Monitoring method based on sleep disorder of patient in neurology department

Abstract

The invention relates to the technical field of medical detection and monitoring, in particular to a neurology patient sleep disorder monitoring method, which comprises the following steps: S1, collecting and preprocessing body signals of a patient; s2, constructing a cross-modal association graph, and forming a heterogeneous graph structure with a time sequence weight and a modal label; s3, performing graph convolution and sparse attention operation on the cross-modal association graph, and executing node screening and edge weight updating; s4, time domain and frequency domain decomposition is carried out, and alignment is carried out based on time and frequency indexes; s5, performing segmentation and boundary calibration on the feature sequence, dividing sleep stages, and indexing and storing cross-modal fragments; s6, retrieving abnormal stage fragments, constructing abnormal feature sub-graphs and executing structure clustering operation; and S7, mapping the sleep disorder category labels and the cross-modal features to generate associated features and output results. According to the method, the cross-modal heterogeneous graph is constructed and the time-frequency features are fused, so that the recognition precision and the analysis depth of the sleep disorder of the patient in the neurology department are improved.

Description

Technical Field

[0001] This invention relates to the field of medical testing and monitoring technology, and in particular to a monitoring method for sleep disorders in neurology patients. Background Technology

[0002] With the accelerating pace of modern life and the rising incidence of neurological diseases, sleep disorders are becoming increasingly widespread and complex. Clinical studies have shown that a large number of patients in traditional Chinese medicine (TCM) neurology departments experience varying degrees of sleep disturbances during disease progression or recovery, such as difficulty falling asleep, frequent awakenings at night, early awakenings, shallow sleep, and nightmares. These sleep abnormalities not only exacerbate the neurological burden of the primary brain disease but may also negatively impact the autonomic nervous system, emotional system, and immune function, interfering with the treatment process. Therefore, accurate and continuous monitoring and analysis of sleep quality in neurology patients has become a crucial aspect of the diagnosis, treatment, and rehabilitation of neurological diseases.

[0003] Existing sleep monitoring methods primarily rely on polysomnography (PSG) or portable sleep devices to collect signals such as electroencephalogram (EEG), electrocardiogram (ECG), respiration, blood oxygenation, and body movement. Time-series analysis is then used to determine sleep structure and disorder types. However, these methods suffer from several problems: First, mainstream models are often based on single-modality or simple signal fusion, making it difficult to accurately reflect the interactions and dynamic evolution between different physiological signals, especially when dealing with the complex clinical characteristics of patients with TCM-related brain diseases, which can easily lead to identification errors. Second, current models generally ignore multidimensional features related to TCM diagnostic theories and lack the ability to perform cross-modal correlation modeling and time-frequency joint analysis of sleep disorder states, resulting in insufficient clinical interpretability of diagnostic results. Third, traditional analysis pathways cannot achieve structured expression and high-dimensional clustering of abnormal sleep patterns, making it difficult to provide individual-specific subtyping and sleep disorder category labeling support.

[0004] Furthermore, existing research has not yet established a technical framework for sleep disorder identification that integrates multi-source signals and label information for neurology scenarios. It also lacks an overall solution that effectively integrates key algorithm chains such as graph neural networks, time-frequency feature decoupling, multi-scale stage identification, and abnormal feature clustering, which limits the depth of application of intelligent diagnosis and treatment systems in TCM clinical scenarios.

[0005] Therefore, how to provide a monitoring method for sleep disorders in neurology patients is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a monitoring method for sleep disorders in neurological patients. This invention fully integrates key technologies such as cross-modal graph neural modeling, time-frequency feature decomposition, structured stage recognition, and abnormal clustering analysis. It describes in detail the entire process from physiological signal acquisition, heterogeneous graph construction, graph convolution and attention fusion, time-frequency coupling analysis to the generation of sleep disorder category labels. It has the advantages of strong integration, high recognition accuracy, clear structural expression, and excellent clinical adaptability.

[0007] A method for monitoring sleep disorders in neurology patients according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect and preprocess the body signals of neurology patients during sleep.

[0009] S2. Construct a cross-modal correlation graph, using different signal sources as nodes and the feature dependencies between different time slices as edges, to form a heterogeneous graph structure with temporal weights and modal labels;

[0010] S3. Perform multi-level graph convolution and sparse attention calculations on the cross-modal association graph, perform modal channel filtering at the node level, perform temporal weight updates at the edge level, and fuse multi-scale subgraph features to generate a cross-modal feature matrix.

[0011] S4. Extract continuous state change patterns using time-domain decomposition units, extract feature spectrum patterns using frequency-domain decomposition units, and align features based on time and frequency indices to generate time-frequency coupled feature sequences.

[0012] S5. Perform sequence segmentation and stage boundary labeling on the time-frequency coupled feature sequence to form the sleep stage division result, and index and store the cross-modal feature segments corresponding to each stage;

[0013] S6. Retrieve stage segments that meet the anomaly determination conditions in cross-modal feature segments, construct an anomaly feature subgraph and perform structural clustering operation, and output the corresponding sleep disorder category label sequence;

[0014] S7. Perform feature mapping between the sleep disorder category label sequence and the corresponding cross-modal feature fragments to generate sleep disorder associated features and output the analysis results.

[0015] Optionally, the body signals include electroencephalogram (EEG) signals, heart rate signals, respiratory rate signals, body movement signals, and blood oxygen saturation signals, and all signals are synchronously collected and stored according to a unified time reference.

[0016] Optionally, the preprocessing includes multi-scale segmentation, bandpass filtering for noise reduction, outlier removal, missing value filling, and amplitude normalization.

[0017] Optionally, S2 specifically includes:

[0018] S21. Align the preprocessed EEG signals, heart rate signals, respiratory rate signals, body movement signals and blood oxygen saturation signals according to a unified time reference, and set the acquisition channel corresponding to each type of signal source as a node to form a node set containing five nodes, and assign a modal label corresponding to the signal type to each node.

[0019] S22. Define the feature dependency relationship between different time slices as directed edges. The starting point and ending point of each edge correspond to the source node and the target node, respectively. Calculate the temporal weight of each edge according to the time slice interval so that each edge can reflect the correlation between the source node signal and the target node signal under time delay.

[0020] S23. Combine the node set, edge set, temporal weight set, and modal label set to generate a heterogeneous graph structure, so that each edge retains its corresponding temporal weight and each node retains its corresponding modal label, thereby forming a cross-modal association graph.

[0021] Optionally, the cross-modal association graph consists of a node set, an edge set, a temporal weight set, and a modal label set. Each node in the node set corresponds to a preprocessed body signal source aligned with a unified time reference. The modal label set records the correspondence between nodes and their respective signal types. Each edge in the edge set connects two different nodes and represents the feature dependency relationship between corresponding time slices. The temporal weight set records the numerical magnitude of each edge at different time slice intervals. All nodes and edges are associated according to temporal weights and modal labels to form a heterogeneous graph structure, which serves as the input data structure for multi-level graph convolution operations and sparse attention calculations.

[0022] Optionally, S3 specifically includes:

[0023] S31. Based on the cross-modal association graph, construct the initial feature representation of each node, summarize all nodes to form a node feature matrix, and perform multi-level graph convolution operation. Perform multi-level graph convolution operation in sequence. In each layer, the node features are weighted and summed based on the feature vectors of neighboring nodes and the attention coefficients, and the updated node feature vector is generated through a nonlinear function.

[0024] S32. Perform modal channel filtering at the node level. Match the corresponding channel weight vector according to the modal label of the node. Element-wise weight the node feature vector in the channel dimension to obtain the filtered node feature vector.

[0025] S33. Perform temporal weight update at the edge level, and weight and fuse the current edge weight with the updated value calculated based on the correlation coefficient of node features according to the balance coefficient to generate a new temporal weight value.

[0026] S34. During the multi-scale sub-graph feature fusion process, perform the following operations:

[0027] At each scale, obtain the node feature matrix after processing through multi-level graph convolution and modal channel filtering;

[0028] The node feature matrices at all scales are concatenated along the feature dimension to form a concatenated matrix with expanded dimension, which is used to integrate local and global structural information at each scale.

[0029] A linear transformation operation is performed on the concatenated feature matrix to unify the dimensions of the high-dimensional concatenation result, resulting in a cross-modal feature matrix.

[0030] Each row of the cross-modal feature matrix is ​​used as the fused node feature representation to describe the feature response and structural embedding of different signal sources in a multi-scale graph structure.

[0031] Optionally, S4 specifically includes:

[0032] S41. Decompose the cross-modal feature matrix into temporal components by time index, and perform weighted operations on the feature vectors of each time point using a preset temporal decomposition kernel function to obtain a temporal component matrix that reflects the feature changes at different time points.

[0033] S42. Input the cross-modal feature matrix into the frequency domain decomposition unit according to the frequency index, and use the fast Fourier transform function to extract the amplitude information of each node at different frequencies to generate the frequency domain component matrix.

[0034] S43. Based on the time index and frequency index, the time-domain component matrix and the frequency-domain component matrix are feature-aligned. By constructing a time-frequency mapping function, the time features and frequency features under the same state are concatenated to generate a time-frequency coupled feature sequence. The time-frequency mapping function performs channel alignment and feature concatenation operations on the time-domain features of the same node at different time points and their corresponding frequency-domain features at different frequencies by constructing a bidirectional mapping relationship between the time index and the frequency index, so as to ensure the consistency of the coupling result in terms of dimension and state semantics.

[0035] Optionally, S5 specifically includes:

[0036] S51. Sort the time-frequency coupled feature sequences into a sequence matrix according to the time index, so that the feature vectors corresponding to each time point are arranged in an orderly manner on the time axis;

[0037] S52. The sequence matrix is ​​segmented using a sliding window with a fixed length and a fixed step size to obtain multiple sets of continuous feature segments, each segment covering the time-frequency features of the corresponding time interval;

[0038] S53. When calibrating the stage boundary, calculate the ratio of the mean difference to the standard deviation of adjacent feature segments as the difference index. When the difference exceeds the preset threshold, mark the position as the stage boundary.

[0039] S54. Based on the calibrated stage boundaries, the time-frequency coupled feature sequence is divided into multiple sleep stages, and the corresponding feature segments are indexed in the cross-modal feature matrix to generate a stage index table to store the cross-modal feature positions of each stage.

[0040] Optionally, S6 specifically includes:

[0041] S61. Based on the sleep stage segmentation results, calculate the anomaly score for the cross-modal feature segments corresponding to each sleep stage. The anomaly score is normalized based on the difference between the mean and standard deviation of the segment features and the stage features.

[0042] S62. When the abnormal score exceeds the preset threshold, the stage is marked as an abnormal stage, and the node set and edge set corresponding to the stage are extracted from the cross-modal feature matrix to generate an abnormal feature subgraph containing nodes, edges, temporal weights and modal labels.

[0043] S63. Convert all abnormal feature subgraphs into structural description vectors, perform clustering operations based on feature similarity to obtain multiple cluster centers, and assign each abnormal feature subgraph to the most similar cluster center to generate the corresponding sleep disorder category label sequence.

[0044] Optionally, S7 specifically includes:

[0045] S71. Link the sleep disorder category label sequence with the stage index table to obtain the cross-modal feature fragment set corresponding to each abnormal stage;

[0046] S72. Concatenate and fuse each sleep disorder category label with the corresponding cross-modal feature fragment set to obtain a sleep disorder association feature vector that reflects the relationship between the category label and the feature content;

[0047] S73. Sort all sleep disorder-related feature vectors by stage number to form a sleep disorder-related feature sequence, and generate an analysis result table containing stage number, sleep disorder category, and cross-modal feature summary.

[0048] The beneficial effects of this invention are:

[0049] First, based on the actual clinical needs of patients with neurological diseases, this invention constructs a cross-modal association graph model with temporal weights and modal labels by collecting multi-source physiological signals such as electroencephalogram, heart rate, respiration, body movement, and blood oxygen saturation. This breaks through the limitations of traditional single-modal monitoring methods and realizes the structured expression and deep fusion of complex interaction relationships between various modal signals during sleep, effectively improving the accuracy and robustness of sleep state modeling for patients with neurological diseases.

[0050] Secondly, this invention introduces a multi-level graph convolution and sparse attention mechanism to perform modal channel screening at the node level and update temporal weights at the edge level. It combines a time-frequency dual-domain decomposition method to perform fine-grained reconstruction of signal features and completes the automatic division of sleep stages through sliding window segmentation and boundary labeling. On this basis, feature subgraphs are constructed for abnormal stages and cluster analysis is performed to output clinically interpretable sleep disorder category labels, which significantly enhances the interpretability and classification recognition ability of the monitoring results.

[0051] Finally, this invention performs feature mapping between sleep disorder category labels and corresponding cross-modal feature fragments, generates sleep disorder-related features, and outputs analysis results, forming a complete intelligent sleep disorder monitoring solution for application scenarios in traditional Chinese medicine neurology. It has comprehensive advantages of deep integration, accurate identification, and strong adaptability, and can provide high-value data support and decision-making basis for personalized intervention and rehabilitation assessment. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a flowchart of a method for monitoring sleep disorders in neurology patients proposed in this invention;

[0054] Figure 2 This is a flowchart of the cross-modal association graph construction and preprocessing process for a monitoring method for sleep disorders in neurology patients proposed in this invention.

[0055] Figure 3 This is a flowchart illustrating the graph convolution and time-frequency coupling of a monitoring method for sleep disorders in neurology patients proposed in this invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0057] refer to Figure 1-3 A method for monitoring sleep disorders in neurology patients includes the following steps:

[0058] S1. Collect and preprocess the body signals of neurology patients during sleep.

[0059] S2. Construct a cross-modal correlation graph, using different signal sources as nodes and the feature dependencies between different time slices as edges, to form a heterogeneous graph structure with temporal weights and modal labels;

[0060] S3. Perform multi-level graph convolution and sparse attention calculations on the cross-modal association graph, perform modal channel filtering at the node level, perform temporal weight updates at the edge level, and fuse multi-scale subgraph features to generate a cross-modal feature matrix.

[0061] S4. Extract continuous state change patterns using time-domain decomposition units, extract feature spectrum patterns using frequency-domain decomposition units, and align features based on time and frequency indices to generate time-frequency coupled feature sequences.

[0062] S5. Perform sequence segmentation and stage boundary labeling on the time-frequency coupled feature sequence to form the sleep stage division result, and index and store the cross-modal feature segments corresponding to each stage;

[0063] S6. Retrieve stage segments that meet the anomaly determination conditions in cross-modal feature segments, construct an anomaly feature subgraph and perform structural clustering operation, and output the corresponding sleep disorder category label sequence;

[0064] S7. Perform feature mapping between the sleep disorder category label sequence and the corresponding cross-modal feature fragments to generate sleep disorder associated features and output the analysis results.

[0065] This invention provides a monitoring method for sleep disorders in neurological patients. The overall process is constructed from multi-source signal acquisition, cross-modal graph modeling, graph neural fusion, time-frequency decoupling, stage division, anomaly identification to disorder label output. It can realize full-cycle, multi-scale, and intelligent identification and classification of sleep states in patients with brain diseases, and improve the accuracy and adaptability of sleep disorder analysis in traditional Chinese medicine clinical scenarios.

[0066] In this embodiment, the body signals include electroencephalogram (EEG) signals, heart rate signals, respiratory rate signals, body movement signals, and blood oxygen saturation signals. During acquisition, a multi-channel physiological signal acquisition unit is connected to EEG acquisition electrodes, ECG electrodes, a breathing belt, a body movement sensor, and a photoelectric blood oxygen sensor, respectively. Before the patient enters sleep, the acquisition channel status is initialized and a unified time reference is set. A synchronous trigger signal controls each sensor channel to start acquiring data simultaneously. During the acquisition process, the raw data stream is continuously recorded according to the set sampling rate, and all signal streams are synchronously marked and cached in real time according to a unified timestamp in the signal acquisition buffer.

[0067] This invention sets EEG signals, heart rate signals, respiratory rate signals, body movement signals, and blood oxygen saturation signals as core physiological parameters, and uses a unified time reference for synchronous acquisition and storage, ensuring the alignment consistency of multi-source data in the time domain and improving the accuracy of subsequent correlation modeling and time series analysis.

[0068] In this embodiment, the preprocessing includes: performing multi-scale segmentation processing on the synchronously acquired signals according to a preset time window; inputting the signal segment of each time window into a bandpass filter to perform filtering and noise reduction operations within a specific frequency band; then performing outlier detection on the filtered signal and removing outlier data points that exceed a set threshold; filling missing data points using an interpolation algorithm; and normalizing the filled signal segments according to the global amplitude range to ensure that all signals remain consistent within the same amplitude range, ultimately generating a standardized signal sequence.

[0069] This invention performs multi-scale segmentation, bandpass filtering, anomaly removal, missing data filling, and normalization on the acquired signals, which enhances the quality and stability of the original data and ensures the reliability and usability of signal feature extraction and fusion in the subsequent modeling stage.

[0070] In this embodiment, S2 specifically includes:

[0071] S21. Let the preprocessed EEG signal, heart rate signal, respiratory rate signal, body movement signal, and blood oxygen saturation signal be X1(t), X2(t), X3(t), X4(t), and X5(t), respectively, where t is a time index aligned to a unified time base. Define the acquisition channel corresponding to each signal source as a node v. i The node set is represented as V = {v1, v2, v3, v4, v5}, and each node is assigned a modal label m. i , where m i For node v i Corresponding signal type identifier;

[0072] S22. Define the feature dependencies between different time slices as directed edges e. ij (τ), where i and j represent the source node and target node numbers respectively, τ is the time slice interval, and the temporal weight w is used to calculate the edge. ij When (τ), the formula is used:

[0073]

[0074] Where L is the number of sampling points within the time slice, X i (t k ) indicates that at time index t k node v i The signal value, Xj (t k +τ) represents node v after a delay of τ. j The signal value;

[0075] S23. Based on the node set V and the edge set E = {e ij (τ)}、Boundary weight set W={w ij (τ)} and the modal label set M={m i Generate a heterogeneous graph structure G = (V, E, W, M), where each edge retains its corresponding temporal weight and each node retains its corresponding modal label, thus forming a cross-modal association graph.

[0076] This invention proposes a strategy for constructing heterogeneous graphs based on a node-edge structure. Signal sources are mapped as nodes, and the dependencies between time slices are mapped as edges. Modal labels and temporal weight attributes are introduced to realize the structural expression of different modalities and multi-time series data, thereby improving the model's ability to perceive cross-signal interactions.

[0077] In this embodiment, the cross-modal association graph consists of a node set, an edge set, a temporal weight set, and a modal label set. Each node in the node set corresponds to a preprocessed body signal source aligned with a unified time base. The modal label set records the correspondence between nodes and their respective signal types. Each edge in the edge set connects two different nodes and represents the feature dependency relationship between corresponding time slices. The temporal weight set records the numerical magnitude of each edge at different time slice intervals. All nodes and edges are associated according to temporal weights and modal labels to form a heterogeneous graph structure, which serves as the input data structure for multi-level graph convolution operations and sparse attention calculations.

[0078] This invention effectively improves the expressive power of dependencies between multi-source signals and the modeling accuracy of graph neural networks by constructing a cross-modal correlation graph that integrates temporal weights and modal labels.

[0079] In this embodiment, S3 specifically includes:

[0080] S31. Based on the cross-modal association graph, construct the initial feature representation of each node, summarize all nodes to form the node feature matrix H(0), and perform multi-level graph convolution operations. The update method for each layer of graph convolution is as follows:

[0081]

[0082] in, Represents the node v at level l. i eigenvectors, Represents the node v at level l-1 j The feature vector, N(i) represents the node v i The neighborhood group, W represents the attention coefficient of the l-th layer. (l) Let be the weight matrix of the l-th layer, and σ(·) be the nonlinear activation function;

[0083] S32. Perform modal channel filtering at the node level, assuming node v i The modal label is m i The corresponding channel weight vector is p i The filtered node feature vectors The formula is:

[0084]

[0085] Where ⊙ represents element-wise multiplication, p i For node v i The channel weight vector corresponding to the modality label;

[0086] S33. Perform time-series weight update at the edge level, assuming edge e ij The time-series weight of (τ) is w ij (τ), the updated time series weights are The calculation formula is:

[0087]

[0088] Where β is the time-series weight balance coefficient. Represents node v i With node v j Correlation coefficient of eigenvectors;

[0089] S34. In multi-scale subgraph feature fusion, the node feature matrices at different scales are concatenated and mapped to a cross-modal feature matrix F. The specific process is as follows:

[0090] Suppose that at each scale *s*, the node feature matrix is ​​processed by graph convolution at layer *l* and modal channel filtering. Then, the feature matrices at all scales are concatenated along the feature dimension to obtain the concatenated feature matrix *H*. cat The calculation method is as follows:

[0091]

[0092] Wherein, Concat(·) represents the concatenation operation along the feature dimension. These represent the node feature matrices at different scales after filtering. d represents the feature dimension after concatenation. s Let S be the feature dimension at scale s, where S is the total number of scales;

[0093] For the concatenated feature matrix H cat By performing a linear mapping, we obtain the cross-modal feature matrix F. The mapping formula is as follows:

[0094] F = H cat ·W f +b f ;

[0095] Among them, W f Let b be the linear transformation weight matrix. f It is the bias vector;

[0096] Each row in F is used as a fused node-level cross-modal feature vector to represent the structural embedding representation of each signal source at different scales.

[0097] This invention introduces multi-level graph convolution and sparse attention mechanisms into graph neural networks, while performing modal channel filtering on node features, performing temporal updates on edge weights, and fusing multi-scale subgraph information to construct a high-quality cross-modal feature matrix, effectively enhancing the model's ability to represent and model complex sleep states.

[0098] In this embodiment, S4 specifically includes:

[0099] S41. Index the cross-modal feature matrix F according to the time index {t1,t2,…,t…} T Perform time-domain decomposition, and let the time-domain decomposition kernel function be K. t (Δt), then node v i At time index t k Temporal components at the location The calculation formula is:

[0100]

[0101] Among them, F i (t u ) represents node v i At time index t u The eigenvector at point Δt = t k -t u This represents the time difference, where T represents the total number of time slices;

[0102] S42. Index the cross-modal feature matrix F according to the frequency index {f1,f2,…,f...} Q The input frequency domain decomposition unit is used to obtain node v through the Fast Fourier Transform (FFT) function. i At frequency index f q Frequency domain components at the location The calculation formula is:

[0103]

[0104] in, Indicates the frequency index f qThe corresponding complex exponential transformation factor is given below, where j is the imaginary unit, |·| represents the complex modulo operation, and Q represents the number of frequency points;

[0105] S43. Based on the time index and frequency index, divide the time-domain component matrix F (t) With frequency domain component matrix F (f) Perform feature alignment, define a time-frequency mapping function, and assign the time index t to the time index t. k With frequency index f q Alignment: The time-frequency mapping function constructs a bidirectional mapping relationship between time indices and frequency indices, performing channel alignment and feature concatenation operations on the time-domain features of the same node at different time points and their corresponding frequency-domain features, ensuring the consistency of the coupling results in terms of dimension and state semantics, and generating a time-frequency coupled feature sequence S:

[0106] s kq =Concat(F (t) (t k ),F (f) (f q ));

[0107] Where Concat(·) represents the concatenation operation along the feature dimension, s kq ∈S is the time index t k With frequency index f q The corresponding time-frequency coupled feature vector.

[0108] This invention extracts continuous state change patterns and feature spectrum patterns through time-domain decomposition and frequency-domain decomposition modules, respectively. Based on time index and frequency index, feature alignment is performed to generate coupled feature sequences with temporal and frequency structures, thereby achieving a deeper understanding and modeling of sleep state changes.

[0109] In this embodiment, S5 specifically includes:

[0110] S51. Obtain the time-frequency coupled feature sequence S, and index S by time index {t1,t2,…,t…} T Arranged into a sequence matrix S T ;

[0111] S52, Using the sliding window function W(·) to apply S T Perform segmentation, let the window length be λ and the step size be δ, then the feature fragment U of the p-th segment. p The calculation formula is:

[0112] U p =W(S) T ,p)={s (p-1)δ+1 ,s (p-1)δ+2 ,…,s (p-1)δ+λ};

[0113] Where p = 1, 2, ..., P, and P is the number of segments;

[0114] S53. When calibrating the stage boundary, calculate the adjacent segment U. p with U p+1 Difference index Δ p The formula is:

[0115]

[0116] Where μ(·) represents the mean vector of the segment features, σ(·) represents the standard deviation vector of the segment features, and ||·||2 represents the Euclidean norm, when Δ p When the value is greater than θ, position p is marked as the stage boundary, and θ is the stage segmentation threshold.

[0117] S54. Based on all stage boundaries, divide the time-frequency coupled feature sequence S into multiple sleep stages, and for each stage Z... r In the cross-modal feature matrix F, the corresponding segment indexes are stored to form a stage index table. in For stage Z r The corresponding set of feature indices.

[0118] This invention uses a sliding window and a difference index to divide time-frequency coupled feature sequences into stages and constructs a stage index table, realizing automatic segmentation and accurate positioning of different sleep stages, providing basic support for subsequent anomaly identification and structural clustering, and enhancing the refinement of the analysis.

[0119] In this embodiment, S6 specifically includes:

[0120] S61. Based on the sleep stage division results, let Z be the stage for each sleep stage. r The corresponding set of cross-modal feature fragments is F. r ;Calculation stage Z r Abnormal score E r The formula is:

[0121]

[0122] Among them, |F r |For stage Z r The number of feature fragments, μ(F) r ) is the mean vector of the stage features, σ(F) r ) is the standard deviation vector of the stage features, and ||·||2 is the Euclidean norm;

[0123] S62, when E r When >γ, stage Z rThe abnormal phase is marked as such, γ is the anomaly detection threshold, and the set of nodes V corresponding to the abnormal phase is extracted from the cross-modal feature matrix F. r With edge set Constructing anomaly feature subgraphs Among them W r Let M be the set of corresponding edge weights. r For the corresponding modal tag set;

[0124] S63. For all sets of abnormal feature subgraphs Perform structural clustering operations, and let the structural description vector of each subgraph be g. r A clustering algorithm based on feature similarity is used to calculate cluster centers, and each subgraph is assigned to the most similar cluster center, thereby generating a corresponding sleep disorder category label sequence. Where R abn y represents the number of abnormal phases. r For stage Z r The corresponding sleep disorder category label.

[0125] This invention calculates the abnormal score for each sleep stage, extracts the abnormal stages and constructs feature subgraphs, performs cluster analysis based on structural similarity, and outputs a sequence of sleep disorder category labels. This achieves a structured expression and automated classification of abnormal sleep behaviors, improving the model's abnormality detection capability and clinical reference value.

[0126] In this embodiment, S7 specifically includes:

[0127] S71. Obtain the sleep disorder category label sequence Where R abn y represents the number of abnormal phases. r For stage Z r The corresponding sleep disorder category labels are from the stage index table. Retrieve the feature index set for each abnormal stage And extract the corresponding feature fragment set from the cross-modal feature matrix F.

[0128] S72, Label sleep disorder categories y r With the corresponding feature fragment set Q r Perform feature mapping to construct a sleep disorder-related feature vector A. r The calculation formula is:

[0129]

[0130] Among them, |Q r | represents the number of feature segments, q represents the q-th feature segment, e(y r(This is a label for the sleep disorder category y) r The embedded vector, Concat(·) represents the concatenation operation along the feature dimension;

[0131] S73, Associate sleep disorder feature vector A r According to stage Z r Arrange the data to generate a sequence of sleep disorder-related features, input it into the analysis unit, and generate an analysis results table containing stage number, sleep disorder category, and cross-modal feature summary.

[0132] This invention performs feature mapping between sleep disorder category labels and cross-modal feature fragments, outputting sleep disorder-related feature sequences, thereby improving the visualization and clinical interpretability of diagnostic results.

[0133] Example 1:

[0134] To verify the feasibility of this invention in practice, it was applied to a neurological clinical monitoring scenario to conduct continuous sleep status assessments on patients exhibiting typical sleep disorder symptoms such as insomnia, vivid dreams, early awakening, nocturnal awakening, and sleep apnea. In this scenario, 35 patients diagnosed with neurology were selected as experimental subjects, including 18 patients with moderate to severe sleep disorders and 17 patients with mild disorders or subjective insomnia. The patient group encompassed clinical manifestations such as mild stroke sequelae, cognitive decline, autonomic nervous system dysfunction, and depression accompanied by insomnia.

[0135] In practical applications, the following steps are taken: First, EEG, heart rate, respiratory rate, body movement, and blood oxygen saturation signals are collected from each patient during sleep. A 5-channel synchronous acquisition device is used for calibration and continuous recording at a unified time reference. The data acquisition period is from 22:00 to 06:00 the next day, with a sampling rate of 128Hz and an average data storage capacity of 420MB per person. Next, multi-scale segmentation processing is performed on the acquired signals. Noise reduction is applied to each channel using a bandpass filter, outliers are removed using a quintile criterion, missing segments are filled using a linear interpolation algorithm, and the amplitudes of various signals are unified using maximum-minimum normalization.

[0136] The preprocessed signal is input into a cross-modal graph construction unit to build a set of nodes and edges. Each channel is defined as a modal node, and the cross-dependency between time slices is defined as an edge. A 5-second segment is extracted using a sliding window, and the correlation weights between edges are calculated and modal labels are recorded. The heterogeneous graph is then input into a graph convolution and sparse attention fusion unit. A 4-layer graph convolutional network extracts fused features. Modal channel filtering is performed at the node level, and weight updates based on correlation coefficients are performed at the edge level. Finally, a 128-dimensional cross-modal feature vector is obtained. Next, time-domain and frequency-domain decomposition operations are performed on the fused feature matrix, and time and frequency indices are aligned to generate a time-frequency coupled feature sequence.

[0137] The feature sequences were segmented into stages using a sliding window segmentation and difference detection algorithm, with an average of 8–10 sleep stages per patient. Abnormalities were scored for each stage; segments exhibiting significant abnormal respiratory fluctuations, abnormal heart rate fluctuations, and high-frequency EEG activity were marked as abnormal state segments, accounting for approximately 15%–35% of all sleep cycles. For these abnormal stages, abnormal feature subgraphs were constructed, and structural description-based clustering analysis was performed, ultimately classifying the abnormal stages into five main sleep disorder types: apnea, heart rate fluctuation, light sleep, hyperkinesia, and mixed.

[0138] To verify the effectiveness of the analysis, a comparison was made with traditional single-modal sleep scoring algorithms (such as single-channel EEG + rule-based threshold judgment). The results show that the present invention significantly improves obstacle recognition accuracy, stage segmentation precision, and sleep structure reconstruction integrity. Specific experimental data are as follows:

[0139] Table 1 Comparison of the Invention Method and Traditional Methods in Sleep Disorder Identification Task

[0140]

[0141] As shown in Table 1, the method of this invention comprehensively outperforms traditional methods in multiple key dimensions such as accuracy, positioning precision, and diagnostic efficiency. It exhibits significant performance advantages, particularly in tasks involving multi-channel signal interaction anomalies, complex state recognition, and structural stage segmentation, validating its practicality and advanced nature in complex clinical scenarios within neurology. This embodiment fully demonstrates the improvement effect of this invention over existing technologies and has promising prospects for widespread application.

[0142] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring sleep disorders in neurological patients, characterized in that, Includes the following steps: S1. Collect and preprocess the body signals of neurology patients during sleep. S2. Construct a cross-modal correlation graph, using different signal sources as nodes and the feature dependencies between different time slices as edges, to form a heterogeneous graph structure with temporal weights and modal labels; S3. Perform multi-level graph convolution and sparse attention calculations on the cross-modal association graph, perform modal channel filtering at the node level, perform temporal weight updates at the edge level, and fuse multi-scale subgraph features to generate a cross-modal feature matrix. S4. Extract continuous state change patterns using time-domain decomposition units, extract feature spectrum patterns using frequency-domain decomposition units, and align features based on time and frequency indices to generate time-frequency coupled feature sequences. S5. Perform sequence segmentation and stage boundary labeling on the time-frequency coupled feature sequence to form the sleep stage division result, and index and store the cross-modal feature segments corresponding to each stage; S6. Retrieve stage segments that meet the anomaly determination conditions in cross-modal feature segments, construct an anomaly feature subgraph and perform structural clustering operation, and output the corresponding sleep disorder category label sequence; S7. Perform feature mapping between the sleep disorder category label sequence and the corresponding cross-modal feature fragments to generate sleep disorder associated features and output the analysis results.

2. The method for monitoring sleep disorders in neurological patients according to claim 1, characterized in that, The body signals include electroencephalogram (EEG) signals, heart rate signals, respiratory rate signals, body movement signals, and blood oxygen saturation signals. All signals are collected and stored synchronously according to a unified time reference.

3. The method for monitoring sleep disorders in neurological patients according to claim 1, characterized in that, The preprocessing includes multi-scale segmentation, bandpass filtering for noise reduction, outlier removal, missing value filling, and amplitude normalization.

4. The method for monitoring sleep disorders in neurological patients according to claim 1, characterized in that, S2 specifically includes: S21. Align the preprocessed EEG signals, heart rate signals, respiratory rate signals, body movement signals and blood oxygen saturation signals according to a unified time reference, and set the acquisition channel corresponding to each type of signal source as a node to form a node set containing five nodes, and assign a modal label corresponding to the signal type to each node. S22. Define the feature dependency relationship between different time slices as directed edges. The starting point and ending point of each edge correspond to the source node and the target node, respectively. Calculate the temporal weight of each edge according to the time slice interval so that each edge can reflect the correlation between the source node signal and the target node signal under time delay. S23. Combine the node set, edge set, temporal weight set, and modal label set to generate a heterogeneous graph structure, so that each edge retains its corresponding temporal weight and each node retains its corresponding modal label, thereby forming a cross-modal association graph.

5. A method for monitoring sleep disorders in neurological patients according to claim 4, characterized in that, The cross-modal association graph consists of a node set, an edge set, a temporal weight set, and a modal label set. Each node in the node set corresponds to a preprocessed body signal source aligned with a unified time reference. The modal label set records the correspondence between nodes and their respective signal types. Each edge in the edge set connects two different nodes and represents the feature dependency relationship between corresponding time slices. The temporal weight set records the numerical magnitude of each edge at different time slice intervals. All nodes and edges are associated according to temporal weights and modal labels to form a heterogeneous graph structure, which serves as the input data structure for multi-level graph convolution operations and sparse attention calculations.

6. The method for monitoring sleep disorders in neurological patients according to claim 1, characterized in that, S3 specifically includes: S31. Based on the cross-modal association graph, construct the initial feature representation of each node, summarize all nodes to form a node feature matrix, and perform multi-level graph convolution operation. Perform multi-level graph convolution operation in sequence. In each layer, the node features are weighted and summed based on the feature vectors of neighboring nodes and the attention coefficients, and the updated node feature vector is generated through a nonlinear function. S32. Perform modal channel filtering at the node level. Match the corresponding channel weight vector according to the modal label of the node. Element-wise weight the node feature vector in the channel dimension to obtain the filtered node feature vector. S33. Perform temporal weight update at the edge level, and weight and fuse the current edge weight with the updated value calculated based on the correlation coefficient of node features according to the balance coefficient to generate a new temporal weight value. S34. During the multi-scale sub-graph feature fusion process, perform the following operations: At each scale, obtain the node feature matrix after processing through multi-level graph convolution and modal channel filtering; The node feature matrices at all scales are concatenated along the feature dimension to form a concatenated matrix with expanded dimension, which is used to integrate local and global structural information at each scale. A linear transformation operation is performed on the concatenated feature matrix to unify the dimensions of the high-dimensional concatenation result, resulting in a cross-modal feature matrix. Each row of the cross-modal feature matrix is ​​used as the fused node feature representation to describe the feature response and structural embedding of different signal sources in a multi-scale graph structure.

7. The method for monitoring sleep disorders in neurological patients according to claim 1, characterized in that, S4 specifically includes: S41. Decompose the cross-modal feature matrix into temporal components by time index, and perform weighted operations on the feature vectors of each time point using a preset temporal decomposition kernel function to obtain a temporal component matrix that reflects the feature changes at different time points. S42. Input the cross-modal feature matrix into the frequency domain decomposition unit according to the frequency index, and use the fast Fourier transform function to extract the amplitude information of each node at different frequencies to generate the frequency domain component matrix. S43. Based on the time index and frequency index, the time-domain component matrix and the frequency-domain component matrix are feature-aligned. By constructing a time-frequency mapping function, the time features and frequency features under the same state are concatenated to generate a time-frequency coupled feature sequence. The time-frequency mapping function performs channel alignment and feature concatenation operations on the time-domain features of the same node at different time points and their corresponding frequency-domain features at different frequencies by constructing a bidirectional mapping relationship between the time index and the frequency index, so as to ensure the consistency of the coupling result in terms of dimension and state semantics.

8. The method for monitoring sleep disorders in neurological patients according to claim 1, characterized in that, S5 specifically includes: S51. Sort the time-frequency coupled feature sequences into a sequence matrix according to the time index, so that the feature vectors corresponding to each time point are arranged in an orderly manner on the time axis; S52. The sequence matrix is ​​segmented using a sliding window with a fixed length and a fixed step size to obtain multiple sets of continuous feature segments, each segment covering the time-frequency features of the corresponding time interval; S53. When calibrating the stage boundary, calculate the ratio of the mean difference to the standard deviation of adjacent feature segments as the difference index. When the difference exceeds the preset threshold, mark the position as the stage boundary. S54. Based on the calibrated stage boundaries, the time-frequency coupled feature sequence is divided into multiple sleep stages, and the corresponding feature segments are indexed in the cross-modal feature matrix to generate a stage index table to store the cross-modal feature positions of each stage.

9. A method for monitoring sleep disorders in neurological patients according to claim 1, characterized in that, S6 specifically includes: S61. Based on the sleep stage segmentation results, calculate the anomaly score for the cross-modal feature segments corresponding to each sleep stage. The anomaly score is normalized based on the difference between the mean and standard deviation of the segment features and the stage features. S62. When the abnormal score exceeds the preset threshold, the stage is marked as an abnormal stage, and the node set and edge set corresponding to the stage are extracted from the cross-modal feature matrix to generate an abnormal feature subgraph containing nodes, edges, temporal weights and modal labels. S63. Convert all abnormal feature subgraphs into structural description vectors, perform clustering operations based on feature similarity to obtain multiple cluster centers, and assign each abnormal feature subgraph to the most similar cluster center to generate the corresponding sleep disorder category label sequence.

10. A method for monitoring sleep disorders in neurological patients according to claim 1, characterized in that, Specifically, S7 includes: S71. Link the sleep disorder category label sequence with the stage index table to obtain the cross-modal feature fragment set corresponding to each abnormal stage; S72. Concatenate and fuse each sleep disorder category label with the corresponding cross-modal feature fragment set to obtain a sleep disorder association feature vector that reflects the relationship between the category label and the feature content; S73. Sort all sleep disorder-related feature vectors by stage number to form a sleep disorder-related feature sequence, and generate an analysis result table containing stage number, sleep disorder category, and cross-modal feature summary.

Images