A Blood Pressure Signal Abnormality Early Warning Method Based on Attention Mechanism
By performing event segment decomposition and phase space reconstruction on blood pressure signals, and combining the improved MTAD-GAT model, the shortcomings of existing technologies in identifying changes in blood pressure waveform structure and dynamic evolution relationships have been addressed, thereby improving the accuracy and stability of blood pressure abnormality early warning.
Patent Information
- Application Number
- CN202611353557.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-02
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies are insufficient to fully characterize the structural changes in blood pressure waveforms within a single beat cycle and the dynamic evolution relationships between different beats, resulting in limited ability to identify complex abnormal blood pressure patterns.
By decomposing blood pressure signals into event segments, constructing a blood pressure event graph, and introducing delayed embedding, an improved MTAD-GAT model of phase space reconstruction attention mechanism is established. This enables joint modeling of the structural relationship of blood pressure events and the dynamic relationship of phase space, and generates counterfactual normal blood pressure waveforms to calculate abnormal deviations.
It significantly improves the accuracy and stability of blood pressure abnormality early warning, can effectively identify complex blood pressure abnormality patterns, and realizes structural perception and dynamic modeling of blood pressure waveforms.
Smart Images

Figure CN123116090A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and medical signal processing technology, and in particular to a method for early warning of abnormal blood pressure signals based on an attention mechanism. Background Technology
[0002] With the rapid development of continuous blood pressure monitoring devices and wearable medical devices, real-time acquisition and long-term monitoring of blood pressure signals are playing an increasingly important role in cardiovascular disease early warning and health management. Regarding the detection of abnormal blood pressure signals, existing technologies typically employ statistical analysis or machine learning-based methods to model blood pressure time series data. For example, they utilize sliding window statistical features, traditional machine learning classification models, or deep neural networks to analyze blood pressure signal trends and identify abnormal fluctuations. However, in practical applications, existing technologies still have many shortcomings.
[0003] Traditional methods for detecting abnormal blood pressure mostly rely on feature extraction and model training based on raw blood pressure time series. This makes it difficult to fully characterize the structural changes of blood pressure waveforms within a single beat cycle and the dynamic evolution relationship between different beats, resulting in limited ability to identify complex abnormal blood pressure patterns. While some deep learning methods can improve the accuracy of abnormality detection to a certain extent, most of them only use local or global statistical features of the time series for modeling, lacking the characterization of the nonlinear dynamics of blood pressure signals and failing to effectively reflect the evolution of blood pressure signals in phase space.
[0004] Therefore, how to provide an attention-based method for early warning of abnormal blood pressure signals is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] One objective of this invention is to propose a blood pressure signal anomaly early warning method based on an attention mechanism. This invention decomposes blood pressure signals into event segments and constructs a blood pressure event graph, introduces delayed embedding to achieve phase space reconstruction, and designs an improved MTAD-GAT model that incorporates an attention mechanism for phase space reconstruction to achieve joint modeling of the structural relationship of blood pressure events and the dynamic relationship of phase space. At the same time, it establishes an individual normal blood pressure state memory bank and generates counterfactual normal blood pressure waveforms. By calculating the waveform structure deviation and trend deviation, an anomaly score is formed, thereby improving the accuracy and stability of blood pressure anomaly early warning.
[0006] A blood pressure signal abnormality early warning method based on an attention mechanism according to an embodiment of the present invention includes the following steps:
[0007] Step 1: Acquire continuous blood pressure signals, preprocess the continuous blood pressure signals, and segment the cardiac cycle according to the position of the blood pressure waveform trough to obtain a single-beat blood pressure segment sequence;
[0008] Step 2: Decompose the single-beat blood pressure fragment sequence into event fragments, extract blood pressure event fragments and corresponding event features, and construct a blood pressure event fragment sequence in chronological order;
[0009] Step 3: Construct a blood pressure event graph using the blood pressure event segments as nodes and the event features as node attributes; establish phase transition relationships within single-beat blood pressure segments and establish similar phase connection relationships between adjacent single-beat blood pressure segments;
[0010] Step 4: Perform delayed embedding processing on the blood pressure event fragment sequence to construct a phase space state vector sequence, and construct a phase space adjacency graph based on the distance relationship between the phase space state vectors;
[0011] Step 5: Based on the blood pressure event graph and the phase space adjacency graph, perform joint graph attention calculation using the improved MTAD-GAT model that introduces a phase space reconstruction attention mechanism to obtain the pulsation state representation vector;
[0012] Step 6: Establish an individual normal blood pressure state memory bank, calculate the similarity between the pulse state representation vector and the historical state representation vector, and determine the counterfactual normal blood pressure waveform based on the similarity;
[0013] Step 7: Compare the current blood pressure waveform with the counterfactual normal blood pressure waveform to calculate the abnormal deviation, and calculate the abnormal score based on the abnormal deviation sequence to generate a blood pressure abnormality warning result.
[0014] Optionally, step one specifically includes:
[0015] The blood pressure of the subject is continuously sampled, and blood pressure time series data is obtained according to a preset sampling frequency. The blood pressure time series data consists of multiple pressure sampling points arranged in chronological order.
[0016] The blood pressure time series data is subjected to signal preprocessing, which includes bandpass filtering of the blood pressure time series data and amplitude normalization of the filtered blood pressure time series data to form a preprocessed blood pressure sequence.
[0017] In the preprocessed blood pressure sequence, the pressure values of adjacent sampling points are compared in chronological order. When the pressure value of a sampling point is less than both the pressure value of the previous sampling point and the pressure value of the next sampling point, the corresponding sampling point is determined as the blood pressure waveform trough point, and the time position corresponding to the blood pressure waveform trough point is recorded.
[0018] The blood pressure data between two adjacent blood pressure waveform troughs are segmented according to time sequence, and the blood pressure data between two adjacent blood pressure waveform troughs is divided into a single beat blood pressure segment, which is then arranged in time sequence to form a single beat blood pressure segment sequence.
[0019] Calculate the duration and pressure amplitude variation range of each single-beat blood pressure segment, and delete single-beat blood pressure segments that exceed the preset duration threshold range and preset pressure amplitude threshold range to form a single-beat blood pressure segment sequence.
[0020] Optionally, step two specifically includes:
[0021] Local slope calculations were performed on the single-beat blood pressure segment sequence in chronological order. The pressure change slope sequence was obtained by calculating the ratio of the pressure difference to the time difference between adjacent sampling points. Based on the direction of change of the pressure change slope sequence, the single-beat blood pressure segment was divided into multiple candidate event segments.
[0022] In single-beat blood pressure segments, key event points of blood pressure waveforms are identified based on the pressure change slope sequence. Key event points of blood pressure waveforms include the inflection point of rapid pressure rise, the systolic peak point, the inflection point of pressure fall, and the diastolic flat point. Blood pressure data between adjacent key event points of blood pressure waveforms are segmented according to time sequence to form multiple blood pressure event segments.
[0023] For each blood pressure event segment, segment structure features are calculated, including event duration, pressure amplitude formed by the difference between the maximum and minimum pressure values, average pressure value of the event segment, and average absolute value of the pressure change slope.
[0024] Segment transition features are calculated for adjacent blood pressure event segments. The segment transition features include the time interval between adjacent event segments, the difference in pressure amplitude, and the ratio of the difference in the average slope of adjacent event segments to the time interval.
[0025] Phase structure features are calculated for each blood pressure event segment. The phase structure features include the ratio of the event segment start time to the duration of the single-beat blood pressure segment, and the ratio of the event segment duration to the duration of the single-beat blood pressure segment.
[0026] The fragment structure features, fragment transition features, and phase structure features are combined to form a blood pressure event feature vector, and then arranged according to the chronological order of the occurrence of blood pressure event fragments to construct a blood pressure event fragment sequence.
[0027] Optionally, step three specifically includes:
[0028] Each blood pressure event segment in the blood pressure event segment sequence is defined as a node in the blood pressure event graph according to the time sequence, and the blood pressure event feature vector is used as the node attribute. The blood pressure event feature vector is composed of segment structure features, segment transition features and phase structure features.
[0029] Within a single-beat blood pressure segment, phase transition relationships between nodes are established according to the temporal order of blood pressure event segments. Two temporally adjacent blood pressure event segment nodes are connected by directed edges to form a phase transition connection structure within the single-beat blood pressure segment.
[0030] Based on the phase ratio value in the phase structure characteristics, blood pressure event segment nodes in the same phase interval in different single-beat blood pressure segments are compared, and when the phase ratio difference is less than a preset threshold, a cross-pulse connection relationship is established between the corresponding nodes to form a similar phase connection structure.
[0031] The phase transition connection structure is merged with similar phase connection structures to form the edge set of the blood pressure event graph, and the adjacency matrix of the blood pressure event graph is generated based on the connection relationship between nodes.
[0032] Generate the blood pressure event graph data structure based on the adjacency matrix and node attributes of the blood pressure event graph.
[0033] Optionally, step four specifically includes:
[0034] Delayed embedding processing is performed on the blood pressure event feature vectors in the blood pressure event segment sequence. Multiple blood pressure event feature vectors at different time intervals are selected according to a preset time delay interval and then concatenated in chronological order to form a phase space state vector.
[0035] Arrange all phase space state vectors in chronological order to form a phase space state vector sequence;
[0036] Calculate the distance between any two phase space state vectors in the phase space state vector sequence, and calculate the phase ratio of the phase space state vectors during the duration of a single beat blood pressure segment;
[0037] The phase space state vectors are sorted by distance value from smallest to largest, and a preset number of nearest neighbor phase space state vectors are selected to form candidate neighborhoods.
[0038] The phase ratio values of the phase space state vectors in the candidate neighborhood are compared. When the phase ratio difference is less than the preset phase threshold, an undirected edge is established between the corresponding phase space state vectors, and a phase space adjacency graph is constructed based on all the undirected edges.
[0039] Optionally, step five specifically includes:
[0040] Input the blood pressure event graph and phase space adjacency graph into the improved MTAD-GAT model;
[0041] The improved MTAD-GAT model sets up a dual-graph embedding module at the input end, and performs linear mapping processing on the node attribute matrix in the blood pressure event graph and the phase space state vector matrix in the phase space adjacency graph respectively to obtain the event node embedding representation and the phase space node embedding representation.
[0042] At the node level, the event node embedding representation and the phase space node embedding representation are concatenated according to their temporal correspondence, and then mapped to a unified dimension through a linear transformation to obtain a joint node representation;
[0043] The joint node representation is input into the backbone structure of the improved MTAD-GAT model. The backbone structure includes a phase transition graph attention branch, a cross-pulsation phase evolution graph attention branch, and a phase space reconstruction attention branch.
[0044] The phase transition graph attention branch uses the phase transition connection relationship in the blood pressure event graph as the edge constraint. It performs graph attention calculation on the joint node representations corresponding to adjacent blood pressure event segments within the same single-beat blood pressure segment, and performs weighted aggregation of adjacent node representations according to attention weights to obtain the phase transition representation.
[0045] The attention branch of the cross-pulse phase evolution graph uses the same phase connection relationship in the blood pressure event graph as the edge constraint. It performs graph attention calculation on the joint node representations corresponding to blood pressure event segments in the same phase interval in adjacent single-pulse blood pressure segments, and performs weighted aggregation on the adjacent node representations according to the attention weight to obtain the cross-pulse phase evolution representation.
[0046] The phase space reconstruction attention branch takes the phase space state vector matrix and the phase space adjacency graph as input, and includes:
[0047] Perform a linear transformation on the phase space state vector matrix to obtain the query vector matrix, key vector matrix, and value vector matrix;
[0048] Extract the set of adjacent phase space state vectors corresponding to each target phase space state vector based on the phase space adjacency graph;
[0049] Calculate the state vector distance between the target phase space state vector and the adjacent phase space state vector, and the difference between the phase ratio value corresponding to the target phase space state vector and the phase ratio value corresponding to the adjacent phase space state vector.
[0050] The basic attention score is obtained by performing a dot product operation between the query vector and the key vector. The phase space reconstruction offset term is obtained based on the distance value of the state vector and the phase ratio difference. The phase space reconstruction offset term is added to the basic attention score to obtain the reconstruction attention score.
[0051] Normalization is performed on the reconstructed attention score to obtain the phase space attention weight, and the value vectors corresponding to the adjacent phase space state vectors are weighted and aggregated according to the phase space attention weight to obtain the phase space reconstruction representation;
[0052] The phase transition representation, the cross-pulsation phase evolution representation, and the phase space reconstruction representation are concatenated and processed by linear mapping to obtain the pulsation state representation vector.
[0053] Optionally, step six specifically includes:
[0054] An individual normal blood pressure state memory bank is established, which consists of the pulsation state representation vector corresponding to the historical single blood pressure segment and the corresponding blood pressure waveform segment, and is stored in chronological order.
[0055] The similarity is calculated between the pulsation state representation vector corresponding to the current single-beat blood pressure segment and the historical state representation vector in the individual's normal blood pressure state memory database to obtain a similarity sequence.
[0056] The historical state representation vectors are sorted according to their similarity, and a set of candidate normal states is formed by selecting the top-ranked historical state representation vectors based on their similarity.
[0057] Extract historical blood pressure waveform segments corresponding to the candidate normal state set, and perform weighted summation of the historical blood pressure waveform segments according to the similarity weights corresponding to the historical state representation vectors to obtain the counterfactual normal blood pressure waveform.
[0058] Optionally, step seven specifically includes:
[0059] Align the current blood pressure waveform with the counterfactual normal blood pressure waveform in chronological order, and establish a one-to-one waveform point pair at the corresponding sampling time position;
[0060] For each pair of waveform points, calculate the absolute value of the pressure difference and arrange them in chronological order to form a pressure deviation sequence;
[0061] The average value of all deviation values in the pressure deviation sequence is calculated, and the maximum deviation value in the pressure deviation sequence is extracted. The average deviation value and the maximum deviation value are weighted and summed according to the preset weight to form the waveform structure deviation.
[0062] Calculate the slope of pressure change between the current blood pressure waveform and the counterfactual normal blood pressure waveform at adjacent sampling points, take the absolute value of the slope difference at the corresponding sampling positions, and average all slope differences to form the waveform trend deviation.
[0063] The deviation of waveform structure and the deviation of waveform trend are weighted and summed according to preset weights to form the abnormal deviation.
[0064] The abnormal deviations corresponding to multiple consecutive single-beat blood pressure segments are accumulated in chronological order to form a cumulative abnormal deviation, and an abnormal score is calculated based on the cumulative abnormal deviation.
[0065] When the abnormal score is greater than the preset abnormal threshold, the blood pressure signal is judged as an abnormal state and a blood pressure abnormality warning result is generated.
[0066] The beneficial effects of this invention are:
[0067] This invention constructs a blood pressure event graph structure by segmenting continuous blood pressure signals into cardiac cycles and decomposing them into event segments. The graph structure uses blood pressure event segments as nodes and event features as node attributes, introducing event-level structural representation based on the temporal characteristics of the blood pressure signal, effectively improving the representational ability of blood pressure waveform structure information. During the graph structure modeling stage, phase transition relationships within a single-beat blood pressure segment and similar phase connection relationships across beats are simultaneously established, enabling the model to simultaneously characterize the structural changes of the blood pressure waveform within a single beat cycle and the dynamic evolution relationships between different beats. In the feature modeling stage, a delayed embedding mechanism is introduced to reconstruct the phase space of the blood pressure event feature sequence, and a phase space adjacency graph is constructed based on the distance and phase ratio relationships between phase space state vectors, thereby introducing a nonlinear dynamic state representation of the blood pressure signal into the model. In terms of model design, a phase space reconstruction attention mechanism is incorporated. The improved MTAD-GAT model, through the collaborative computation of phase transition graph attention branches, cross-pulse phase evolution graph attention branches, and phase space reconstruction attention branches, achieves joint modeling of the structural relationship of blood pressure events and the dynamic relationship of phase space, significantly enhancing the model's ability to express complex blood pressure change patterns. In the anomaly detection stage, an individual's normal blood pressure state memory is established. By calculating the similarity between the current pulsation state representation vector and the historical normal state representation vector, a counterfactual normal blood pressure waveform is constructed. Based on this, an anomaly deviation calculation mechanism is designed, consisting of waveform structure deviation and waveform trend deviation. This allows anomaly detection to no longer rely on a single prediction error but instead combines individual normal state for difference analysis, thereby significantly improving the accuracy and stability of blood pressure anomaly early warning and realizing structural perception, dynamic modeling, and individualized intelligent early warning of abnormal blood pressure fluctuations. Attached Figure Description
[0068] 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:
[0069] Figure 1 This is a schematic diagram of the overall process of an attention-based blood pressure signal abnormality early warning method proposed in this invention;
[0070] Figure 2This is a schematic diagram of the blood pressure event graph construction process in the blood pressure signal abnormality early warning method based on the attention mechanism proposed in this invention;
[0071] Figure 3 This is a schematic diagram of the improved MTAD-GAT model structure in the blood pressure signal abnormality early warning method based on the attention mechanism proposed in this invention. Detailed Implementation
[0072] 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.
[0073] refer to Figures 1-3 A blood pressure signal abnormality early warning method based on attention mechanism includes the following steps:
[0074] Step 1: Acquire continuous blood pressure signals, preprocess the continuous blood pressure signals, and segment the cardiac cycle according to the position of the blood pressure waveform trough to obtain a single-beat blood pressure segment sequence;
[0075] Step 2: Decompose the single-beat blood pressure fragment sequence into event fragments, extract blood pressure event fragments and corresponding event features, and construct a blood pressure event fragment sequence in chronological order;
[0076] Step 3: Construct a blood pressure event graph using blood pressure event segments as nodes and event features as node attributes. Establish phase transition relationships within single-beat blood pressure segments and establish similar phase connection relationships between adjacent single-beat blood pressure segments.
[0077] Step 4: Perform delayed embedding processing on the blood pressure event fragment sequence to construct a phase space state vector sequence, and construct a phase space adjacency graph based on the distance relationship between the phase space state vectors;
[0078] Step 5: Based on the blood pressure event graph and the phase space adjacency graph, perform joint graph attention calculation using the improved MTAD-GAT model that introduces a phase space reconstruction attention mechanism to obtain the pulsation state representation vector;
[0079] Step 6: Establish an individual normal blood pressure state memory bank, calculate the similarity between the pulse state representation vector and the historical state representation vector, and determine the counterfactual normal blood pressure waveform based on the similarity;
[0080] Step 7: Compare the current blood pressure waveform with the counterfactual normal blood pressure waveform to calculate the abnormal deviation, and calculate the abnormal score based on the abnormal deviation sequence to generate a blood pressure abnormality warning result.
[0081] In this embodiment, step one specifically includes:
[0082] The blood pressure of the subject is continuously sampled, and blood pressure time series data is obtained according to the preset sampling frequency. The blood pressure time series data consists of multiple pressure sampling points arranged in chronological order.
[0083] Signal preprocessing is performed on the blood pressure time series data. The signal preprocessing includes bandpass filtering of the blood pressure time series data and amplitude normalization of the filtered blood pressure time series data to form a preprocessed blood pressure series.
[0084] In the preprocessed blood pressure sequence, the pressure values of adjacent sampling points are compared in chronological order. When the pressure value of a sampling point is less than both the pressure value of the previous sampling point and the pressure value of the next sampling point, the corresponding sampling point is determined as the blood pressure waveform trough point, and the time position corresponding to the blood pressure waveform trough point is recorded.
[0085] The blood pressure data between two adjacent blood pressure waveform troughs are segmented according to time sequence, and the blood pressure data between two adjacent blood pressure waveform troughs is divided into a single beat blood pressure segment, which is then arranged in time sequence to form a single beat blood pressure segment sequence.
[0086] Calculate the duration and pressure amplitude variation range of each single-beat blood pressure segment, and delete single-beat blood pressure segments that exceed the preset duration threshold range and preset pressure amplitude threshold range to form a single-beat blood pressure segment sequence.
[0087] In this implementation, the blood pressure acquisition device uses an arterial pressure sensor or a non-invasive continuous blood pressure monitoring device for signal acquisition, with the blood pressure signal sampling frequency set to 125Hz–250Hz. Bandpass filtering uses a digital bandpass filter with a filtering frequency range of 0.5Hz–20Hz to remove low-frequency baseline drift and high-frequency noise. Amplitude normalization uses a linear normalization method based on minimum and maximum values to convert the blood pressure time series data to an amplitude range of 0–1. Blood pressure waveform troughs are determined by comparing the pressure values of three consecutive sampling points; when the pressure value at the intermediate sampling point is lower than the pressure values at the preceding and following sampling points, the intermediate sampling point is marked as a blood pressure waveform trough. The threshold range for the duration of a single-stroke blood pressure segment is set to 0.4s–1.5s, corresponding to a heart rate range of 40–150 beats / minute; the threshold range for pressure amplitude change is set to 10mmHg–120mmHg. When the duration or pressure amplitude change of a single-stroke blood pressure segment exceeds the threshold range, the corresponding single-stroke blood pressure segment is deleted from the single-stroke blood pressure segment sequence.
[0088] In this embodiment, step two specifically involves:
[0089] Local slope calculations were performed on the single-beat blood pressure segment sequence in chronological order. The pressure change slope sequence was obtained by calculating the ratio of the pressure difference to the time difference between adjacent sampling points. Based on the direction of change of the pressure change slope sequence, the single-beat blood pressure segment was divided into multiple candidate event segments.
[0090] In single-beat blood pressure segments, key event points of blood pressure waveforms are identified based on the pressure change slope sequence. Key event points of blood pressure waveforms include the inflection point of rapid pressure rise, the systolic peak point, the inflection point of pressure fall, and the diastolic flat point. Blood pressure data between adjacent key event points of blood pressure waveforms are segmented according to time sequence to form multiple blood pressure event segments.
[0091] For each blood pressure event segment, segment structure features are calculated, including event duration, pressure amplitude formed by the difference between the maximum and minimum pressure values, average pressure value of the event segment, and average absolute value of the pressure change slope.
[0092] Segment transition features are calculated for adjacent blood pressure event segments. These features include the time interval between adjacent event segments, the difference in pressure amplitude, and the ratio of the difference in the average slope of adjacent event segments to the time interval.
[0093] Phase structure features are calculated for each blood pressure event segment. The phase structure features include the ratio of the event segment start time to the duration of the single-beat blood pressure segment, and the ratio of the event segment duration to the duration of the single-beat blood pressure segment.
[0094] The fragment structure features, fragment transition features, and phase structure features are combined to form a blood pressure event feature vector, and then arranged according to the chronological order of the occurrence of blood pressure event fragments to construct a blood pressure event fragment sequence.
[0095] In this implementation, the sampling frequency of the single-stroke blood pressure segment is maintained at 125Hz–250Hz. The local slope is calculated by dividing the pressure difference between adjacent sampling points by the time difference, where the time interval between adjacent sampling points is 0.004s–0.008s. Segments with a pressure change slope greater than 5 mmHg / s are identified as rapid pressure rise segments, and segments with an absolute pressure change slope less than 2 mmHg / s are identified as flat segments. The systolic peak point is determined by searching for the local maximum pressure value in the single-stroke blood pressure segment; the pressure drop inflection point is determined by the position where the slope changes from a positive value to a negative value. The duration of the blood pressure event segment is generally between 0.05s and 0.6s. The pressure amplitude is obtained by subtracting the minimum pressure value of the segment from the maximum pressure value of the segment. The average pressure value of the event segment is calculated by summing the pressure values of all sampling points within the segment and dividing by the number of sampling points. The time interval between adjacent event segments is generally between 0.01s and 0.3s. The proportion of the event segment start time in the duration of the single-stroke blood pressure segment is usually between 0 and 1.
[0096] In this embodiment, step three specifically includes:
[0097] Each blood pressure event segment in the blood pressure event segment sequence is defined as a node in the blood pressure event graph according to the time sequence, and the blood pressure event feature vector is used as the node attribute. The blood pressure event feature vector is composed of segment structure features, segment transition features and phase structure features.
[0098] Within a single-beat blood pressure segment, phase transition relationships between nodes are established according to the temporal order of blood pressure event segments. Two temporally adjacent blood pressure event segment nodes are connected by directed edges to form a phase transition connection structure within the single-beat blood pressure segment.
[0099] Based on the phase ratio value in the phase structure characteristics, blood pressure event segment nodes in the same phase interval in different single-beat blood pressure segments are compared, and when the phase ratio difference is less than a preset threshold, a cross-pulse connection relationship is established between the corresponding nodes to form a similar phase connection structure.
[0100] The phase transition connection structure is merged with similar phase connection structures to form the edge set of the blood pressure event graph, and the adjacency matrix of the blood pressure event graph is generated based on the connection relationship between nodes.
[0101] Generate the blood pressure event graph data structure based on the adjacency matrix and node attributes of the blood pressure event graph.
[0102] In this implementation, the nodes of the blood pressure event graph are composed of each blood pressure event segment in the blood pressure event segment sequence. Node attributes are represented by blood pressure event feature vectors, with the feature vector dimension set to 8-12 dimensions. Phase transition relationships within a single-beat blood pressure segment are established using directed edges according to the temporal order of the blood pressure event segments. Each single-beat blood pressure segment typically contains 3-8 blood pressure event segments. The phase structure feature is represented by the ratio of the event segment start time to the single-beat blood pressure segment duration, with a phase ratio value ranging from 0 to 1. Cross-pulse phase connections are established by comparing the phase ratio values of blood pressure event segments in different single-beat blood pressure segments. When the difference between two phase ratio values is less than 0.10, a cross-pulse connection edge is established between the corresponding nodes. The adjacency matrix of the blood pressure event graph is represented using a binary matrix. Nodes with connections have matrix elements with a value of 1, while nodes without connections have matrix elements with a value of 0.
[0103] In this embodiment, step four specifically includes:
[0104] Delayed embedding processing is performed on the blood pressure event feature vectors in the blood pressure event segment sequence. Multiple blood pressure event feature vectors at different time intervals are selected according to a preset time delay interval and then concatenated in chronological order to form a phase space state vector.
[0105] Arrange all phase space state vectors in chronological order to form a phase space state vector sequence;
[0106] Calculate the distance between any two phase space state vectors in the phase space state vector sequence, and calculate the phase ratio of the phase space state vectors during the duration of a single beat blood pressure segment;
[0107] The phase space state vectors are sorted by distance value from smallest to largest, and a preset number of nearest neighbor phase space state vectors are selected to form candidate neighborhoods.
[0108] The phase ratio values of the phase space state vectors in the candidate neighborhood are compared. When the phase ratio difference is less than the preset phase threshold, an undirected edge is established between the corresponding phase space state vectors, and a phase space adjacency graph is constructed based on all the undirected edges.
[0109] In this implementation, the feature vector dimension of blood pressure events is set to 8–12 dimensions. The time delay interval in the delayed embedding process is set to 2–5 blood pressure event segments, and the delayed embedding dimension is set to 3–6. Phase space state vectors are formed by sequentially concatenating state vectors from adjacent time positions. The distance between phase space state vectors is calculated using Euclidean distance, which is obtained by summing the squares of the differences between corresponding feature components and then taking the square root. During neighborhood selection, the vectors are sorted in ascending order of distance value, and the top 5–10 nearest neighbor state vectors are selected as candidate neighbors. The phase ratio value is obtained by the ratio of the start time of a blood pressure event segment to the duration of a single-beat blood pressure segment, with a value ranging from 0 to 1. When the phase ratio difference between two phase space state vectors is less than 0.10, a connection is established between the corresponding nodes to construct the phase space adjacency graph.
[0110] In this embodiment, step five specifically includes:
[0111] Input the blood pressure event graph and phase space adjacency graph into the improved MTAD-GAT model;
[0112] The improved MTAD-GAT model sets up a dual-graph embedding module at the input end, which performs linear mapping processing on the node attribute matrix in the blood pressure event graph and the phase space state vector matrix in the phase space adjacency graph respectively, to obtain the event node embedding representation and the phase space node embedding representation.
[0113] At the node level, the event node embedding representation and the phase space node embedding representation are concatenated according to their temporal correspondence, and then mapped to a unified dimension through a linear transformation to obtain a joint node representation;
[0114] The joint node representation is input into the backbone structure of the improved MTAD-GAT model. The backbone structure includes a phase transition graph attention branch, a cross-pulsation phase evolution graph attention branch, and a phase space reconstruction attention branch.
[0115] The phase transition graph attention branch uses the phase transition connection relationship in the blood pressure event graph as the edge constraint. It performs graph attention calculation on the joint node representations corresponding to adjacent blood pressure event segments within the same single-beat blood pressure segment, and performs weighted aggregation of adjacent node representations according to attention weights to obtain the phase transition representation.
[0116] The attention branch of the cross-pulse phase evolution graph uses the same phase connection relationship in the blood pressure event graph as the edge constraint. It performs graph attention calculation on the joint node representations corresponding to blood pressure event segments in the same phase interval in adjacent single-pulse blood pressure segments, and performs weighted aggregation on the adjacent node representations according to the attention weight to obtain the cross-pulse phase evolution representation.
[0117] The phase space reconstruction attention branch takes the phase space state vector matrix and the phase space adjacency graph as input, and includes:
[0118] Perform a linear transformation on the phase space state vector matrix to obtain the query vector matrix, key vector matrix, and value vector matrix;
[0119] Extract the set of adjacent phase space state vectors corresponding to each target phase space state vector based on the phase space adjacency graph;
[0120] Calculate the state vector distance between the target phase space state vector and the adjacent phase space state vector, and the difference between the phase ratio value corresponding to the target phase space state vector and the phase ratio value corresponding to the adjacent phase space state vector.
[0121] The basic attention score is obtained by performing a dot product operation between the query vector and the key vector. The phase space reconstruction offset term is obtained based on the distance value of the state vector and the phase ratio difference. The phase space reconstruction offset term is added to the basic attention score to obtain the reconstruction attention score.
[0122] Normalization is performed on the reconstructed attention score to obtain the phase space attention weight, and the value vectors corresponding to the adjacent phase space state vectors are weighted and aggregated according to the phase space attention weight to obtain the phase space reconstruction representation;
[0123] The phase transition representation, the cross-pulsation phase evolution representation, and the phase space reconstruction representation are concatenated and processed by linear mapping to obtain the pulsation state representation vector.
[0124] In this implementation, the dual-graph embedding module of the improved MTAD-GAT model maps the blood pressure event graph node attribute matrix and the phase space state vector matrix to a unified feature space, with the embedding dimension set to 64–128. The main structure of the model consists of 3–6 stacked graph attention network layers, each containing 4–8 attention heads. The distance between phase space state vectors is calculated using Euclidean distance. The phase ratio is obtained by the ratio of the start time of the blood pressure event segment to the duration of a single-beat blood pressure segment, with a value ranging from 0 to 1. The phase space reconstruction offset is obtained by a linear combination of the state vector distance value and the phase ratio difference, where the distance weight coefficient is set to 0.5–0.8 and the phase weight coefficient is set to 0.2–0.5. The reconstruction attention score is obtained by summing the basic attention score and the phase space reconstruction offset, and then normalized using the Softmax function to obtain the phase space attention weight.
[0125] The improved MTAD-GAT model maintains the same overall network structure as the MTAD-GAT model, both employing a time-series anomaly detection framework based on graph attention networks. The model inputs are time-series feature representations, which are described through a graph structure to depict the relationships between features. A graph attention mechanism is used to weight and aggregate information from neighboring nodes. The model structure includes an input embedding layer, a graph attention computation layer, and a feature fusion layer. A multi-head attention mechanism is used to learn the representation of the associated features between nodes, thereby obtaining a high-dimensional node representation that can characterize changes in the time-series state for subsequent anomaly detection.
[0126] Based on the MTAD-GAT model, a phase space reconstruction attention mechanism is introduced, and a phase space adjacency graph is constructed. Specifically, delayed embedding is first performed on the blood pressure event feature sequence to construct a phase space state vector sequence, and a phase space adjacency graph is established based on the distance relationship between the state vectors. In the graph attention calculation process, a phase space reconstruction offset term is introduced on the basis of the original attention score calculation. The phase space reconstruction offset term is jointly determined by the distance value and the phase ratio difference between the phase space state vectors, and is superimposed on the basic attention score to form the phase space reconstruction attention weight.
[0127] By introducing a phase space reconstruction attention mechanism into the MTAD-GAT model, the model considers not only the topological connections between blood pressure events but also the evolution of blood pressure dynamic states in phase space when aggregating node information. This enhances the model's ability to express the nonlinear dynamic changes in blood pressure signals. This improvement enhances the model's ability to identify abnormal blood pressure fluctuation patterns, enabling it to obtain stable state representations even in complex blood pressure signal environments, thereby improving the accuracy and stability of blood pressure anomaly warnings.
[0128] In this embodiment, step six specifically includes:
[0129] An individual normal blood pressure state memory bank is established, which consists of the pulsation state representation vector corresponding to the historical single blood pressure segment and the corresponding blood pressure waveform segment, and is stored in chronological order.
[0130] The similarity is calculated between the pulsation state representation vector corresponding to the current single-beat blood pressure segment and the historical state representation vector in the individual's normal blood pressure state memory database to obtain a similarity sequence.
[0131] The historical state representation vectors are sorted according to their similarity, and a set of candidate normal states is formed by selecting the top-ranked historical state representation vectors based on their similarity.
[0132] Extract historical blood pressure waveform segments corresponding to the candidate normal state set, and perform weighted summation of the historical blood pressure waveform segments according to the similarity weights corresponding to the historical state representation vectors to obtain the counterfactual normal blood pressure waveform.
[0133] In this implementation, the individual normal blood pressure state memory bank is constructed from continuously collected historical blood pressure data, with the number of historical single-beat blood pressure segments stored in the memory bank set to 300-1000. The dimension of the pulse state representation vector is set to 64-128. Similarity calculation uses the cosine similarity method, obtaining the similarity value by dividing the dot product of two state representation vectors by the product of their magnitudes, with similarity values ranging from -1 to 1. Historical state representation vectors are sorted according to similarity, and the top 5-10 similarity vectors are selected as the candidate normal state set. The length of the corresponding historical blood pressure waveform segment is consistent with the length of the current single-beat blood pressure segment. The similarity weight is obtained by normalizing the candidate state similarity values, ensuring the sum of all weights equals 1, and then weighted summing the historical blood pressure waveform segments according to the normalized weights to obtain the counterfactual normal blood pressure waveform.
[0134] In this embodiment, step seven specifically includes:
[0135] Align the current blood pressure waveform with the counterfactual normal blood pressure waveform in chronological order, and establish a one-to-one waveform point pair at the corresponding sampling time position;
[0136] For each pair of waveform points, calculate the absolute value of the pressure difference and arrange them in chronological order to form a pressure deviation sequence;
[0137] The average value of all deviation values in the pressure deviation sequence is calculated, and the maximum deviation value in the pressure deviation sequence is extracted. The average deviation value and the maximum deviation value are weighted and summed according to the preset weight to form the waveform structure deviation.
[0138] Calculate the slope of pressure change between the current blood pressure waveform and the counterfactual normal blood pressure waveform at adjacent sampling points, take the absolute value of the slope difference at the corresponding sampling positions, and average all slope differences to form the waveform trend deviation.
[0139] The deviation of waveform structure and the deviation of waveform trend are weighted and summed according to preset weights to form the abnormal deviation.
[0140] The abnormal deviations corresponding to multiple consecutive single-beat blood pressure segments are accumulated in chronological order to form a cumulative abnormal deviation, and an abnormal score is calculated based on the cumulative abnormal deviation.
[0141] When the abnormal score is greater than the preset abnormal threshold, the blood pressure signal is judged as an abnormal state and a blood pressure abnormality warning result is generated.
[0142] In this implementation, the sampling frequency of the current blood pressure waveform is consistent with that of the counterfactual normal blood pressure waveform, set at 125Hz–250Hz, and the number of sampling points for a single-beat blood pressure segment is 80–300. In the waveform structure deviation calculation, the average deviation value is weighted at 0.6–0.8, and the maximum deviation value is weighted at 0.2–0.4. The pressure change slope is calculated by dividing the pressure difference between adjacent sampling points by the sampling time interval, which is 0.004s–0.008s. The waveform trend deviation is represented by the average of the absolute values of all slope differences. In the abnormal deviation calculation, the waveform structure deviation weight is set at 0.5–0.7, and the waveform trend deviation weight is set at 0.3–0.5. The cumulative number of consecutive single-beat blood pressure segments is set at 5–12. The abnormal score is calculated by normalizing the cumulative abnormal deviation; when the abnormal score is greater than 0.65–0.80, the blood pressure signal is determined to be in an abnormal state.
[0143] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous blood pressure monitoring system in a hospital's cardiovascular monitoring center to perform anomaly early warning analysis on long-term ambulatory blood pressure monitoring data. The hospital deployed continuous blood pressure monitoring equipment in its intensive care unit and hypertension clinic to record patients' blood pressure changes in real time. In actual clinical environments, blood pressure signals are affected by various factors such as patient movement, sensor contact changes, equipment noise, and individual physiological differences, leading to unstable fluctuations, local abnormal abrupt changes, and periodic waveform distortions in blood pressure waveforms. Traditional blood pressure anomaly detection methods typically rely on threshold judgments based on blood pressure time series or simple statistical features for anomaly identification. These methods struggle to effectively distinguish between genuine abnormal fluctuations and short-term noise interference, and lack fine-grained analysis capabilities for changes in the internal structure of blood pressure waveforms, resulting in high false alarm rates and insufficient anomaly detection stability. Therefore, a method for anomaly early warning in continuous blood pressure monitoring systems is needed that can simultaneously characterize the structural changes and dynamic evolution of blood pressure waveforms to improve the accuracy and stability of blood pressure anomaly detection.
[0144] In this scenario, the method of the present invention first receives blood pressure signal data collected by a continuous blood pressure monitoring device. The monitoring device continuously samples the patient's blood pressure at a sampling frequency of 200 Hz, with each blood pressure signal sequence recorded for 30 minutes to 2 hours. The collected blood pressure time series is first processed by bandpass filtering and amplitude normalization to eliminate device noise and baseline drift. Subsequently, the trough positions in the blood pressure waveform are identified, and the blood pressure data between adjacent troughs is used as a complete single-beat blood pressure segment, thereby forming a continuous single-beat blood pressure segment sequence. In actual data, a patient generates approximately 60 to 90 single-beat blood pressure segments per minute, with each segment containing approximately 120 to 240 sampling points.
[0145] After obtaining the single-beat blood pressure segment sequence, each segment is decomposed into event segments. Key structural points in the blood pressure waveform are identified by calculating the pressure change slope between adjacent sampling points, including rapid rise inflection points, systolic peak points, fall inflection points, and diastolic plateau points. Based on these key points, each single-beat blood pressure segment is divided into multiple blood pressure event segments, and structural and transition features are extracted for each event segment, such as event duration, pressure amplitude, mean pressure value, and the relationship between adjacent events. Subsequently, a blood pressure event graph is constructed using blood pressure event segments as nodes, establishing phase transition connections within the same beat and similar phase connections between different beats, enabling the model to capture the structural change patterns of the blood pressure waveform.
[0146] After the blood pressure event graph is constructed, delayed embedding is performed on the blood pressure event feature sequences to concatenate the event feature vectors into a phase space state vector, thereby obtaining the dynamic evolution trajectory of the blood pressure signal in phase space. A phase space adjacency graph is then established based on the distance and phase ratio relationships between the phase space state vectors. Subsequently, the blood pressure event graph and the phase space adjacency graph are input into the improved MTAD-GAT model. Joint calculations are performed through a phase transfer attention branch, a cross-pulse phase evolution attention branch, and a phase space reconstruction attention branch to obtain the pulsation state representation vector corresponding to each single-beat blood pressure segment.
[0147] After obtaining the pulse state representation vector, this invention further establishes an individual normal blood pressure state memory bank. This memory bank consists of the patient's historical stable blood pressure state data, recording the state representation vectors and corresponding blood pressure waveforms corresponding to multiple historical single-pulse blood pressure segments. When a new blood pressure segment arrives, the system calculates the similarity between the current state representation vector and the historical normal state vector, and selects several historical states with the highest similarity as reference samples. Based on these reference samples, a counterfactual normal blood pressure waveform is reconstructed, which is the normal blood pressure waveform that should theoretically appear under the current physiological conditions. Subsequently, the current blood pressure waveform and the counterfactual normal blood pressure waveform are aligned point by point. By calculating the pressure deviation, structural deviation, and trend deviation, the abnormal deviation is obtained, and an abnormal score is calculated within multiple continuous pulse ranges, thereby completing the blood pressure abnormality warning.
[0148] To verify the beneficial effects of this invention, data from 120 patients in a hospital's intensive care unit continuous blood pressure monitoring database were selected as experimental data. Approximately 78% of these patients had normal blood pressure, while about 22% exhibited varying degrees of blood pressure abnormalities. The method of this invention was compared with three common methods: a traditional statistical threshold detection method, an LSTM-based time-series anomaly detection method, and a blood pressure anomaly detection method based on a standard GAT model. The experimental results are shown in Table 1.
[0149] Table 1. Performance Comparison of Different Blood Pressure Abnormality Detection Methods
[0150] Comparison indicators Statistical threshold detection method LSTM time series detection method GAT Blood Pressure Abnormality Detection Method Method of the present invention Number of data samples 12000 12000 12000 12000 Anomaly detection accuracy (%) 78.6 85.3 89.8 94.6 False alarm rate (%) 18.2 12.7 9.4 5.8 Average warning lead time (s) 3.1 4.6 5.2 7.4 Stability rating 0.62 0.71 0.79 0.88
[0151] As shown in Table 1, the method of this invention achieves an anomaly identification accuracy of 94.6%, which is approximately 16% higher than traditional statistical thresholding methods, approximately 9% higher than LSTM methods, and approximately 4.8% higher than ordinary GAT models. Regarding the false alarm rate, the method of this invention reduces it to 5.8%, significantly lower than other methods. Furthermore, in terms of anomaly warning lead time, the method of this invention can identify abnormal blood pressure fluctuations on average 7.4 seconds in advance, demonstrating better warning capabilities compared to traditional methods. In addition, due to the introduction of a phase space reconstruction attention mechanism and an individual normal blood pressure state memory database, the method of this invention has stronger adaptability among different patients and exhibits higher detection stability in long-term monitoring scenarios.
[0152] The results of the above implementation show that the method of the present invention, by combining blood pressure event structure modeling, phase space dynamic modeling and the improved MTAD-GAT attention mechanism, effectively expresses the complex change patterns of blood pressure signals, and achieves more stable and reliable anomaly detection results by constructing counterfactual normal blood pressure waveforms, thereby significantly improving the overall performance of the blood pressure anomaly early warning system and verifying the feasibility and practical value of the method of the present invention in actual medical monitoring scenarios.
[0153] 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. An attention mechanism-based blood pressure signal anomaly warning method, characterized in that, Includes the following steps: Step 1: Acquire continuous blood pressure signals, preprocess the continuous blood pressure signals, and segment the cardiac cycle according to the position of the blood pressure waveform trough to obtain a single-beat blood pressure segment sequence; Step 2: Decompose the single-beat blood pressure fragment sequence into event fragments, extract blood pressure event fragments and corresponding event features, and construct a blood pressure event fragment sequence in chronological order; Step 3: Construct a blood pressure event graph using the blood pressure event segments as nodes and the event features as node attributes; establish phase transition relationships within single-beat blood pressure segments and establish similar phase connection relationships between adjacent single-beat blood pressure segments; Step 4: Perform delayed embedding processing on the blood pressure event fragment sequence to construct a phase space state vector sequence, and construct a phase space adjacency graph based on the distance relationship between the phase space state vectors; Step 5: Based on the blood pressure event graph and the phase space adjacency graph, perform joint graph attention calculation using the improved MTAD-GAT model that introduces a phase space reconstruction attention mechanism to obtain the pulsation state representation vector; Step 6: Establish an individual normal blood pressure state memory bank, calculate the similarity between the pulse state representation vector and the historical state representation vector, and determine the counterfactual normal blood pressure waveform based on the similarity; Step 7: Compare the current blood pressure waveform with the counterfactual normal blood pressure waveform to calculate the abnormal deviation, and calculate the abnormal score based on the abnormal deviation sequence to generate a blood pressure abnormality warning result.
2. The blood pressure signal abnormality early warning method based on attention mechanism according to claim 1, characterized in that, Step one specifically involves: The blood pressure of the subject is continuously sampled, and blood pressure time series data is obtained according to a preset sampling frequency. The blood pressure time series data consists of multiple pressure sampling points arranged in chronological order. The blood pressure time series data is subjected to signal preprocessing, which includes bandpass filtering of the blood pressure time series data and amplitude normalization of the filtered blood pressure time series data to form a preprocessed blood pressure sequence. In the preprocessed blood pressure sequence, the pressure values of adjacent sampling points are compared in chronological order. When the pressure value of a sampling point is less than both the pressure value of the previous sampling point and the pressure value of the next sampling point, the corresponding sampling point is determined as the blood pressure waveform trough point, and the time position corresponding to the blood pressure waveform trough point is recorded. The blood pressure data between two adjacent blood pressure waveform troughs are segmented according to time sequence, and the blood pressure data between two adjacent blood pressure waveform troughs is divided into a single beat blood pressure segment, which is then arranged in time sequence to form a single beat blood pressure segment sequence. Calculate the duration and pressure amplitude variation range of each single-beat blood pressure segment, and delete single-beat blood pressure segments that exceed the preset duration threshold range and preset pressure amplitude threshold range.
3. The blood pressure signal abnormality early warning method based on attention mechanism according to claim 1, characterized in that, Step two specifically involves: Local slope calculations were performed on the single-beat blood pressure segment sequence in chronological order. The pressure change slope sequence was obtained by calculating the ratio of the pressure difference to the time difference between adjacent sampling points. Based on the direction of change of the pressure change slope sequence, the single-beat blood pressure segment was divided into multiple candidate event segments. In single-beat blood pressure segments, key event points of blood pressure waveforms are identified based on the pressure change slope sequence. Key event points of blood pressure waveforms include the inflection point of rapid pressure rise, the systolic peak point, the inflection point of pressure fall, and the diastolic flat point. Blood pressure data between adjacent key event points of blood pressure waveforms are segmented according to time sequence to form multiple blood pressure event segments. For each blood pressure event segment, segment structure features are calculated, including event duration, pressure amplitude formed by the difference between the maximum and minimum pressure values, average pressure value of the event segment, and average absolute value of the pressure change slope. Segment transition features are calculated for adjacent blood pressure event segments. The segment transition features include the time interval between adjacent event segments, the difference in pressure amplitude, and the ratio of the difference in the average slope of adjacent event segments to the time interval. Phase structure features are calculated for each blood pressure event segment. The phase structure features include the ratio of the event segment start time to the duration of the single-beat blood pressure segment, and the ratio of the event segment duration to the duration of the single-beat blood pressure segment. The fragment structure features, fragment transition features, and phase structure features are combined to form a blood pressure event feature vector, and then arranged according to the chronological order of the occurrence of blood pressure event fragments to construct a blood pressure event fragment sequence.
4. The blood pressure signal abnormality early warning method based on attention mechanism according to claim 1, characterized in that, Step three specifically involves: Each blood pressure event segment in the blood pressure event segment sequence is defined as a node in the blood pressure event graph according to the time sequence, and the blood pressure event feature vector is used as the node attribute. The blood pressure event feature vector is composed of segment structure features, segment transition features and phase structure features. Within a single-beat blood pressure segment, phase transition relationships between nodes are established according to the temporal order of blood pressure event segments. Two temporally adjacent blood pressure event segment nodes are connected by directed edges to form a phase transition connection structure within the single-beat blood pressure segment. Based on the phase ratio value in the phase structure characteristics, blood pressure event segment nodes in the same phase interval in different single-beat blood pressure segments are compared, and when the phase ratio difference is less than a preset threshold, a cross-pulse connection relationship is established between the corresponding nodes to form a similar phase connection structure. The phase transition connection structure is merged with similar phase connection structures to form the edge set of the blood pressure event graph, and the adjacency matrix of the blood pressure event graph is generated based on the connection relationship between nodes. Generate the blood pressure event graph data structure based on the adjacency matrix and node attributes of the blood pressure event graph.
5. The blood pressure signal abnormality early warning method based on attention mechanism according to claim 1, characterized in that, Step four specifically involves: Delayed embedding processing is performed on the blood pressure event feature vectors in the blood pressure event segment sequence. Multiple blood pressure event feature vectors at different time intervals are selected according to a preset time delay interval and then concatenated in chronological order to form a phase space state vector. Arrange all phase space state vectors in chronological order to form a phase space state vector sequence; Calculate the distance between any two phase space state vectors in the phase space state vector sequence, and calculate the phase ratio of the phase space state vectors during the duration of a single beat blood pressure segment; The phase space state vectors are sorted by distance value from smallest to largest, and a preset number of nearest neighbor phase space state vectors are selected to form candidate neighborhoods. The phase ratio values of the phase space state vectors in the candidate neighborhood are compared. When the phase ratio difference is less than the preset phase threshold, an undirected edge is established between the corresponding phase space state vectors, and a phase space adjacency graph is constructed based on all the undirected edges.
6. The blood pressure signal abnormality early warning method based on attention mechanism according to claim 1, characterized in that, Step five specifically involves: Input the blood pressure event graph and phase space adjacency graph into the improved MTAD-GAT model; The improved MTAD-GAT model sets up a dual-graph embedding module at the input end, and performs linear mapping processing on the node attribute matrix in the blood pressure event graph and the phase space state vector matrix in the phase space adjacency graph respectively to obtain the event node embedding representation and the phase space node embedding representation. At the node level, the event node embedding representation and the phase space node embedding representation are concatenated according to their temporal correspondence, and then mapped to a unified dimension through a linear transformation to obtain a joint node representation; The joint node representation is input into the backbone structure of the improved MTAD-GAT model. The backbone structure includes a phase transition graph attention branch, a cross-pulsation phase evolution graph attention branch, and a phase space reconstruction attention branch. The phase transition graph attention branch uses the phase transition connection relationship in the blood pressure event graph as the edge constraint. It performs graph attention calculation on the joint node representations corresponding to adjacent blood pressure event segments within the same single-beat blood pressure segment, and performs weighted aggregation of adjacent node representations according to attention weights to obtain the phase transition representation. The attention branch of the cross-pulse phase evolution graph uses the same phase connection relationship in the blood pressure event graph as the edge constraint. It performs graph attention calculation on the joint node representations corresponding to blood pressure event segments in the same phase interval in adjacent single-pulse blood pressure segments, and performs weighted aggregation on the adjacent node representations according to the attention weight to obtain the cross-pulse phase evolution representation. The phase space reconstruction attention branch takes the phase space state vector matrix and the phase space adjacency graph as input, and includes: Perform a linear transformation on the phase space state vector matrix to obtain the query vector matrix, key vector matrix, and value vector matrix; Extract the set of adjacent phase space state vectors corresponding to each target phase space state vector based on the phase space adjacency graph; Calculate the state vector distance between the target phase space state vector and the adjacent phase space state vector, and the difference between the phase ratio value corresponding to the target phase space state vector and the phase ratio value corresponding to the adjacent phase space state vector. The basic attention score is obtained by performing a dot product operation between the query vector and the key vector. The phase space reconstruction offset term is obtained based on the distance value of the state vector and the phase ratio difference. The phase space reconstruction offset term is added to the basic attention score to obtain the reconstruction attention score. Normalization is performed on the reconstructed attention score to obtain the phase space attention weight, and the value vectors corresponding to the adjacent phase space state vectors are weighted and aggregated according to the phase space attention weight to obtain the phase space reconstruction representation; The phase transition representation, the cross-pulsation phase evolution representation, and the phase space reconstruction representation are concatenated and processed by linear mapping to obtain the pulsation state representation vector.
7. The blood pressure signal abnormality early warning method based on attention mechanism according to claim 1, characterized in that, Step six specifically involves: An individual normal blood pressure state memory bank is established, which consists of the pulsation state representation vector corresponding to the historical single blood pressure segment and the corresponding blood pressure waveform segment, and is stored in chronological order. The similarity is calculated between the pulsation state representation vector corresponding to the current single-beat blood pressure segment and the historical state representation vector in the individual's normal blood pressure state memory database to obtain a similarity sequence. The historical state representation vectors are sorted according to their similarity, and a set of candidate normal states is formed by selecting the top-ranked historical state representation vectors based on their similarity. Extract historical blood pressure waveform segments corresponding to the candidate normal state set, and perform weighted summation of the historical blood pressure waveform segments according to the similarity weights corresponding to the historical state representation vectors to obtain the counterfactual normal blood pressure waveform.
8. The blood pressure signal abnormality early warning method based on attention mechanism according to claim 1, characterized in that, Step seven specifically involves: Align the current blood pressure waveform with the counterfactual normal blood pressure waveform in chronological order, and establish a one-to-one waveform point pair at the corresponding sampling time position; For each pair of waveform points, calculate the absolute value of the pressure difference and arrange them in chronological order to form a pressure deviation sequence; The average value of all deviation values in the pressure deviation sequence is calculated, and the maximum deviation value in the pressure deviation sequence is extracted. The average deviation value and the maximum deviation value are weighted and summed according to the preset weight to form the waveform structure deviation. Calculate the slope of pressure change between the current blood pressure waveform and the counterfactual normal blood pressure waveform at adjacent sampling points, take the absolute value of the slope difference at the corresponding sampling positions, and average all slope differences to form the waveform trend deviation. The deviation of waveform structure and the deviation of waveform trend are weighted and summed according to preset weights to form the abnormal deviation. The abnormal deviations corresponding to multiple consecutive single-beat blood pressure segments are accumulated in chronological order to form a cumulative abnormal deviation, and an abnormal score is calculated based on the cumulative abnormal deviation. When the abnormal score is greater than the preset abnormal threshold, the blood pressure signal is judged as an abnormal state and a blood pressure abnormality warning result is generated.
