A sleep disorder auxiliary diagnosis method based on physiological signals and a storage medium

Abstract

The application discloses a sleep disorder auxiliary diagnosis method based on physiological signals and a storage medium. Multi-channel physiological signal data collected in advance is received and preprocessed, and a multi-channel physiological signal data set is constructed. A signal feature vector is extracted, signal structure retention, rhythm continuity and disturbance density are calculated, and a channel quality factor is constructed. A cross-channel coupling proportional transmission structure is established, and a reconstructed signal feature vector is obtained. An abnormal response feature is constructed, a signal response time delay correction amount is calculated, and consistency calibration processing is performed. A local abnormality driving structure is constructed, a short-time state recursive mechanism and a long-time structure evolution mechanism are combined, a double-level time sequence structure feature is established, a sleep state auxiliary analysis model is constructed, and a sleep state auxiliary analysis result is obtained. The storage medium stores a computer program, and the computer program is executed by a processor to realize the steps of the foregoing method.

Description

Technical Field

[0001] This invention relates to the field of physiological signal processing technology, and in particular to an auxiliary diagnostic method and storage medium for sleep disorders based on physiological signals. Background Technology

[0002] Currently, with the development of sleep monitoring technology, multi-channel physiological signals are widely used in sleep stage analysis and sleep disorder diagnosis. Polysomnography (PSG) technology simultaneously collects EEG, ECG, respiration, blood oxygenation, and electromyography signals to continuously record the neural activity state, cardiovascular changes, and respiratory regulation process during human sleep, thus providing objective evidence for sleep quality assessment and sleep disorder identification. However, due to the characteristics of physiological signals such as noise interference, significant individual differences, and complex time series changes, and the differences in expressive power and change patterns among different signal channels, traditional analysis methods still have certain limitations in signal feature extraction, stage determination, and multi-channel information synergistic utilization, making it difficult to maintain stability and diagnostic accuracy in complex sleep monitoring scenarios.

[0003] Publication No. CN113397560A discloses a training method for a sleep stage recognition model, which achieves sleep stage recognition by performing spectrum conversion on EEG signals and extracting a set of brain wave feature values ​​to train a neural network model. Publication No. CN106709469A discloses an automatic sleep stage classification method based on multiple features of EEG and EMG, which achieves sleep stage classification by extracting multi-band features of EEG and EMG signals and inputting them into a support vector machine model.

[0004] However, existing technologies do not fully consider the dynamic changes of different signal channels within a continuous time window and their correlation and evolution characteristics across time windows during multi-channel physiological signal processing. They neglect the state transfer characteristics of signals in time series and the structural connections in the process of coordinated changes between channels. At the same time, they lack effective modeling methods for the stable expression of signals in complex noise environments. As a result, existing methods have problems such as insufficient feature expression, incomplete information utilization, insufficient time series continuity, and limited stability and accuracy of diagnostic results in multi-channel complex sleep monitoring scenarios. Summary of the Invention

[0005] This invention proposes an auxiliary diagnostic method and storage medium for sleep disorders based on physiological signals. Addressing the problems of significant signal fluctuations, insufficient inter-channel correlation, inconsistent abnormal response characteristics, and inadequate continuous time-series analysis capabilities in sleep monitoring of multi-channel physiological signals, this invention proposes a multi-stage processing mechanism integrating channel quality factor calculation, cross-channel coupling ratio transfer, abnormal response consistency calibration, and two-level time-series feature modeling. This mechanism includes receiving pre-collected multi-channel physiological signal data and constructing a multi-channel physiological signal dataset; extracting signal feature vectors from each signal channel based on the dataset and calculating signal structure preservation, rhythm continuity, and perturbation density to construct channel quality factors; establishing a cross-channel coupling ratio transfer structure based on the correlation between channel quality factors and signal feature vectors and generating reconstructed signal feature vectors; constructing abnormal response features based on the reconstructed signal feature vectors and calculating signal response delay correction for consistency calibration; and constructing a local abnormality driving structure based on the calibrated abnormal response features and establishing two-level time-series structural features by combining short-term state recursion and long-term structural evolution mechanisms. This achieves unified modeling of multi-channel physiological signal characteristics and their temporal evolution patterns.

[0006] A method for assisting in the diagnosis of sleep disorders based on physiological signals, the specific method is as follows: S1. Receive pre-collected multi-channel physiological signal data, preprocess the multi-channel physiological signal data, and construct a multi-channel physiological signal dataset; S2. Based on the multi-channel physiological signal dataset, extract the signal feature vectors of each signal channel, calculate the signal structure preservation, rhythm continuity, and perturbation density, and construct the channel quality factor. S3. Based on the correlation between the channel quality factor and the signal feature vector, establish a cross-channel coupling ratio transfer structure, and perform weighted fusion of the signal features of adjacent time windows and the current time window to obtain the reconstructed signal feature vector. S4. Based on the reconstructed signal feature vector, construct the abnormal response features, calculate the signal response delay correction amount based on the correlation between the abnormal response features, and perform consistency calibration on the abnormal response features. S5. Based on the calibrated abnormal response characteristics, construct a local abnormality driving structure, and combine the short-time state recursion mechanism and the long-time structure evolution mechanism to establish a two-level temporal structure feature. S6. Construct a sleep state auxiliary analysis model, input a multi-channel physiological signal dataset, and proceed through steps S2 to S5 in sequence to obtain the sleep state auxiliary analysis results.

[0007] Furthermore, to achieve the above objectives, the present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the aforementioned sleep disorder auxiliary diagnosis method based on physiological signals.

[0008] Preferably, for the construction of a multi-channel physiological signal dataset, firstly, pre-collected multi-channel physiological signal data is received, including electroencephalogram (EEG), electrocardiogram (ECG), respiratory signal, blood oxygenation signal, and electromyogram (EMG). Simultaneously, the time index information, channel number, channel distribution information of the sleep monitoring device, sleep stage identifier, and monitoring status information corresponding to each signal channel are recorded. Then, the multi-channel physiological signal data undergoes format standardization processing, and filtering is used to remove baseline drift components and high-frequency interference components. Next, amplitude normalization processing is performed on the physiological signal data of each signal channel, and abrupt fluctuation segments are smoothed to improve the continuity and stability of the multi-channel physiological signal data. Finally, the pre-processed multi-channel physiological signal data is bound and organized with the time index information, channel number, channel distribution information, sleep stage identifier, and physiological signal sequence to construct a multi-channel physiological signal dataset containing the channel number, time index information, sleep stage identifier, and physiological signal sequence.

[0009] Preferably, during sleep monitoring, multi-channel physiological signals exhibit significant continuous changes in time series, substantial differences in response between different signal channels, and complex local fluctuation characteristics. The original physiological signals at a single time point are insufficient to fully characterize the changes within the current time window and their continuous relationship with adjacent time periods. This invention divides the physiological signal sequences of each signal channel into time windows and extracts the corresponding signal feature vectors for each signal channel. It further calculates the signal structure preservation, rhythm continuity, and perturbation density, and constructs channel quality factors. This provides a unified data foundation for the subsequent establishment of cross-channel coupling ratio transfer structures, consistency calibration of abnormal response characteristics, and construction of two-level temporal structure features.

[0010] Furthermore, in step S2, based on the multi-channel physiological signal dataset, the physiological signal sequences of each signal channel are divided into time windows, and the signal feature vectors corresponding to each signal channel are extracted based on the corresponding physiological signal sequences, linear matrices, and bias vectors; based on the signal feature vectors, amplitude change adjustment coefficients, and stability terms, the signal structure preservation of each signal channel within the corresponding time window is calculated. Based on the signal feature vector, reference feature vector, and stability term, calculate the rhythmic continuity of each signal channel within the corresponding time window; Based on the signal feature vector and stability term, calculate the perturbation density of each signal channel within the corresponding time window; based on the signal structure preservation quantity, rhythm continuity quantity, and perturbation density quantity, and combined with the weighting coefficients corresponding to the rhythm continuity quantity, the weighting coefficients corresponding to the signal structure preservation quantity, the weighting coefficients corresponding to the perturbation density quantity, and the stability term, construct the channel quality factor.

[0011] Furthermore, addressing the problem that the single time window features of multi-channel physiological signals during sleep monitoring cannot fully represent the state of each signal channel and the differences between different channels, this invention proposes a cross-channel coupling modeling mechanism based on channel quality factors and signal feature vectors. First, within each time window, the signal feature vectors corresponding to each signal channel are extracted, and the signal structure preservation, rhythm continuity, and perturbation density are calculated to construct the channel quality factor. Then, based on the correlation between signal channels, the channel coupling ratio is calculated, and a cross-channel coupling ratio transfer structure is constructed. Under this structure, the signal feature vectors of adjacent time windows and the current time window are weighted and fused to calculate the aggregated feature representation of each signal channel, and finally, the reconstructed signal feature vector is generated. This reconstructed feature can uniformly represent the change state of multi-channel signals within continuous time windows and the correlation features between channels.

[0012] Preferably, in the process of multi-channel physiological signal processing, the signal features within a single time window and the aggregated features between channels are difficult to fully reflect the time delay relationship and dynamic changes of abnormal response features. This invention calculates abnormal response features based on the reconstructed signal feature vector and calculates the signal response time delay correction based on the correlation between abnormal response features. Then, it performs consistency calibration on the abnormal response features, thereby realizing unified modeling of multi-channel abnormal states and continuous correction across time windows, providing a data foundation for subsequent local abnormality-driven structure construction, short-term state recursion mechanism and long-term structure evolution mechanism.

[0013] Furthermore, in step S3, based on the correlation between the channel quality factor and the signal feature vector, and combining the channel quality factor, the components of each dimension of the signal feature vector, the correlation base value, the stability term, the total number of signal channels, and the dimension of the signal feature vector, the coupling ratio between signal channels is calculated, and a cross-channel coupling ratio transfer structure is constructed. Under the cross-channel coupling ratio transfer structure, the weighted coefficients corresponding to the previous time window, the current time window, and the next time window, as well as the coupling ratios and signal feature vectors in the previous, current, and next time windows, are combined to perform weighted fusion of the signal features of adjacent time windows and the current time window, and calculate the aggregated feature representation of each signal channel. Based on the aggregated feature representation, the reconstructed signal feature vector is generated by combining the reconstruction mapping matrix, the reconstruction bias vector, and the signal feature vector dimension.

[0014] Furthermore, in multi-channel physiological signal processing, the reconstructed signal feature vectors not only exhibit differences in the degree of abnormal changes within a continuous time window, but also show inconsistent response intensities and asynchronous response times when abnormal states occur in different signal channels. It is difficult to stably characterize the true evolution of multi-channel abnormal states in the time series based solely on local features within the current time window. Therefore, this invention constructs abnormal response features based on the reconstructed signal feature vectors and calculates the signal response delay correction based on the correlation between abnormal response features. Subsequently, time alignment and consistency calibration are performed on the abnormal response features to improve the consistency of expression of abnormal response features across time windows and signal channels, enhance the stability and distinguishability of abnormal state representation, and provide an input basis for subsequent local abnormality-driven structure construction, short-term state recursion mechanism, and long-term structure evolution mechanism.

[0015] Preferably, in multi-channel physiological signal processing, the calibrated abnormal response features still exhibit local variation in intensity within a continuous time window. The abnormal states within different time windows are non-uniform in terms of variation amplitude and location. Therefore, this invention constructs a local abnormality driving structure based on the calibrated abnormal response features and combines the local driving weights corresponding to the previous time window and the next time window to jointly characterize the abnormal changes within the current time window. This enhances the model's ability to perceive local abnormality trends, ensuring that the relationship between abnormal states in adjacent time windows is stably preserved, and providing a structural basis for the execution of subsequent short-term state recursion mechanisms.

[0016] Furthermore, in step S4, based on the reconstructed signal feature vector and the component relationships of the reconstructed signal feature vectors corresponding to the previous time window, the abnormal response characteristics of each signal channel within the current time window are calculated. Based on the correlation between abnormal response features, combined with the preset time delay search range, time window offset, abnormal response feature dimension and stability term, the signal response time delay correction amount between different signal channels is calculated. Based on the signal response delay correction amount, the abnormal response features are time-aligned to obtain the time-aligned abnormal response features. Based on the original abnormal response characteristics, the time-aligned abnormal response characteristics, the signal response delay correction, the total number of signal channels, the dimensions of the abnormal response characteristics, and the stability terms, the abnormal response characteristics are subjected to consistency calibration.

[0017] Furthermore, in multi-channel physiological signal processing, local abnormal driving structures exhibit short-term propagation differences within continuous time windows. The transmission speed and intensity of abnormal changes in different signal channels are not consistent between adjacent time windows. Therefore, this invention implements a short-term state recursion mechanism based on the local abnormal driving structure and the corresponding short-term state recursion result within the previous time window, combined with a short-term state recursion adjustment coefficient, to characterize the transmission relationship of abnormal states between adjacent time windows and to recursively model the continuous evolution process of local abnormal changes within a short time range, thereby improving the model's ability to characterize the trend of short-term abnormal state changes.

[0018] Preferably, in multi-channel physiological signal processing, the short-term state recursion results exhibit cumulative evolutionary differences within a continuous time window. Although abnormal state changes within a short time range can reflect local evolutionary trends, they are still difficult to fully describe the continuous change patterns over a longer time range. Therefore, this invention, based on the results of the short-term state recursion mechanism, combines the long-term cumulative structural information up to the current time window, the structural evolution intensity of the short-term state recursion results on the time axis, and the long-term structural evolution adjustment coefficient to execute a long-term structural evolution mechanism, thereby establishing a two-level temporal structure feature, enabling the joint representation of short-term abnormal change information and long-term evolutionary information under a unified structure.

[0019] Furthermore, in step S5, a local anomaly driving structure is constructed based on the calibrated anomaly response characteristics, combined with the local driving weights corresponding to the previous time window and the local driving weights corresponding to the next time window. Based on the local anomaly driving structure, combined with the corresponding short-time state recursion result and short-time state recursion adjustment coefficient in the previous time window, the short-time state recursion mechanism is executed. Based on the results of the short-term state recursion mechanism, combined with the long-term cumulative structural information up to the current time window, the structural evolution intensity of the short-term state recursion results on the time axis, and the long-term structural evolution adjustment coefficient, the long-term structural evolution mechanism is executed to establish a two-level temporal structure feature.

[0020] Furthermore, in multi-channel physiological signal processing, after acquiring the local abnormal driving structure, this invention first establishes the short-term change relationship of the abnormal state within a continuous time window by combining a short-term state recursion mechanism, so that the local abnormal changes of each signal channel between adjacent time windows can form continuous recursion results. Then, based on the short-term state recursion results, a two-level temporal structure feature is established by combining a long-term structure evolution mechanism, thereby uniformly representing the local change law of the abnormal state in a short time range and the cumulative evolution law in a long time range, so that the model can simultaneously take into account the sensitivity of local abnormal changes and the overall temporal stability.

[0021] Preferably, in the process of multi-channel physiological signal analysis, the present invention strengthens the consistent expression of abnormal response features in the time series by jointly modeling the local abnormal driving structure, the short-term state recursion mechanism and the long-term structural evolution mechanism. This enables the local abnormal change information and the long-term evolution information to be synergistically integrated in a unified two-level temporal structure feature, thereby improving the model's ability to characterize the abnormal state change patterns in multi-channel physiological signals and further improving the stability and accuracy of sleep disorder auxiliary diagnosis results.

[0022] Preferably, in step S6, addressing the issues of significant signal quality fluctuations, complex abnormal response characteristics, and difficulty in stably representing the state evolution patterns within continuous time windows in multi-channel physiological signals during sleep monitoring, this invention constructs a sleep disorder auxiliary diagnostic model based on physiological signals. This model sequentially executes the following stages: signal feature vector extraction and channel quality factor construction; cross-channel coupling ratio transfer and reconstructed signal feature vector generation; abnormal response feature construction and consistency calibration; and establishment of local abnormal driving structures and two-level temporal structure features. First, a multi-channel physiological signal dataset is input. Signal feature vectors are extracted based on the physiological signal sequences of each signal channel, and signal structure is calculated. The channel quality factor is constructed by maintaining the quantity, rhythm continuity quantity, and perturbation density quantity. Then, a cross-channel coupling ratio transfer structure is established based on the correlation between the channel quality factor and the signal feature vector. The signal features of adjacent time windows and the current time window are weighted and fused to generate the reconstructed signal feature vector. Next, the abnormal response features are calculated based on the reconstructed signal feature vector, and the signal response delay correction quantity is calculated based on the correlation between the abnormal response features to complete the consistency calibration process. Then, a local abnormality driving structure is constructed based on the calibrated abnormal response features, and a two-level temporal structure feature is established by combining the short-time state recursion mechanism and the long-time structure evolution mechanism. Finally, the auxiliary diagnostic results of sleep disorders are obtained.

[0023] Compared with existing technologies, this invention focuses on the time-series evolution characteristics of multi-channel physiological signals during sleep monitoring. It constructs a modeling mechanism oriented towards signal channel correlation and continuous time window features, proposes signal feature extraction and channel quality assessment strategies, and calculates channel quality factors by combining signal structure preservation, rhythm continuity, and perturbation density. This achieves a complete characterization of the feature expression of multi-channel physiological signals within each time window. A cross-channel coupling ratio transfer structure is established across time windows, and the signal feature vectors of adjacent time windows and the current time window are weighted and fused to generate reconstructed signal feature vectors to capture the dynamic correlation between channels. Subsequently, based on the reconstructed... The abnormal response features of each signal channel are calculated from the signal feature vectors, and the signal response delay correction is calculated based on the correlation between the abnormal response features to complete the consistency calibration process, thereby enhancing the distinguishability and stability of abnormal patterns. Next, a local abnormality driving structure is constructed based on the calibrated abnormal response features, and instantaneous state change information is extracted by combining a short-time state recursion mechanism. Finally, a long-time structure evolution mechanism is executed based on the short-time state recursion results to generate a two-level temporal structure feature. By integrating the above stages end-to-end, accurate modeling of the state change law of multi-channel physiological signals within a continuous time window and stable output of sleep state auxiliary analysis results are achieved. Attached Figure Description

[0024] Figure 1 This is a flowchart of an auxiliary diagnostic method for sleep disorders based on physiological signals provided by the present invention.

[0025] Figure 2 This is a structural diagram of the channel quality factor provided by the present invention.

[0026] Figure 3 This is a structural diagram of the cross-channel coupling ratio transfer structure provided by the present invention.

[0027] Figure 4 This is a structural diagram of the construction of abnormal response features and the consistency calibration process provided by the present invention.

[0028] Figure 5 This is a structural diagram of the construction of local anomaly-driven structure and two-level temporal structure features provided by the present invention.

[0029] Figure 6 This is a comparison diagram of the original physiological signal and the processed physiological signal provided by the present invention.

[0030] Figure 7 This is a comparison chart of sleep state auxiliary analysis results provided by the present invention. Detailed Implementation

[0031] This invention proposes an auxiliary diagnostic method for sleep disorders based on physiological signals. Addressing the problems of significant signal-noise interference, insufficient utilization of inter-channel correlation, inconsistent abnormal response characteristics, and difficulty in stably representing the state evolution pattern within a continuous time window in multi-channel physiological signals during sleep monitoring, this invention proposes a multi-stage processing mechanism that integrates channel quality factor calculation, cross-channel coupling ratio transfer, abnormal response consistency calibration, and two-level temporal structure feature modeling. This includes preprocessing pre-acquired multi-channel physiological signal data to construct a multi-channel physiological signal dataset; extracting signal feature vectors for each signal channel from the multi-channel physiological signal dataset; and calculating signal structure preservation, rhythm continuity, and perturbation density to construct the channel structure. The system establishes a cross-channel coupling ratio transfer structure based on the correlation between channel quality factors and signal feature vectors. It also performs weighted fusion of signal features from adjacent time windows and the current time window to generate reconstructed signal feature vectors. Based on the reconstructed signal feature vectors, it constructs abnormal response features and calculates signal response delay correction based on the correlation between abnormal response features for consistency calibration. Based on the calibrated abnormal response features, it constructs a local abnormality driving structure and establishes a two-level temporal structure feature by combining short-term state recursion mechanism and long-term structure evolution mechanism. This enables unified modeling of multi-channel physiological signal features and their state evolution laws within continuous time windows, and outputs sleep state auxiliary analysis results.

[0032] Please see Figure 1 As shown in the figure, a sleep disorder auxiliary diagnosis method based on physiological signals is described in the embodiments of this application, and the specific steps are as follows.

[0033] S1. Receive pre-collected multi-channel physiological signal data, preprocess the multi-channel physiological signal data, and construct a multi-channel physiological signal dataset.

[0034] The construction process of the multi-channel physiological signal dataset includes three stages: data reception, signal preprocessing, and data structuring. In the data reception stage, pre-collected multi-channel physiological signal data from normal sleep samples and samples exhibiting sleep disorders are received. This multi-channel physiological signal data includes EEG signals, ECG signals, respiration signals, blood oxygenation signals, and electromyography (EMG) signals. The physiological signal sequence of each signal channel over time is recorded, along with corresponding time index information, signal channel numbers, and channel distribution information of the sleep monitoring device. The sleep stage identifier and monitoring status information under each time index are also simultaneously labeled to ensure the alignment consistency of the multi-channel physiological signals in the time dimension. In the signal preprocessing stage, bandpass filtering is performed on the EEG signals (0.5Hz to 45Hz), the ECG signals (0.5Hz to 40Hz), and the respiration signals (0.1Hz to 10Hz). Bandpass filtering is also performed on the EMG signals. The signals undergo bandpass filtering. Based on this, the physiological signals from each channel are uniformly resampled to ensure consistency across channels. Subsequently, amplitude normalization is performed on each channel signal, mapping the signal amplitude to the [-1, 1] interval. For signal segments with abrupt fluctuations, a sliding window smoothing process is applied, with a sliding window length of 1 s and a step size of 0.25 s, to improve signal continuity and stability. In the data organization stage, the preprocessed multi-channel physiological signals are divided into fixed time windows with a length of 30 s and a step size of 15 s. Within each time window, complete sequence data of each signal channel is retained. This data is then uniformly bound and organized with corresponding time index information, signal channel number, channel distribution information, sleep stage identifiers, and physiological signal sequences to construct a multi-channel physiological signal dataset containing signal channel numbers, time index information, sleep stage identifiers, and physiological signal sequences. This dataset serves as the input basis for subsequent signal feature vector extraction and channel quality factor calculation.

[0035] S2. Based on the multi-channel physiological signal dataset, extract the signal feature vectors of each signal channel, calculate the signal structure retention, rhythm continuity, and perturbation density, and construct the channel quality factor.

[0036] Furthermore, in step S2, a structure diagram of the channel quality factor is constructed, the process of which is as follows: Figure 2 As shown, the specific steps for constructing the structure diagram of the channel quality factor are as follows.

[0037] S21. Based on the multi-channel physiological signal dataset, the physiological signal sequences of each signal channel are divided into time windows, and the signal feature vectors corresponding to each signal channel are extracted within each time window. The mathematical model is as follows: ; in, This represents the physiological signal sequence corresponding to the m-th signal channel within the t-th time window. This represents the linear matrix corresponding to the m-th signal channel. This represents the bias vector corresponding to the m-th signal channel; in this embodiment, the uniform resampling frequency is set to 100Hz, the time window length is set to 30s, and the signal feature vector dimension is set to 64. In this embodiment, The calculation formula is: ; in, For matrix elements, Number of sampling points within the time window; in this embodiment, =3000; In this embodiment, The calculation formula is: ; in, These are vector elements.

[0038] S22. Based on the signal feature vector, calculate the signal structure preservation of each signal channel within the corresponding time window. The mathematical model is as follows: ; in, Represents the signal eigenvector The j-th dimension component, k represents the amplitude change adjustment coefficient, and e represents the stability term. In this embodiment, = 10 -8 k = 1.5.

[0039] S23. Based on the signal feature vector, calculate the rhythmic continuity of each signal channel within the corresponding time window. The mathematical model is as follows: ; in, This represents the reference feature vector corresponding to the m-th signal channel. This represents the j-th dimension component of the reference eigenvector z. Indicates a stable term; in this embodiment, =10 -8 Reference feature vector Take the signal feature vector corresponding to the m-th signal channel within the previous time window, i.e. = Correspondingly = When t=1, let = .

[0040] S24. Based on the signal feature vector, calculate the disturbance density of each signal channel within the corresponding time window. The mathematical model is as follows: ; in, Represents the signal eigenvector The j-th dimension component, Indicates a stable term; in this embodiment, =10 -8 .

[0041] S25. Based on the combination calculation of signal structure preservation, rhythm continuity, and perturbation density, the channel quality factor is constructed, and its mathematical model is as follows: ; in, This represents the weighting coefficient corresponding to the continuous quantity of the rhythm. This represents the weighting coefficients corresponding to the signal structure preservation parameters. This represents the weighting coefficient corresponding to the disturbance density. Indicates a stable term; in this embodiment, =0.40, =0.35, =0.25, = 10 -8 .

[0042] S3. Based on the correlation between the channel quality factor and the signal feature vector, establish a cross-channel coupling ratio transfer structure, and perform weighted fusion of the signal features of adjacent time windows and the current time window to obtain the reconstructed signal feature vector.

[0043] Furthermore, in step S3, a structural diagram of the cross-channel coupling proportional transfer structure is established, and its process is as follows: Figure 3 As shown in the diagram, the specific steps for establishing the cross-channel coupling proportional transfer structure are as follows.

[0044] S31. Based on the correlation between the channel quality factor and the signal eigenvector, a cross-channel coupling proportional transfer structure is established, and its mathematical model is as follows: ; in, This represents the channel quality factor of the u-th signal channel within the t-th time window. This represents the k-th dimension component in the signal feature vector corresponding to the u-th signal channel within the t-th time window. The k-th dimension component of the signal feature vector corresponding to the m-th signal channel within the t-th time window. Indicates the baseline value of the degree of correlation. The term represents stability, M represents the total number of signal channels, and B represents the dimension of the signal feature vector; in this embodiment, =0.10, = 10 -8 .

[0045] S32. Weighted fusion of signal features from adjacent time windows and the current time window is performed to calculate the aggregated feature representation of each signal channel. The mathematical model is as follows: ; in, This represents the weighting coefficient corresponding to the previous time window. This represents the weighting coefficient corresponding to the current time window. This represents the weighting coefficient corresponding to the next time window. , , Let represent the coupling ratio from the u-th signal channel to the m-th signal channel within the previous time window, the current time window, and the next time window, respectively. , , These represent the signal feature vectors corresponding to the u-th signal channel within the previous time window, the current time window, and the next time window, respectively; in this embodiment... =0.25, =0.50, =0.25.

[0046] S33. Based on the weighted fusion result, the reconstructed signal feature vector is obtained, and its mathematical model is: ; in, This represents the reconstruction mapping matrix corresponding to the m-th signal channel. This represents the reconstructed bias vector corresponding to the m-th signal channel. Let B represent the aggregated feature representation of the m-th signal channel within the t-th time window, and let B represent the dimension of the signal feature vector; in this embodiment, when B=64, , .

[0047] S4. Based on the reconstructed signal feature vector, construct abnormal response features, calculate the signal response delay correction amount based on the correlation between abnormal response features, and perform consistency calibration on the abnormal response features.

[0048] Furthermore, in step S4, a structure diagram for constructing abnormal response characteristics and performing consistency calibration is generated, and the process is as follows: Figure 4 As shown, the specific steps for constructing the abnormal response features and performing consistency calibration are as follows.

[0049] S41, Based on the reconstructed signal feature vector, anomaly response features are constructed, and their mathematical model is as follows: ; in, Let k be the k-th dimension component of the reconstructed signal feature vector corresponding to the m-th signal channel within the t-th time window.

[0050] S42. The mathematical model for calculating the signal response delay correction based on the correlation between abnormal response characteristics is as follows: ; in, Indicates the preset time delay search range. Here, represents the time window offset, D represents the dimension of the abnormal response feature, and represents the stability term. In this embodiment... =2, = 10 -8 .

[0051] S43. Based on the signal response delay correction, the abnormal response features corresponding to the u-th signal channel are time-aligned to obtain the time-aligned abnormal response features. The mathematical model for this is: ; in, This indicates that the u-th signal channel is in the... The k-th dimension component of the abnormal response features corresponding to each time window. This indicates that the u-th signal channel is in the... The k-th dimension component of the abnormal response features corresponding to each time window. Indicates the signal response delay correction amount The integer part, Indicates the signal response delay correction amount The decimal part, and satisfying , In this embodiment, when season ,when season .

[0052] S44. Based on the signal response delay correction, a consistency calibration process is performed on the abnormal response characteristics. The mathematical model is as follows: ; in, This represents the correction amount based on the signal response delay from the u-th signal channel to the m-th signal channel. Let k be the k-th dimension component of the original abnormal response feature corresponding to the m-th signal channel within the t-th time window. The k-th dimension component is obtained by time alignment of the abnormal response feature corresponding to the u-th signal channel, where M is the total number of signal channels and D represents the dimension of the abnormal response feature. It is a stable term; in this embodiment, = 10 -8 .

[0053] S5. Based on the calibrated abnormal response characteristics, construct a local abnormality driving structure, and combine the short-time state recursion mechanism and the long-time structure evolution mechanism to establish a two-level temporal structure feature.

[0054] Furthermore, in step S5, a structure diagram of the local anomaly-driven structure and the two-level temporal structure features is constructed, and the process is as follows: Figure 5 As shown, the specific steps for constructing the structure diagram of the local anomaly-driven structure and the two-level temporal structure features are as follows.

[0055] S51. Based on the calibrated abnormal response characteristics, a local abnormality driving structure is constructed, and its mathematical model is as follows: ; in, This represents the k-th dimension component in the calibrated abnormal response feature corresponding to the m-th signal channel within the t-th time window. This represents the local driving weights corresponding to the previous time window. This represents the local driving weight corresponding to the next time window; in this embodiment, = 0.25, = 0.25, when t=1, let = When t=T, = .

[0056] S52. Based on the local anomaly-driven structure, a short-time state recursion mechanism is executed, the mathematical model of which is: ; in, This represents the k-th dimension component in the local anomaly driving structure corresponding to the m-th signal channel within the t-th time window. This represents the k-th dimension component in the short-time state recursion result corresponding to the m-th signal channel within the previous time window. This represents the short-time state recursive adjustment coefficient; in this embodiment, =0.60, when t=1, let = .

[0057] S53. Based on the results of the short-time state recursion mechanism, the long-time structure evolution mechanism is executed to establish a two-level temporal structure feature, the mathematical model of which is: ; in, This represents the long-term cumulative structure information up to the t-th time window. This represents the intensity of structural evolution along the time axis as a result of short-time state recursion. This represents the long-term structural evolution adjustment coefficient; in this embodiment, =0.35, when t=1, let = .

[0058] Furthermore, in step S6, the sleep state auxiliary analysis model of this invention is implemented using the Python programming language and trained and inferred based on the PyTorch deep learning framework. The model input is a joint feature tensor composed of the reconstructed signal feature vector output from step S3, the abnormal response features after consistency calibration output from step S4, and the two-level temporal structure features output from step S5. This input data is organized into a T×M×H form according to continuous time windows, where T... T represents the number of consecutive time windows, M represents the number of physiological signal channels, and H represents the feature dimension corresponding to each signal channel within a single time window. In this embodiment, T is set to 20, M to 4, and H to 32. During training, all samples are divided into training, validation, and test sets in an 8:1:1 ratio, with a batch size of 12 and 180 training epochs. Parameter updates employ the AdamW optimization strategy, with an initial learning rate of 0.0005 and a weight decay coefficient of 0.0001. After linear warm-up in the first 10 training epochs, the learning rate is gradually decayed to 0.00001 using cosine annealing to reduce parameter fluctuations in the early stages of training and improve training performance. Later convergence stability; to suppress the impact of gradient surges during backpropagation on model training, the gradient pruning threshold is set to 5.0; the optimization objective of model training consists of constraints on the reconstructed signal feature vector, alignment constraints between abnormal response features and consistency calibration, continuous constraints on the two-level temporal structure features, and output supervision terms, with weights of 0.20, 0.20, 0.25, and 0.35 respectively; after each training epoch, the joint loss value is calculated using the validation set, and the model parameters corresponding to the minimum joint loss value are used as the final deployment parameters to ensure the stability and generalization ability of the model in sleep-state assisted analysis tasks.

[0059] In an optional embodiment, the present invention also provides a computer-readable storage medium storing a sleep state auxiliary analysis program for executing S1 to S6. The sleep state auxiliary analysis program includes computer-executable instructions. When the computer-executable instructions are executed by a processor, the program receives pre-collected multi-channel physiological signal data and constructs a multi-channel physiological signal dataset. It then sequentially executes signal feature vector extraction, channel quality factor construction, cross-channel coupling ratio transfer, abnormal response consistency calibration, two-level temporal structure feature establishment, and sleep state auxiliary analysis model output processing, thereby realizing the steps in the aforementioned sleep disorder auxiliary diagnosis method based on physiological signals.

[0060] Furthermore, in step S6, the multi-channel physiological signal dataset is input into the constructed sleep state-assisted analysis model for processing, and the experimental results are as follows: Figure 6 and Figure 7 As shown; Figure 6 This is a comparison chart of raw and processed physiological signals, showing the changes in EEG, ECG, and respiratory signals over time. The horizontal axis represents time, and the vertical axis represents normalized amplitude. The raw signals characterize the fluctuation characteristics of each signal channel in its unprocessed state, while the processed signals characterize the signal expression results after preprocessing, channel quality factor construction, cross-channel coupling ratio transfer structure modeling, abnormal response feature consistency calibration, and the establishment of a two-level temporal structure feature. Figure 6 It can be seen that, after processing, the signal maintains the main change trends of each signal channel, while suppressing local mutation interference and significantly enhancing the overall continuity and stability. This indicates that the present invention can effectively preserve key change information in multi-channel physiological signals and improve the recognizability of subsequent analysis inputs. Figure 7 This is a comparison chart of sleep state auxiliary analysis results. The horizontal axis represents the test samples, and the vertical axis represents the state probability. The bar chart represents the output probability corresponding to each test sample. The dots represent the reference category, and the dashed lines represent the discrimination threshold. Figure 7 It can be seen that the output probability of most abnormal state samples is higher than the discrimination threshold, while the output probability of most normal state samples is lower than the discrimination threshold. Only a few boundary samples are close to the discrimination threshold. This indicates that the sleep state auxiliary analysis model constructed in this invention can effectively analyze the sleep state change characteristics based on multi-channel physiological signal processing. It also verifies the synergistic effect and application effect of channel quality factor, reconstructed signal feature vector, abnormal response characteristics after consistency calibration, and two-level temporal structure features in the sleep state auxiliary analysis task.

[0061] The above description is only a preferred embodiment of the present invention. Those skilled in the art can make several substitutions, adjustments and combinations without departing from the concept of the present invention. These modifications and improvements should all be considered within the scope of protection of the present invention.

Claims

1. A method for auxiliary diagnosis of sleep disorders based on physiological signals, characterized in that, Includes the following steps: S1. Receive pre-collected multi-channel physiological signal data, preprocess the multi-channel physiological signal data, and construct a multi-channel physiological signal dataset; S2. Based on the multi-channel physiological signal dataset, extract the signal feature vectors of each signal channel, calculate the signal structure preservation, rhythm continuity, and perturbation density, and construct the channel quality factor. S3. Based on the correlation between the channel quality factor and the signal feature vector, establish a cross-channel coupling ratio transfer structure, and perform weighted fusion of the signal features of adjacent time windows and the current time window to obtain the reconstructed signal feature vector. S4. Based on the reconstructed signal feature vector, construct the abnormal response features, calculate the signal response delay correction amount based on the correlation between the abnormal response features, and perform consistency calibration on the abnormal response features. S5. Based on the calibrated abnormal response characteristics, construct a local abnormality driving structure, and combine the short-time state recursion mechanism and the long-time structure evolution mechanism to establish a two-level temporal structure feature. S6. Construct a sleep state auxiliary analysis model, input a multi-channel physiological signal dataset, and proceed through steps S2 to S5 in sequence to obtain the sleep state auxiliary analysis results.

2. The sleep disorder auxiliary diagnosis method based on physiological signals according to claim 1, characterized in that, The multi-channel physiological signal data includes data items corresponding to electroencephalogram (EEG), electrocardiogram (ECG), respiratory signal, blood oxygenation signal, and electromyogram (EMG), and includes the physiological signal sequence of each signal channel changing over time. The multi-channel physiological signal data also includes time index information, signal channel number, and channel distribution information of the sleep monitoring device corresponding to the physiological signal. The multi-channel physiological signal data also includes sleep stage identifiers and monitoring status information under the corresponding time index. Based on the multi-channel physiological signal data, time index information, signal channel number, channel distribution information, and sleep stage identifier, a multi-channel physiological signal dataset containing signal channel number, time index information, sleep stage identifier, and physiological signal sequence is constructed.

3. The sleep disorder auxiliary diagnosis method based on physiological signals according to claim 2, characterized in that, Based on the multi-channel physiological signal dataset, the physiological signal sequences of each signal channel are divided into time windows, and the signal feature vectors corresponding to each signal channel are extracted based on the corresponding physiological signal sequences, linear matrices and bias vectors. Based on the signal feature vector, amplitude variation adjustment coefficient, and stability term, calculate the signal structure preservation of each signal channel within the corresponding time window; Based on the signal feature vector, reference feature vector, and stability term, calculate the rhythmic continuity of each signal channel within the corresponding time window; Based on the signal feature vector and stability term, calculate the perturbation density of each signal channel within the corresponding time window; based on the signal structure preservation quantity, rhythm continuity quantity, and perturbation density quantity, and combined with the weighting coefficients corresponding to the rhythm continuity quantity, the weighting coefficients corresponding to the signal structure preservation quantity, the weighting coefficients corresponding to the perturbation density quantity, and the stability term, construct the channel quality factor.

4. The sleep disorder auxiliary diagnosis method based on physiological signals according to claim 3, characterized in that, Based on the correlation between the channel quality factor and the signal feature vector, and combining the channel quality factor, each component of the signal feature vector, the correlation base value, the stability term, the total number of signal channels, and the dimension of the signal feature vector, the coupling ratio between signal channels is calculated, and a cross-channel coupling ratio transfer structure is constructed. Under the cross-channel coupling ratio transfer structure, the weighted coefficients corresponding to the previous time window, the current time window, and the next time window, as well as the coupling ratios and signal feature vectors in the previous, current, and next time windows, are combined to perform weighted fusion of the signal features of adjacent time windows and the current time window, and calculate the aggregated feature representation of each signal channel. Based on the aggregated feature representation, the reconstructed signal feature vector is generated by combining the reconstruction mapping matrix, the reconstruction bias vector, and the signal feature vector dimension.

5. The sleep disorder auxiliary diagnosis method based on physiological signals according to claim 4, characterized in that, Based on the reconstructed signal feature vector, and combined with the component relationships of the reconstructed signal feature vectors corresponding to the previous time window, the current time window, and the next time window, the abnormal response characteristics of each signal channel within the current time window are calculated. Based on the correlation between abnormal response features, combined with the preset time delay search range, time window offset, abnormal response feature dimension and stability term, the signal response time delay correction amount between different signal channels is calculated. Based on the signal response delay correction amount, the abnormal response features are time-aligned to obtain the time-aligned abnormal response features. Based on the original abnormal response characteristics, the time-aligned abnormal response characteristics, the signal response delay correction, the total number of signal channels, the dimensions of the abnormal response characteristics, and the stability terms, the abnormal response characteristics are subjected to consistency calibration.

6. The sleep disorder auxiliary diagnosis method based on physiological signals according to claim 5, characterized in that, Based on the calibrated abnormal response characteristics, and combined with the local driving weights corresponding to the previous time window and the next time window, a local abnormal driving structure is constructed. Based on the local anomaly driving structure, combined with the corresponding short-time state recursion result and short-time state recursion adjustment coefficient in the previous time window, the short-time state recursion mechanism is executed. Based on the results of the short-term state recursion mechanism, combined with the long-term cumulative structural information up to the current time window, the structural evolution intensity of the short-term state recursion results on the time axis, and the long-term structural evolution adjustment coefficient, the long-term structural evolution mechanism is executed to establish a two-level temporal structure feature.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of a sleep disorder auxiliary diagnosis method based on physiological signals as described in any one of claims 1 to 6.

Images