A method and system for intelligent quality control of impulse oscillometry lung function test

Abstract

The application provides a kind of pulse oscillation lung function test intelligent quality control method and system, belong to medical instrument data processing, biomedical signal processing and artificial intelligence technical field, this method includes: synchronous acquisition oral cavity pressure, respiratory flow and facial strain / vibration signal;Three-way signal is analyzed in time-frequency to construct three-dimensional time-frequency information tensor;The tensor is input into deep learning model to carry out artifact classification.The application fundamentally solves the problem that traditional method cannot distinguish cough, swallowing, air leakage and other artifacts by fusing facial strain as an orthogonal information source, realizes intelligent quality control from "phenomenon level" to "cause level", and can provide closed-loop real-time feedback, significantly improves the automation degree and result reliability of IOS detection.

Description

Technical Field

[0001] This invention belongs to the fields of medical device data processing, biomedical signal processing and artificial intelligence technology, and specifically relates to an intelligent quality control method and system for pulse oscillation lung function testing. Background Technology

[0002] Impulse oscillometric (IOS) is a non-invasive, non-effort-dependent method for testing lung function, requiring only that the patient maintain calm breathing while on the device. This advantage makes it ideal for children, the elderly, or patients who cannot cooperate with traditional forced spirometer tests.

[0003] However, the results of IOS testing are extremely sensitive to the patient's cooperation during the testing process. Unconscious behaviors of the subject during the testing process, such as coughing, swallowing, glottal closure (breath-holding), speaking, or poor sealing of the instrument's mouthpiece (air leakage), can seriously interfere with the collected pressure and flow signals, leading to distortion of respiratory impedance (Zrs) calculations and thus affecting the accuracy of clinical diagnosis.

[0004] Currently, mainstream iOS devices have the following major deficiencies in quality control (QC):

[0005] 1. Reliance on manual interpretation: Existing quality control protocols primarily rely on physicians who, after testing, visually observe the pressure-flow curve and manually select respiratory signal segments they subjectively deem "qualified" for analysis. This method heavily depends on the operator's experience and subjectivity, resulting in low efficiency and a lack of consistency.

[0006] 2. Weak automatic detection capability: Existing equipment has a low level of intelligent quality control. Even if there are some automatic anomaly detection algorithms, they usually only rely on a single signal source (e.g., only analyze sudden changes in pressure or flow signals), and can only achieve "phenomenon-level" detection (i.e., "signal anomaly"), but cannot achieve "cause-level" classification (i.e., "why is it abnormal").

[0007] 3. Artifact Identification Confusion: It is extremely difficult to distinguish different types of artifacts using only pressure or flow signals. For example, a slight cough or a leak in the mouthpiece can both manifest as a sudden shift in the signal, which a single-source signal cannot effectively classify. In particular, behaviors such as "swallowing" or "glottis closure" may appear very weak in pressure / flow signals, but their impact on impedance calculations can be fatal.

[0008] 4. Inability to detect air leaks: Current technology has difficulty accurately detecting air leaks outside the instrument, which is one of the main reasons for IOS detection failure.

[0009] Therefore, there is an urgent clinical need for an intelligent quality control method and system for pulse oscillation lung function testing that can automatically, objectively, and in real time detect and accurately classify different abnormal respiratory behaviors, in order to solve the bottleneck problems of poor repeatability and difficulty in interpretation of current IOS test results. Summary of the Invention

[0010] To overcome the aforementioned deficiencies of the prior art, the present invention aims to provide an automatic, objective, and high-precision intelligent quality control method and system for pulse oscillation lung function testing. This system can not only detect abnormal respiratory signals in real time, but also accurately classify the causes of abnormalities, thereby achieving closed-loop real-time feedback guidance and high-fidelity automatic data purification.

[0011] To achieve the above objectives, the present invention provides an intelligent quality control method for pulse oscillation pulmonary function testing, comprising:

[0012] A smart quality control method for pulse oscillation pulmonary function testing includes the following steps:

[0013] (S1) Simultaneously collect oral pressure signals, respiratory flow signals, and facial strain or vibration signals of the subject during the IOS test;

[0014] (S2) Perform time-frequency joint analysis on the oral pressure signal, respiratory flow signal, and facial strain or vibration signal respectively to generate their respective two-dimensional time-frequency diagrams;

[0015] (S3) Stack the respective two-dimensional time-frequency graphs along the modal dimension to construct a three-dimensional time-frequency information tensor;

[0016] (S4) Input the three-dimensional time-frequency information tensor into a pre-trained deep learning model to perform feature extraction and classification, and output the classification label of the current signal segment;

[0017] (S5) Perform the corresponding quality control operation according to the classification label.

[0018] Furthermore, the facial strain or vibration signal in (S1) is acquired by a sensor placed on the skin surface of the subject's cheek, jaw, or throat; the sensor is preferably a piezoelectric thin film sensor or a microelectromechanical system accelerometer, used to capture mechanical vibration or surface strain caused by coughing, swallowing, air leakage, or speaking.

[0019] Furthermore, the time-frequency joint analysis in (S2) is implemented using short-time Fourier transform or continuous wavelet transform.

[0020] Furthermore, the deep learning model in (S4) is a convolutional neural network containing a large kernel attention module, which is used to capture the contextual dependencies in the three-dimensional time-frequency information tensor.

[0021] Furthermore, the classification labels in (S4) include at least “normal tidal breathing” and at least one artifact label selected from “coughing”, “swallowing”, “air leakage”, “glottic closure / breath-holding”, and “speaking”.

[0022] Furthermore, the quality control operation in (S5) includes at least one of the following:

[0023] (a) Real-time feedback: When the classification label is an artifact label, provide “cause-level” visual or voice prompts to the operator and / or subject;

[0024] (b) Intelligent interception and purification: Automatically remove all signal segments classified as artifact labels and use only signal segments classified as "normal tidal breathing" to calculate IOS lung function parameters.

[0025] The present invention also provides an intelligent quality control system for pulse oscillation lung function testing, comprising:

[0026] The multimodal signal synchronous acquisition module is used to synchronously acquire oral pressure signals, respiratory flow signals, and facial strain or vibration signals of the subject during the IOS test.

[0027] The signal conditioning and preprocessing module is used to perform digital cleaning, synchronous sampling, and filtering on the raw signal;

[0028] The multidimensional time-frequency tensor construction module is used to perform time-frequency joint analysis on each signal to generate a two-dimensional time-frequency graph, and stack it along the modal dimension to construct a three-dimensional time-frequency information tensor;

[0029] The physiological behavior tracing and classification module is used to input the three-dimensional time-frequency information tensor into a pre-trained deep learning model, extract features, and output physiological behavior tracing and classification labels.

[0030] The closed-loop quality control and feedback execution module is used to perform automated quality control processing based on the classification labels.

[0031] Furthermore, the facial strain or vibration sensor in the multimodal signal synchronous acquisition module is a piezoelectric thin film sensor or a MEMS accelerometer, deployed on the subject's cheek, jaw, or throat.

[0032] Furthermore, the deep learning model in the physiological behavior tracing and classification module is a convolutional neural network that includes a large kernel attention module.

[0033] Furthermore, the closed-loop quality control and feedback execution module includes:

[0034] The subject feedback unit provides real-time error correction instructions when artifact labels are detected.

[0035] The intelligent purification unit is used to automatically remove signal data marked as artifacts and send the purified data to the IOS parameter calculation unit.

[0036] Compared with the prior art, the technical solution disclosed in this invention has the following beneficial effects:

[0037] This invention fundamentally solves the problem of traditional methods failing to distinguish artifacts such as coughing, swallowing, and air leakage by integrating facial strain as an orthogonal information source. It achieves intelligent quality control from the "phenomenon level" to the "cause level," and provides closed-loop real-time feedback, significantly improving the automation and reliability of IOS detection. Specifically:

[0038] 1. Multimodal fusion to resolve artifact confusion: The greatest innovation of this invention lies in the introduction of "facial strain" as an orthogonal information source. Traditional methods, relying solely on pressure / flow rate, cannot distinguish between "coughing" and "air leakage." In this invention, "swallowing" produces strong features on the strain sensor, while "air leakage" may cause baseline shifts in the strain signal. This multimodal fusion fundamentally solves the ambiguity problem of single-source signals.

[0039] 2. "Cause-level" quality control, not "phenomenon-level" quality control: This invention not only detects anomalies (phenomena), but also classifies the causes of anomalies. This enables the system to provide actionable and specific feedback (such as "Do not swallow" rather than "Signal abnormality"), greatly improving detection efficiency and success rate.

[0040] 3. Closed-loop real-time feedback: Based on "cause-level" classification, the system can provide subjects with immediate and specific error correction instructions (such as "please wrap the mouthpiece tightly"), allowing subjects to dynamically correct their behavior during the test, which significantly improves the first-time success rate of the test.

[0041] 4. Objectivity and Automation: The method of this invention completely replaces subjective manual interpretation that relies on experience, realizing a standardized, objective, and automated IOS quality control process, which significantly improves the reliability of IOS test results and its clinical application value.

[0042] 5. High-fidelity data processing: By constructing a 3D time-frequency tensor and using an advanced AI model that includes a convolutional neural network with a large kernel attention module, the complex correlation information between time, frequency and modality required to distinguish different artifacts is preserved to the maximum extent, which is superior to traditional methods that rely on manual extraction of statistical features. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic flowchart of an intelligent quality control method for pulse oscillation lung function testing provided by an embodiment of the present invention;

[0045] Figure 2 A respiratory signal spectrum diagram provided for an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of an abnormal respiratory flow segment provided in an embodiment of the present invention;

[0047] Figure 4 A schematic diagram of the main nodes of the core process provided in the embodiments of the present invention; Detailed Implementation

[0048] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] like Figures 1-4 As shown, this embodiment of the invention provides an intelligent quality control method for pulse oscillation pulmonary function testing, comprising the following steps:

[0051] (S1) Simultaneously collect oral pressure signals, respiratory flow signals, and facial strain or vibration signals of the subject during the IOS test;

[0052] In the present invention embodiment (S1), the facial strain or vibration signal is acquired by a sensor placed on the skin surface of the subject's cheek, jaw or throat; the sensor is preferably a piezoelectric thin film sensor or a microelectromechanical system accelerometer, used to capture mechanical vibration or surface strain caused by coughing, swallowing, air leakage or speaking.

[0053] In practice, the piezoelectric thin-film sensor is gently attached to the underside of the subject's jaw or the side of the throat. Its excellent flexibility and skin compatibility allow it to sensitively detect the contraction and strain of the laryngeal muscles during swallowing, as well as the high-frequency vibrations generated by coughing. The MEMS accelerometer, on the other hand, can be fixed to the cheek near the corner of the mouth. By detecting minute changes in facial vibration acceleration, it effectively captures the transmission of vocal cord vibrations during speech and the localized airflow disturbances caused by air leakage. The output signals of both sensors are processed by a low-noise preamplifier circuit to ensure the signal-to-noise ratio of the original signal, providing a high-quality data foundation for subsequent analysis.

[0054] Specifically, the following steps are included:

[0055] S11, Multimodal Deployment: Subjects are instructed to attach facial strain sensors (S) to their cheeks, jaws, or throats and hold a mouthpiece with integrated pressure (P) and flow (F) sensors in their mouths.

[0056] S12, Synchronous signal acquisition: A high-speed multi-channel ADC (analog-to-digital converter) is used to synchronously sample the signals from the P, F, and S sensors to obtain the three original time series signals P(t), F(t), and S(t).

[0057] S13, Signal preprocessing: Perform anti-aliasing filtering, signal amplification, digital baseline calibration, and preliminary time-frequency adaptive filtering on the above signal.

[0058] (S2) Perform time-frequency joint analysis on oral pressure signal, respiratory flow signal and facial strain or vibration signal respectively to generate their respective two-dimensional time-frequency diagrams.

[0059] In this embodiment of the invention, the time-frequency joint analysis in (S2) is implemented using short-time Fourier transform or continuous wavelet transform.

[0060] Specifically, the following steps are included:

[0061] S21, Time window division: The system extracts an overlapping time window of a predetermined length (e.g., 512ms) in steps of predetermined time (e.g., 128ms).

[0062] S22, Time-Frequency Analysis and Processing: Perform STFT on the three signal data within each of the above time windows, and output three small-sized 2D time-frequency graphs.

[0063] (S3) Stack the two-dimensional time-frequency graphs along the modal dimension to construct a three-dimensional time-frequency information tensor.

[0064] Dynamic tensor construction is performed by stacking and normalizing three 2D time-frequency graphs to construct a 3D time-frequency information tensor. Synthesize a multidimensional feature matrix containing pressure, flow rate, and strain information (such as...) ).

[0065] Specifically, the following steps are included:

[0066] S31, Stacking: This involves creating a two-dimensional time-frequency spectrogram of the three signals. Stacking is performed along a new "modal / channel" dimension.

[0067] S32, Normalization: Normalize the generated tensor using Z-score or Min-Max to eliminate dimensional differences between different sensors.

[0068] S33, Output: Finally, a three-dimensional time-frequency information tensor (PF-Tensor) is generated, with dimensions of... This tensor contains information on three modes—pressure, flow rate, and surface strain—at any time t and any frequency f.

[0069] (S4) Input the three-dimensional time-frequency information tensor into a pre-trained deep learning model, perform feature extraction and classification, and output the classification label of the current signal segment.

[0070] In this embodiment of the invention, the deep learning model in (S4) is a convolutional neural network containing a large kernel attention module, which is used to capture the contextual dependencies in the three-dimensional time-frequency information tensor.

[0071] The classification labels in (S4) include at least “normal tidal breathing” and at least one artifact label selected from “coughing”, “swallowing”, “air leakage”, “glottic closure / breath-holding”, and “speaking”.

[0072] Specifically, the following steps are included:

[0073] S41, Feature Extraction: VGG (or its lightweight variant) is used as the backbone for feature extraction to extract local texture and hierarchical features from the time-frequency map.

[0074] S42, Classification Output: The model’s Softmax classifier accurately classifies the input 3D tensor into one of the predefined labels, such as: “[1] normal tidal breathing”, “[2] cough”, “[3] swallowing”, “[4] air leakage”, “[5] glottal closure / breath-holding”, “[6] speaking”, “[7] other artifacts”.

[0075] (S5) Perform the corresponding quality control operations according to the classification label.

[0076] The quality control operation in embodiment (S5) of the present invention includes at least one of the following:

[0077] (a) Real-time feedback: When the classification label is an artifact label, provide “cause-level” visual or verbal cues to the operator and / or subject;

[0078] (b) Intelligent interception and purification: Automatically remove all signal segments classified as artifact labels and use only signal segments classified as "normal tidal breathing" to calculate IOS lung function parameters.

[0079] In a preferred embodiment of the present invention, an intelligent quality control system for pulse oscillation lung function testing is provided, the system specifically comprising:

[0080] The multimodal signal synchronous acquisition module is used to simultaneously acquire three physiological signals during IOS testing using an oral pressure sensor, a respiratory flow sensor, and a facial strain sensor. Specifically, the facial strain or vibration sensor can be a piezoelectric thin film sensor or a MEMS accelerometer, which is physically attached to the skin surface of the subject's cheek, jaw, or throat.

[0081] The signal conditioning and preprocessing module is used to perform high-frequency synchronous sampling, anti-aliasing filtering and digital baseline calibration on the above three signals to ensure that the characteristic components of transient artifacts are captured.

[0082] The multidimensional time-frequency tensor construction module is used to perform short-time Fourier transform on each preprocessed signal and stack the resulting two-dimensional time-frequency spectrogram along the channel dimension to construct a three-dimensional time-frequency information tensor that integrates spatial and modal information.

[0083] The physiological behavior tracing and classification module is based on a deep learning model that includes a large kernel attention (LKA) module. This module extracts deep features from the three-dimensional time-frequency tensor to accurately traceive and classify typical physiological behaviors such as coughing, swallowing, and air leakage, and outputs corresponding status labels.

[0084] A closed-loop quality control and feedback execution module is used to achieve intelligent intervention based on classification labels. This module is further subdivided into a subject feedback unit and an intelligent purification unit: the former provides real-time error correction guidance via voice or interface when abnormal physiological behaviors are detected; the latter automatically removes invalid signal segments that are interfered with based on classification results, ensuring that all data entering the IOS parameter calculation unit meets clinical quality control requirements. In a preferred embodiment of the present invention, the intelligent quality control system serves as the core of the portable IOS pulmonary function instrument, achieving high-precision measurement through deep coupling of hardware and software. The system specifically consists of the following five precisely cooperating functional modules:

[0085] 1. Multimodal signal synchronous acquisition module

[0086] The multimodal signal synchronous acquisition module (corresponding to the hardware sensing layer) is used to acquire multi-source physiological signals from the subject in real time. In this embodiment, the module innovatively adopts a three-channel heterogeneous sensor configuration as the cornerstone of high-quality data acquisition:

[0087] Oral pressure sensor: Employs a high-precision miniature differential pressure sensor to detect the oscillating pressure signal P within the subject's oral cavity.

[0088] Breathing flow sensor: By measuring the pressure difference across a specially designed air resistance structure, the breathing flow signal F of the subject can be calculated in real time.

[0089] Facial strain sensor: As a key innovation of this solution, this sensor uses a piezoelectric sensor or a microelectromechanical system (MEMS) accelerometer, physically attached to the skin surface of the subject's cheek, jaw, or throat. Its core function is to capture mechanical vibrations or skin strain caused by physiological artifacts such as coughing, swallowing, air leakage, or speaking, thereby providing high signal-to-noise ratio orthogonal verification information independent of the airflow channel.

[0090] 2. Signal Conditioning and Preprocessing Module

[0091] The signal conditioning and preprocessing module (corresponding to the firmware / algorithm layer) is used to digitally clean and synchronize the original signal.

[0092] In this embodiment, the module employs a high-speed multi-channel analog-to-digital converter (ADC) to perform rigorous synchronous sampling (sampling frequency ≥ 2048Hz) on the aforementioned P, F, and S signals to ensure complete capture of the high-frequency components of transient artifacts. The acquired data undergoes anti-aliasing filtering, signal amplification, digital baseline calibration, and time-frequency adaptive filtering to build a standardized data foundation for subsequent AI inference.

[0093] 3. Multidimensional Time-Frequency Tensor Construction Module

[0094] The multidimensional time-frequency tensor construction module (corresponding to the software preprocessing layer) is used to upgrade one-dimensional time-series signals to an information-rich three-dimensional feature space.

[0095] In this embodiment, the system first performs a short-time Fourier transform (STFT) on the preprocessed three signals P(t), F(t), and S(t) using a time-frequency analysis unit, and employs an optimized window function to balance the time and frequency resolution. Subsequently, a tensor stacking unit stacks the generated three two-dimensional time-frequency maps along the "modal channel" dimension, and after normalization, finally generates a three-dimensional time-frequency information tensor (PF-Tensor). This tensor simultaneously couples the characteristic information of three modes—pressure, flow rate, and surface strain—at any given time and frequency.

[0096] In the construction of the multidimensional time-frequency tensor, the parameter indexing of the STFT has a specific physical meaning. Since the stimulated frequency of the IOS test is usually distributed between 5Hz and 35Hz, in order to accurately separate physiological artifacts (which usually contain high-frequency components), the window length selected in this embodiment is L=512 sampling points, and the overlap rate is O=75%.

[0097] The specific algorithm formula is as follows:

[0098]

[0099] Where x(n) represents the preprocessed P, F, and S signals, and w is the Hann window function. The feature matrix generated by this algorithm can effectively separate the high-frequency airflow noise caused by air leakage from the fundamental frequency component of normal breathing on the frequency axis. The 3×128×128 tensor formed after normalizing the three signals not only contains amplitude information but also implicitly contains the phase difference characteristics between pressure and flow rate, which plays a decisive role in identifying glottal closure (abrupt phase change).

[0100] 4. Physiological Behavior Tracing and Classification Module

[0101] The physiological behavior tracing and classification module (corresponding to the core AI engine layer), as the core decision-making unit of the system, is responsible for real-time decoding and causal tracing of the time-frequency tensor.

[0102] In this embodiment, the module employs a custom convolutional neural network architecture (LKA_VGG). This architecture uses a lightweight VGG as the feature extraction backbone and embeds a Large Kernel Attention (LKA) module. The LKA module can effectively capture a wide range of contextual dependencies in the time-frequency tensor, such as identifying the energy burst features of cough artifacts across the entire frequency band. Finally, a classifier accurately maps the input 3D tensor to predefined physiological behavior labels, including but not limited to normal tidal breathing, coughing, swallowing, air leakage, glottal closure, and speaking.

[0103] The specific hierarchical structure design of the LKA_VGG model is as follows:

[0104] Feature extraction layer: Contains 4 groups of convolutional blocks, each group consisting of two 3×3 convolutional layers followed by a max pooling layer with a stride of 2, with the number of channels being 64, 128, 256, and 512 respectively.

[0105] LKA core layer: A large kernel attention module is embedded after the 3rd and 4th convolutional blocks. This module consists of a 5×5 depthwise convolution, a 7×7 depthwise dilated convolution, and a 1×1 pointwise convolution.

[0106] Logical Connection: This design allows the model to achieve a receptive field of up to 13×13 without significantly increasing computational load, enabling simultaneous observation of the temporal evolution of pressure fluctuations and facial muscle strain within a respiratory cycle. This distinguishes the model from traditional methods that rely solely on threshold judgments, allowing it to identify complex "speaking and breathing" behaviors.

[0107] During the training process of the physiological behavior tracing and classification module, the input three-dimensional time-frequency tensor is not an independent feature superposition. This invention establishes a collaborative constraint mechanism based on physiological fluid dynamics and biomechanics:

[0108] Leakage-related features: When a leak occurs in the mouthpiece, the flow rate F decreases, and the facial strain signal S captures a specific low-frequency drift (0.5Hz-2Hz) caused by relaxation of the cheek muscles. The model captures this cross-modal coherence through a large kernel attention module (LKA).

[0109] Training parameter configuration: In this embodiment, the deep learning model uses the cross-entropy loss function and introduces focus loss to address the extreme imbalance between 'normal breathing' and 'transient artifact' samples during testing. The initial learning rate is set to 1×10⁻⁶. -4 The Adam optimizer was used, and the training set was iterated for 200 epochs. The accuracy on the validation set remained stable at over 96.5%.

[0110] 5. Closed-loop quality control and feedback execution module

[0111] The closed-loop quality control and feedback execution module (corresponding to the application interaction layer) is used to perform automated quality control processing based on waveform feature extraction.

[0112] In this embodiment, the system implements a closed-loop process from identification to intervention:

[0113] Real-time status guidance: On the operation interface, the system overlays preset classification prompts (such as "air leak") on the corresponding position of the breathing waveform according to the classification results.

[0114] Cause-based interactive feedback: The system executes precise voice prompts based on the classification results of physiological behaviors. For example, when an air leak is detected, it prompts "Please cover the mouthpiece with your lips"; when a glottis is detected to be closed, it prompts "Please relax your breathing", thereby achieving real-time correction of the subject's behavior.

[0115] Intelligent data purification: Based on a preset feature screening model, the system automatically identifies and removes abnormal data segments that do not conform to standard respiratory envelope characteristics. This mechanism completely replaces the subjectivity of manual screening, ensuring that all data segments used to calculate the final IOS parameters are clinically valid.

[0116] In a preferred embodiment of the present invention, the method flow of the intelligent quality control system is designed as a streaming pipeline that runs in real time on an iOS device, and the specific implementation steps are as follows:

[0117] First, system initialization and sensor deployment are performed. Before the test begins, subjects are instructed to attach facial strain sensors to their cheeks, jaws, or throats, and hold a mouthpiece integrating pressure and flow sensors in their mouths. The multimodal signal synchronous acquisition module is activated and performs sensor self-tests and baseline calibration to ensure the initial zero points of the three heterogeneous signals are consistent.

[0118] Secondly, synchronous acquisition and time window buffering of multi-channel signals are performed. Once the IOS test is initiated, the signal conditioning and preprocessing module uses a high-speed ADC to simultaneously acquire three raw time-series signals: pressure P(t), flow rate F(t), and facial strain S(t) at a sampling rate of 2048Hz. The acquired data is fed in real-time into a FIFO (First-In-First-Out) circular buffer for continuous streaming computation.

[0119] Subsequently, time windowing and real-time time-frequency transformation are performed. The system extracts overlapping time window data from the buffer with a preset step size (e.g., 128ms). The multidimensional time-frequency tensor construction module performs short-time Fourier transform (STFT) on the three signals within each time window, outputting three two-dimensional time-frequency graphs reflecting different physical dimensions.

[0120] In this embodiment, the specific mathematical implementation of the real-time time-frequency transformation follows the following definition of Short-Time Fourier Transform (STFT), and parameter indexing is provided for the physical characteristics of IOS testing:

[0121]

[0122] Where x(n) represents the synchronously acquired discrete signals P(t), F(t), or S(t), respectively; w is the optimized Hann window function, with a window length N of 1024 points (corresponding to a 500ms time window at a sampling rate of Fs=2048Hz) to ensure that the frequency resolution is sufficient to distinguish the fundamental frequency Δf≈2Hz of the IOS; the effective calculation range of the variable f is limited to 5Hz to 35Hz and its higher harmonic bands. This range covers the core excitation frequency generated by the pulse oscillator. Through this transformation, transient artifacts in the time domain (such as sudden impulses caused by coughing) are mapped to energy bursts in a specific frequency band.

[0123] Next, the dynamic construction of the multidimensional feature tensor is completed. The system stacks the three two-dimensional time-frequency images along the channel dimension and performs normalization processing to construct a three-dimensional time-frequency information tensor that couples pressure, flow rate and facial mechanical vibration information, realizing the dimensionality upgrade from one-dimensional time series signal to multidimensional feature matrix.

[0124] When constructing the three-dimensional time-frequency information tensor T, the system performs cross-modal nonlinear normalization and channel stacking, which is mathematically expressed as follows:

[0125]

[0126] Norm(·) employs the Z-score normalization algorithm to eliminate dimensional differences between pressure (kPa), flow rate (L / s), and facial strain (g or με). The generated tensor T∈R 3×F×T Not only are the three signals aligned spatially, but the phase correlation features of respiratory dynamics (P, F) and biomechanics (S) are also mathematically coupled. For example, when "glottic closure" occurs, the tensor exhibits reverse truncation in the P and F channels, while the S channel exhibits characteristic silence in the laryngeal frequency band. This multidimensional tensor feature is the fundamental basis for the subsequent "causal-level attribution" of the LKA_VGG model.

[0127] Subsequently, the process moves to the physiological behavior pattern discrimination and source classification stage. The physiological behavior source classification module feeds the constructed 3D tensor into the pre-loaded LKA_VGG deep learning model. Through the model's forward propagation calculation, the system outputs in real time the probability distribution of whether the time window belongs to "normal breathing" or various "physiological artifacts," and takes the item with the highest probability as the classification label for that time period.

[0128] The LKA_VGG model captures long-range dependencies in the time-frequency tensor through a large-kernel attention mechanism. The generation logic of its attention map is as follows:

[0129]

[0130] This formula simulates a 13×13 receptive field through decomposition of the convolution. Physically, this design allows the model to simultaneously perceive the global time-frequency evolution within a complete respiratory cycle (approximately 2-4 seconds). Compared to traditional 3×3 convolutions, the LKA module in this embodiment can effectively capture the temporal synchronicity between the persistent oscillations caused by the "speaking" behavior in the S channel and the irregular envelope in the F channel, thereby achieving high-precision physiological behavior tracing.

[0131] Finally, closed-loop quality control feedback and data purification processes are executed. The closed-loop quality control and feedback execution module drives the closed-loop logic in real time based on the classification tags:

[0132] 1. Real-time feedback: If the system identifies artifact labels such as "air leakage" or "swallowing", it immediately triggers voice error correction instructions and interface alarms to guide the subject to correct their behavior.

[0133] 2. Data Tagging and Aggregation: The system performs quality tagging on this time window within the internal data stream.

[0134] 3. Intelligent purification and parameter calculation: After the test, the system automatically retrieves all data segments marked as "normal tidal breathing" and splices them into a clean dataset. The system only uses this clean dataset to calculate the final IOS diagnostic parameters, thus eliminating the interference of artifacts on clinical diagnostic results.

[0135] Through the above steps, this invention achieves fully automated, high-precision, real-time intelligent quality control of the entire testing process. Clinical comparative experiments and technical efficacy verification are as follows:

[0136] 1. Experimental Design and Grouping

[0137] To verify the clinical effectiveness of the "feature recognition-based closed-loop quality control system" described in this invention, this study recruited a total of 120 participants and obtained their informed consent. The participants were randomly assigned to the following three groups:

[0138] Experimental group (Group A, n=40): included patients diagnosed with asthma or COPD, using the IOS system with real-time feedback and automatic purification functions described in this invention.

[0139] Traditional group (Group B, n=40): Includes patients of the same type, using standard IOS devices, with qualified waveforms manually selected by experienced clinicians.

[0140] Healthy control group (Group C, n=40): Includes healthy volunteers, used to establish a baseline for normal respiratory characteristics.

[0141] 2. Experimental Procedure and Interference Handling

[0142] Each subject underwent three repeated tests. During the experiment, the system captured and recorded the following abnormal breathing behaviors:

[0143] Real-time identification: In the experimental group, the system triggered "cause-level" feedback alarms a total of 156 times. Among them, 42 alarms were triggered for "air leakage", 68 for "swallowing", and 46 for "glottis closure".

[0144] Problem resolution rate: Upon receiving the voice prompt, 92% of the participants immediately corrected their breathing behavior within the current test cycle without having to restart the test process.

[0145] 3. Comparative Analysis of Test Results

[0146] After the experiment was completed, the processing efficiency and accuracy of the three sets of data were quantitatively analyzed, and the results are shown in the table below:

[0147] Evaluation indicators Experimental group (this invention) Traditional group (manual screening) Improvement / Enhancement Status Average time per test 180s 450s 270s Data discard rate 4.2% 21.5% Significantly reduced the burden of repeated testing. Consistency of Results (CV Value) <5% 12.8% Automatic purification ensures parameter stability. Impedance parameter R5 accuracy Highly correlated with the gold standard (r=0.98) Moderately correlated with the gold standard (r=0.85) Reduces numerical bias caused by artifacts.

[0148] 4. Experimental Conclusions

[0149] As can be seen from the above comparative experiments, the present invention effectively solves the following technical problems through feature-driven closed-loop feedback:

[0150] The reproducibility problem has been solved: the automated purification logic eliminates the subjectivity of manual screening, making the diagnostic results obtained by different operators highly consistent.

[0151] Improved clinical efficiency: Real-time "cause-level" feedback significantly reduced invalid tests due to improper subject cooperation and shortened outpatient waiting time.

[0152] Data purity is ensured: the automatic truncation algorithm can accurately identify minute, invisible swallowing or glottal closure interferences, ensuring that the basic data used for impedance calculation is "pure breathing" in a physical sense, thereby improving the sensitivity of diagnosis.

[0153] 5. Supplementary Experiments: Ablation Study Validation

[0154] To further verify the necessity of the facial strain signal S, this study conducted ablation comparisons. When only P and F signals were used as inputs, the system's accuracy in identifying "hidden air leaks" (small leaks that are difficult to detect with the naked eye) was only 72%; however, after introducing the facial strain signal S and fusing it using three-dimensional tensors, the accuracy improved to 94%. Experiments demonstrate that the facial strain signal provides crucial physical boundary conditions for the model, effectively preventing the model from misclassifying "deep breathing" as "obstructive air leaks."

Claims

1. A smart quality control method for pulse oscillation pulmonary function testing, characterized in that, Includes the following steps: (S1) Simultaneously collect oral pressure signals, respiratory flow signals, and facial strain or vibration signals of the subject during the IOS test; (S2) Perform time-frequency joint analysis on the oral pressure signal, respiratory flow signal, and facial strain or vibration signal respectively to generate their respective two-dimensional time-frequency diagrams; (S3) Stack the respective two-dimensional time-frequency graphs along the modal dimension to construct a three-dimensional time-frequency information tensor; (S4) Input the three-dimensional time-frequency information tensor into a pre-trained deep learning model to perform feature extraction and classification, and output the classification label of the current signal segment; (S5) Perform the corresponding quality control operation according to the classification label.

2. The method according to claim 1, characterized in that, The facial strain or vibration signal in (S1) is acquired by one or more sensors placed on the skin surface of the subject's cheek, jaw, or throat to capture surface strain or mechanical vibration signals caused by coughing, swallowing, air leakage, or speaking.

3. The method according to claim 2, characterized in that, The sensor is a piezoelectric thin-film sensor or a microelectromechanical system accelerometer.

4. The method according to claim 1, characterized in that, In step (S3), the two-dimensional time-frequency graphs are stacked and normalized to construct a three-dimensional time-frequency information tensor. The three-dimensional time-frequency information tensor contains three modal information of pressure, flow rate and surface strain at any time t and any frequency f.

5. The method according to claim 1, characterized in that, The deep learning model in (S4) is a convolutional neural network containing a large kernel attention module, which is used to capture the contextual dependencies in the three-dimensional time-frequency information tensor.

6. The method according to claim 1, characterized in that, The classification labels in (S4) include at least "normal tidal breathing" and at least one artifact label selected from "cough", "swallowing", "air leakage", "glottic closure / breath-holding", and "speaking".

7. The method according to claims 1 to 6, characterized in that, The quality control operation in (S5) includes at least one of the following: (a) Real-time feedback: When the classification label is an artifact label, provide the operator and / or subject with a "cause-level" visual or voice prompt; (b) Intelligent interception and purification: Automatically remove all signal segments classified as artifact labels and use only signal segments classified as "normal tidal breathing" to calculate IOS lung function parameters.

8. An intelligent quality control system for pulse oscillation lung function testing, characterized in that, include: The multimodal signal synchronous acquisition module is used to synchronously acquire oral pressure signals, respiratory flow signals, and facial strain or vibration signals of the subject during the IOS test. The signal conditioning and preprocessing module is used to clean and synchronously sample the acquired signals to provide a standard data basis for subsequent processing. The multidimensional time-frequency tensor construction module is used to perform time-frequency joint analysis on the oral pressure signal, respiratory flow signal, and facial strain or vibration signal, and stack the generated two-dimensional time-frequency graphs along the modal dimension to construct a three-dimensional time-frequency information tensor. The physiological behavior tracing and classification module is used to input the three-dimensional time-frequency information tensor into a pre-trained deep learning model, perform feature extraction and classification, and output the classification label of the current signal segment; The closed-loop quality control and feedback execution module is used to perform corresponding quality control operations based on the classification labels.

9. The system according to claim 8, characterized in that, The facial strain or vibration sensor in the multimodal signal synchronous acquisition module is a piezoelectric thin film sensor or a MEMS accelerometer, deployed on the subject's cheek, jaw, or throat.

10. The system according to claim 8, characterized in that, The deep learning model in the extraction and classification module is a convolutional neural network that includes a large kernel attention module.

11. The system according to claim 8, characterized in that, The execution module includes: The subject feedback unit provides real-time error correction instructions when artifact labels are detected. The intelligent purification unit is used to automatically remove signal data marked as artifacts and send the purified data to the IOS parameter calculation unit.

Images