A respiratory pattern anomaly recognition method and system based on time series data mining

Abstract

The present application relates to the field of respiratory mode anomaly detection, and proposes a respiratory mode anomaly recognition method and system based on time series data mining, which comprises: collecting original respiratory signals and preprocessing to obtain respiratory time series data; using a multi-scale feature extraction network to process the respiratory time series data to obtain a multi-scale respiratory feature vector; using a generative adversarial network to perform sample augmentation on the multi-scale respiratory feature vector and labeled samples to obtain an augmented training sample set; using a decision system to train a model on the augmented training sample set and perform anomaly detection on the respiratory feature vector to be detected to obtain an anomaly detection result and an uncertainty evaluation result; and based on the anomaly detection result and the uncertainty evaluation, combining clinical risk assessment to output graded early warning information. The present application ensures that the feature extraction process is highly matched with the individual respiratory cycle difference, and realizes high-precision detection, dynamic risk assessment and graded early warning of respiratory mode anomalies.

Description

Technical Field

[0001] This invention relates to the field of respiratory pattern anomaly detection, and in particular to a method and system for respiratory pattern anomaly identification based on time-series data mining. Background Technology

[0002] Respiratory pattern anomaly identification refers to the technical methods used to identify and diagnose various respiratory diseases by analyzing and processing respiratory time-series data. With the increasing incidence of chronic respiratory diseases, respiratory pattern anomaly identification has become a crucial part of clinical diagnosis. Traditional respiratory anomaly detection mainly relies on doctors' clinical experience and simple physiological indicator monitoring; however, this method suffers from high subjectivity and limited detection accuracy. Compared with traditional methods, respiratory pattern anomaly identification methods based on time-series data mining can extract deep-level feature patterns from a large amount of respiratory signals, identify subtle abnormal changes, and have higher detection accuracy and stronger objectivity. These methods typically include multiple stages such as signal preprocessing, feature extraction, pattern recognition, and anomaly detection, and can simultaneously process respiratory feature information in both the time and frequency domains, exhibiting greater adaptability, higher detection sensitivity, and a wider range of applications. However, respiratory pattern anomaly identification based on time-series data mining also faces technical challenges, the key being how to effectively extract multi-scale features of respiratory signals, handle sample imbalance, and achieve clinically interpretable anomaly detection results.

[0003] In existing technologies, respiratory pattern abnormality identification mainly employs fixed-window convolutional neural networks and traditional machine learning methods for feature extraction and abnormality detection, achieving basic respiratory abnormality identification functions. However, the fixed-window feature extraction method cannot adapt to the differences in respiratory cycles among different individuals, leading to inaccurate feature extraction and consequently affecting the accuracy and robustness of abnormality detection, resulting in poor performance in practical clinical applications. Summary of the Invention

[0004] In view of this, the present invention proposes a respiratory pattern anomaly identification method and system based on time-series data mining, which solves the problem that the existing fixed-window feature extraction method cannot adapt to the differences in respiratory cycles of different individuals, resulting in inaccurate feature extraction, which in turn affects the accuracy and robustness of anomaly detection and has poor performance in actual clinical applications.

[0005] The technical solution of this invention is implemented as follows: On one hand, this invention provides a method for identifying respiratory pattern anomalies based on time-series data mining, comprising the following steps:

[0006] Raw respiratory signals are acquired, and signal denoising, baseline correction, and segmentation are performed on the raw respiratory signals to obtain respiratory time series data;

[0007] A multi-scale feature extraction network that is adaptive to the respiratory cycle is used to process respiratory time series data to obtain multi-scale respiratory feature vectors;

[0008] The multi-scale respiratory feature vectors and labeled samples are amplified using a generative adversarial network guided by respiratory pathology to obtain an amplified training sample set.

[0009] The amplified training sample set is used to train the model through a decision system for respiratory clinical risk perception, and anomaly detection is performed on the respiratory feature vectors to be detected to obtain anomaly detection results and uncertainty assessment results.

[0010] Based on the abnormal detection results and uncertainty assessment, combined with clinical risk assessment, graded early warning information is output.

[0011] Based on the above technical solutions, preferably, the step of using a respiratory cycle adaptive multi-scale feature extraction network to process respiratory time-series data to obtain multi-scale respiratory feature vectors includes:

[0012] The multi-scale feature extraction network includes a respiratory cycle adaptive convolutional layer, a respiratory phase-sensitive attention mechanism, and a respiratory-specific spectral analysis module.

[0013] The respiratory cycle adaptive convolutional layer receives respiratory time series data, detects changes in the respiratory cycle in real time through a cycle-aware mechanism, dynamically adjusts the size of the convolutional window, and outputs time-domain features with synchronized cycle.

[0014] The breathing phase-sensitive attention mechanism receives the time-domain features, decomposes the breathing signal into three phases—inspiratory, expiratory, and transition phases—through a phase encoder, adaptively assigns attention weights to each phase, and outputs phase-weighted features.

[0015] The respiratory-specific spectrum analysis module receives respiratory time-series data, performs wavelet transform using a dedicated wavelet basis function for respiratory signals, extracts frequency domain features through a non-uniform frequency band division strategy, and fuses them with the phase-weighted features to output a multi-scale respiratory feature vector.

[0016] Based on the above technical solutions, preferably, the cycle sensing mechanism receives respiratory timing data through a respiratory cycle prediction model, estimates the current respiratory cycle length in real time, dynamically adjusts the convolution kernel size according to the cycle sensitivity factor, and outputs convolution features synchronized with the respiratory cycle.

[0017] The phase encoder receives respiratory time-series data, uses a respiratory phase decomposition algorithm to decompose the respiratory signal into a phase vector representation, and combines prior knowledge of respiratory disease phase sensitivity to allocate phase weights through a multi-head attention structure to output phase differentiation features.

[0018] Based on the above technical solutions, preferably, the step of using a respiratory pathology-guided generative adversarial network to amplify the multi-scale respiratory feature vectors and labeled samples to obtain an amplified training sample set includes:

[0019] The generative adversarial network includes a pathology-guided generator, a hierarchical comparative learning framework, and a clinical validation feedback mechanism.

[0020] A generator guided by pathological knowledge receives the multi-scale respiratory feature vectors and labeled samples, and combines them with prior knowledge of respiratory diseases to generate candidate amplification samples, thus obtaining a preliminary amplification sample set.

[0021] The preliminary amplified sample set is quality-assessed and screened using a hierarchical contrastive learning framework, and the generated samples are clinically validated using a clinical validation feedback mechanism to obtain the amplified training sample set.

[0022] Based on the above technical solutions, preferably, the pathology knowledge-guided generator specifically includes:

[0023] It receives multi-scale respiratory feature vectors and corresponding disease labeling information, extracts feature distribution patterns of different disease types through a respiratory pathology knowledge base, adopts a conditional generative adversarial network architecture, uses prior disease knowledge as a conditional constraint input generator to generate candidate amplified samples that conform to the feature distribution of specific diseases, and outputs a preliminary amplified sample set with disease labels.

[0024] The hierarchical comparative learning framework and clinical validation feedback mechanism specifically include:

[0025] The hierarchical contrastive learning framework receives the initial augmented sample set, evaluates the similarity between the sample and the real sample through sample-level contrastive learning, evaluates the feature consistency of the generated sample through feature-level contrastive learning, and outputs the quality assessment result. The clinical validation feedback mechanism receives the quality assessment result, evaluates the medical effectiveness of the generated sample through clinical expert annotation and clinical indicator validation, selects samples that meet clinical standards, and outputs the augmented training sample set.

[0026] Based on the above technical solutions, preferably, the step of training the model on the amplified training sample set using the respiratory clinical risk perception decision system, and performing anomaly detection on the respiratory feature vectors to be detected to obtain anomaly detection results and uncertainty assessment results includes:

[0027] The decision-making system includes risk-weighted Bayesian inference of respiratory diseases, dynamic threshold adjustment of physiological state perception, and clinical evolution trajectory analysis.

[0028] By receiving and amplifying the training sample set through the risk-weighted Bayesian inference of respiratory diseases, a risk weight matrix of respiratory diseases is constructed. The clinical risk level of different disease types is used as the prior probability weight to adjust the Bayesian inference process based on risk perception, and the risk-weighted disease probability distribution is output.

[0029] The system adjusts the received respiratory feature vector and patient physiological state parameters through the dynamic threshold of the physiological state perception, analyzes the patient's current age, gender, underlying diseases and other physiological state information based on the physiological state assessment module, dynamically adjusts the abnormal detection threshold, and outputs an individualized detection threshold.

[0030] The clinical evolution trajectory analysis receives historical respiratory data and current test results, constructs an individualized respiratory pattern evolution trajectory model, analyzes the development trend and evolution path of respiratory abnormalities, and outputs abnormality detection results and uncertainty assessment results.

[0031] Based on the above technical solutions, preferably, the dynamic threshold adjustment of the respiratory disease risk-weighted Bayesian inference and physiological state perception specifically includes:

[0032] The respiratory disease risk weight matrix is ​​constructed through statistical analysis of clinical data. It uses the clinical severity, treatment urgency, and misdiagnosis risk of different respiratory diseases as weighting factors, and performs risk perception weighting adjustment on the prior probability and likelihood function of Bayesian inference to output a disease probability distribution based on clinical risk.

[0033] The physiological state assessment module receives the patient's age, gender, BMI, medical history, and current medication status. It calculates an individualized physiological state score based on a multi-dimensional physiological state scoring algorithm, and dynamically calculates a threshold adjustment coefficient in conjunction with the respiratory baseline level. The detection threshold is then adjusted individually, and a personalized detection threshold adapted to the patient's physiological state is output.

[0034] Based on the above technical solutions, preferably, the acquisition of raw respiratory signals, the denoising, baseline correction, and segmentation of the raw respiratory signals to obtain respiratory time-series data include:

[0035] The raw respiratory signals collected are evaluated for signal quality. Based on the evaluation results, the corresponding adaptive denoising algorithm is selected to perform signal denoising and baseline correction to obtain purified respiratory signals.

[0036] The respiratory cycle of the purified respiratory signal is detected, and intelligent segmentation is performed based on the detected respiratory cycle information to obtain preprocessed respiratory time sequence data.

[0037] Based on the above technical solutions, preferably, the step of outputting graded early warning information based on the abnormal detection results and uncertainty assessment, combined with clinical risk assessment, includes:

[0038] The graded early warning information includes the abnormality type, severity, and confidence level. It receives the abnormality detection results and uncertainty assessment results, and performs risk level assessment by combining patient basic information and respiratory disease risk database through the clinical risk assessment module to obtain a comprehensive risk assessment result.

[0039] Based on the comprehensive risk assessment results, the anomaly type, severity, and confidence level are determined by the graded early warning generation module, and graded early warning information is generated according to the preset graded standards.

[0040] On the other hand, the present invention also provides a respiratory pattern anomaly identification system based on time-series data mining, the system comprising:

[0041] The respiratory data acquisition module is used to acquire raw respiratory signals, perform signal denoising, baseline correction and segmentation on the raw respiratory signals to obtain respiratory time series data;

[0042] The respiratory feature extraction module is used to process respiratory time series data using a respiratory cycle adaptive multi-scale feature extraction network to obtain multi-scale respiratory feature vectors.

[0043] The training sample amplification module is used to amplify the multi-scale respiratory feature vectors and labeled samples through a generative adversarial network guided by respiratory pathology, so as to obtain an amplified training sample set.

[0044] The respiratory abnormality detection module is used to train the model on the amplified training sample set through the respiratory clinical risk perception decision system, and to perform abnormality detection on the respiratory feature vector to be detected, so as to obtain the abnormality detection result and the uncertainty assessment result.

[0045] The anomaly detection and assessment module is used to output graded early warning information based on the anomaly detection results and uncertainty assessment, combined with clinical risk assessment.

[0046] The breathing pattern anomaly identification method and system based on time-series data mining of the present invention has the following advantages over the prior art:

[0047] (1) By integrating the respiratory cycle adaptive multi-scale feature extraction network, the respiratory pathology-guided generative adversarial network and the respiratory clinical risk perception decision system, the feature extraction window is dynamically adjusted based on the respiratory cycle adaptive mechanism. The sample amplification method guided by pathological knowledge is adopted, and the clinical risk perception and uncertainty assessment mechanism are introduced to realize adaptive feature extraction and multi-scale feature fusion of respiratory time series data. This ensures that the feature extraction process is highly matched with the individual respiratory cycle differences, and realizes high-precision detection, dynamic risk assessment and graded early warning of respiratory pattern abnormalities.

[0048] (2) By integrating the respiratory cycle adaptive convolutional layer, the respiratory phase-sensitive attention mechanism and the respiratory-specific spectrum analysis module, the size of the convolutional window is dynamically adjusted based on the cycle perception mechanism, the respiratory signal is decomposed by the phase encoder, and a special wavelet basis function and non-uniform frequency band division strategy for the respiratory signal are introduced. This enables the collaborative extraction of the time domain, frequency domain and phase features of the respiratory time series data. While ensuring that the feature extraction process is highly matched with the individual respiratory cycle changes, it also achieves accurate multi-scale feature representation, adaptive feature fusion and high-precision feature extraction of the respiratory pattern.

[0049] (3) By integrating Bayesian inference of respiratory disease risk weighting, dynamic threshold adjustment of physiological state perception and clinical evolution trajectory analysis, the prior probability of Bayesian inference is adjusted based on the respiratory disease risk weight matrix, the detection threshold is dynamically adjusted using the physiological state assessment module, and an individualized respiratory pattern evolution trajectory model is introduced to realize differentiated treatment of clinical risks of different disease types and precise adjustment of individualized detection thresholds, as well as risk perception and dynamic threshold optimization of abnormal respiratory pattern detection. Attached Figure Description

[0050] 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. 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.

[0051] Figure 1 This is a flowchart of a breathing pattern anomaly identification method based on time-series data mining according to the present invention. Detailed Implementation

[0052] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0053] Please see Figure 1 This invention provides a method for identifying respiratory pattern anomalies based on time-series data mining, comprising the following steps:

[0054] Raw respiratory signals are acquired, and signal denoising, baseline correction, and segmentation are performed on the raw respiratory signals to obtain preprocessed respiratory time series data.

[0055] A multi-scale feature extraction network that is adaptive to the respiratory cycle is used to process respiratory time series data to obtain multi-scale respiratory feature vectors;

[0056] The multi-scale respiratory feature vectors and labeled samples are amplified using a generative adversarial network guided by respiratory pathology to obtain an amplified training sample set.

[0057] The amplified training sample set is used to train the model through a decision system for respiratory clinical risk perception, and anomaly detection is performed on the respiratory feature vectors to be detected to obtain anomaly detection results and uncertainty assessment results.

[0058] Based on the abnormality detection results and uncertainty assessment, combined with clinical risk assessment, graded early warning information is output, which includes the abnormality type, severity, and confidence level.

[0059] Specifically, this embodiment integrates a respiratory cycle adaptive multi-scale feature extraction network, a respiratory pathology-guided generative adversarial network, and a respiratory clinical risk perception decision system. Based on the respiratory cycle adaptive mechanism, it dynamically adjusts the feature extraction window, adopts a pathology knowledge-guided sample amplification method, and introduces clinical risk perception and uncertainty assessment mechanisms. This enables adaptive feature extraction and multi-scale feature fusion of respiratory time series data, ensuring that the feature extraction process is highly matched with the individual respiratory cycle differences. This achieves high-precision detection, dynamic risk assessment, and graded early warning of respiratory pattern abnormalities.

[0060] The process involves acquiring raw respiratory signals, performing signal denoising, baseline correction, and segmentation on the raw respiratory signals to obtain respiratory time-series data, including:

[0061] The raw respiratory signals collected are evaluated for signal quality. Based on the evaluation results, the corresponding adaptive denoising algorithm is selected to perform signal denoising and baseline correction to obtain purified respiratory signals.

[0062] In one specific embodiment, the signal quality assessment and adaptive denoising processing include:

[0063] The signal-to-noise ratio, baseline drift, and artifact density of the original respiratory signal are calculated, and the filtering parameters are adaptively selected based on the signal quality assessment results. When the signal-to-noise ratio is lower than the set threshold, a multi-scale denoising algorithm based on wavelet transform is used. When the baseline drift is severe, an adaptive baseline correction algorithm based on polynomial fitting is used. When motion artifacts are present, an artifact removal algorithm based on independent component analysis is used.

[0064] The respiratory cycle of the purified respiratory signal is detected, and intelligent segmentation is performed based on the detected respiratory cycle information to obtain preprocessed respiratory time sequence data.

[0065] In one specific embodiment, the respiratory cycle detection and intelligent segmentation processing includes:

[0066] The respiratory cycle is identified by a combination of peak detection and frequency domain analysis. The main respiratory frequency is determined by calculating the autocorrelation function of the respiratory signal. Based on the detected respiratory cycle length, the segment window length is set to 3-5 respiratory cycles. An overlapping segmentation strategy is adopted with an overlap rate of 50% to ensure that each segment contains complete respiratory cycle information.

[0067] Specifically, this embodiment integrates intelligent segmentation processing of signal quality assessment, adaptive denoising algorithm selection, and respiratory cycle detection. Based on the quality assessment results of signal-to-noise ratio, baseline drift degree, and artifact density, it adaptively selects filtering parameters and employs multi-scale denoising, polynomial fitting baseline correction, and independent component analysis artifact removal algorithms. It also introduces a cycle detection method combining peak detection and frequency domain analysis to achieve personalized preprocessing of respiratory signals of different quality and real-time accurate segmentation based on the respiratory cycle. This ensures that the preprocessing process is highly matched with the quality characteristics of the original signal, achieving high-quality purification, accurate cycle detection, and optimal segmentation processing of respiratory time series data.

[0068] The respiratory time-series data is processed using a respiratory cycle-adaptive multi-scale feature extraction network to obtain multi-scale respiratory feature vectors, including:

[0069] The multi-scale feature extraction network includes a respiratory cycle adaptive convolutional layer, a respiratory phase-sensitive attention mechanism, and a respiratory-specific spectral analysis module.

[0070] The respiratory cycle adaptive convolutional layer receives respiratory time series data, detects changes in the respiratory cycle in real time through a cycle-aware mechanism, dynamically adjusts the size of the convolutional window, and outputs time-domain features with synchronized cycle.

[0071] The breathing phase-sensitive attention mechanism receives the time-domain features, decomposes the breathing signal into three phases—inspiratory, expiratory, and transition phases—through a phase encoder, adaptively assigns attention weights to each phase, and outputs phase-weighted features.

[0072] The respiratory-specific spectrum analysis module receives respiratory time-series data, performs wavelet transform using a dedicated wavelet basis function for respiratory signals, extracts frequency domain features through a non-uniform frequency band division strategy, and fuses them with the phase-weighted features to output a multi-scale respiratory feature vector.

[0073] In one specific embodiment, the formula for dynamically adjusting the convolution window that adapts to the respiratory cycle is:

[0074] ;

[0075] in, For a moment The dynamic convolution window size, The base convolution window size (usually 3-5 sampling points). This is the periodicity sensitivity coefficient, used to control the degree to which periodic variations affect the window size. For real-time predicted respiratory cycle length, The historical average respiratory cycle It is an adaptive function based on respiratory variability. The respiratory variability coefficient is used to represent the regularity of the respiratory cycle. This is the variability impact coefficient, used to control the intensity of the impact of variability on window adjustment.

[0076] The formula for calculating the attention weights that are sensitive to the breathing phase is as follows:

[0077] ;

[0078] ;

[0079] in, For position Arrive at the location Breathing phase-sensitive attention weights For position The query vector, For position The key vector, Let be the dimension of the key vector. Given the length of the input sequence, This is a phase bias term used to introduce respiratory phase information. The phase sensitivity coefficient is used to adjust the intensity of the influence of phase information. This represents the currently detected respiratory cycle length. and Positions and The corresponding time point, This is a disease-specific phase shift, used to represent the phase characteristic shift amount of different diseases.

[0080] The cycle sensing mechanism receives respiratory time series data through a respiratory cycle prediction model, estimates the current respiratory cycle length in real time, dynamically adjusts the convolution kernel size according to the cycle sensitivity factor, and outputs convolution features synchronized with the respiratory cycle.

[0081] The phase encoder receives respiratory time-series data, uses a respiratory phase decomposition algorithm to decompose the respiratory signal into a phase vector representation, and combines prior knowledge of respiratory disease phase sensitivity to allocate phase weights through a multi-head attention structure to output phase differentiation features.

[0082] Specifically, this embodiment integrates a respiratory cycle adaptive convolutional layer, a respiratory phase-sensitive attention mechanism, and a respiratory-specific spectrum analysis module. Based on the cycle-aware mechanism, it dynamically adjusts the size of the convolutional window, uses a phase encoder to decompose the respiratory signal, and introduces a dedicated wavelet basis function for the respiratory signal and a non-uniform frequency band division strategy. This enables the collaborative extraction of time-domain, frequency-domain, and phase features of respiratory time-series data. While ensuring that the feature extraction process is highly matched with the individual respiratory cycle changes, it achieves accurate multi-scale feature representation, adaptive feature fusion, and high-precision feature extraction of the respiratory pattern.

[0083] The multi-scale respiratory feature vectors and labeled samples are augmented using a generative adversarial network guided by respiratory pathology to obtain an augmented training sample set, including:

[0084] The generative adversarial network includes a pathology-guided generator, a hierarchical comparative learning framework, and a clinical validation feedback mechanism.

[0085] A generator guided by pathological knowledge receives the multi-scale respiratory feature vectors and labeled samples, and combines them with prior knowledge of respiratory diseases to generate candidate amplification samples, thus obtaining a preliminary amplification sample set.

[0086] In one specific embodiment, the pathology knowledge-guided generator specifically includes:

[0087] It receives multi-scale respiratory feature vectors and corresponding disease labeling information, extracts feature distribution patterns of different disease types through a respiratory pathology knowledge base, and adopts a conditional generative adversarial network architecture. It uses prior disease knowledge as a conditional constraint input generator to generate candidate amplified samples that conform to the feature distribution of specific diseases, and outputs a preliminary amplified sample set with disease labels.

[0088] The preliminary amplified sample set is quality-assessed and screened using a hierarchical contrastive learning framework, and the generated samples are clinically validated using a clinical validation feedback mechanism to obtain the amplified training sample set.

[0089] In one specific embodiment, the hierarchical contrastive learning framework and clinical validation feedback mechanism specifically include:

[0090] The hierarchical contrastive learning framework receives the initial augmented sample set, evaluates the similarity between the sample and the real sample through sample-level contrastive learning, evaluates the feature consistency of the generated sample through feature-level contrastive learning, and outputs the quality assessment result. The clinical validation feedback mechanism receives the quality assessment result, evaluates the medical effectiveness of the generated sample through clinical expert annotation and clinical indicator validation, selects samples that meet clinical standards, and outputs the augmented training sample set.

[0091] Specifically, this embodiment integrates a pathology-guided generator, a hierarchical comparative learning framework, and a clinical validation feedback mechanism. Based on prior knowledge constraints of respiratory diseases, it generates candidate samples that conform to the distribution of specific disease characteristics. It uses sample-level and feature-level comparative learning to evaluate the quality of generated samples and introduces clinical expert annotation and clinical indicator validation mechanisms. This enables targeted sample generation for respiratory disease characteristics and sample screening based on clinical effectiveness, ensuring a high degree of consistency and medical effectiveness between the amplified samples and real clinical samples. It achieves high-quality amplification of respiratory abnormality samples, clinical effectiveness validation, and optimization of the training dataset.

[0092] The decision-making system for respiratory clinical risk perception trains a model on the amplified training sample set and performs anomaly detection on the respiratory feature vectors to be detected, obtaining anomaly detection results and uncertainty assessment results, including:

[0093] The decision-making system includes risk-weighted Bayesian inference of respiratory diseases, dynamic threshold adjustment of physiological state perception, and clinical evolution trajectory analysis.

[0094] By receiving and amplifying the training sample set through the risk-weighted Bayesian inference of respiratory diseases, a risk weight matrix of respiratory diseases is constructed. The clinical risk level of different disease types is used as the prior probability weight to adjust the Bayesian inference process based on risk perception, and the risk-weighted disease probability distribution is output.

[0095] The system adjusts the received respiratory feature vector and patient physiological state parameters through the dynamic threshold of the physiological state perception, analyzes the patient's current age, gender, underlying diseases and other physiological state information based on the physiological state assessment module, dynamically adjusts the abnormal detection threshold, and outputs an individualized detection threshold.

[0096] The clinical evolution trajectory analysis receives historical respiratory data and current test results, constructs an individualized respiratory pattern evolution trajectory model, analyzes the development trend and evolution path of respiratory abnormalities, and outputs abnormality detection results and uncertainty assessment results.

[0097] In one specific embodiment, the Bayesian inference formula for risk-weighted respiratory diseases is as follows:

[0098] ;

[0099] ;

[0100] in, Given observation features Lower disease The posterior probability, For disease Observational characteristics under conditions The likelihood probability, For disease The prior probability (based on epidemiological data). For disease The overall risk weight, This is a risk sensitivity parameter used to adjust the degree of influence of risk weights in decision-making. This represents the total number of disease categories. This is a weighting coefficient for clinical severity. As a weighting factor for the urgency of treatment, This is a weighting coefficient for the risk of misdiagnosis. For disease Clinical severity (0-1 standardized dimension). Treatment urgency index (0-1 standardized dimension). The risk coefficient for misdiagnosis (0-1 standardized dimension).

[0101] The dynamic threshold formula for the physiological state perception is:

[0102] ;

[0103] ;

[0104] ;

[0105] in, This is the dynamically adjusted anomaly detection threshold. The basic detection threshold (representing the standard for healthy individuals). The total number of physiological state factors (including physiological state factors such as age and comorbidities). For the first The weighting coefficients of each physiological state factor. For the first A rating function for each physiological state. For the first A physiological state parameter value, The patient's age (in years). The number of underlying diseases is given, 65 represents the threshold age for the elderly population, 1.2, 0.2, and 0.1 are empirical coefficients of the age function, and 0.15 is the coefficient for the impact of comorbidities. The hyperbolic tangent function is used to smooth out the effects of age. It is a natural logarithm function used for nonlinear modeling of the number of comorbidities.

[0106] Specifically, when Time (age factor):

[0107] ,in, (Patient's age)

[0108] ;

[0109] when Time (comorbidity factors):

[0110] ,in, (Number of underlying diseases)

[0111] ;

[0112] In this embodiment, the specific calculation formula for the dynamically adjusted anomaly detection threshold is as follows:

[0113] ;

[0114] The weighting coefficient for the age factor. These are the weighting coefficients for comorbidity factors. The patient's age (in years). The number of underlying diseases is given, 65 is the threshold age for the elderly population, 1.2, 0.2, and 0.1 are the empirical coefficients of the age function, and 0.15 is the coefficient of comorbidity impact.

[0115] The risk-weighted Bayesian inference of respiratory diseases and the dynamic threshold adjustment of physiological state perception specifically include:

[0116] The respiratory disease risk weight matrix is ​​constructed through statistical analysis of clinical data. It uses the clinical severity, treatment urgency, and misdiagnosis risk of different respiratory diseases such as asthma, COPD, and pneumonia as weighting factors. The prior probability and likelihood function of Bayesian inference are adjusted by risk perception weighting, and the disease probability distribution based on clinical risk is output.

[0117] The physiological state assessment module receives the patient's age, gender, BMI, medical history, and current medication status. It calculates an individualized physiological state score based on a multi-dimensional physiological state scoring algorithm, and dynamically calculates a threshold adjustment coefficient in conjunction with the respiratory baseline level. The detection threshold is then adjusted individually, and a personalized detection threshold adapted to the patient's physiological state is output.

[0118] Specifically, this embodiment integrates Bayesian inference with respiratory disease risk weighting, dynamic threshold adjustment for physiological state perception, and clinical evolution trajectory analysis. Based on the respiratory disease risk weight matrix, it adjusts the prior probability of Bayesian inference, uses a physiological state assessment module to dynamically adjust the detection threshold, and introduces an individualized respiratory pattern evolution trajectory model. This enables differentiated processing of clinical risks for different disease types, precise adjustment of individualized detection thresholds, risk perception for abnormal respiratory pattern detection, and dynamic threshold optimization.

[0119] Based on the abnormal detection results and uncertainty assessment, combined with clinical risk assessment, the system outputs tiered early warning information, including:

[0120] The graded early warning information includes the type of abnormality, severity, and confidence level. It receives the abnormality detection results and uncertainty assessment results, and conducts a risk level assessment by combining the patient's basic information and a respiratory disease risk database through the clinical risk assessment module to obtain a comprehensive risk assessment result.

[0121] In one specific embodiment, the clinical risk assessment module receives abnormal detection results, uncertainty assessment results, and patient basic information. It matches the patient's age, gender, past medical history, and other information through a respiratory disease risk database, uses a multi-dimensional risk assessment algorithm to quantify the clinical risk of abnormal detection results, calculates a comprehensive risk score by combining the uncertainty assessment results, and outputs a comprehensive risk assessment result that includes risk level and risk factors.

[0122] Based on the comprehensive risk assessment results, the anomaly type, severity, and confidence level are determined by the graded early warning generation module, and graded early warning information is generated according to the preset graded standards.

[0123] In one specific embodiment, the graded early warning generation module receives the comprehensive risk assessment results, determines the specific abnormality type based on feature pattern matching using an abnormality type identification algorithm, maps the risk score to three severity levels (mild, moderate, and severe) using a severity quantification algorithm, calculates the final confidence level based on the model prediction confidence level and the consistency of clinical risk using a confidence calculation algorithm, and outputs graded early warning information containing abnormality type, severity, and confidence level.

[0124] Specifically, this embodiment integrates a clinical risk assessment module, an abnormality type identification algorithm, and a graded early warning generation module. Based on a respiratory disease risk database, it matches patient basic information to perform multi-dimensional risk quantification. It uses feature pattern matching to determine the abnormality type and severity mapping algorithm, and introduces a confidence calculation algorithm combined with model prediction confidence and clinical risk consistency assessment. This achieves the transformation of abnormality detection results into clinical value and personalized risk assessment, ensuring that the early warning information is highly matched with the patient's clinical condition and disease risk. It realizes accurate type identification, graded risk assessment, and personalized early warning output for respiratory pattern abnormalities.

[0125] The present invention also provides a respiratory pattern anomaly identification system based on time-series data mining, the system comprising:

[0126] The respiratory data acquisition module is used to acquire raw respiratory signals, perform signal denoising, baseline correction and segmentation on the raw respiratory signals to obtain respiratory time series data;

[0127] The respiratory feature extraction module is used to process respiratory time series data using a respiratory cycle adaptive multi-scale feature extraction network to obtain multi-scale respiratory feature vectors.

[0128] The training sample amplification module is used to amplify the multi-scale respiratory feature vectors and labeled samples through a generative adversarial network guided by respiratory pathology, so as to obtain an amplified training sample set.

[0129] The respiratory abnormality detection module is used to train the model on the amplified training sample set through the respiratory clinical risk perception decision system, and to perform abnormality detection on the respiratory feature vector to be detected, so as to obtain the abnormality detection result and the uncertainty assessment result.

[0130] The anomaly detection and assessment module is used to output graded early warning information based on the anomaly detection results and uncertainty assessment, combined with clinical risk assessment.

[0131] Specifically, this embodiment presents a respiratory pattern anomaly recognition system based on time-series data mining. By integrating a respiratory data acquisition module, a respiratory feature extraction module, a training sample amplification module, a respiratory anomaly detection module, and an anomaly detection evaluation module, it achieves end-to-end respiratory anomaly detection based on a multi-module collaborative signal processing flow. It adopts an integrated architecture of adaptive data preprocessing, multi-scale feature extraction, sample amplification, and risk perception decision-making, and introduces inter-module data flow optimization and system-level performance monitoring mechanisms. This enables fully automated processing from raw respiratory signals to graded early warning information, ensuring that the functions of each module are highly matched with clinical application needs. This achieves systematic integration, intelligent detection, and clinical application of respiratory pattern anomaly recognition.

[0132] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying respiratory pattern anomalies based on time-series data mining, characterized in that, Includes the following steps: Raw respiratory signals are acquired, and signal denoising, baseline correction, and segmentation are performed on the raw respiratory signals to obtain respiratory time series data; A multi-scale feature extraction network that is adaptive to the respiratory cycle is used to process respiratory time series data to obtain multi-scale respiratory feature vectors; The respiratory time-series data is processed using a respiratory cycle-adaptive multi-scale feature extraction network to obtain multi-scale respiratory feature vectors, including: The multi-scale feature extraction network includes a respiratory cycle adaptive convolutional layer, a respiratory phase-sensitive attention mechanism, and a respiratory-specific spectral analysis module. The respiratory cycle adaptive convolutional layer receives respiratory time series data, detects changes in the respiratory cycle in real time through a cycle-aware mechanism, dynamically adjusts the size of the convolutional window, and outputs time-domain features with synchronized cycle. The respiratory phase-sensitive attention mechanism receives the temporal features, decomposes the respiratory signal into three phases—inspiratory, expiratory, and transitional phases—through a phase encoder, adaptively assigns attention weights to each phase, and outputs phase-weighted features. The respiratory-specific spectrum analysis module receives respiratory time-series data, performs wavelet transform using a dedicated wavelet basis function for respiratory signals, extracts frequency domain features through a non-uniform frequency band division strategy, and fuses them with the phase-weighted features to output a multi-scale respiratory feature vector. The multi-scale respiratory feature vectors and labeled samples are amplified using a generative adversarial network guided by respiratory pathology to obtain an amplified training sample set. The amplified training sample set is used to train a model through a respiratory clinical risk perception decision system, and anomaly detection is performed on the respiratory feature vectors to be detected to obtain anomaly detection results and uncertainty assessment results. The decision-making system for respiratory clinical risk perception trains a model on the amplified training sample set and performs anomaly detection on the respiratory feature vectors to be detected, obtaining anomaly detection results and uncertainty assessment results, including: The decision-making system includes risk-weighted Bayesian inference of respiratory diseases, dynamic threshold adjustment of physiological state perception, and clinical evolution trajectory analysis. By receiving and amplifying the training sample set through the risk-weighted Bayesian inference of respiratory diseases, a risk weight matrix of respiratory diseases is constructed. The clinical risk level of different disease types is used as the prior probability weight to adjust the Bayesian inference process based on risk perception, and the risk-weighted disease probability distribution is output. The system adjusts the received respiratory feature vector and patient physiological state parameters through the dynamic threshold of the physiological state perception, analyzes the patient's current age, gender, and underlying disease physiological state information based on the physiological state assessment module, dynamically adjusts the abnormal detection threshold, and outputs an individualized detection threshold. The clinical evolution trajectory analysis receives historical respiratory data and current test results, constructs an individualized respiratory pattern evolution trajectory model, analyzes the development trend and evolution path of respiratory abnormalities, and outputs abnormality detection results and uncertainty assessment results. Based on the abnormal detection results and uncertainty assessment, combined with clinical risk assessment, graded early warning information is output.

2. The method for identifying respiratory pattern anomalies based on time-series data mining as described in claim 1, characterized in that, The cycle sensing mechanism receives respiratory time series data through a respiratory cycle prediction model, estimates the current respiratory cycle length in real time, dynamically adjusts the convolution kernel size according to the cycle sensitivity factor, and outputs convolution features synchronized with the respiratory cycle. The phase encoder receives respiratory time-series data, uses a respiratory phase decomposition algorithm to decompose the respiratory signal into a phase vector representation, and combines prior knowledge of respiratory disease phase sensitivity to allocate phase weights through a multi-head attention structure to output phase differentiation features.

3. The method for identifying respiratory pattern anomalies based on time-series data mining as described in claim 1, characterized in that, The multi-scale respiratory feature vectors and labeled samples are augmented using a generative adversarial network guided by respiratory pathology to obtain an augmented training sample set, including: The generative adversarial network includes a pathology-guided generator, a hierarchical comparative learning framework, and a clinical validation feedback mechanism. A generator guided by pathological knowledge receives the multi-scale respiratory feature vectors and labeled samples, and combines them with prior knowledge of respiratory diseases to generate candidate amplification samples, thus obtaining a preliminary amplification sample set. The preliminary amplified sample set is quality-assessed and screened using a hierarchical contrastive learning framework, and the generated samples are clinically validated using a clinical validation feedback mechanism to obtain the amplified training sample set.

4. The method for identifying respiratory pattern anomalies based on time-series data mining as described in claim 3, characterized in that, The pathology-guided generator specifically includes: It receives multi-scale respiratory feature vectors and corresponding disease labeling information, extracts feature distribution patterns of different disease types through a respiratory pathology knowledge base, adopts a conditional generative adversarial network architecture, uses prior disease knowledge as a conditional constraint input generator to generate candidate amplified samples that conform to the feature distribution of specific diseases, and outputs a preliminary amplified sample set with disease labels. The hierarchical comparative learning framework and clinical validation feedback mechanism specifically include: The hierarchical contrastive learning framework receives the initial augmented sample set, evaluates the similarity between the sample and the real sample through sample-level contrastive learning, evaluates the feature consistency of the generated sample through feature-level contrastive learning, and outputs the quality assessment result. The clinical validation feedback mechanism receives the quality assessment result, evaluates the medical effectiveness of the generated sample through clinical expert annotation and clinical indicator validation, selects samples that meet clinical standards, and outputs the augmented training sample set.

5. The method for identifying respiratory pattern anomalies based on time-series data mining as described in claim 1, characterized in that, The risk-weighted Bayesian inference of respiratory diseases and the dynamic threshold adjustment of physiological state perception specifically include: The respiratory disease risk weight matrix is ​​constructed through statistical analysis of clinical data. It uses the clinical severity, treatment urgency, and misdiagnosis risk of different respiratory diseases as weighting factors, and performs risk perception weighting adjustment on the prior probability and likelihood function of Bayesian inference to output a disease probability distribution based on clinical risk. The physiological state assessment module receives the patient's age, gender, BMI, medical history, and current medication status. It calculates an individualized physiological state score based on a multi-dimensional physiological state scoring algorithm, and dynamically calculates a threshold adjustment coefficient in conjunction with the respiratory baseline level. The detection threshold is then adjusted individually, and a personalized detection threshold adapted to the patient's physiological state is output.

6. The method for identifying respiratory pattern anomalies based on time-series data mining as described in claim 1, characterized in that, The process involves acquiring raw respiratory signals, performing signal denoising, baseline correction, and segmentation on the raw respiratory signals to obtain respiratory time-series data, including: The raw respiratory signals collected are evaluated for signal quality. Based on the evaluation results, the corresponding adaptive denoising algorithm is selected to perform signal denoising and baseline correction to obtain purified respiratory signals. The respiratory cycle of the purified respiratory signal is detected, and intelligent segmentation is performed based on the detected respiratory cycle information to obtain preprocessed respiratory time sequence data.

7. The method for identifying respiratory pattern anomalies based on time-series data mining as described in claim 1, characterized in that, Based on the abnormal detection results and uncertainty assessment, combined with clinical risk assessment, the system outputs tiered early warning information, including: The graded early warning information includes the abnormality type, severity, and confidence level. It receives the abnormality detection results and uncertainty assessment results, and performs risk level assessment by combining patient basic information and respiratory disease risk database through the clinical risk assessment module to obtain a comprehensive risk assessment result. Based on the comprehensive risk assessment results, the anomaly type, severity, and confidence level are determined by the graded early warning generation module, and graded early warning information is generated according to the preset graded standards.

8. A respiratory pattern anomaly identification system based on time-series data mining, used to execute the respiratory pattern anomaly identification method based on time-series data mining as described in any one of claims 1-7, characterized in that, The system includes: The respiratory data acquisition module is used to acquire raw respiratory signals, perform signal denoising, baseline correction and segmentation on the raw respiratory signals to obtain respiratory time series data; The respiratory feature extraction module is used to process respiratory time series data using a respiratory cycle adaptive multi-scale feature extraction network to obtain multi-scale respiratory feature vectors. The training sample amplification module is used to amplify the multi-scale respiratory feature vectors and labeled samples through a generative adversarial network guided by respiratory pathology, so as to obtain an amplified training sample set. The respiratory abnormality detection module is used to train the model on the amplified training sample set through the respiratory clinical risk perception decision system, and to perform abnormality detection on the respiratory feature vector to be detected, so as to obtain the abnormality detection result and the uncertainty assessment result. The anomaly detection and assessment module is used to output graded early warning information based on the anomaly detection results and uncertainty assessment, combined with clinical risk assessment.

Images