Interpretable Respiratory Event Detection Method Based on Prior Guidance Graph Structure Learning
By employing a priori guided graph structure learning approach, the problems of multi-scale temporal information and clinical interpretability in OSA diagnosis were solved, achieving high-precision and visualized respiratory event detection, and improving the model's generalization ability and physician trust.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JIANGXI UNIVERSITY OF FINANCE AND ECONOMICS
- Filing Date
- 2026-03-16
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies in OSA diagnosis neglect multi-scale temporal information, lack adaptive and constrained graph structure learning, and lack clinical interpretability, resulting in poor model generalization ability and low physician trust.
A priori guided graph structure learning approach is adopted, which uses notch filtering and bandpass filtering to denoise multimodal physiological signals, combines segment-level and global-level graph learning to construct an objective function, and uses clinical prior knowledge to guide graph structure learning for feature extraction and prediction.
It effectively captures multi-scale signal features, provides visualized physiological evidence, enhances the interpretability of the model and its generalization ability among different subjects, and improves detection accuracy and physician trust.
Smart Images
Figure CN121867708B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence-assisted diagnostic technology, and in particular to an interpretable respiratory event detection method based on prior guided graph structure learning. Background Technology
[0002] Sleep apnea-hypopnea syndrome (OSA) is a common sleep-disordered breathing condition characterized by repeated collapse of the upper airway during sleep, leading to apnea or hypopnea, which in turn causes intermittent hypoxemia and sleep fragmentation. OSA is closely associated with serious complications such as cardiovascular disease, metabolic disorders, and cognitive impairment. Currently, polysomnography (PSG) is the "gold standard" for diagnosing OSA, analyzing multiple physiological signals including electroencephalography (EEG), electrooculography (EOG), electromyography (EMG), nasal and oral airflow, chest and abdominal respiratory movements, and blood oxygen saturation. However, traditional manual PSG interpretation is time-consuming, labor-intensive, and relies heavily on physician experience, making it difficult to meet the needs of large-scale screening.
[0003] With the development of deep learning, automatic detection methods based on convolutional neural networks (CNNs) or recurrent neural networks (RNNs) have been widely used. However, existing technologies have the following main drawbacks:
[0004] (1) Ignoring multi-scale temporal information: Physiological signals (such as airflow and blood oxygen) are usually asynchronous. Single-scale feature extraction is difficult to capture instantaneous changes within a segment (such as sudden stop of airflow) and long-range delay effects across segments (such as subsequent decrease in blood oxygen).
[0005] (2) Lack of adaptive and constrained graph structure learning: Existing graph neural network (GNN) methods mostly rely on predefined static graphs or completely free data-driven graphs. The former cannot adapt to individual differences among different subjects, while the latter is prone to learning false connections that violate physiological common sense under noise interference, resulting in poor model generalization ability.
[0006] (3) Insufficient clinical interpretability: Most “black box” models cannot explain the basis for judgment and are difficult to gain the trust of doctors. Summary of the Invention
[0007] In view of the above, the main objective of this invention is to propose an interpretable respiratory event detection method based on prior guidance graph structure learning, so as to solve the above-mentioned technical problems.
[0008] This invention proposes an interpretable respiratory event detection method based on prior guidance graph structure learning, the method comprising the following steps:
[0009] Step 1: Input the multimodal physiological signals into the detection model, and perform denoising processing on the multimodal physiological signals through notch filtering and bandpass filtering to obtain denoised signal data; perform normalization operation on the denoised signal data to obtain preprocessed multimodal signal data.
[0010] The preprocessed multimodal signal data is divided into a preset number of fixed-length, non-overlapping time segments to obtain a time segment sequence;
[0011] Step 2: Perform segment-level graph learning and feature extraction on the time segment sequence to obtain the segment-level overall representation and segment adjacency matrix;
[0012] Step 3: Utilize fragment-level global representations to perform global-level learning and long-range dependency extraction to obtain global-level global representations and global adjacency matrices;
[0013] Step 4: Construct an objective function using the fragment adjacency matrix and the global adjacency matrix; use the objective function to update the detection model, obtaining the updated detection model;
[0014] Step 5: Input the multimodal physiological signals into the updated detection model, execute steps 1 to 3, and obtain the updated segment-level global representation and the updated global-level global representation;
[0015] By combining the updated fragment-level global representation and the updated global-level global representation, a fused feature vector is obtained by fusing local detail features and global dependency features.
[0016] The fused feature vector is input into a multilayer perceptron classifier to obtain the prediction result.
[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0018] 1. This invention introduces a prior-guided graph structure learning mechanism and uses common medical knowledge (such as the continuity of physiological signals and specific coupling relationships between modalities) as constraints to effectively filter out random noise and spurious associations in the data.
[0019] 2. This invention can intuitively demonstrate the physiological basis of model decisions through visualized fragment-level and global-level adjacency matrices. For example, the model can clearly present pathological features such as "when obstructive apnea occurs, chest and abdominal respiratory movements are enhanced but airflow signals are interrupted (fragment-level decoupling)" and "airflow events lead to a delayed decrease in blood oxygen (global-level causal chain)." This visualization result, which aligns with clinical understanding, greatly enhances doctors' trust in AI-assisted diagnostic systems and helps in the rapid localization of lesions.
[0020] 3. This invention, through a multi-scale graph learning architecture (combining fragment-level micro-interactions and global-level macro-dependencies), can simultaneously capture short-term signal waveform distortions and long-term evolution patterns spanning multiple time segments. Compared to CNNs that only focus on local features or RNNs that only focus on temporal sequences, this invention can more comprehensively and completely characterize the spatiotemporal dynamics of breathing events. Attached Figure Description
[0021] Figure 1 This is a flowchart of the interpretable respiratory event detection method based on prior guided graph structure learning proposed in this invention;
[0022] Figure 2 This is a schematic diagram of the overall process of the interpretable respiratory event detection method based on prior guided graph structure learning proposed in this invention. Detailed Implementation
[0023] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0024] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.
[0025] Please see Figure 1 This invention proposes an interpretable respiratory event detection method based on prior guidance graph structure learning, which includes the following steps:
[0026] Step 1: Input the multimodal physiological signals into the detection model, and perform denoising processing on the multimodal physiological signals through notch filtering and bandpass filtering to obtain denoised signal data; perform normalization operation on the denoised signal data to obtain preprocessed multimodal signal data.
[0027] The preprocessed multimodal signal data is divided into a preset number of fixed-length, non-overlapping time segments to obtain a time segment sequence.
[0028] Further, in step 1, multimodal polysomnography (PSG) physiological signals of the subject are acquired, specifically including blood oxygen saturation (SpO2), nasal airflow, and thoracic effort signals. After denoising (e.g., notch filtering, bandpass filtering) and normalization processing, the multimodal physiological signals are divided into... A non-overlapping time segment;
[0029] Specifically, the first Each segment is represented as (In this invention) ),in, The number of sampling points for each segment, and the overall representation of a single sample sequence is: .
[0030] Step 2: Perform segment-level graph learning and feature extraction on the time segment sequence to obtain the segment-level overall representation and segment adjacency matrix;
[0031] In step 2, segment-level graph learning and feature extraction are performed on the time segment sequence to obtain the segment-level overall representation and the segment adjacency matrix. The specific steps are as follows:
[0032] Each time segment in the time segment sequence is used as input, and the original signal within each time segment is converted into a high-dimensional feature representation through a parameter-shared encoder to obtain modality-level node features;
[0033] By utilizing modality-level node features, an adaptive structure learner is used to learn the connection relationships between modalities within a segment in order to obtain the segment adjacency matrix.
[0034] A fragment-level graph is constructed based on the fragment adjacency matrix and modal-level node features;
[0035] By utilizing the fragment adjacency matrix and modal-level node features, message passing is performed on the fragment-level graph structure to update node representations and obtain the final updated hidden state of the nodes.
[0036] Perform a graph-level readout operation on the final updated node hidden state to obtain the fragment embedding representation;
[0037] By embedding all segments into time series representations using one-dimensional convolution, a segment-level overall representation is obtained.
[0038] In the process of using modality-level node features to learn the connection relationships between modalities within a segment through an adaptive structure learner to obtain the segment adjacency matrix, the following relationship exists:
[0039] ;
[0040] in, Indicates the first The segment adjacency matrix of segments, This represents the symmetry treatment that makes the diagonal zero. This indicates the processing of the Sigmoid function. This indicates that the flat structure is reconstructed into a matrix. Represents a nonlinear transformation. This indicates that the node features are flattened. Represents modal-level node features;
[0041] In the process of using the fragment adjacency matrix and modal-level node features to perform message passing on the fragment-level graph structure to update node representations and obtain the final updated hidden state of the nodes, the following relationship exists:
[0042] ;
[0043] in, Indicates the first The segment in the first The hidden state of the layer's nodes Indicates the first Linear transformation processing of layers, Indicates the first Learnable parameters of the layer Indicates the first The segment in the first The hidden state of the layer's nodes, and ; This represents the layer index of a graph neural network. This represents the weighted neighborhood aggregation result based on the weighted adjacency matrix;
[0044] During the graph-level readout operation on the finally updated node hidden state to obtain the fragment embedding representation, the following relationship exists:
[0045] ;
[0046] in, Indicates the first Fragment embedding representation of a segment, This indicates a graph-level readout operation. Indicates the first The final updated hidden state of the node for each fragment. express A real vector space.
[0047] It should be noted that the subscripts in this invention The symbol indicates that it belongs to the "segment-level".
[0048] Furthermore, in this step, the relational expression for constructing the fragment-level graph is:
[0049] ;
[0050] in, Represents a fragment-level graph; It represents a set of nodes, and each node in the set corresponds one-to-one with a mode; Denotes the set of edges, by The structure was obtained.
[0051] Step 3: Utilize fragment-level global representations to perform global-level learning and long-range dependency extraction to obtain global-level global representations and global adjacency matrices;
[0052] In step 3, global learning and long-range dependency extraction are performed using the fragment-level global representation to obtain the global global representation and the global adjacency matrix. The specific steps are as follows:
[0053] Based on the fragment-level global representation, adjacent fragments are paired non-overlappingly and spliced and linearly projected at each modality-level node dimension to obtain global node features;
[0054] The global node features are input into the data-driven branch, and the data-driven adjacency matrix is obtained by multi-head scaling dot product attention calculation.
[0055] In the prior-guided branch, a prior adjacency matrix is constructed using clinical prior knowledge, and a learnable gating matrix is introduced to obtain the physiological constraint adjacency matrix.
[0056] The data-driven adjacency matrix and the physiologically constrained adjacency matrix are fused by convex combination to obtain the global adjacency matrix;
[0057] A global weighted undirected graph is constructed based on global node features and a global adjacency matrix. Feature propagation and aggregation are performed on the global weighted undirected graph using the global adjacency matrix and global node features to extract global long-range dependency feature representations and obtain a global overall representation.
[0058] In the process of inputting global node features into the data-driven branch and calculating the data-driven adjacency matrix through multi-head scaling dot product attention, the following relationship exists:
[0059] ;
[0060] in, This represents a data-driven adjacency matrix. Indicates the total number of attention heads. The first node is obtained by projecting global node features. A query matrix with attention heads The first node is obtained by projecting global node features. Transpose of the query matrix of each attention head; This represents the feature dimension of each attention head, which is the ratio of the global node feature dimension to the number of attention heads. This indicates a normalization operation;
[0061] In the prior-guided branch, the process of constructing a prior adjacency matrix using clinical prior knowledge and introducing a learnable gating matrix to obtain the physiological constraint adjacency matrix corresponds to the following relationship:
[0062] ;
[0063] in, Represents the physiological constraint adjacency matrix. Represents the prior adjacency matrix. This represents a learnable gated matrix. This represents the Hadamard element-wise product;
[0064] In the process of fusing the data-driven adjacency matrix and the physiologically constrained adjacency matrix through convex combination to obtain the global adjacency matrix, the following relationship exists:
[0065] ;
[0066] in, Represents the global adjacency matrix; Represents the fusion coefficient, and .
[0067] Specifically, this step learns the global adjacency matrix through a two-branch mechanism that includes a data-driven branch and a priori-guided branch. The data-driven branch uses a self-attention mechanism to mine potential cross-segment associations, while the priori-guided branch uses clinical prior knowledge and gating mechanisms to introduce physiological constraints. The two branches are fused to obtain the global adjacency matrix and extract global long-range dependency feature representations.
[0068] It should be noted that the subscripts in this invention The symbol indicates that it belongs to the "global level".
[0069] Furthermore, in this step, the relational expression for constructing the globally weighted undirected graph is:
[0070] ;
[0071] in, Represents a globally weighted undirected graph; This represents the global set of nodes, which is obtained by pairing adjacent time segments non-overlapping and generating nodes according to modality. Denotes the set of edges, by Build it; This represents the characteristics of global nodes.
[0072] Step 4: Construct an objective function using the fragment adjacency matrix and the global adjacency matrix; use the objective function to update the detection model, obtaining the updated detection model;
[0073] In step 4, an objective function is constructed using the fragment adjacency matrix and the global adjacency matrix. This objective function is then used to update the detection model, resulting in the updated detection model. The objective function consists of fragment-level sparse regularization, global-level sparse regularization, global graph smoothing regularization, prior consistency regularization, and the standard cross-entropy classification loss.
[0074] Physiological interactions are typically sparse across different scales, with only a few key modalities coupling and dominating respiratory events. By imposing an L1 penalty on the adjacency matrix, the model is forced to eliminate redundant connections. The expressions for the fragment-level sparse regularization term and the global-level sparse regularization term in the corresponding process are as follows:
[0075] ;
[0076] ;
[0077] in, This indicates a fragment-level sparse regularization term. Indicates the total number of time segments. Indicates the number of modes. Indicates the first Nodes in the segment adjacency matrix of each segment With nodes Edge weights between them This represents a global-level sparse regularization term. Represents the nodes in the global adjacency matrix With nodes Edge weights between them Indicates the number of nodes in the global graph;
[0078] Long-range physiological dependencies should remain stable within their relevant temporal contexts and should not undergo abrupt changes. A smoothing penalty based on the Laplacian matrix is used to encourage the smooth evolution of node connection patterns. The expression for the global graph smoothing regularization term is:
[0079] ;
[0080] in, This represents the global graph smoothing regularization term. Represents the global graph Laplacian matrix. The degree matrix representing the global adjacency matrix. Represents the trace of a matrix. Denotes the Frobenius norm. This represents the transpose of the global adjacency matrix. This represents the transpose of the global graph Laplacian matrix;
[0081] To ensure that the learned global structure does not deviate from clinical consensus, a penalty is applied to the deviation between the global adjacency matrix and the prior adjacency matrix. This ensures that while the model utilizes data-driven learning, it is strictly constrained by medical common sense. The expression for the prior consistency regularization term is:
[0082] ;
[0083] in, This indicates a priori consistency regularization;
[0084] By constructing a complex structure regularization objective function to balance prediction accuracy and the interpretability of physiological mechanisms, it is defined as a weighted combination of classification loss and a four-term graph structure regularization term. The expression of the objective function is as follows:
[0085] ;
[0086] in, Describe the objective function. This represents the standard cross-entropy classification loss. The weight coefficients represent the fragment-level sparse regularization terms. Represents the weight coefficients of the global sparse regularization term. This represents the weight coefficients of the global graph smoothing regularization term. This represents the weight coefficient of the prior consistency regularization term.
[0087] It should be noted that this invention employs the Adam optimization algorithm, minimizing the total loss function through backpropagation. The invention synchronously updates all learnable parameters, including those of the encoder, adaptive structure learner, and classifier. Through multiple regularization mechanisms, this invention not only prevents overfitting of the model on small clinical sample data but also significantly enhances the model's generalization ability across different individual subjects.
[0088] This invention improves detection accuracy in low signal-to-noise ratio environments by explicitly modeling multi-scale graph structures and introducing clinical prior guidance, and provides a "white-box" decision-making basis through a visualized adjacency matrix.
[0089] Step 5: Input the multimodal physiological signals into the updated detection model, execute steps 1 to 3, and obtain the updated segment-level global representation and the updated global-level global representation;
[0090] By combining the updated fragment-level global representation and the updated global-level global representation, a fused feature vector is obtained by fusing local detail features and global dependency features.
[0091] The fused feature vector is input into a multilayer perceptron classifier to obtain the prediction result;
[0092] In step 5, using the updated fragment-level global representation and the updated global-level global representation, a fused feature vector is obtained by fusing local detail features and global dependency features, with the following relationship:
[0093] ;
[0094] in, Represents the fused feature vector. This represents the updated fragment-level overall representation. This represents the updated global representation. This indicates a concatenation operation, which is the operation of joining two representation vectors along the feature dimension to form a higher-dimensional fused representation;
[0095] In the process of inputting the fused feature vector into a multilayer perceptron classifier to obtain the prediction result, the following relationship exists:
[0096] ;
[0097] in, Indicates the prediction result. This means selecting the category with the highest score from the classifier's output as the prediction result.
[0098] For details, please refer to Figure 2Multimodal physiological signals are segmented and preprocessed over time before being input into a shared encoder to extract modality-level node features. Based on these node features, the signals are then fed into a graph learning module, which includes fragment-level graph learning and global-level graph learning. Fragment-level graph learning sequentially constructs fragment graphs and learns fragment adjacency matrices, performs Graph Neural Network (GNN) message passing, and reads out graphs to obtain fragment-level representations. Global-level graph learning sequentially constructs global graphs (stitching and projection), learns prior-guided adjacency matrices, learns data-driven adjacency matrices, and fuses adjacency matrices to obtain a global adjacency matrix. Based on GNN message passing and graph readout, a global-level representation is obtained. The fragment-level and global-level representations are then fused, and the fused features are input into a Multi-Layer Perceptron (MLP) classifier to output prediction results.
[0099] Furthermore, the final result obtained in this step The probability distribution matrix contains the predicted probability for each category (normal, hypoventilation, and apnea), and the category with the highest probability value is selected as the final predicted respiratory event type.
[0100] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0101] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0102] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.
Claims
1. An interpretable respiratory event detection method based on prior guided graph structure learning, characterized in that, The method includes the following steps: Step 1: Input the multimodal physiological signals into the detection model, and perform denoising processing on the multimodal physiological signals through notch filtering and bandpass filtering to obtain denoised signal data; perform normalization operation on the denoised signal data to obtain preprocessed multimodal signal data. The preprocessed multimodal signal data is divided into a preset number of fixed-length, non-overlapping time segments to obtain a time segment sequence; Step 2: Perform segment-level graph learning and feature extraction on the time segment sequence to obtain the segment-level overall representation and segment adjacency matrix; Step 3: Utilize fragment-level global representations to perform global-level learning and long-range dependency extraction to obtain global-level global representations and global adjacency matrices; Step 4: Construct an objective function using the fragment adjacency matrix and the global adjacency matrix; use the objective function to update the detection model, obtaining the updated detection model; Step 5: Input the multimodal physiological signals into the updated detection model, execute steps 1 to 3, and obtain the updated segment-level global representation and the updated global-level global representation; By combining the updated fragment-level global representation and the updated global-level global representation, a fused feature vector is obtained by fusing local detail features and global dependency features. The fused feature vector is input into a multilayer perceptron classifier to obtain the prediction result.
2. The interpretable respiratory event detection method based on prior guided graph structure learning according to claim 1, characterized in that, In step 2, segment-level graph learning and feature extraction are performed on the time segment sequence to obtain the segment-level overall representation and the segment adjacency matrix. The specific steps are as follows: Each time segment in the time segment sequence is used as input, and the original signal within each time segment is converted into a high-dimensional feature representation through a parameter-shared encoder to obtain modality-level node features; By utilizing modality-level node features, an adaptive structure learner is used to learn the connection relationships between modalities within a segment in order to obtain the segment adjacency matrix. A fragment-level graph is constructed based on the fragment adjacency matrix and modal-level node features; By utilizing the fragment adjacency matrix and modal-level node features, message passing is performed on the fragment-level graph structure to update node representations and obtain the final updated hidden state of the nodes. Perform a graph-level readout operation on the final updated node hidden state to obtain the fragment embedding representation; By embedding all segments into one-dimensional convolution and aggregating the time series representations, a segment-level overall representation is obtained.
3. The interpretable respiratory event detection method based on prior guided graph structure learning according to claim 2, characterized in that, In the process of using modality-level node features to learn the connection relationships between modalities within a segment through an adaptive structure learner to obtain the segment adjacency matrix, the following relationship exists: ； in, Indicates the first The segment adjacency matrix of segments, This represents the symmetry treatment that makes the diagonal zero. This indicates the processing of the Sigmoid function. This indicates that the flat structure is reconstructed into a matrix. Represents a nonlinear transformation. This indicates that the node features are flattened. Represents modal-level node features; In the process of using the fragment adjacency matrix and modal-level node features to perform message passing on the fragment-level graph structure to update node representations and obtain the final updated hidden state of the nodes, the following relationship exists: ； in, Indicates the first The segment in the first The hidden state of the layer's nodes Indicates the first Linear transformation processing of layers, Indicates the first Learnable parameters of the layer Indicates the first The segment in the first The hidden state of the nodes in the layer.
4. The interpretable respiratory event detection method based on prior guided graph structure learning according to claim 3, characterized in that, During the graph-level readout operation on the finally updated node hidden state to obtain the fragment embedding representation, the following relationship exists: ； in, Indicates the first Fragment embedding representation of a segment, This indicates a graph-level readout operation. Indicates the first The final updated hidden state of the node for each fragment. express A real vector space.
5. The interpretable respiratory event detection method based on prior guided graph structure learning according to claim 4, characterized in that, In step 3, global learning and long-range dependency extraction are performed using fragment-level global representations to obtain a global global representation and a global adjacency matrix. The specific steps are as follows: Based on the fragment-level global representation, adjacent fragments are paired non-overlappingly and spliced and linearly projected at each modality-level node dimension to obtain global node features; The global node features are input into the data-driven branch, and the data-driven adjacency matrix is obtained by multi-head scaling dot product attention calculation. In the prior-guided branch, a prior adjacency matrix is constructed using clinical prior knowledge, and a learnable gating matrix is introduced to obtain the physiological constraint adjacency matrix. The data-driven adjacency matrix and the physiologically constrained adjacency matrix are fused by convex combination to obtain the global adjacency matrix; A globally weighted undirected graph is constructed based on global node features and a global adjacency matrix; By utilizing the global adjacency matrix and global node features, feature propagation and aggregation are performed on the globally weighted undirected graph to extract global long-range dependency feature representations, resulting in a global-level overall representation.
6. The interpretable respiratory event detection method based on prior guided graph structure learning according to claim 5, characterized in that, In the process of inputting global node features into the data-driven branch and calculating the data-driven adjacency matrix through multi-head scaling dot product attention, the following relationship exists: ； in, This represents a data-driven adjacency matrix. Indicates the total number of attention heads. The first node is obtained by projecting global node features. A query matrix with attention heads The first node is obtained by projecting global node features. Transpose of the query matrix of each attention head This represents the feature dimension of each attention head. This indicates a normalization operation; In the prior-guided branch, the process of constructing a prior adjacency matrix using clinical prior knowledge and introducing a learnable gating matrix to obtain the physiological constraint adjacency matrix corresponds to the following relationship: ； in, Represents the physiological constraint adjacency matrix. Represents the prior adjacency matrix. This represents a learnable gated matrix. This represents the element-wise product of Hadamard.
7. The interpretable respiratory event detection method based on prior guided graph structure learning according to claim 6, characterized in that, In the process of fusing the data-driven adjacency matrix and the physiologically constrained adjacency matrix through convex combination to obtain the global adjacency matrix, the following relationship exists: ； in, Represents the global adjacency matrix. This represents the fusion coefficient.
8. The interpretable respiratory event detection method based on prior guided graph structure learning according to claim 7, characterized in that, In step 4, an objective function is constructed using the fragment adjacency matrix and the global adjacency matrix. This objective function is then used to update the detection model, resulting in the updated detection model. The objective function consists of fragment-level sparse regularization, global-level sparse regularization, global graph smoothing regularization, prior consistency regularization, and standard cross-entropy classification loss. The expression for the fragment-level sparse regularization is as follows: ； in, This indicates a fragment-level sparse regularization term. Indicates the total number of time segments. Indicates the number of modes. Indicates the first Nodes in the segment adjacency matrix of each segment With nodes Edge weights between them; The expression for the global sparse regularization term is: ； in, This represents a global-level sparse regularization term. Represents the nodes in the global adjacency matrix With nodes Edge weights between them This indicates the number of nodes in the global graph.
9. The interpretable respiratory event detection method based on prior guided graph structure learning according to claim 8, characterized in that, In constructing the objective function using the fragment adjacency matrix and the global adjacency matrix, the expression for the global graph smoothing regularization term is: ； in, This represents the global graph smoothing regularization term. Represents the global graph Laplacian matrix. The degree matrix representing the global adjacency matrix. Represents the trace of a matrix. Describing the Frobenius norm, This represents the transpose of the global adjacency matrix. This represents the transpose of the global graph Laplacian matrix; The expression for the prior consistency regularization term is: ； in, This indicates a priori consistency regularization; The expression for the objective function is: ； in, Describe the objective function. This represents the standard cross-entropy classification loss. The weight coefficients represent the fragment-level sparse regularization terms. Represents the weight coefficients of the global sparse regularization term. This represents the weight coefficients of the global graph smoothing regularization term. This represents the weight coefficient of the prior consistency regularization term.
10. The interpretable respiratory event detection method based on prior guided graph structure learning according to claim 9, characterized in that, In step 5, using the updated fragment-level global representation and the updated global-level global representation, a fused feature vector is obtained by fusing local detail features and global dependency features, with the following relationship: ； in, Represents the fused feature vector. This represents the updated fragment-level overall representation. This represents the updated global representation. Indicates a splicing operation; In the process of inputting the fused feature vector into a multilayer perceptron classifier to obtain the prediction result, the following relationship exists: ； in, Indicates the prediction result. This means selecting the category with the highest score from the classifier's output as the prediction result.