow for a wider range of cycle fluctuations.

[0090] This means that the number of data collection cycles, the length of data collection time, or the cycle screening criteria are dynamically adjusted based on the lung function status of the subjects (lung function-related disease information / age group) to improve the reliability of prediction.

[0091] Please see Figure 2The present invention also proposes a lung compliance prediction model training system, comprising: The sample data acquisition module is used to acquire sample data, which includes: parameter curves measured during the subject's breathing process, and corresponding lung compliance values. The parameter curves include: (1) The parameter curves used to reflect the respiratory snapshot include: inspiratory flow rate-volume curve and expiratory flow rate-volume curve; (2) The parameter curves used to reflect the respiratory accumulation include: inspiratory time-volume curve and expiratory time-volume curve. A prediction model training module is used to train a preset prediction model using the sample data to obtain a prediction model for lung compliance; wherein the prediction model includes: a feature extraction module, a global feature aggregation module, and a regression prediction module; wherein the prediction model training module includes: The feature information extraction unit is used to extract feature information from the parameter curve through the feature extraction module. The feature information includes: (1) the rate of change and / or the magnitude of change of at least one curve segment; (2) the smoothness or fluctuation of at least one curve segment; the parameter curve can be divided into at least two curve segments. The feature vector integration unit is used to integrate the feature information through the global feature aggregation module to form a feature vector; wherein, one of the feature vectors is used to characterize at least one breathing test, and the breathing test includes at least one breathing cycle; A mapping relationship learning unit is used to learn the mapping relationship between the feature vector and the lung compliance prediction value through the regression prediction module; The prediction error evaluation unit is used to determine whether the prediction model has converged based on the model prediction error. If it has, the training ends. The model prediction error is the difference between the predicted lung compliance value and the corresponding lung compliance value.

[0092] Alternatively, in other embodiments, the present invention provides a method for applying a lung compliance prediction model, the method comprising: providing a lung compliance prediction model, the lung compliance prediction model being trained by a lung compliance prediction model training method as described in any embodiment of the present invention; Obtain the measured parameter curves of the object under test; The measured parameter curve data is input into the lung compliance prediction model, and the lung compliance prediction model outputs the corresponding lung compliance value.

[0093] Preferably, in this embodiment, the parameter curve is marked with weight information, which includes: curve weight and / or segment weight. The curve weight is the overall influence weight of a parameter curve, and the segment weight is the local influence weight of a local region of the parameter curve. Correspondingly, the preprocessing of the measured parameter curve includes the following steps: A noise score is generated for the parameter curve using a noise assessment method. If the noise score is greater than the preset first score, the parameter curve is marked as a discard curve; When the noise score is greater than a preset second score and less than or equal to the first score, a curve weight reduction suggestion is generated for the parameter curve, and the curve weight reduction suggestion is used to reduce the curve weight.

[0094] In some embodiments, a higher noise score may indicate a greater likelihood of noise in the parameter curve.

[0095] In other words, the present invention also proposes a preprocessing mechanism for parameter curves, which can perform noise reduction on parameter curves to eliminate noise caused by abnormal conditions such as sudden changes, abnormal peaks or long plateaus that do not conform to physiological meaning, thereby avoiding interference from abnormal noise on prediction and causing the prediction results to be distorted.

[0096] For example, each parametric curve (also simply referred to as a curve) is labeled with an influence weight. When the prediction model learns and integrates the feature information of each curve, it will simultaneously consider the influence weight of each curve. A larger influence weight allows for a more appropriate increase in the proportion of that curve's influence in the final prediction conclusion. In other words, a larger influence weight also reflects, to some extent, the relatively higher reliability of the data for that curve.

[0097] Furthermore, in this embodiment, to distinguish the differences in weight adjustment under different stages / scenarios, two different weight concepts are defined: curve weight and segment weight. Curve weight refers to the overall influence weight of a complete curve, while segment weight refers to the local influence weight set for a local area of ​​a curve.

[0098] In this embodiment, the curve weight reduction suggestion refers to reducing the overall weight (i.e., curve weight) of a parametric curve (such as an inhalation velocity-volume curve).

[0099] Preferably, in some embodiments, when generating a weighting suggestion for the curve, the method further includes the step of: Obtain the physiological state of the subject to be tested. The physiological state may include: age, gender, medical history (especially the prevalence of lung function-related diseases), or other information that can reflect the state of lung function.

[0100] In some embodiments, the lung function level of the subject can be predicted based on at least one physiological state (such as age) and a preset grading rule. For example, if the subject is older (e.g., older than a preset age), the lung function level can be determined to be lower.

[0101] The lung function level of the test subject is preliminarily predicted based on the physiological state. The higher the lung function level, the stronger the lung function integrity of the test subject, or the better the lung health status of the test subject. When the lung function level is greater than the set first function level, the curve weighting suggestion is accepted. That is to say, in this embodiment, when the target is a young adult with adequate physical fitness, the curve weighting suggestion given at this time can be accepted.

[0102] Preferably, when the lung function level is less than or equal to the first function level, the curve weighting suggestion is not accepted.

[0103] In other words, in this embodiment, when the target group is the elderly, those with a history of lung disease, or young children, a conservative denoising strategy will be adopted, that is, the curve will not be weighted for the time being, and the prediction model will be allowed to treat it as regular data.

[0104] In this embodiment, a curve marked as a discard curve indicates that the data corresponding to that curve can be deleted.

[0105] Preferably, in some embodiments, when the noise score is greater than a preset third score and less than or equal to the second score, a segment weighting suggestion is generated for the curve. The segment weighting suggestion is used to generate segment weights for at least one curve segment or to increase the segment weights of at least one curve segment.

[0106] In this embodiment, when the noise is relatively small, local weight adjustment of the curve (i.e., adjusting segment weights) is allowed.

[0107] Exemplarily, in this embodiment, the step further includes: Obtain physiological analysis points; The parameter curve is divided into at least two curve segments based on the physiological analysis points; Select at least one curve segment as the segment of interest; In response to the segment weighting suggestion, the segment weight corresponding to the segment of interest is increased.

[0108] Among them, physiological analysis points can be obtained based on the patient's own situation (such as physiological state).

[0109] For example, different physiological analysis points can be used for patients with different lung function levels.

[0110] In some embodiments, physiological analysis points are boundary points used to divide the parametric curve, which can typically divide the parametric curve into rapid exhalation segment, rapid exhalation segment, plateau segment, etc.

[0111] In some embodiments, the focus segment can be a preset default segment, or it can be obtained based on the patient's own situation (such as physiological state).

[0112] For example, during the testing process, patients may experience physiological disturbances such as coughing, swallowing, sighing, slight body movement, and unstable breathing rhythm due to discomfort or limited cooperation. At the same time, the testing equipment may be affected by factors such as sensor baseline drift, pipeline airflow disturbance, poor contact, and environmental electromagnetic interference, which may cause a large number of abnormal noise signals such as spikes, jumps, local oscillations, and baseline shifts in the collected respiratory curve.

[0113] Especially in scenarios such as clinical intensive care, post-anesthesia recovery, or emergency mechanical ventilation, doctors often need to continuously and rapidly monitor patients' lung compliance in order to promptly detect deterioration in lung function, adjust ventilator parameters, or guide weaning. In these scenarios, data acquisition is frequent and the amount of data is large, making repeated calibration inefficient and impractical.

[0114] The dataset used in this invention covers different types of subjects, including but not limited to: healthy individuals, COPD patients, asthma patients, and people of different age groups. The respiratory curves of different types of subjects may show significant differences, such as prolonged expiratory time, decreased peak flow rate, and flat or irregular curves. This invention, through a weighting adjustment mechanism, enables the lung compliance prediction model to adapt to different groups and different breathing patterns.

[0115] In some embodiments, the present invention helps to achieve differentiated processing of noisy data and abnormal pathological data by introducing a weighted training mechanism based on data quality (e.g., marking parametric curves with large noise scores as discarded curves) and physiological labels (e.g., determining whether to accept curve weighting suggestions based on physiological state).

[0116] By generating weighting suggestions for noisy data with unstable signals or abnormal acquisition, the interference of noise on model training can be effectively reduced, while fully preserving the individualized physiological performance of the curves. This allows the model to fully learn the real respiratory characteristics of different patients, avoiding the loss of effective information due to excessive noise reduction and improving the model's generalization ability in complex clinical scenarios. In other words, this invention provides a restrictive weighted training mechanism that achieves a good balance between suppressing noise interference and preserving the personalized characteristics of the data.

[0117] In other words, this restrictive weighted training mechanism helps avoid misjudging individual differences as anomalous data and weakening them, thus preserving their key pathological features. By generating weighting suggestions, deweighting suggestions, or labeling the curves, the weights of the curves can be adjusted more carefully to control the impact of the noise reduction process on the original data.

[0118] From another perspective, different individuals and different physical conditions can lead to significant differences in parameter curves. This invention chooses to adjust the weights from a more macroscopic perspective, marking or adjusting the weights of some obvious noise, avoiding premature group differentiation, and thus avoiding excessive difficulty or misjudgment caused by group differentiation.

[0119] It should be understood that this weighting mechanism can significantly improve the predictive accuracy and stability of the model in different populations (including healthy people and sick people), verifying the good applicability of the present invention in complex clinical scenarios.

[0120] To address this, this invention proposes a restrictive parametric curve denoising process. In non-invasive lung compliance prediction scenarios, by employing a more conservative denoising strategy on the parametric curves (such as discarding high-noise curves and reducing the weight of medium-noise curves), the basic accuracy of the prediction results can be ensured through denoising, while the data processing pressure of the model can be reduced through appropriate data reduction. This achieves a better balance between prediction accuracy and data processing pressure, thereby quickly outputting a more accurate lung compliance prediction result and supporting doctors to make better subsequent decisions.

[0121] Furthermore, this invention employs a local weighting mechanism for curve segments, which is a further refinement of the overall restrictive noise reduction strategy. Specifically, it applies weighted adjustments to the segments of interest within the parametric curve, while retaining the original weights of the remaining segments.

[0122] In other words, segment weighting eliminates the need for overall curve refitting or reconstruction; weight adjustments are performed only on local segments, significantly reducing the computational and memory requirements of data preprocessing. This mechanism helps avoid high computational overhead, enabling the model to maintain a fast response even in continuous monitoring scenarios.

[0123] In some embodiments, the steps further include: Recommended settings for ventilator parameters are generated based on the lung compliance values.

[0124] In some embodiments, the input data originates from various parameter curves acquired by a spirometer. For each spirometer test, the following four types of curve data are treated as a complete input unit: inspiratory flow rate-volume curve; expiratory flow rate-volume curve; inspiratory time-volume curve; and expiratory time-volume curve. Each curve is uniformly resampled or interpolated to the same length during the preprocessing stage, thereby forming a time-series signal with a consistent structure. Subsequently, the four curves are combined in a predetermined order to form a multi-channel one-dimensional signal representation, where each channel corresponds to a parameter curve type.

[0125] During the model sample data construction process, it is preferable to perform quality screening and preprocessing on the parameter curve data collected by the spirometer to ensure that the data used for model training and prediction has basic integrity and usability.

[0126] Specifically, in actual spirometer testing, insufficient subject cooperation, measurement interruptions, equipment vibration, or sensor malfunctions may lead to significant incompleteness or abnormalities in some parameter curves. To address these issues, current methods involve manual screening or cleaning of sample data during the data preprocessing stage, including, but not limited to, the following: When the parameter curve has obvious time or volume intervals missing, making it impossible to form a complete inhalation or exhalation process, or when the curve has obvious abrupt changes, abnormal peaks, or long plateaus (e.g., the flow rate or volume signal remains unchanged for a long time), the curve sample can be removed. The purpose of the above screening and cleaning process is mainly to exclude obviously invalid or severely distorted measurement data, rather than to perform fine manual correction or human intervention on the parameter curves, so as to ensure that the model can still learn the real differences in lung function.

[0127] In an embodiment of the present invention, the lung compliance prediction model employs an end-to-end deep learning regression model. This model can be broadly termed a one-dimensional convolutional neural network prediction model based on multi-channel parameter curves.

[0128] Specifically, the model structure can be divided into the following functional modules, each of which adopts the following types or forms (the following names are for illustrative purposes only and do not constitute a limitation): 1. Feature extraction module: The feature extraction module employs a one-dimensional convolutional neural network (1D-CNN). This module automatically extracts temporal features reflecting the dynamic changes in the respiratory process by performing one-dimensional convolution operations on multi-channel parametric curve signals. This type of network is suitable for processing continuous time or volume sequence signals and can effectively capture local change patterns and overall trend features in parametric curves.

[0129] 2. Global Feature Aggregation Module: After completing curve-level feature extraction, a global feature aggregation module is set up to integrate features extracted from different time points. This module adopts a global average pooling feature aggregation structure to obtain a centralized representation of the overall features of a single breath test.

[0130] In some embodiments, the global feature aggregation module includes at least one of global average pooling, max pooling, or a feature weighting structure based on an attention mechanism.

[0131] 3. Regression Prediction Module: The regression prediction module employs a fully connected neural network to output corresponding lung compliance estimates based on the converged global features. This module primarily performs numerical regression, enabling the model to output continuous lung compliance prediction results.

[0132] The model types and module structures described above are implementation examples. This invention is not limited to a specific number of network layers or parameter settings, but rather aims to achieve an end-to-end mapping from spirometer parameter curves to lung compliance estimation results through the combination of the above modules.

[0133] Example 2 This invention also proposes a method for training a lung compliance prediction model, comprising the following steps: S200, acquire sample data, the sample data including: parameter curves measured during the subject's breathing process, and corresponding lung compliance values, the parameter curves including: (1) The parameter curves used to reflect the respiratory snapshot include: inspiratory flow rate-volume curve and expiratory flow rate-volume curve; (2) The parameter curves used to reflect the respiratory accumulation include: inspiratory time-volume curve and expiratory time-volume curve. S201, using the sample data to train a preset prediction model to obtain a prediction model for lung compliance; wherein, the prediction model includes: a feature extraction module, a global feature aggregation module, and a regression prediction module; wherein, S201 includes the following steps: S2011, The feature extraction module extracts feature information from the parameter curve, the feature information including a first type of feature information (or may be called the first type of feature) and a second type of feature information (or may be called the second type of feature). Among them, the first type of feature information refers to the self-data features of the curve itself, which are derived from a single parametric curve; the first type of feature information includes: (1) the rate of change and / or the magnitude of change of at least one curve segment; (2) the smoothness or fluctuation of at least one curve segment; the parametric curve can be divided into at least two curve segments; Among them, the second type of feature information refers to the trend difference of different parameter curves in the same interval (such as a certain time range or a certain capacity range); S2012, the feature information is integrated by the global feature aggregation module to form a feature vector; wherein, one feature vector is used to characterize at least one breathing test, and the breathing test includes at least one breathing cycle; S2013, The regression prediction module learns the mapping relationship between the feature vector and the lung compliance prediction value; S2014, determine whether the prediction model has converged based on the model prediction error. If so, end the training. The model prediction error is the difference between the predicted lung compliance value and the corresponding lung compliance value.

[0134] For example, the same interval can refer to the same time interval and / or the same capacity interval.

[0135] In some embodiments, the first type of feature (the curve's self-data) originates from the individual parametric curve itself and belongs to the curve's self-data features, including but not limited to the overall steepness and smoothness, the distribution characteristics of the rate of volume change during inhalation and exhalation, local inflection points, plateau segments, and overall trends of change. These features are used to reflect the mechanical response characteristics of the lungs to gas inhalation and exhalation during a single breath.

[0136] In some embodiments, the second type of feature (i.e., comparative data between curves) refers to the ability of different types of parameter curves (e.g., inspiratory flow rate-volume curve, expiratory flow rate-volume curve, inspiratory time-volume curve, and expiratory time-volume curve) to reflect the trend differences (or response differences) of the lungs in different physical quantity dimensions under the same volume range or the same respiratory stage in the same spirometer test.

[0137] In some embodiments, trend difference can refer to a difference in a type of feature information.

[0138] For example, trend differences can refer to differences in the steepness or flatness of the inspiratory flow rate-volume curve and the inspiratory time-volume curve within the same volume range.

[0139] A complete respiratory cycle includes an inhalation process and an exhalation process. However, in this invention, the second type of feature mainly emphasizes the contrast between different curve types, rather than the contrast between different respiratory cycles. That is, the second type of feature, the contrast feature, is used to characterize the coordinated variation characteristics of the same respiratory process under different measurement dimensions.

[0140] In some embodiments, the feature information further includes a first type of feature information (or may be referred to as a first type of feature) and / or a second type of feature information (or may be referred to as a second type of feature). The first type of feature information refers to the self-data characteristics of a single parameter curve itself; the second type of feature information refers to the response differences of different parameter curves across different physical quantity dimensions.

[0141] For example, the inspiratory time-volume curve can be compared with the expiratory time-volume curve at the same time point or the same respiratory stage (such as the end of inspiration or the mid-expiratory point) to analyze the patient's lung health status.

[0142] It should be understood that dividing respiratory features into first-class features (the curve itself) and second-class features (comparison between curves) enables multi-level and in-depth analysis of respiratory data, thereby providing more sufficient and reliable feature support for non-invasive lung compliance testing.

[0143] In other words, in this embodiment, the progressive analysis from intra-curve features to inter-curve correlation features allows for a more systematic and in-depth mining of respiratory data. This hierarchical processing approach helps to uncover the inherent physiological logic between multi-dimensional respiratory signals, enabling the model to learn the mapping relationship between respiratory mechanics and lung compliance more comprehensively and accurately.

[0144] In some embodiments, the present invention provides a training system for a lung compliance prediction model, used to implement the lung compliance prediction model training method steps described in any embodiment of the present invention.

[0145] In some embodiments, the present invention provides a lung compliance prediction model training system, comprising: The sample data acquisition module is used to acquire sample data, which includes: the parameter curves of the object and the corresponding lung compliance values. The parameter curve data includes: inspiratory flow rate-volume curve, expiratory flow rate-volume curve, inspiratory time-volume curve, and expiratory time-volume curve. A prediction model training module is used to train a preset prediction model using the parameter curve data and the corresponding lung compliance values ​​to obtain a prediction model for lung compliance. The prediction model includes a feature extraction module, a global feature aggregation module, and a regression prediction module. During training, the prediction model outputs predicted lung compliance values ​​based on the parameter curve data, and updates the trainable parameters in the model using an error backpropagation mechanism based on the error between the predicted lung compliance values ​​and the corresponding lung compliance values. The trainable parameters include the convolutional layer parameters in the feature extraction module and the fully connected layer parameters in the regression prediction module. The prediction model training module includes: A feature information extraction unit is used to extract feature information from the parameter curve through the feature extraction module; The feature vector integration unit is used to integrate the feature information through the global feature aggregation module to form a feature vector; wherein, a feature vector is used to characterize a breathing test, which contains at least one breathing cycle; The model parameter update unit is used to update the model parameters based on the difference between the lung compliance prediction value and the corresponding lung compliance value, by using the regression prediction module to output the mapping relationship of the lung compliance prediction value based on the feature vector.

[0146] In the operating room, lung compliance prediction can predict changes in lung mechanics due to changes in body position or the establishment of pneumoperitoneum after anesthesia induction, assisting anesthesiologists in preventing intraoperative hypoxemia. In emergency and pre-hospital transport scenarios, rapid non-invasive prediction can guide emergency personnel to implement appropriate respiratory support as early as possible in resource-limited environments. In addition, in respiratory rehabilitation and outpatient assessment, lung compliance prediction can also be used to dynamically monitor the disease progression of patients with conditions such as pulmonary fibrosis, providing quantitative evidence for adjusting treatment plans.

[0147] In some embodiments, abnormal fluctuations or signal distortions can be identified and assigned lower weights by calculating the rate of change and stability index of the respiratory curve within a set time window, thereby reducing their interference with model parameter updates.

[0148] In some embodiments, patient data with disease labels (such as COPD, asthma, etc.) may be misjudged as abnormal data because their respiratory curves may exhibit atypical characteristics (such as prolonged expiration, decreased flow rate, etc.). Therefore, this invention introduces label information during the weight calculation process to compensate for or independently set weights for this type of data, thereby ensuring its effective participation in model training.

[0149] In other words, a weighted training mechanism based on sample quality and physiological characteristics can be used to improve the model's adaptability to complex respiratory curves. Without this mechanism, the abnormal curve shape of the patient population data leads to certain errors in model prediction; after introducing weight adjustment, the model can correctly identify the physiological significance of this type of data, thereby further improving prediction accuracy.

[0150] Table 1 shows the changes in prediction results for simulation experiments using different model strategies, under the same real-world lung compliance data. Among them: Baseline is the original model output without the weighting mechanism; Weighted-1 is the prediction result after introducing a noisy data weighting mechanism in the simulation experiment, showing that abnormal bias samples are significantly corrected; Weighted-2 is the result after further simulation experiments with the introduction of data weight optimization strategies (including adaptation processing for special samples), and the predicted values ​​further converge to the true values.

[0151] Table 1: Where: MAE (Mean Absolute Error): represents the average "absolute error" between the predicted and actual values. The unit is consistent with lung compliance.

[0152] If MAE = 0.15, it means the model's average prediction error is approximately 0.15 units. RMSE (Root Mean Square Error): This is an error metric more sensitive to large errors (e.g., severe prediction bias) and better reflects model stability. MAE is the mean error. RMSE indicates the presence of severe errors. R² (Coefficient of Determination): Measures the model's ability to interpret data. A value of 1 indicates perfect prediction, 0 indicates random guessing, and <0 indicates worse than random guessing.

[0153] Based on existing experimental data, without introducing a weighting mechanism, the model's MAE is 0.658, RMSE is 0.864, and R² is 0.633. Through simulation experiments introducing a data quality-based weighting mechanism to reduce the weight of noisy data, the model performance is significantly improved, with MAE decreasing to approximately 0.54, RMSE decreasing to approximately 0.71, and R² increasing to approximately 0.72.

[0154] Furthermore, after introducing a weighted compensation mechanism for the affected population, the model simulation experiment showed that the predictive ability was further enhanced, with MAE reduced to about 0.48, RMSE reduced to about 0.65, and R² increased to about 0.78.

[0155] The results above demonstrate that the weighting mechanism can effectively improve the model's prediction accuracy and cross-population applicability.

[0156] In some embodiments, this application also provides a schematic block diagram of the structure of a computer device, please see... Figure 3 Computer programs can be used in situations such as Figure 3 It runs on the computer device shown. Figure 3 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The memory may include non-volatile storage media and internal memory. The non-volatile storage media may store an operating system and computer programs. The computer programs include program instructions that, when executed, cause the processor to perform arbitrary methods. The processor provides computational and control capabilities to support the operation of the entire computer device. The internal memory provides an environment for the execution of the computer programs in the non-volatile storage media; when executed by the processor, these programs cause the processor to perform arbitrary methods. The network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 3 The structures shown are merely block diagrams of a portion of the structure related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. It should be understood that the processor may be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0157] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0158] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a computer terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0159] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for training a lung compliance prediction model, characterized in that, Including the following steps: S100, acquire sample data, the sample data including: parameter curves measured during the subject's breathing process, and corresponding lung compliance values, the parameter curves including: (1) The parameter curves used to reflect the respiratory snapshot include: inspiratory flow rate-volume curve and expiratory flow rate-volume curve; (2) The parameter curves used to reflect the respiratory accumulation include: inspiratory time-volume curve and expiratory time-volume curve. S101, the preset prediction model is trained using the sample data to obtain a prediction model for lung compliance; wherein, the prediction model includes: a feature extraction module, a global feature aggregation module, and a regression prediction module; wherein, S101 includes the following steps: S1011, the feature extraction module extracts feature information from the parameter curve, the feature information including: (1) the rate of change and / or magnitude of change of at least one curve segment; (2) the smoothness or fluctuation of at least one curve segment; the parameter curve can be divided into at least two curve segments; S1012, the feature information is integrated by the global feature aggregation module to form a feature vector; wherein, one feature vector is used to characterize at least one breathing test, and the breathing test includes at least one breathing cycle; S1013, The regression prediction module learns the mapping relationship between the feature vector and the lung compliance prediction value; S1014, determine whether the prediction model has converged based on the model prediction error. If so, end the training. The model prediction error is the difference between the predicted lung compliance value and the corresponding lung compliance value.

2. The method according to claim 1, characterized in that, S1011 includes: The parameter curves are aligned according to a uniform scale, and a multi-channel input is constructed so that different parameter curves correspond to each other at the same physiological stage, thereby forming a multi-channel respiratory data sequence. The feature extraction module performs feature extraction on the respiratory data sequence to extract feature information; And / or, the feature extraction module includes: at least one one-dimensional convolutional layer; And / or, the feature extraction module employs a one-dimensional convolutional neural network, and the regression prediction module employs a fully connected neural network.

3. The method according to claim 2, characterized in that, S1011 also includes: The parameter curves are preprocessed; the preprocessing methods include resampling or interpolation.

4. The method according to claim 3, characterized in that, The parameter curves are obtained through a pulmonary function testing device, which includes a spirometer, a volume plethysmometer, or a sensor device for measuring respiratory flow and volume.

5. A method for applying a lung compliance prediction model, characterized in that, The application method includes: A lung compliance prediction model is provided, wherein the lung compliance prediction model is trained by a lung compliance prediction model training method as described in any one of claims 1-4; Obtain the measured parameter curves of the object under test; After preprocessing the measured parameter curves, they are input into the lung compliance prediction model, which then outputs the corresponding lung compliance values.

6. The application method of the lung compliance prediction model according to claim 5, characterized in that, The parameter curve is labeled with weight information, which includes: curve weight and / or segment weight. The curve weight is the overall influence weight of a parameter curve, and the segment weight is the local influence weight of a local region of the parameter curve. Correspondingly, the preprocessing of the measured parameter curve includes the following steps: A noise score is generated for the parameter curve using a noise assessment method. If the noise score is greater than the preset first score, the parameter curve is marked as a discard curve; When the noise score is greater than a preset second score and less than or equal to the first score, a curve weight reduction suggestion is generated for the parameter curve, and the curve weight reduction suggestion is used to reduce the curve weight.

7. The application method of the lung compliance prediction model according to claim 6, characterized in that, When generating the curve weighting suggestion, the following steps are also included: To obtain the physiological state of the subject under test; Predict the lung function level of the subject based on the physiological state; If the lung function level is greater than the set first function level, then the curve weighting suggestion is accepted.

8. The application method of the lung compliance prediction model according to claim 6, characterized in that, include: When the noise score is greater than a preset third score and less than or equal to the second score, a segment weighting suggestion is generated for the curve. The segment weighting suggestion is used to generate segment weights for at least one curve segment or to increase the segment weights of at least one curve segment.

9. The application method of the lung compliance prediction model according to claim 5, characterized in that, It also includes the following steps: Recommended settings for ventilator parameters are generated based on the lung compliance values.

10. A training system for a lung compliance prediction model, characterized in that, include: The sample data acquisition module is used to acquire sample data, which includes: parameter curves measured during the subject's breathing process, and corresponding lung compliance values. The parameter curves include: (1) The parameter curves used to reflect the respiratory snapshot include: inspiratory flow rate-volume curve and expiratory flow rate-volume curve; (2) The parameter curves used to reflect the respiratory accumulation include: inspiratory time-volume curve and expiratory time-volume curve. A prediction model training module is used to train a preset prediction model using the sample data to obtain a prediction model for lung compliance; wherein the prediction model includes: a feature extraction module, a global feature aggregation module, and a regression prediction module; wherein the prediction model training module includes: The feature information extraction unit is used to extract feature information from the parameter curve through the feature extraction module. The feature information includes: (1) the rate of change and / or the magnitude of change of at least one curve segment; (2) the smoothness or fluctuation of at least one curve segment; the parameter curve can be divided into at least two curve segments. The feature vector integration unit is used to integrate the feature information through the global feature aggregation module to form a feature vector; wherein, one of the feature vectors is used to characterize at least one breathing test, and the breathing test includes at least one breathing cycle; A mapping relationship learning unit is used to learn the mapping relationship between the feature vector and the lung compliance prediction value through the regression prediction module; The prediction error evaluation unit is used to determine whether the prediction model has converged based on the model prediction error. If it has, the training ends. The model prediction error is the difference between the predicted lung compliance value and the corresponding lung compliance value.

Images