A Method and System for Detecting Pediatric Respiratory Abnormalities Based on Physiological Signals

Abstract

The present invention discloses a method and system for detecting abnormal breathing in children based on physiological signals, which relates to the field of medical information processing technology, including collecting multimodal physiological signals, using wavelet decomposition and soft threshold function to eliminate noise, combining morphological opening and closing operations to output a preprocessed signal group; calculating the time domain features, frequency domain features and nonlinear features of the preprocessed signal group, and organizing the features into a three-layer pyramid structure, using an attention mechanism to analyze the time series similarity, and outputting a fused feature sequence; performing preliminary anomaly detection on the fused feature sequence, and classifying it based on age grouping thresholds, extracting context feature vectors of windows before and after the abnormal point, and performing secondary discrimination to achieve accurate identification and graded warning of respiratory abnormalities. The present invention constructs a respiratory monitoring system with high reliability, strong adaptability and intelligent learning ability, which provides comprehensive technical support for early identification, real-time warning and intelligent diagnosis of respiratory diseases in children.

Description

Technical Field

[0001] The present invention relates to the technical field of medical information processing, and in particular to a method and system for detecting abnormal breathing in children based on physiological signals. Background Art

[0002] With the rapid development of medical Internet of Things technology, pediatric respiratory monitoring systems are playing an increasingly important role in clinical applications. Existing respiratory monitoring technologies have the following limitations: Traditional monitoring methods mostly rely on a single physiological signal, such as only collecting respiratory movement or blood oxygen saturation data, which is easily disturbed by external factors such as body movement and environmental noise, resulting in unstable monitoring results.

[0003] Existing feature extraction methods are often limited to simple time domain indicator analysis, such as basic parameters such as respiratory rate and respiratory depth. They fail to fully tap the rich information contained in respiratory signals and find it difficult to accurately reflect the overall functional status of the respiratory system. Conventional anomaly detection algorithms mostly use fixed threshold judgments, lack consideration of the differences in respiratory characteristics of children of different age groups, and fail to effectively utilize the temporal correlation of respiratory signals, resulting in a high false alarm rate.

[0004] Existing systems generally lack adaptive learning capabilities, are unable to continuously optimize detection models based on accumulated monitoring data, and are difficult to adapt to individual differences and changes in disease conditions. These technical bottlenecks have seriously restricted the clinical application of pediatric respiratory monitoring systems, and there is an urgent need to develop new intelligent respiratory monitoring technologies to improve the reliability, accuracy, and adaptability of the system. Summary of the invention

[0005] In view of the problems existing in the existing pediatric respiratory abnormality detection method based on physiological signals, the present invention is proposed.

[0006] Therefore, the problem to be solved by the present invention is how to achieve accurate identification and reliable warning of abnormal breathing in children through intelligent fusion analysis and adaptive feature learning of physiological signals.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] In the first aspect, an embodiment of the present invention provides a method for detecting abnormal breathing in children based on physiological signals, which includes: collecting multimodal physiological signals, performing preliminary filtering through a second-order Butterworth bandpass filter, eliminating noise using wavelet decomposition and a soft threshold function, removing baseline drift in combination with morphological opening and closing operations, and outputting a preprocessed signal group; segmenting the preprocessed signal group using a sliding window, calculating time domain features, obtaining frequency domain features through wavelet packet transform, calculating nonlinear features using phase space reconstruction, and organizing the features into a three-layer pyramid structure, using an attention mechanism to achieve feature weighted fusion, combining the DTW algorithm to analyze time series similarity, and outputting a fused feature sequence; inputting the fused feature sequence into a random forest classifier for preliminary anomaly detection, classifying based on age grouping thresholds, extracting context feature vectors of windows before and after the abnormal point, using an SVM classifier for secondary discrimination, and continuously optimizing model parameters through a gradient descent method to achieve accurate identification and graded warning of respiratory abnormalities.

[0009] As a preferred embodiment of the method for detecting abnormal breathing in children based on physiological signals of the present invention, the multimodal physiological signals include respiratory movement signals, electrocardiographic signals, blood oxygen saturation signals and body movement signals; the collected initial physiological signal group is input into a bandpass filter group to filter out high-frequency noise to obtain a first filtered signal; the first filtered signal is subjected to wavelet decomposition, and the transformation formula is expressed as:

[0010] ;

[0011] in, Indicates signal In scale and translation The wavelet transform coefficients under is the original input time domain signal, is the wavelet basis function, represents the energy normalization factor, Representing the scaling and translation terms of the wavelet basis function; discretizing the continuous wavelet transform formula, extracting the high-frequency coefficients and low-frequency coefficients of the first filtered signal, and obtaining a wavelet coefficient matrix.

[0012] As a preferred solution of the method for detecting abnormal breathing in children based on physiological signals of the present invention, based on the wavelet coefficient matrix, the high-frequency coefficients are shrunk by a soft threshold function, and the soft threshold function is expressed as:

[0013] ;

[0014] ;

[0015] in, represents the high frequency coefficient, is the threshold parameter, express The symbol function of is the noise standard deviation, Indicates the length of the signal, is the output value of the soft threshold function, which represents the wavelet coefficient after threshold processing; the denoising coefficient matrix is ​​reconstructed by wavelet, and the reconstruction formula is expressed as:

[0016] ;

[0017] in, is the scale function; is the wavelet function; is the maximum number of decomposition levels, is the reconstructed time domain signal, is the low-frequency approximation coefficient, is the detail coefficient, is a translation parameter; a second filtered signal is obtained by performing wavelet reconstruction on the noise reduction coefficient matrix; the second filtered signal is processed by using morphological opening and closing operations to output a preprocessed signal group, and the opening operation is expressed as:

[0018] ;

[0019] The closing operation is expressed as:

[0020] ;

[0021] in, represents the respiratory signal after wavelet denoising, Represents a structural element, represents the erosion operation, Represents the dilation operation.

[0022] As a preferred solution of the method for detecting abnormal breathing in children based on physiological signals of the present invention, the preprocessed signal group is segmented using a sliding window method, and the segmented preprocessed signal is expressed as:

[0023] ;

[0024] ;

[0025] in, Indicates signal fragments, Indicates the total length of the signal, represents the window length, represents the sliding step length, represents the time variable within the window, Indicates the total number of segments; identifies feature points through the peak-valley detection algorithm and extracts the peak point set of each signal segment , valley point set And zero-crossing set , expressed as:

[0026] ;

[0027] ;

[0028] ;

[0029] in, represents the set of all peak points, Indicates The time position of the peak point, Indicates signal exist The first-order derivative at a point, Indicates signal exist The second derivative at the point, represents all points that meet the conditions; represents the set of all valley points, Indicates The time position of the valley point, Indicates signal exist The first-order derivative at a point, Indicates signal exist The second derivative at the point; represents the set of all zero-crossing points, Indicates The time position of the zero crossing point, Indicates signal exist The value at the point, Indicates signal exist The first derivative at the point; according to the peak point set and valley point set Calculate the time domain feature vector of the respiratory signal , time domain feature vector Including respiratory rate , breathing depth Inspiratory time and exhalation time ; The respiratory rate It is expressed as:

[0030] ;

[0031] ;

[0032] in, Respiratory rate, Represents the set of time intervals between adjacent peak points, Indicates The time position of the peak point, Indicates The time position of the peak point, Indicates the peak point number, from 1 to ; Indicates the total number of peak points; the breathing depth It is expressed as:

[0033] ;

[0034] in, Indicates The amplitude of the peak point, Indicates The amplitude of the valley point, Indicates the total number of peak points or valley points; the inspiratory time It is expressed as:

[0035] ;

[0036] ;

[0037] The exhalation time It is expressed as:

[0038] ;

[0039] ;

[0040] in, represents the average inspiratory time, represents the average exhalation time, represents the time interval from the peak point to the valley point, i.e. the inspiratory phase, represents the time interval from the valley point to the next peak point, i.e., the exhalation phase. Indicates The time position of the valley point, Indicates The time position of the peak point, Indicates The time position of the peak point; by the respiratory rate Reflects the speed of breathing rhythm and judges breathing abnormalities; Reflects the ventilation volume of a single breath and evaluates respiratory efficiency. When abnormal values ​​appear, it indicates respiratory dysfunction; Inspiratory time and exhalation time Reflects the timing structure of breathing and assesses the coordination of breathing.

[0041] As a preferred solution of the method for detecting abnormal breathing in children based on physiological signals of the present invention, wavelet packet transform is used to perform multi-scale decomposition on signal segments to obtain energy distribution characteristics of different frequency bands, including frequency band energy ratio Sum band power spectrum entropy , the decomposition of the wavelet packet transform at each layer produces approximate coefficients and detail coefficients, which are expressed as:

[0042] ;

[0043] ;

[0044] in, Indicates Tier The approximation coefficient of the nodes, represents the decomposition level and the depth of the wavelet packet tree; Indicates the node number, indicating the The first nodes, Represents the discrete point number of the time series, represents the index of the filter coefficient, represents the low-pass filter coefficient, represents the coefficient of the parent node in the previous layer, Indicates Tier The detail coefficient of each node, represents the high-pass filter coefficient; the frequency band energy ratio It is expressed as:

[0045] ; ;

[0046] in, Indicates Tier The energy value of the frequency band, Indicates the decomposition level, represents the frequency band index, is the square of the wavelet packet approximation coefficient, representing the instantaneous energy; Indicates The relative energy ratio of the frequency bands, represents the maximum number of decomposition levels, represents the band count index, Indicates the total number of frequency bands in the last layer; the power spectrum entropy of the frequency bands It is expressed as:

[0047] ;

[0048] in, Indicates The relative energy ratio of the frequency bands, Indicates the maximum number of decomposition levels.

[0049] As a preferred embodiment of the method for detecting abnormal breathing in children based on physiological signals of the present invention, the phase space reconstruction technique is used to map the signal fragments into a high-dimensional phase space, and the nonlinear feature vector is calculated. , the nonlinear eigenvector Includes the maximum Lyapunov exponent , correlation dimension And the sample entropy ; Using the time delay method to construct dimensional phase space, expressed as:

[0050] ;

[0051] ;

[0052] ;

[0053] in, represents the reconstructed phase space vector, represents the original time series, Indicates time delay, represents the embedding dimension, represents the time series index, represents the autocorrelation function, Indicates the delay value, represents the sequence length, represents the mean of the series, represents the condition that minimizes the autocorrelation function value; the maximum Lyapunov exponent , expressed as:

[0054] ;

[0055] ;

[0056] ;

[0057] in, represents the maximum Lyapunov exponent, represents the total evolution time, Indicates the number of tracks, represents the distance after the phase space orbital evolution, represents the initial distance of the phase space orbit, represents the reference orbit point, represents the nearest neighbor point, represents the evolution time step; the correlation dimension It is expressed as:

[0058] ;

[0059] in, represents the correlation dimension, represents the distance threshold, represents the associated integral; further, the associated integral It is expressed as:

[0060] ;

[0061] in, represents the number of phase space points, express Step function, Represents the distance between two points in phase space; the sample entropy It is expressed as:

[0062] ;

[0063] in, represents the sample entropy, Indicates the pattern length, represents the similarity tolerance, The dimension is The number of matches when The dimension is The number of matches at .

[0064] As a preferred embodiment of the method for detecting abnormal breathing in children based on physiological signals of the present invention, , frequency domain feature vector and the nonlinear eigenvector , construct a three-layer feature pyramid, expressed as:

[0065] ;

[0066] ;

[0067] ;

[0068] in, is the bottom layer, containing time domain features; Represents the total number of time domain features; is the middle layer, containing frequency domain features; Represents the total number of frequency domain features; is the top layer, containing nonlinear features; Represents the total number of nonlinear features; for each layer of features , through three learnable weight matrices , , Perform linear transformation to obtain the query matrix , key matrix Sum Matrix , map the original features into a space suitable for calculating attention weights, and then calculate and The similarity is divided by the scaling factor to avoid the gradient vanishing problem, and The function is normalized and finally compared with the value matrix Multiply to get the attention-weighted feature representation; the feature weighted fusion process is expressed as:

[0069] ;

[0070] ;

[0071] in, is the final fused feature vector, is the feature level index, Indicates Layer feature vector, Indicates The weight coefficient of the layer feature, Indicates The importance score of the layer feature; the DTW algorithm is used to analyze the temporal relationship between multiple signal segments, and the cumulative distance matrix between two respiratory signals is calculated, which is expressed as:

[0072] ;

[0073] in, Indicates starting point To point The minimum cumulative distance, Indicates that two sequences are at position and The distance metric at Indicates the cumulative distance in the upward direction, Indicates the cumulative distance to the left. represents the cumulative distance in the diagonal direction, Select the minimum value in the three directions; convert the distance into a similarity measure, expressed as:

[0074] ;

[0075] in, For the and The similarity of the signal segments, is the scale parameter; the features are combined with the similarity information to obtain a feature sequence containing a temporal relationship , expressed as:

[0076] ;

[0077] in, represents the final feature sequence, Indicates The feature vector of the signal segment, Indicates The similarity vector between a signal segment and other segments, , Indicates the total number of signal fragments.

[0078] As a preferred embodiment of the method for detecting abnormal breathing in children based on physiological signals of the present invention, the characteristic sequence Input a pre-trained random forest classifier, which contains Decision trees are used to classify feature sequences, and the abnormal probability value is calculated according to the voting results of each decision tree, which is expressed as:

[0079] ;

[0080] in, is the number of decision trees, For the The classification results of the decision tree are is the abnormal probability value; according to the abnormal probability value and age group threshold table Perform preliminary screening, where the age group threshold is calculated as follows:

[0081] ;

[0082] in, represents the age group index; For the The mean of normal breathing indexes in the group, For the The standard deviation of normal breathing indicators in the group; according to the abnormal probability value and age group threshold table Perform preliminary screening, when the abnormal probability value When the value exceeds the normal range of the corresponding age group, the abnormal mark sequence at the corresponding time Mark as 1, otherwise mark as 0, and get the filtered abnormal mark sequence ; For the abnormal marker sequence after screening For each abnormal point in , extract the feature sequence of the 10 time windows before and after it, and calculate the context feature vector, which is expressed as:

[0083] ;

[0084] in, Indicates The characteristic sequence of 10 time windows before and after the moment The mean of Represents the feature sequence within the same time range The standard deviation of represents the average rate of change of the feature sequence, Represents the context feature vector; according to the context feature vector Train the support vector machine classifier, use the radial basis kernel function for quadratic discrimination, and add abnormal samples to the training set In the training set, the training set is used every 24 hours. Update the classifier parameters, expressed as:

[0085] ;

[0086] in, is the learning rate, is the cross entropy loss function, represents the current model parameters, is the updated model parameter; by calculating the abnormal probability at each time point , generating anomaly tag sequences , and then obtain the context feature vector , the final discrimination result is obtained through the SVM classifier .

[0087] As a preferred embodiment of the method for detecting abnormal breathing in children based on physiological signals of the present invention, in the process of distinguishing > and = 1, it means that the breathing index at this moment is initially abnormal, the breathing frequency exceeds the normal range of the age group, the breathing depth is abnormally shallow or too deep, and the breathing ratio may be unbalanced, and a second judgment confirmation is required; if the above conditions are met and the context feature vector show If it exceeds the normal range, it means that there is persistent respiratory abnormality, abnormal average respiratory level, possible apnea, and risk of respiratory failure, and the trend of change needs to be continuously monitored; if If it increases, it means that the breathing rhythm is unstable, the breathing pattern is disordered, there may be difficulty breathing, and intermittent apnea, and the fluctuation needs to be observed; if If the SVM classifier outputs > 0, it is confirmed as a real anomaly, an alarm is triggered immediately, the time of the anomaly is recorded, and the anomaly sample is added to the training set , update the model parameters after 24 hours, generate an abnormal event report, notify medical staff to pay attention, and activate the emergency response mechanism; if multiple consecutive time windows are judged as abnormal, it means that there is persistent respiratory dysfunction and there may be serious respiratory diseases, which require emergency intervention from medical staff. The alarm level will be raised and respiratory support may be required. An emergency treatment plan needs to be formulated and transfer to the intensive care unit should be considered.

[0088] In the second aspect, an embodiment of the present invention provides a pediatric respiratory abnormality detection system based on physiological signals, which includes: an acquisition module, which is used to acquire multimodal physiological signals, perform preliminary filtering through a second-order Butterworth bandpass filter, use wavelet decomposition and soft threshold function to eliminate noise, combine morphological opening and closing operations to remove baseline drift, and output a preprocessed signal group; a processing module, which is used to segment the preprocessed signal group using a sliding window, calculate time domain features, obtain frequency domain features through wavelet packet transform, use phase space reconstruction to calculate nonlinear features, and organize the features into a three-layer pyramid structure, use an attention mechanism to achieve feature weighted fusion, combine the DTW algorithm to analyze time series similarity, and output a fused feature sequence; an early warning module, which is used to input the fused feature sequence into a random forest classifier for preliminary anomaly detection, classify based on age grouping thresholds, extract context feature vectors of windows before and after the abnormal point, use an SVM classifier for secondary discrimination, and continuously optimize model parameters through the gradient descent method to achieve accurate identification and graded early warning of respiratory abnormalities.

[0089] The beneficial effects of the present invention are as follows: the present invention improves the anti-interference ability of the system through multi-source information fusion, enhances the expression ability of features by adopting multi-scale feature extraction, improves the adaptability of feature fusion by dynamic weight allocation based on attention mechanism, reduces the false alarm rate by dual discrimination mechanism, and forms a reliable, stable and self-learning respiratory monitoring system as a whole, which provides effective technical support for early warning and intelligent diagnosis of children's respiratory diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0091] Figure 1 The figure is a flow chart of a method for detecting abnormal breathing in children based on physiological signals. DETAILED DESCRIPTION

[0092] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0093] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0094] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0095] Example 1, reference Figure 1 , which is the first embodiment of the present invention, provides a method for detecting abnormal breathing in children based on physiological signals, comprising:

[0096] S1: Collect multimodal physiological signals, perform preliminary filtering through a second-order Butterworth bandpass filter, use wavelet decomposition and soft threshold function to eliminate noise, combine morphological opening and closing operations to remove baseline drift, and output a preprocessed signal group.

[0097] The piezoelectric sensor is used to collect chest and abdominal respiratory motion signals, the ECG electrode is used to collect ECG signals, the reflective blood oxygen sensor is used to collect blood oxygen saturation signals, and the triaxial acceleration sensor is used to collect body motion signals to obtain an initial physiological signal group;

[0098] Among them, the initial physiological signal group includes respiratory signals , ECG signal , blood oxygen signal and body motion signals . Respiratory signal The sampling frequency is set to 100Hz, and the ECG signal The sampling frequency is set to 250Hz, and the blood oxygen signal The sampling frequency is set to 60Hz, and the body motion signal The sampling frequency is set to 50 Hz, and the digital sampling is performed through a high-precision analog-to-digital converter with a sampling accuracy of 16 bits.

[0099] The initial physiological signal group is input into the bandpass filter group. Since the respiratory frequency of children of different age groups is significantly different, the passband width of the bandpass filter group is set according to the respiratory frequency range corresponding to the age group of children. The bandpass filter adopts a second-order Butterworth filter, and its transfer function is It is expressed as:

[0100] ;

[0101] in, is the signal frequency, is the low cutoff frequency, is a high cutoff frequency, is the filter order and =2.

[0102] For children aged 0 to 6 months =0.3Hz, =0.8Hz; Children aged 6 to 12 months =0.25Hz, =0.7Hz; Children aged 1 to 3 years =0.22Hz, =0.6Hz; children aged 3 to 6 years =0.2Hz, =0.5Hz, children aged 6 to 12 years =0.18Hz, =0.4Hz.

[0103] The high-frequency noise in the initial physiological signal group is filtered out by the band-pass filter group to obtain a first filtered signal.

[0104] The second-order Butterworth filter has the maximally flat passband response characteristic and can effectively maintain the amplitude characteristics of the signal. If a higher order is selected, although a steeper roll-off characteristic can be obtained, it will cause greater phase distortion and affect the time domain characteristics of the respiratory signal. The second-order filter has a smaller group delay and can better maintain the phase characteristics of the signal. For physiological signals such as respiratory signals that need to be monitored in real time, smaller phase distortion helps to accurately reflect the occurrence time of respiratory events.

[0105] The first filtered signal is decomposed by wavelet, and the db4 wavelet basis function is selected, and a 5-layer decomposition scale is set. The db4 wavelet basis function has good orthogonality and compact support, and is suitable for time-frequency analysis of physiological signals. The continuous wavelet transform formula is expressed as:

[0106] ;

[0107] in, Indicates signal In scale and translation The wavelet transform coefficients under is the original input time domain signal, is the wavelet basis function, represents the energy normalization factor, Represents the scaling and translation term of the wavelet basis function.

[0108] Discretize the continuous wavelet transform formula and express it as:

[0109] ;

[0110] in, represents discrete wavelet transform coefficients, is the number of decomposition layers, used to control the frequency resolution, is the time series position, used to control time positioning, is a discrete input signal, is the discrete time series index.

[0111] The high-frequency coefficients and the low-frequency coefficients of the first filtered signal are extracted to obtain a wavelet coefficient matrix.

[0112] According to the wavelet coefficient matrix, the high-frequency coefficients are shrunk by the soft threshold function, which is expressed as:

[0113] ;

[0114] ;

[0115] in, represents the high frequency coefficient, is the threshold parameter, express The symbol function of is the noise standard deviation, Indicates the length of the signal, is the output value of the soft threshold function, which represents the wavelet coefficient after threshold processing.

[0116] By retaining the low-frequency coefficients, the noise reduction coefficient matrix is ​​obtained; the soft threshold function has better continuity than the hard threshold function and can reduce mutations in the reconstructed signal.

[0117] The noise reduction coefficient matrix is ​​reconstructed by wavelet, and the reconstruction formula is expressed as:

[0118] ;

[0119] in, is a scaling function used to reconstruct the low-frequency components of the signal; is a wavelet function used to reconstruct the high-frequency components of the signal; is the maximum number of decomposition levels and =5, is the reconstructed time domain signal, is the low-frequency approximation coefficient, is the detail coefficient, is the translation parameter. The wavelet reconstruction process ensures the integrity of the reconstructed signal through the linear combination of orthogonal wavelet basis functions.

[0120] The second filtered signal is obtained by performing wavelet reconstruction on the noise reduction coefficient matrix.

[0121] The second filtered signal is processed using morphological opening and closing operations, and the opening operation is expressed as:

[0122] ;

[0123] The closing operation is expressed as:

[0124] ;

[0125] in, represents the respiratory signal after wavelet denoising, Represents a structural element, represents the erosion operation, Represents the dilation operation.

[0126] The structural element uses a flat structural element, and its length is set to 1 / 4 of the sampling period of the second filtered signal. An opening operation is first performed to remove the burrs at the signal peaks, and then a closing operation is performed to fill the depressions at the signal valleys to eliminate the baseline drift of the second filtered signal and obtain a preprocessed signal group.

[0127] S2: The preprocessed signal is segmented by sliding window, the time domain features are calculated, the frequency domain features are obtained by wavelet packet transform, the nonlinear features are calculated by phase space reconstruction, and the features are organized into a three-layer pyramid structure. The attention mechanism is used to realize feature weighted fusion, and the DTW algorithm is combined to analyze the time series similarity and output the fused feature sequence.

[0128] In order to capture the dynamic characteristics of the respiratory signal, the sliding window method is used to segment the preprocessed signal. and sliding step length , which not only ensures that the signal segment contains a complete breathing cycle, but also reflects the dynamic changes of the breathing pattern. The segmented preprocessed signal is expressed as:

[0129] ;

[0130] ;

[0131] in, Indicates signal fragments, Indicates the total length of the signal, represents the window length, represents the sliding step length, represents the time variable within the window, Indicates the total number of fragments, Indicates rounding down.

[0132] Identify feature points through the peak-valley detection algorithm and extract the peak point set of each signal segment , valley point set And zero-crossing set , expressed as:

[0133] ;

[0134] ;

[0135] ;

[0136] in, represents the set of all peak points, Indicates The time position of the peak point, Indicates signal exist The first-order derivative at a point, Indicates signal exist The second derivative at the point, represents all points that meet the conditions; represents the set of all valley points, Indicates The time position of the valley point, Indicates signal exist The first-order derivative at a point, Indicates signal exist The second derivative at the point; represents the set of all zero-crossing points, Indicates The time position of the zero crossing point, Indicates signal exist The value at the point, Indicates signal exist The first derivative at a point.

[0137] By peak point set Determine the maximum inhalation position of breathing, calculate the respiratory rate, and evaluate the respiratory amplitude; Determine the maximum exhalation position of breathing, calculate the breathing depth, and evaluate the baseline drift of breathing; Distinguish the inspiratory and expiratory phases, count the respiratory cycles, and assess the symmetry of breathing.

[0138] According to the peak point set and valley point set Calculate the time domain feature vector of the respiratory signal , time domain feature vector Including respiratory rate , breathing depth Inspiratory time and exhalation time .

[0139] Respiratory rate It is expressed as:

[0140] ;

[0141] ;

[0142] in, Respiratory rate, Represents the set of time intervals between adjacent peak points, Indicates The time position of the peak point, Indicates The time position of the peak point, Indicates the peak point number, from 1 to ; Indicates the total number of peak points.

[0143] Breathing depth It is expressed as:

[0144] ;

[0145] in, Indicates The amplitude of the peak point, Indicates The amplitude of the valley point, Indicates the total number of peak points (or valley points).

[0146] Inspiratory time It is expressed as:

[0147] ;

[0148] ;

[0149] Exhalation time It is expressed as:

[0150] ;

[0151] ;

[0152] in, represents the average inspiratory time, represents the average exhalation time, represents the time interval from the peak point to the valley point, i.e. the inspiratory phase, represents the time interval from the valley point to the next peak point, i.e., the exhalation phase. Indicates The time position of the valley point, Indicates The time position of the peak point, Indicates The time position of the peak point; by the respiratory rate Reflects the speed of breathing rhythm and judges breathing abnormalities; Reflects the ventilation volume of a single breath and evaluates respiratory efficiency. When abnormal values ​​appear, it indicates respiratory dysfunction; Inspiratory time and exhalation time Reflects the time structure of breathing and assesses the coordination of breathing. and An abnormal ratio may indicate airway obstruction.

[0153] Wavelet packet transform is used to perform multi-scale decomposition of signal segments to obtain energy distribution characteristics of different frequency bands, including frequency band energy ratio Sum band power spectrum entropy , wavelet packet transform has better time-frequency localization ability than traditional Fourier transform, and the decomposition of each layer produces approximate coefficients and detail coefficients, which are expressed as:

[0154] ;

[0155] ;

[0156] in, Indicates Tier The approximation coefficient of the nodes, represents the decomposition level and the depth of the wavelet packet tree; Indicates the node number, indicating the The first nodes, Represents the discrete point number of the time series, represents the index of the filter coefficient, represents the low-pass filter coefficient, represents the coefficient of the parent node in the previous layer, Indicates Tier The detail coefficient of each node, Represents the high-pass filter coefficients.

[0157] Band Energy Ratio It is expressed as:

[0158] ;

[0159] ;

[0160] in, Indicates Tier The energy value of the frequency band, Indicates the decomposition level, represents the frequency band index, is the square of the wavelet packet approximation coefficient, representing the instantaneous energy. Indicates The relative energy ratio of the frequency bands, represents the maximum number of decomposition levels, represents the band count index, Indicates the total number of bands in the last layer.

[0161] By relative energy ratio The relative contribution of each frequency band is evaluated to eliminate the influence of absolute differences in signal intensity and facilitate comparison between different samples.

[0162] Band power spectrum entropy It is expressed as:

[0163] ;

[0164] in, Indicates The relative energy ratio of the frequency bands, Indicates the maximum number of decomposition levels.

[0165] By frequency band spectral entropy It reflects the complexity of respiratory signals, evaluates the regularity of breathing patterns, and helps identify pathological conditions.

[0166] Use phase space reconstruction technology to map signal fragments into high-dimensional phase space and calculate nonlinear eigenvectors , the nonlinear eigenvector Includes the maximum Lyapunov exponent , correlation dimension And the sample entropy .

[0167] The nonlinear dynamic characteristics of the respiratory signal are revealed by phase space reconstruction, which can reflect the complexity and stability of the respiratory system. dimensional phase space, expressed as:

[0168] ;

[0169] ;

[0170] ;

[0171] in, represents the reconstructed phase space vector, represents the original time series, Indicates time delay, represents the embedding dimension, represents the time series index, represents the autocorrelation function, Indicates the delay value, represents the sequence length, represents the mean of the series, represents the condition that minimizes the autocorrelation function value.

[0172] Furthermore, the maximum Lyapunov exponent is calculated , expressed as:

[0173] ;

[0174] ;

[0175] ;

[0176] in, represents the maximum Lyapunov exponent, represents the total evolution time, Indicates the number of tracks, represents the distance after the phase space orbital evolution, represents the initial distance of the phase space orbit, represents the reference orbit point, represents the nearest neighbor point, represents the evolution time step.

[0177] Correlation Dimension It is expressed as:

[0178] ;

[0179] in, represents the correlation dimension, represents the distance threshold, represents the correlation integral.

[0180] Furthermore, the correlation score It is expressed as:

[0181] ;

[0182] in, represents the number of phase space points, express Step function, Represents the distance between two points in phase space.

[0183] Sample Entropy It is expressed as:

[0184] ;

[0185] in, represents the sample entropy, Indicates the pattern length, represents the similarity tolerance, The dimension is The number of matches when The dimension is The number of matches at .

[0186] Through phase space reconstruction, the dynamic characteristics of breathing are revealed, abnormal breathing patterns are identified, and the stability of the respiratory system is evaluated; through the Lyapunov index, the stability of the respiratory system is evaluated, respiratory disorders are predicted, and chaotic behavior is identified; through the correlation dimension, the complexity of respiratory regulation is evaluated, pathological conditions are identified, and respiratory pattern conversion is analyzed; through sample entropy, the regularity of breathing is evaluated, abnormal breathing patterns are identified, and changes in respiratory function are predicted. The complex dynamic characteristics of respiratory signals are described from different angles through nonlinear eigenvectors, providing an important basis for respiratory function evaluation and abnormality detection.

[0187] Construct a three-layer feature pyramid structure and transform the time domain feature vector , frequency domain feature vector and the nonlinear eigenvector They are placed at the bottom, middle and top levels of the pyramid respectively.

[0188] Based on the representation ability and computational complexity of the features, a three-layer pyramid structure is constructed, which is expressed as:

[0189] ;

[0190] ;

[0191] ;

[0192] in, is the bottom layer, containing time domain features; Represents the total number of time domain features; is the middle layer, containing frequency domain features; Represents the total number of frequency domain features; is the top layer, containing nonlinear features; Represents the total number of nonlinear features.

[0193] For each layer of features , through three learnable weight matrices , , Perform linear transformation to obtain the query matrix , key matrix Sum Matrix , the purpose is to map the original features into a space suitable for calculating attention weights. Then calculate and The similarity is divided by the scaling factor to avoid the gradient vanishing problem, and The function is normalized and finally compared with the value matrix Multiply them together to get the attention-weighted feature representation.

[0194] The feature weighted fusion process is expressed as:

[0195] ;

[0196] ;

[0197] in, is the final fused feature vector, is the feature level index (1, 2, 3 correspond to the time domain, frequency domain, and nonlinear feature layers respectively), Indicates Layer feature vector, Indicates The weight coefficient of the layer feature, Indicates Importance scores of layer features.

[0198] The dynamic time warping algorithm is used to calculate the fusion feature vector of multiple signal segments in the preprocessed signal group. The temporal similarity matrix between , output feature sequence .

[0199] The DTW algorithm is used to analyze the timing relationship between multiple signal segments and calculate the cumulative distance matrix between two respiratory signals, which is expressed as:

[0200] ;

[0201] in, Indicates starting point To point The minimum cumulative distance, Indicates that two sequences are at position and The distance metric at Indicates the cumulative distance in the upward direction, Indicates the cumulative distance to the left. represents the cumulative distance in the diagonal direction, Select the minimum value in the three directions.

[0202] Convert the distance to a similarity measure, expressed as:

[0203] ;

[0204] in, For the and The similarity of the signal segments, is the scale parameter.

[0205] Combine the features with the similarity information to obtain a feature sequence containing temporal relationships , expressed as:

[0206] ;

[0207] in, represents the final feature sequence, Indicates The feature vector of the signal segment, Indicates The similarity vector between a signal segment and other segments, , Indicates the total number of signal fragments.

[0208] Through multi-scale feature extraction, attention mechanism fusion and temporal relationship analysis, a complete feature representation containing multi-level signal features and temporal correlation information is obtained, which provides a good foundation for subsequent analysis and processing.

[0209] S3: The fused features are input into the random forest classifier for preliminary anomaly detection, and classification is performed based on the age grouping threshold. The context feature vector of the window before and after the anomaly point is extracted, and the SVM classifier is used for secondary discrimination. The model parameters are continuously optimized through the gradient descent method to achieve accurate identification of respiratory abnormalities and graded warning.

[0210] The feature sequence Input a pre-trained random forest classifier, which contains Decision trees are used to classify feature sequences, and the abnormal probability value is calculated according to the voting results of each decision tree, which is expressed as:

[0211] ;

[0212] in, is the number of decision trees, For the The classification results of the decision tree are is the abnormal probability value.

[0213] According to the abnormal probability value and age group threshold table Perform preliminary screening, where the age group threshold is calculated as follows:

[0214] ;

[0215] in, represents the age group index, =1,2,3,4,5, corresponding to 0-6 months, 6-12 months, 1-3 years, 3-6 years, and 6-12 years respectively; For the The mean of normal breathing indexes in the group, For the The standard deviation of normal breathing indicators in the group.

[0216] According to the abnormal probability value and age group threshold table Perform preliminary screening, when the abnormal probability value When the value exceeds the normal range of the corresponding age group, the abnormal mark sequence at the corresponding time Mark as 1, otherwise mark as 0, and get the filtered abnormal mark sequence .

[0217] After screening, abnormal marker sequences For each abnormal point in , extract the feature sequence of the 10 time windows before and after it, and calculate the context feature vector, which is expressed as:

[0218] ;

[0219] in, Indicates The characteristic sequence of 10 time windows before and after the moment The mean of Represents the feature sequence within the same time range The standard deviation of represents the average rate of change of the feature sequence, Represents the context feature vector.

[0220] According to the context feature vector Train the support vector machine classifier, use the radial basis kernel function for quadratic discrimination, and add abnormal samples to the training set In the training set, the training set is used every 24 hours. Update the classifier parameters, expressed as:

[0221] ;

[0222] in, is the learning rate, is the cross entropy loss function, represents the current model parameters, is the cross entropy loss function, are the updated model parameters.

[0223] By calculating the abnormal probability at each time point , generating anomaly tag sequences , and then obtain the context feature vector , the final discrimination result is obtained through the SVM classifier .

[0224] In the process of identification, if > and = 1, it means that the respiratory indicators at that moment are initially abnormal. The respiratory rate may exceed the normal range for the age group, the breathing depth may be abnormally shallow or too deep, and the breathing ratio may be unbalanced, and a secondary judgment and confirmation is required.

[0225] If the above conditions are met and the context feature vector show If it is beyond the normal range, it means that there is persistent respiratory abnormality, abnormal average respiratory level, possible apnea, and possible risk of respiratory failure, and the changing trend needs to be continuously monitored.

[0226] like If it increases significantly, it means that the respiratory rhythm is unstable, the breathing pattern is disordered, there may be difficulty breathing, and there may be intermittent apnea, and the fluctuations need to be observed.

[0227] like A sudden increase indicates a sudden change in breathing pattern and breathing rhythm, which may lead to acute breathing difficulties and airway obstruction, requiring immediate attention.

[0228] If the SVM classifier outputs > 0, it is confirmed as a real anomaly, an alarm is triggered immediately, the time of the anomaly is recorded, and the anomaly sample is added to the training set , update the model parameters after 24 hours, generate an abnormal event report, notify medical staff to pay attention, and activate the emergency response mechanism.

[0229] If multiple consecutive time windows are judged as abnormal, it means that there is persistent respiratory dysfunction and there may be serious respiratory diseases, which require emergency intervention from medical personnel. The alarm level will be raised and respiratory support may be required. A detailed medical examination is recommended, an emergency treatment plan needs to be developed, and transfer to the intensive care unit should be considered.

[0230] In summary, the present invention improves the anti-interference ability of the system through multi-source information fusion, adopts multi-scale feature extraction to enhance the expressiveness of features, improves the adaptability of feature fusion through dynamic weight allocation based on attention mechanism, and reduces the false alarm rate through dual discrimination mechanism. As a whole, a reliable, stable and self-learning respiratory monitoring system is formed, which provides effective technical support for the early warning and intelligent diagnosis of children's respiratory diseases. Overall, the present invention constructs a respiratory monitoring system with high reliability, strong adaptability and intelligent learning ability, which provides comprehensive technical support for the early identification, real-time warning and intelligent diagnosis of children's respiratory diseases, and has important clinical application value for improving the diagnosis and treatment level of children's respiratory diseases.

[0231] This embodiment further provides a pediatric respiratory abnormality detection system based on physiological signals, including:

[0232] The acquisition module is used to collect multimodal physiological signals, perform preliminary filtering through a second-order Butterworth bandpass filter, eliminate noise using wavelet decomposition and soft threshold function, remove baseline drift in combination with morphological opening and closing operations, and output a preprocessed signal group;

[0233] The processing module is used to segment the preprocessed signal group using a sliding window, calculate the time domain features, obtain the frequency domain features through wavelet packet transform, calculate the nonlinear features using phase space reconstruction, and organize the features into a three-layer pyramid structure. The attention mechanism is used to achieve feature weighted fusion, and the DTW algorithm is combined to analyze the time series similarity, and output the fused feature sequence;

[0234] The early warning module is used to input the fused feature sequence into the random forest classifier for preliminary anomaly detection, classify based on the age grouping threshold, extract the context feature vector of the window before and after the anomaly point, use the SVM classifier for secondary discrimination, and continuously optimize the model parameters through the gradient descent method to achieve accurate identification of respiratory abnormalities and graded early warning.

[0235] This embodiment also provides a computer device, which is suitable for the case of a pediatric respiratory abnormality detection method based on physiological signals, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the pediatric respiratory abnormality detection method based on physiological signals as proposed in the above embodiment.

[0236] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0237] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the method for detecting abnormal breathing in children based on physiological signals as proposed in the above embodiment is implemented.

[0238] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be found in the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0239] Example 2: This example provides a method for detecting abnormal breathing in children based on physiological signals. In order to verify the beneficial effects of the present invention, a scientific demonstration is carried out through simulation experiments.

[0240] 100 cases of children's respiratory monitoring data (50 cases of normal breathing and 50 cases of abnormal breathing) were selected and analyzed using the traditional single signal method, the basic feature fusion method and the method of the present invention. The experiment was carried out under the same hardware environment, using 5-fold cross validation, and the performance indicators included accuracy, sensitivity, specificity, false alarm rate and calculation time. The specific experimental results are shown in Table 1:

[0241] Table 1 Comparative test results

[0242] Evaluation indicators Traditional single signal approach Basic feature fusion method Method of the present invention Accuracy (%) 82.5 88.7 95.8 Sensitivity(%) 80.3 86.5 94.2 Specificity(%) 84.7 89.2 96.5 False alarm rate (%) 15.3 10.8 3.5 Computation time (ms) 85 156 198

[0243] By comparing the experimental results, it can be seen that the method of the present invention is superior to the traditional methods in various key performance indicators, among which the accuracy rate reaches 95.8%, which is 13.3 percentage points higher than the traditional single signal method and 7.1 percentage points higher than the basic feature fusion method; in terms of sensitivity and specificity, the method of the present invention reaches 94.2% and 96.5% respectively, which are significantly higher than the other two methods; especially in terms of false alarm rate control, the method of the present invention reduces it to 3.5%, which is a significant improvement compared with 15.3% of the traditional single signal method and 10.8% of the basic feature fusion method.

[0244] Although the calculation time of the method of the present invention is relatively long, considering the performance improvement it brings, this calculation overhead is completely acceptable, and it can still meet the needs of real-time monitoring in practical applications. These data fully confirm the significant advantages of the present invention in improving monitoring accuracy and reliability, indicating that it has good application prospects in the field of children's respiratory monitoring.

[0245] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for detecting abnormal breathing in children based on physiological signals, characterized in that: include: Collect multimodal physiological signals, perform preliminary filtering through a second-order Butterworth bandpass filter, use wavelet decomposition and soft threshold function to eliminate noise, combine morphological opening and closing operations to remove baseline drift, and output a preprocessed signal group; The preprocessed signal group is segmented by a sliding window, the time domain features are calculated, the frequency domain features are obtained by wavelet packet transform, the nonlinear features are calculated by phase space reconstruction, and the features are organized into a three-layer pyramid structure. The attention mechanism is used to realize feature weighted fusion, the DTW algorithm is combined to analyze the time series similarity, and the fused feature sequence is output; The fused feature sequence is input into a random forest classifier for preliminary anomaly detection, classification is performed based on age grouping thresholds, context feature vectors of windows before and after the anomaly point are extracted, a SVM classifier is used for secondary discrimination, and model parameters are continuously optimized through a gradient descent method to achieve accurate identification of respiratory abnormalities and graded warnings; The multimodal physiological signals include respiratory motion signals, electrocardiogram signals, blood oxygen saturation signals and body motion signals; Inputting the collected initial physiological signal group into a bandpass filter group to filter out high-frequency noise to obtain a first filtered signal; The first filtered signal is subjected to wavelet decomposition, and the transformation formula is expressed as: ； in, Indicates signal In scale and translation The wavelet transform coefficients under is the original input time domain signal, is the wavelet basis function, represents the energy normalization factor, Represents the scaling and translation term of the wavelet basis function; Discretizing the continuous wavelet transform formula, extracting high-frequency coefficients and low-frequency coefficients of the first filtered signal, and obtaining a wavelet coefficient matrix; Based on the wavelet coefficient matrix, the high frequency coefficients are shrunk by a soft threshold function, and the soft threshold function is expressed as: ； ； in, represents the high frequency coefficient, is the threshold parameter, express The symbol function of is the noise standard deviation, Indicates the length of the signal, is the output value of the soft threshold function, which represents the wavelet coefficient after threshold processing; The noise reduction coefficient matrix is ​​reconstructed by wavelet, and the reconstruction formula is expressed as: ； in, Scaling function; is the wavelet function; is the maximum number of decomposition levels, is the reconstructed time domain signal, is the low-frequency approximation coefficient, is the detail coefficient, is the translation parameter; A second filtered signal is obtained by performing wavelet reconstruction on the noise reduction coefficient matrix; The second filtered signal is processed by using morphological opening and closing operations to output a preprocessed signal group. The opening operation is expressed as: ； The closing operation is expressed as: ； in, represents the respiratory signal after wavelet denoising, Represents a structural element, represents the erosion operation, represents the dilation operation; Based on the preprocessed signal group, the sliding window method is used to segment, and the segmented preprocessed signal is expressed as: ； ； in, Indicates signal fragments, Indicates the total length of the signal, represents the window length, represents the sliding step length, represents the time variable within the window, represents the total number of fragments; Identify feature points through the peak-valley detection algorithm and extract the peak point set of each signal segment , valley point set And zero-crossing set , expressed as: ； ； ； in, represents the set of all peak points, Indicates The time position of the peak point, Indicates signal exist The first-order derivative at a point, Indicates signal exist The second derivative at the point, represents all points that meet the conditions; represents the set of all valley points, Indicates The time position of the valley point, Indicates signal exist The first-order derivative at a point, Indicates signal exist The second derivative at the point; represents the set of all zero-crossing points, Indicates The time position of the zero crossing point, Indicates signal exist The value at the point, Indicates signal exist The first derivative at a point; According to the peak point set and valley point set Calculate the time domain feature vector of the respiratory signal , time domain feature vector Including respiratory rate , breathing depth Inspiratory time and exhalation time ; The breathing rate It is expressed as: ； ； in, Respiratory rate, Represents the time interval set between adjacent peak points, Indicates The time position of the peak point, Indicates The time position of the peak point, Indicates the peak point number, from 1 to ; Indicates the total number of peak points; The breathing depth It is expressed as: ； in, Indicates The amplitude of the peak point, Indicates The amplitude of the valley point, Indicates the total number of peak or valley points; The inspiratory time It is expressed as: ； ； The exhalation time It is expressed as: ； ； in, represents the average inspiratory time, represents the average exhalation time, represents the time interval from the peak point to the valley point, i.e. the inspiratory phase, represents the time interval from the valley point to the next peak point, i.e., the exhalation phase. Indicates The time position of the valley point, Indicates The time position of the peak point, Indicates The time position of the peak point; By breathing rate Reflects the speed of breathing rhythm and judges breathing abnormalities; Reflects the ventilation volume of a single breath and evaluates respiratory efficiency. When abnormal values ​​appear, it indicates respiratory dysfunction; Inspiratory time and exhalation time Reflect the time structure of breathing and assess the coordination of breathing; Wavelet packet transform is used to perform multi-scale decomposition of signal segments to obtain energy distribution characteristics of different frequency bands, including frequency band energy ratio Sum-band power spectrum entropy , the decomposition of the wavelet packet transform at each layer produces approximate coefficients and detail coefficients, which are expressed as: ； ； in, Indicates Tier The approximation coefficient of the nodes, represents the decomposition level, indicating the depth of the wavelet packet tree; Indicates the node number, indicating the The first nodes, Represents the discrete point number of the time series, represents the index of the filter coefficient, represents the low-pass filter coefficient, represents the coefficient of the parent node in the previous layer, Indicates Tier The detail coefficient of each node, represents the high-pass filter coefficient; The frequency band energy ratio It is expressed as: ； ； in, Indicates Tier The energy value of the frequency band, Indicates the decomposition level, represents the frequency band index, is the square of the wavelet packet approximation coefficient, representing the instantaneous energy; Indicates The relative energy ratio of the frequency bands, represents the maximum number of decomposition levels, represents the band count index, Indicates the total number of frequency bands in the last layer; The frequency band power spectrum entropy It is expressed as: ； in, Indicates The relative energy ratio of the frequency bands, Indicates the maximum number of decomposition levels; The signal fragments are mapped into high-dimensional phase space using phase space reconstruction technology to calculate the nonlinear eigenvector , the nonlinear eigenvector Includes the maximum Lyapunov exponent , correlation dimension And the sample entropy ; Using the time delay method to construct dimensional phase space, expressed as: ； ； ； in, represents the reconstructed phase space vector, represents the original time series, Indicates time delay, represents the embedding dimension, represents the time series index, represents the autocorrelation function, Indicates the delay value, represents the sequence length, represents the mean of the series, represents the condition that minimizes the autocorrelation function value; The maximum Lyapunov exponent , expressed as: ； ； ； in, represents the maximum Lyapunov exponent, represents the total evolution time, represents the number of tracks, represents the distance after the phase space orbital evolution, represents the initial distance of the phase space orbit, represents the reference orbit point, represents the nearest neighbor point, represents the evolution time step; The correlation dimension It is expressed as: ； in, represents the correlation dimension, represents the distance threshold, represents the association integral; Furthermore, the correlation score It is expressed as: ； in, represents the number of phase space points, express Step function, represents the distance between two points in phase space; The sample entropy It is expressed as: ； in, represents the sample entropy, Indicates the pattern length, represents the similarity tolerance, The dimension is The number of matches when The dimension is The number of matches when Based on the time domain feature vector , frequency domain feature vector and the nonlinear eigenvector , construct a three-layer feature pyramid, expressed as: ； ； ； in, is the bottom layer, containing time domain features; Represents the total number of time domain features; is the middle layer, containing frequency domain features; Represents the total number of frequency domain features; is the top layer, containing nonlinear features; Represents the total number of nonlinear features; For each layer of features , through three learnable weight matrices , , Perform linear transformation to obtain the query matrix , key matrix Sum Matrix , map the original features into a space suitable for calculating attention weights, and then calculate and The similarity is divided by the scaling factor to avoid the gradient vanishing problem, and The function is normalized and finally compared with the value matrix Multiply to get the attention-weighted feature representation; ； ； in, is the final fused feature vector, is the feature level index, Indicates Layer feature vector, Indicates The weight coefficient of the layer feature, Indicates Importance scores of layer features; The DTW algorithm is used to analyze the timing relationship between multiple signal segments and calculate the cumulative distance matrix between two respiratory signals, which is expressed as: ； in, Indicates starting point To point The minimum cumulative distance, Indicates that two sequences are at position and The distance metric at Indicates the cumulative distance in the upward direction, Indicates the cumulative distance to the left. represents the cumulative distance in the diagonal direction, Select the minimum value in the three directions; Convert the distance to a similarity measure, expressed as: ； in, For the and The similarity of the signal segments, is the scale parameter; Combine the features with the similarity information to obtain a feature sequence containing temporal relationships , expressed as: ； in, represents the final feature sequence, Indicates The feature vector of the signal segment, Indicates The similarity vector between a signal segment and other segments, , Indicates the total number of signal fragments.

2. The method for detecting abnormal breathing in children based on physiological signals according to claim 1, characterized in that: The feature sequence Input a pre-trained random forest classifier, which contains Decision trees are used to classify feature sequences, and the abnormal probability value is calculated according to the voting results of each decision tree, which is expressed as: ； in, is the number of decision trees, For the The classification results of the decision tree are is the abnormal probability value; According to the abnormal probability value and age group threshold table Perform preliminary screening, where the age group threshold is calculated as follows: ； in, represents the age group index; For the The mean of normal breathing indexes in the group, For the Standard deviation of normal respiratory index in the group; According to the abnormal probability value and age group threshold table Perform preliminary screening, when the abnormal probability value When the value exceeds the normal range of the corresponding age group, the abnormal mark sequence at the corresponding time Mark as 1, otherwise mark as 0, and get the filtered abnormal mark sequence ; After screening, abnormal marker sequences For each abnormal point in , extract the feature sequence of the 10 time windows before and after it, and calculate the context feature vector, which is expressed as: ； in, Indicates The characteristic sequence of 10 time windows before and after the moment The mean of Represents the feature sequence within the same time range The standard deviation of represents the average rate of change of the feature sequence, represents the context feature vector; According to the context feature vector Train the support vector machine classifier, use the radial basis kernel function for quadratic discrimination, and add abnormal samples to the training set In the training set, the training set is used every 24 hours. Update the classifier parameters, expressed as: ； in, is the learning rate, is the cross entropy loss function, represents the current model parameters, are the updated model parameters; By calculating the abnormal probability at each time point , generating anomaly marker sequences , and then obtain the context feature vector , the final discrimination result is obtained through SVM classifier .

3. The method for detecting abnormal breathing in children based on physiological signals according to claim 2, characterized in that: In the process of identification, if > and = 1, it means that the respiratory index at that moment is initially abnormal, the respiratory rate exceeds the normal range of the age group, the breathing depth is abnormally shallow or too deep, and the breathing ratio may be unbalanced, and a second judgment confirmation is required; If the above conditions are met and the context feature vector show If it is beyond the normal range, it means that there is persistent respiratory abnormality, abnormal average respiratory level, possible apnea, and risk of respiratory failure, and the trend of change needs to be continuously monitored; like If it increases, it means that the breathing rhythm is unstable, the breathing pattern is disordered, there may be difficulty breathing, intermittent apnea, and the fluctuation needs to be observed; like If it is elevated, it means that the breathing pattern has changed suddenly, the breathing rhythm has changed suddenly, there may be acute breathing difficulties, and airway obstruction may occur, which requires immediate attention; If the SVM classifier outputs > 0, it is confirmed as a real anomaly, an alarm is triggered immediately, the time of the anomaly is recorded, and the anomaly sample is added to the training set , after 24 hours, the model parameters are updated, an abnormal event report is generated, medical staff are notified to pay attention, and the emergency response mechanism is activated; If multiple consecutive time windows are judged as abnormal, it means that there is persistent respiratory dysfunction and possible serious respiratory disease, which requires emergency intervention from medical staff. The alarm level will be raised and respiratory support may be required. An emergency treatment plan needs to be developed and transfer to the intensive care unit should be considered.

4. A pediatric breathing abnormality detection system based on physiological signals, based on the pediatric breathing abnormality detection method based on physiological signals according to any one of claims 1 to 3, characterized in that: include: The acquisition module is used to collect multimodal physiological signals, perform preliminary filtering through a second-order Butterworth bandpass filter, eliminate noise using wavelet decomposition and soft threshold function, remove baseline drift in combination with morphological opening and closing operations, and output a preprocessed signal group; The processing module is used to segment the preprocessed signal group using a sliding window, calculate the time domain features, obtain the frequency domain features through wavelet packet transform, calculate the nonlinear features using phase space reconstruction, and organize the features into a three-layer pyramid structure. The attention mechanism is used to achieve feature weighted fusion, and the DTW algorithm is combined to analyze the time series similarity, and output the fused feature sequence; The early warning module is used to input the fused feature sequence into the random forest classifier for preliminary anomaly detection, classify based on the age grouping threshold, extract the context feature vector of the window before and after the anomaly point, use the SVM classifier for secondary discrimination, and continuously optimize the model parameters through the gradient descent method to achieve accurate identification of respiratory abnormalities and graded early warning.

Images