A blood pressure trend analysis method, system and device based on time segment reconstruction

Abstract

This application discloses a method, system, and apparatus for blood pressure trend analysis based on time segment reconstruction, used to achieve quantitative analysis of blood pressure trends across multiple time dimensions. The method includes: acquiring blood pressure time-series data and removing artifacts from the data; segmenting the data to obtain multiple time segments with different time resolutions and extracting segment feature vectors; calculating the temporal similarity between segment feature vectors using a preset dynamic time warping algorithm and dividing them into several pattern clusters using a preset density clustering algorithm; constructing a directed dynamic weighted graph using time segments as nodes and temporal similarity and the similarity between pattern clusters as weighting factors; searching for the optimal path using a preset search algorithm and reconstructing the data to obtain a continuous blood pressure sequence; and performing linear fitting on the continuous blood pressure sequence within analysis windows of different time lengths to generate blood pressure trend indicators for different time lengths.

Description

Technical Field

[0001] This application relates to the field of blood pressure analysis technology, and in particular to a method, system and device for blood pressure trend analysis based on time segment reconstruction. Background Technology

[0002] As the healthcare field upgrades towards precision and intelligence, ambulatory blood pressure monitoring has become an important tool for hypertension diagnosis, cardiovascular disease risk assessment, and chronic disease management. Blood pressure, as a key physiological indicator reflecting the function of the human circulatory system, not only provides crucial information for clinical assessment of conditions and adjustment of treatment plans, but also helps healthy individuals anticipate potential cardiovascular risks. Therefore, blood pressure trend analysis technology is increasingly widely used in clinical medicine and family health management.

[0003] In existing technologies, a fixed-time-dimensional analysis method is typically used to analyze blood pressure trends. Specifically, after acquiring blood pressure data, the data is first preprocessed to remove obvious abnormal interference. Then, feature indicators of the fixed-time dimension are extracted from the processed blood pressure data, and trend calculations are performed using an analytical model to generate trend analysis results for the corresponding dimension.

[0004] However, since blood pressure changes exhibit different characteristics and patterns at different time dimensions, this fixed time dimension can often only reflect the blood pressure situation at a single time dimension, and cannot fully cover various change characteristics. It lacks multi-dimensional quantitative analysis capabilities and is difficult to meet the differentiated needs for blood pressure trend analysis and the need for higher precision trend analysis in different scenarios. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a method, system, and apparatus for blood pressure trend analysis based on time segment reconstruction.

[0006] The technical solution provided in this application is described below:

[0007] The first aspect of this application provides a blood pressure trend analysis method based on time segment reconstruction, the blood pressure trend analysis method comprising: Acquire time-series blood pressure data including systolic blood pressure, diastolic blood pressure and corresponding timestamps, and remove artifacts from the blood pressure time-series data; The blood pressure time series data after artifact removal is segmented using a preset multi-scale sliding window to obtain multiple time segments with different time resolutions, and the segment feature vectors in each time segment are extracted by a preset graph node embedding generation model. The temporal similarity between each segment feature vector is calculated using a preset dynamic time warping algorithm, and the segment feature vectors are divided into several pattern clusters based on the temporal similarity using a preset density clustering algorithm. Using the time segments as nodes and the temporal similarity and the similarity between each pattern cluster as weighting factors for the nodes, a directed dynamic weight graph is constructed. A preset search algorithm is used to search for the optimal path in the directed dynamic weight graph, and the time segment is reconstructed based on the optimal path to obtain a continuous blood pressure sequence; The continuous blood pressure sequence is linearly fitted within analysis windows of different time lengths, and blood pressure trend indicators for different time lengths are generated based on the trend slope obtained from the linear fitting.

[0008] Optionally, the preset graph node embedding generation model includes a feature extraction layer, a pooling layer, and a fully connected layer; The step of extracting segment feature vectors from each time segment using a preset graph node embedding generation model includes: The time segment is input into the feature extraction layer, and local temporal feature maps are extracted from the time segment through the feature extraction layer. The local temporal feature maps are then sent to the pooling layer. The local temporal feature map is selected according to the local saliency screening mechanism of the pooling layer, and the salient features obtained after feature selection are sent to the fully connected layer. The fully connected layer maps the salient features from the high-dimensional space to the low-dimensional vector space to obtain the fragment feature vector.

[0009] Optionally, the multiple time segments with different time resolutions include instantaneous time segments, short-time time segments, medium-time time segments, and long-time time segments; Before inputting the time segment into the feature extraction layer, the blood pressure trend analysis method further includes: The instantaneous time segment, the short-time time segment, the medium-time time segment, and the long-time time segment are subjected to time sequence alignment preprocessing. The time-series alignment preprocessing for the instantaneous time segment, the short-time time segment, the medium-time time segment, and the long-time time segment includes: The missing values ​​of the instantaneous time segment are processed by the neighborhood data completion method, and the missing values ​​of the short time segment, the medium time segment and the long time segment are processed by the linear time series interpolation method. Using the processed long-duration time segment's time axis as a reference axis, the time axes of the short-duration time segment, the medium-duration time segment, and the instantaneous time segment are time-series scaled and aligned based on the reference axis.

[0010] Optionally, the step of inputting the time segment into the feature extraction layer and extracting local temporal feature maps from the time segment through the feature extraction layer includes: The time segment is input into the temporal convolutional attention module of the feature extraction layer; The local neighborhood features of the time segment are extracted using the temporal convolution kernel built into the temporal convolutional attention module; The local neighborhood features are weighted according to the temporal attention mechanism in the temporal convolutional attention module to obtain a local temporal feature map.

[0011] Optionally, the step of performing feature selection on the local temporal feature map according to the local saliency screening mechanism of the pooling layer, and sending the salient features obtained after feature selection to the fully connected layer, includes: The temporal saliency score of each feature point in the local temporal feature map is calculated based on the local saliency screening mechanism of the pooling layer. Feature points whose temporal significance scores are higher than a preset score threshold are selected as a set of highly significant feature points; The set of highly salient feature points is fused, and the resulting salient features are then fed to the fully connected layer.

[0012] Optionally, the preset multi-scale sliding window includes a preset instantaneous scale sliding window, a preset short-time scale sliding window, a preset medium-time scale sliding window, and a preset long-time scale sliding window; The blood pressure time-series data after artifact removal is segmented using a preset multi-scale sliding window to obtain multiple time segments with different time resolutions, including: The blood pressure time series data after artifact removal is traversed along the time axis using preset instantaneous scale sliding windows, preset short-time scale sliding windows, preset medium-time scale sliding windows, and preset long-time scale sliding windows, respectively. The blood pressure time series data in the traversal is truncated according to the window size and sliding step size of the sliding window at each scale to obtain the blood pressure data subsequence corresponding to each scale. The blood pressure data subsequences corresponding to each scale are timestamped and then encapsulated according to the traversal order of the sliding window at each scale to obtain multiple time segments with different time resolutions.

[0013] Optionally, the step of performing linear fitting on the continuous blood pressure sequence within analysis windows of different time lengths, and generating blood pressure trend indicators for different time lengths based on the trend slope obtained from the linear fitting, includes: Calculate the trend slope of the continuous blood pressure series corresponding to analysis windows of different time lengths; The trend slopes are sorted in chronological order, and each trend slope is classified according to a preset trend threshold to obtain a classified trend slope sequence. Based on the classified trend slope sequence, the blood pressure pattern, duration, and frequency of occurrence are statistically analyzed for each time length, and a blood pressure trend index is generated based on the blood pressure pattern, duration, and frequency of occurrence.

[0014] Optionally, the step of constructing a directed dynamic weighted graph using the time segment as a node and the temporal similarity and the similarity between each pattern cluster as weighting factors for the node includes: Each time segment is initialized as a node, and the temporal similarity between any two nodes is obtained. At the same time, the similarity between any two pattern clusters is calculated according to the pattern cluster to which each node belongs. The weighting factor between each node is calculated based on the temporal similarity and the similarity between the pattern clusters using a preset weighted fusion function. The weighting factor is the dynamic weight of the directed edge connecting two nodes. A directed dynamic weight graph is constructed based on the nodes and the corresponding weighting factors.

[0015] Optionally, after searching for the optimal path in the directed dynamic weight graph using a preset search algorithm, the blood pressure trend analysis method further includes: Based on the temporal sequence of each node in the directed dynamic weight graph, the temporal continuity index of the optimal path is calculated in combination with the preset path continuity condition. Determine whether the time continuity index is greater than or equal to a preset index threshold; If so, then the step of reconstructing the time segment based on the optimal path is performed; If not, the constraint parameters of the preset search algorithm are adjusted, and the step of searching for the optimal path in the directed dynamic weight graph is re-executed using the adjusted preset search algorithm.

[0016] A second aspect of this application provides a blood pressure trend analysis system based on time segment reconstruction, the blood pressure trend analysis system comprising: The removal unit is used to acquire blood pressure time-series data including systolic blood pressure, diastolic blood pressure and corresponding timestamps, and to remove artifacts from the blood pressure time-series data. The extraction unit is used to segment the blood pressure time series data after artifact removal using a preset multi-scale sliding window to obtain multiple time segments with different time resolutions, and to extract the segment feature vectors in each time segment through a preset graph node embedding generation model. The partitioning unit is used to calculate the temporal similarity between each segment feature vector using a preset dynamic time warping algorithm, and to partition the segment feature vectors into several pattern clusters based on the temporal similarity using a preset density clustering algorithm. A construction unit is used to construct a directed dynamic weighted graph with the time segment as a node and the temporal similarity and the similarity between each pattern cluster as weighting factors for the node. The reconstruction unit is used to search for the optimal path in the directed dynamic weight graph using a preset search algorithm, and to reconstruct the time segment based on the optimal path to obtain a continuous blood pressure sequence. The generation unit is used to perform linear fitting on the continuous blood pressure sequence within analysis windows of different time lengths, and generate blood pressure trend indicators for different time lengths based on the trend slope obtained from the linear fitting.

[0017] Optionally, the preset graph node embedding generation model includes a feature extraction layer, a pooling layer, and a fully connected layer; The step of extracting segment feature vectors from each time segment using a preset graph node embedding generation model includes: The time segment is input into the feature extraction layer, and local temporal feature maps are extracted from the time segment through the feature extraction layer. The local temporal feature maps are then sent to the pooling layer. The local temporal feature map is selected according to the local saliency screening mechanism of the pooling layer, and the salient features obtained after feature selection are sent to the fully connected layer. The fully connected layer maps the salient features from the high-dimensional space to the low-dimensional vector space to obtain the fragment feature vector.

[0018] Optionally, the multiple time segments with different time resolutions include instantaneous time segments, short-time time segments, medium-time time segments, and long-time time segments; Before inputting the time segment into the feature extraction layer, the blood pressure trend analysis method further includes: The instantaneous time segment, the short-time time segment, the medium-time time segment, and the long-time time segment are subjected to time sequence alignment preprocessing. The time-series alignment preprocessing for the instantaneous time segment, the short-time time segment, the medium-time time segment, and the long-time time segment includes: The missing values ​​of the instantaneous time segment are processed by the neighborhood data completion method, and the missing values ​​of the short time segment, the medium time segment and the long time segment are processed by the linear time series interpolation method. Using the processed long-duration time segment's time axis as a reference axis, the time axes of the short-duration time segment, the medium-duration time segment, and the instantaneous time segment are time-series scaled and aligned based on the reference axis.

[0019] Optionally, the extraction unit is specifically used for: The time segment is input into the temporal convolutional attention module of the feature extraction layer; The local neighborhood features of the time segment are extracted using the temporal convolution kernel built into the temporal convolutional attention module; The local neighborhood features are weighted according to the temporal attention mechanism in the temporal convolutional attention module to obtain a local temporal feature map.

[0020] Optionally, the extraction unit is specifically used for: The temporal saliency score of each feature point in the local temporal feature map is calculated based on the local saliency screening mechanism of the pooling layer. Feature points whose temporal significance scores are higher than a preset score threshold are selected as a set of highly significant feature points; The set of highly salient feature points is fused, and the resulting salient features are then fed to the fully connected layer.

[0021] Optionally, the preset multi-scale sliding window includes a preset instantaneous scale sliding window, a preset short-time scale sliding window, a preset medium-time scale sliding window, and a preset long-time scale sliding window; The blood pressure time-series data after artifact removal is segmented using a preset multi-scale sliding window to obtain multiple time segments with different time resolutions, including: The blood pressure time series data after artifact removal is traversed along the time axis using preset instantaneous scale sliding windows, preset short-time scale sliding windows, preset medium-time scale sliding windows, and preset long-time scale sliding windows, respectively. The blood pressure time series data in the traversal is truncated according to the window size and sliding step size of the sliding window at each scale to obtain the blood pressure data subsequence corresponding to each scale. The blood pressure data subsequences corresponding to each scale are timestamped and then encapsulated according to the traversal order of the sliding window at each scale to obtain multiple time segments with different time resolutions.

[0022] Optionally, the generation unit is specifically used for: Calculate the trend slope of the continuous blood pressure series corresponding to analysis windows of different time lengths; The trend slopes are sorted in chronological order, and each trend slope is classified according to a preset trend threshold to obtain a classified trend slope sequence. Based on the classified trend slope sequence, the blood pressure pattern, duration, and frequency of occurrence are statistically analyzed for each time length, and a blood pressure trend index is generated based on the blood pressure pattern, duration, and frequency of occurrence.

[0023] Optionally, the building unit is specifically used for: Each time segment is initialized as a node, and the temporal similarity between any two nodes is obtained. At the same time, the similarity between any two pattern clusters is calculated according to the pattern cluster to which each node belongs. The weighting factor between each node is calculated based on the temporal similarity and the similarity between the pattern clusters using a preset weighted fusion function. The weighting factor is the dynamic weight of the directed edge connecting two nodes. A directed dynamic weight graph is constructed based on the nodes and the corresponding weighting factors.

[0024] Optionally, a judgment unit may also be included, specifically used for: Based on the temporal sequence of each node in the directed dynamic weight graph, the temporal continuity index of the optimal path is calculated in combination with the preset path continuity condition. Determine whether the time continuity index is greater than or equal to a preset index threshold; If so, then the step of reconstructing the time segment based on the optimal path is performed; If not, the constraint parameters of the preset search algorithm are adjusted, and the step of searching for the optimal path in the directed dynamic weight graph is re-executed using the adjusted preset search algorithm.

[0025] A third aspect of this application provides a blood pressure trend analysis device based on time segment reconstruction, the blood pressure trend analysis device comprising: Processor, memory, input / output units, and bus; The processor is connected to the memory, the input / output unit, and the bus; The memory stores a program, which the processor invokes to execute the first aspect and any optional blood pressure trend analysis method within the first aspect.

[0026] The fourth aspect of this application provides a computer-readable storage medium storing a program that, when executed on a computer, performs the first aspect and any optional blood pressure trend analysis method of the first aspect.

[0027] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: 1. After acquiring blood pressure time series data, artifact removal was performed on the blood pressure time series data to ensure the validity of the data. Then, a preset multi-scale sliding window was used to segment the data to obtain time segments with different time resolutions, breaking through the limitations of a single fixed time dimension and providing a basis for multi-dimensional analysis. 2. Input each time segment into the preset graph node embedding generation model, and extract segment feature vectors through the preset graph node embedding generation model. This can not only accurately capture the subtle changes and potential patterns of blood pressure at different time scales, but also achieve the structured representation of features. 3. The temporal similarity between feature vectors is calculated by a preset dynamic time warping algorithm, and the pattern clusters are divided by density clustering algorithm. Then, a directed dynamic weight graph is constructed with time segments as nodes and temporal similarity and the similarity between each pattern cluster as weighting factors. This realizes the ordered association of multi-scale features and pattern classification, and solves the problem of feature confusion in different time dimensions. 4. Reconstruct continuous blood pressure sequences using a preset search algorithm, complete the temporal correlation between fragmented features, and then obtain blood pressure trend indicators at different time lengths through linear fitting of sliding windows of different lengths, so as to realize the quantitative output of blood pressure trends in multiple time dimensions and meet the differentiated and high-precision analysis needs of different scenarios. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 A schematic flowchart of an embodiment of the blood pressure trend analysis method provided in this application; Figure 2 A flowchart illustrating one implementation of step S102 in the blood pressure trend analysis method provided in this application; Figure 3 A flowchart illustrating another implementation of step S102 in the blood pressure trend analysis method provided in this application; Figure 4 A flowchart illustrating one implementation of step S104 in the blood pressure trend analysis method provided in this application; Figure 5 A flowchart illustrating an optional embodiment of the blood pressure trend analysis method provided in this application after step S105; Figure 6 A flowchart illustrating one implementation of step S106 in the blood pressure trend analysis method provided in this application; Figure 7 A schematic diagram of an embodiment of the blood pressure trend analysis system provided in this application; Figure 8 This is a schematic diagram of an embodiment of the blood pressure trend analysis device provided in this application. Detailed Implementation

[0030] This application provides a blood pressure trend analysis method based on time segment reconstruction, which can realize quantitative analysis of blood pressure trends in multiple time dimensions. In this embodiment, the execution subject of the blood pressure trend analysis method based on time segment reconstruction is not limited to a specific type of medical data processing device, health analysis terminal, or blood pressure monitoring and control system. The method can be executed by any hardware, software, or hardware / software combination system capable of blood pressure time series data acquisition and preprocessing, multi-scale feature extraction, algorithm model operation, and trend analysis calculation, such as medical-grade health monitoring terminals, general-purpose computers, distributed medical data processing servers, edge computing nodes, cloud computing health analysis platforms, and the built-in processing unit of intelligent blood pressure monitors.

[0031] The blood pressure trend analysis method based on time segment reconstruction described in this embodiment can also be implemented through the embedded processing program built into a wearable blood pressure monitoring device, the software module of a medical data analysis system, or by controlling a remote data processing unit through a medical-specific communication interface. Regardless of whether the specific executing entity is a single local processing terminal directly connected to the blood pressure acquisition device, multiple parallel collaborative functional units, or a scalable distributed medical and health data analysis system, this method can be operated according to the step sequence and logic described in the embodiments below.

[0032] Blood pressure time series data refers to a collection of blood pressure-related data collected and recorded in chronological order, including instantaneous values ​​of systolic and diastolic blood pressure and the corresponding data collection timestamps.

[0033] Artifact removal: refers to the preprocessing operation of identifying and deleting interference data that is not the true change in blood pressure in blood pressure time series data. Its main function is to remove invalid interference data, ensure the authenticity and validity of blood pressure time series data, and avoid artifacts affecting the accuracy of subsequent analysis results.

[0034] Preset Dynamic Time Warping Algorithm (DTW): This refers to a pre-defined algorithm used to calculate the temporal similarity between two feature vector segments. Its main advantage is that it can adapt to feature vector segments of different lengths and time resolutions. By dynamically adjusting the time axis, it can accurately calculate the similarity of blood pressure change patterns corresponding to two feature vector segments, providing a basis for pattern clustering.

[0035] Preset Density Clustering Algorithm: This refers to a pre-defined algorithm that clusters feature vectors of segments based on temporal similarity. It groups feature vectors with high temporal similarity into a class based on the temporal similarity between feature vectors of segments, forming several pattern clusters. This enables the classification and summarization of blood pressure change patterns and solves the problem of feature confusion across different time dimensions.

[0036] Please see Figure 1 This application first provides an embodiment of a blood pressure trend analysis method based on time segment reconstruction, which includes: S101. Obtain blood pressure time-series data including systolic blood pressure, diastolic blood pressure and corresponding timestamps, and remove artifacts from the blood pressure time-series data. In this embodiment, blood pressure time-series data including systolic blood pressure value, diastolic blood pressure value and corresponding collection timestamp are collected to ensure that the collected data covers the complete monitoring period, such as the full-time data of 24-hour ambulatory blood pressure monitoring.

[0037] Because blood pressure time-series data contains artifacts that can severely interfere with subsequent multi-scale analysis and feature extraction, artifact removal is performed on the acquired blood pressure time-series data simultaneously with data collection. This process identifies abnormal artifacts caused by factors such as equipment malfunction or strenuous human movement, such as sudden spikes in systolic blood pressure exceeding 200 mmHg without a reasonable cause. After identification, these artifacts that do not reflect the true blood pressure status are precisely removed, ensuring the continuity and accuracy of the data remaining after removal.

[0038] S102. The blood pressure time series data after artifact removal is segmented by a preset multi-scale sliding window to obtain multiple time segments with different time resolutions, and the segment feature vectors in each time segment are extracted by a preset graph node embedding generation model. In this embodiment, multi-scale segmentation and feature extraction are used to overcome the limitations of a single time dimension and achieve structured representation of blood pressure data. Specifically, a preset multi-scale sliding window is used to slide and segment the blood pressure time series data after artifact removal to obtain time segments covering multiple scales, where each scale of the preset multi-scale sliding window corresponds to a different time resolution.

[0039] Subsequently, all time segments are input into a preset graph node embedding generation model. The spatiotemporal convolutional network in the preset graph node embedding generation model is used to capture the change features and temporal correlations within each time segment and transform them into structured segment feature vectors.

[0040] See Figure 2 The following provides a specific implementation of step S102. In this implementation, the preset graph node embedding generation model includes a feature extraction layer, a pooling layer, and a fully connected layer. Multiple time segments with different time resolutions include instantaneous time segments, short-term time segments, medium-term time segments, and long-term time segments. This implementation includes: S10211 Perform time-series alignment preprocessing on instantaneous time segments, short-time time segments, medium-time time segments, and long-time time segments; In this embodiment, by handling differential missing values ​​and aligning with the time axis benchmark, the time benchmark of multi-scale time segments is unified, clearing data obstacles for subsequent feature extraction. The differential missing value handling method adapts to the sampling characteristics of different segments, neighborhood completion preserves the fluctuation characteristics of instantaneous segments, linear interpolation ensures the trend continuity of long-term segments, and time axis scaling alignment fundamentally unifies the time benchmark of multi-scale segments.

[0041] Specifically, differentiated missing value processing is performed for segments with different time resolutions. For instantaneous time segments with high sampling frequency and sensitive to data fluctuations, a neighborhood data completion method is used, selecting the mean of the effective blood pressure data before and after the missing value for completion. For example, if a 1-minute data point is missing in an instantaneous time segment, the mean of systolic and diastolic blood pressure for the preceding and following 30 seconds is used to fill the missing value. For short-term, medium-term, and long-term time segments with relatively low sampling frequency, a linear time series interpolation method is used, fitting the blood pressure value of the missing period based on the effective data points on both sides of the missing value to ensure data continuity.

[0042] After handling missing values, the time axis of a long-term time segment is selected as the baseline axis. This baseline axis covers the entire monitoring period and its time dimension is adapted to long-term trend analysis. Then, based on this baseline axis, the time axes of short-term, medium-term, and instantaneous time segments are time-series scaled and aligned, and the time axes of non-baseline segments are stretched or compressed to ensure that the timestamps of all time segments correspond one-to-one and eliminate the time dimension differences caused by different resolutions.

[0043] S10212, Input the time segment into the temporal convolutional attention module of the feature extraction layer; After preprocessing, the time segments aligned with the time sequence are input into the temporal convolutional attention module of the feature extraction layer in the preset graph node embedding generation model. The temporal convolutional attention module can accurately capture the temporal correlation of the data.

[0044] Specifically, the four types of time segments that have completed temporal alignment need to be organized into tensor data structures that can be recognized by the temporal convolutional attention module. Each tensor contains systolic and diastolic blood pressure values ​​and corresponding alignment timestamp information. Then, the tensor data of the four types of time segments are input in parallel into the temporal convolutional attention module of the feature extraction layer to ensure that the temporal convolutional attention module can receive multi-scale time segment data simultaneously.

[0045] S10213. Extract local neighborhood features of time segments using the temporal convolution kernel built into the temporal convolution attention module; In this embodiment, the temporal convolutional attention module has built-in temporal convolutional kernels of various sizes, such as 1×3, 1×5, and 1×7 one-dimensional convolutional kernels.

[0046] In practical applications, these convolutional kernels are controlled to slide along the time dimension on the input time segment. During the sliding process, each convolutional kernel performs convolution operations on the local neighborhood data within its coverage area, capturing the correlation features between adjacent sampling points. For example, a 1×3 convolutional kernel extracts the blood pressure fluctuation features of three consecutive sampling points in an instantaneous time segment, while a 1×7 convolutional kernel extracts the trend features of seven consecutive sampling points in a long-term time segment. After the capture is completed, the local neighborhood features extracted by different convolutional kernels are integrated to form a multi-dimensional set of local features.

[0047] S10214. Based on the temporal attention mechanism in the temporal convolutional attention module, the local neighborhood features are weighted to obtain a local temporal feature map, and the local temporal feature map is sent to the pooling layer. In this embodiment, the temporal attention mechanism in the temporal convolutional attention module can analyze the importance of local neighborhood features. For example, features of sudden increases or decreases in blood pressure are more important than features of steady fluctuations. In this way, important information can be highlighted and redundant information can be suppressed through weight allocation.

[0048] In practical applications, attention weights are calculated for each local neighborhood feature based on its importance. More important local neighborhood features are assigned greater weights. For example, the feature weight for a sudden spike in blood pressure from 120 mmHg to 160 mmHg in a short time segment is higher than that for a steady fluctuation from 120-125 mmHg. Subsequently, each local neighborhood feature is multiplied by its corresponding weight to complete the weighted fusion process, generating a local temporal feature map, which is then fed into the pooling layer.

[0049] S10215. Calculate the temporal saliency score of each feature point in the local temporal feature map based on the local saliency screening mechanism of the pooling layer; In this embodiment, the pooling layer receives local temporal feature maps from the temporal convolutional attention module. Based on the local saliency screening mechanism of the pooling layer, the saliency of each feature point in the local temporal feature map is analyzed from three dimensions. The first dimension is the magnitude of blood pressure change corresponding to the feature point, and the larger the magnitude, the higher the saliency. The second dimension is the frequency of change of the feature point, and sudden fluctuations are more significant than stable features. The third dimension is the degree of difference between the feature point and the surrounding neighborhood features, and the greater the difference, the higher the saliency.

[0050] After the analysis is completed, the results of the analysis in three dimensions are combined to generate a time-series significance score of 0-100 for each feature point. This time-series significance score can objectively quantify the importance of the feature point and provide a quantitative basis for subsequent analysis.

[0051] S10216. Select feature points with temporal significance scores higher than a preset score threshold as a set of highly significant feature points; In this embodiment, all feature points in the local temporal feature map are traversed, and the temporal saliency score of each feature point is extracted one by one. Then, the temporal saliency score of each feature point is compared with the corresponding preset score threshold. Feature points with temporal saliency scores higher than the preset score threshold are marked as high saliency feature points. All marked high saliency feature points are integrated to obtain a set of high saliency feature points.

[0052] S10217. Perform feature fusion on the set of highly significant feature points and send the fused significant features to the fully connected layer.

[0053] In this embodiment, a weighted splicing fusion method is adopted. Based on the analysis weight of different time segments, the feature points in the set of highly significant feature points are spliced ​​in chronological order. At the same time, higher fusion weights are assigned to key feature points. For example, the blood pressure peak feature weight of an instantaneous time segment is twice that of ordinary features.

[0054] Subsequently, the spliced ​​feature data is normalized to eliminate numerical differences between different feature dimensions, generating a salient feature tensor, which is then fed into a fully connected layer for subsequent low-dimensional mapping.

[0055] S10218. The salient features are mapped from the high-dimensional space to the low-dimensional vector space through the fully connected layer to obtain the segment feature vector.

[0056] The multi-layer transformation of fully connected layers can achieve dimensionality reduction and structured representation of features, simplifying the data structure while retaining key information. Therefore, in this embodiment, the high-dimensional salient features are mapped to low-dimensional fragment feature vectors through fully connected layers to achieve structured representation of features.

[0057] Specifically, the system first receives a salient feature tensor from the pooling layer, which contains key multi-scale blood pressure features. Then, the first layer of neurons in the fully connected layer performs a linear transformation on the salient feature tensor, mapping the features to an intermediate-dimensional space. The non-linear activation function of the fully connected layer then enhances the non-linear expressive power of the features. Next, multiple layers of neurons in the fully connected layer perform successive transformations, gradually compressing the feature dimensions. For example, the 256-dimensional salient feature tensor is compressed into a 32-dimensional low-dimensional vector, which is then output as the segment feature vector representing the core features of the time segment.

[0058] See Figure 3 The following provides a specific implementation method for another step S102, in which the preset multi-scale sliding window includes a preset instantaneous scale sliding window, a preset short-time scale sliding window, a preset medium-time scale sliding window, and a preset long-time scale sliding window. This implementation method includes: S10221. Use preset instantaneous scale sliding window, preset short time scale sliding window, preset medium time scale sliding window and preset long time scale sliding window respectively to slide along the time axis of blood pressure time series data after artifact removal. In this embodiment, four different scale sliding windows are used in parallel to comprehensively cover the time-series blood pressure data after artifact removal. Different scale sliding windows correspond to different time resolutions, and parallel sliding traversal can simultaneously capture the instantaneous fluctuations, short-term changes, medium-term trends, and long-term patterns of blood pressure, providing data support for subsequent multi-dimensional analysis.

[0059] Specifically, the preset multi-scale sliding windows are instantaneous scale sliding window, short-time scale sliding window, medium-time scale sliding window and long-time scale sliding window. These four types of windows are simultaneously aligned with the start timestamp of the blood pressure time series data, and are traversed synchronously along the time direction of the data with the time axis as the reference, ensuring that each sliding window can cover the complete blood pressure monitoring period during the sliding process and that no data is missed.

[0060] It is important to note that the stability of the sliding window needs to be maintained during the traversal process to avoid data truncation deviations caused by window offset.

[0061] S10222. Based on the window size and sliding step size of each scale sliding window, the blood pressure time series data in the traversal is truncated to obtain the blood pressure data subsequence corresponding to each scale. In this embodiment, blood pressure data subsequences of corresponding resolutions are extracted according to the window size and sliding step length of each scale sliding window. For example, the instantaneous scale sliding window size is preset to 1 minute and the sliding step length is preset to 30 seconds; the short-term scale sliding window size is preset to 5 minutes and the sliding step length is preset to 1 minute; the medium-term scale sliding window size is preset to 15 minutes and the sliding step length is preset to 5 minutes; and the long-term scale sliding window size is preset to 30 minutes and the sliding step length is preset to 10 minutes.

[0062] During the sliding window traversal at various scales, when the sliding window moves to a certain position, all blood pressure data within that time period are extracted according to the built-in window size, including systolic and diastolic blood pressure values ​​and corresponding timestamps. Each window completes one extraction operation and generates a blood pressure data subsequence at the corresponding scale for each sliding step, until all scales of the sliding window have been extracted. Subsequently, all extracted blood pressure data subsequences are labeled, clearly defining the scale type and corresponding time interval of each subsequence to ensure traceability.

[0063] S10223. The blood pressure data subsequences corresponding to each scale are timestamped and then encapsulated according to the traversal order of the sliding window at each scale to obtain multiple time segments with different time resolutions.

[0064] In this embodiment, structured multi-resolution time segments are generated through timestamp calibration and sequence encapsulation. Timestamp calibration eliminates the temporal bias of subsequences at different scales, while sequential encapsulation preserves the temporal logic of the blood pressure data.

[0065] Specifically, the timestamp of a preset instantaneous scale sliding window is first selected as the reference time axis. The timestamps of the blood pressure data subsequences extracted by the preset short-time scale sliding window, the preset medium-time scale sliding window, and the preset long-time scale sliding window are calibrated. For blood pressure data subsequences with mismatched timestamps, linear interpolation is used to supplement the blood pressure data at missing time points, or data at repeated time points are merged by downsampling, to ensure that blood pressure data subsequences at all scales can be aligned to a unified time reference.

[0066] Furthermore, according to the traversal order of the sliding windows at each scale, the blood pressure data subsequences within the same time interval are integrated and encapsulated. Each encapsulated data packet contains multi-scale blood pressure data for the corresponding time interval. Finally, all encapsulated data packets are marked as time segments with different time resolutions.

[0067] S103. Calculate the temporal similarity between the feature vectors of each segment using a preset dynamic time warping algorithm, and divide the feature vectors of the segments into several pattern clusters based on temporal similarity using a preset density clustering algorithm. After extracting the feature vectors of each time segment, a preset dynamic time warping algorithm is called to calculate the temporal similarity between the feature vectors of each segment. The higher the temporal similarity, the closer the blood pressure change patterns are to the feature vectors of the segments.

[0068] Subsequently, all the calculated temporal similarities are input into a preset density clustering algorithm. The preset density clustering algorithm will divide the segment feature vectors into several pattern clusters according to the density distribution of temporal similarities. The blood pressure change patterns corresponding to the segment feature vectors within the same pattern cluster are highly similar.

[0069] S104. Construct a directed dynamic weighted graph using time segments as nodes and temporal similarity and the similarity between each pattern cluster as weighting factors for the nodes. The node-based processing in this embodiment can integrate fragmented time segments, while the weighting factor quantifies the degree of correlation between segments. By constructing a directed dynamic weight graph through nodes and weighting factors, the ordered correlation of fragmented features is achieved.

[0070] Specifically, each time segment is treated as an independent node in a directed dynamic weighted graph. The collection time sequence corresponding to each node is defined. The calculated temporal similarity between segments and the similarity between each pattern cluster are used as weighting factors to connect nodes. The magnitude of the weighting factor represents the degree of correlation between nodes. The higher the similarity, the larger the weighting factor.

[0071] Furthermore, according to the acquisition order of time segments, directed connections are established between nodes to form a complete directed dynamic weight graph. Connections with higher weights in the directed dynamic weight graph represent closely related blood pressure change patterns in the corresponding time segments.

[0072] See Figure 4 The following provides a specific implementation method for step S104, which includes: S1041. Initialize each time segment as a node, obtain the temporal similarity between any two nodes, and calculate the similarity between any two pattern clusters according to the pattern cluster to which each node belongs. In this embodiment, each time segment with different time resolution is first initialized as an independent node in a directed dynamic weight graph. Each node is assigned a unique identifier and the pattern cluster information to which each node belongs is marked. The temporal similarity between any two nodes calculated by the dynamic time warping algorithm is retrieved. The temporal similarity is the correlation quantification result at the node individual level.

[0073] Furthermore, based on the cluster affiliation of each node, the similarity between any two pattern clusters is calculated. By statistically analyzing the temporal similarity between nodes within one cluster and all nodes within another cluster, the inter-cluster similarity is obtained, ensuring that the calculation of all pattern cluster pairs is covered.

[0074] S1042. Calculate the weighting factor between each node based on temporal similarity and the similarity between pattern clusters using a preset weighted fusion function. The weighting factor is the dynamic weight of the directed edge connecting two nodes. In this embodiment, the dual similarity is integrated through weighted fusion, and the weighting factor between nodes is calculated as the dynamic weight of the directed edge. The preset weighted fusion function can take into account the dual association characteristics of individual nodes and cluster levels. The dynamic weight reflects the difference in the fusion results of different node pairs, which conforms to the actual association pattern of blood pressure data. For example, the temporal similarity between node A and node B is 0.7 and the inter-cluster similarity is 0.6. After fusion, the weighting factor is obtained as the weight of the edge between the two.

[0075] Specifically, the temporal similarity between any two nodes and the similarity between the clusters to which these two nodes belong are retrieved. The two sets of data are matched one-to-one by node and used as input to a preset weighted fusion function. Subsequently, the preset weighted fusion function assigns appropriate fusion weights to the temporal similarity between nodes and the similarity between clusters according to the needs of blood pressure trend analysis. The temporal similarity between nodes is assigned a higher weight to highlight the actual correlation characteristics of individual nodes. Then, the two similarity indicators are fused and calculated according to the operation rules of the preset weighted fusion function to obtain the weighting factor corresponding to each node. This weighting factor is the dynamic weight of the directed edge connecting the two nodes.

[0076] S1043. Construct a directed dynamic weighted graph based on nodes and their corresponding weighting factors.

[0077] In this embodiment, a directed dynamic weighted graph is constructed based on nodes and dynamic weighting factors. The directionality ensures that the temporal order of blood pressure data is not disrupted, while the dynamic weights quantify the degree of correlation between nodes. The combination of these two aspects allows the graph model to accurately reflect the ordered correlation of fragmented time segments. For example, nodes for 6:00-6:05, 6:05-6:10, and 6:10-6:15 are sequentially directed according to time, and weighting factors of 0.75 and 0.82 are matched to each edge.

[0078] Specifically, all initialized independent nodes are used as the foundation for the graph construction. The blood pressure data collection time sequence corresponding to all nodes is arranged, pointing from nodes with earlier collection times to nodes with later collection times, ensuring that the directedness of the graph conforms to the blood pressure time sequence logic. Subsequently, the weighting factor of each node pair is matched one by one to the corresponding two nodes, serving as the dynamic weight of the directed edge connecting the node pair. The weight value is positively correlated with the degree of correlation between the nodes. Finally, all nodes with unique identifiers, directed edges pointing in chronological order, and the corresponding weighting factors of the edges are integrated to form a directed dynamic weight graph, which clearly presents the node distribution, edge direction, and weight differences.

[0079] S105. The optimal path is searched in the directed dynamic weight graph using a preset search algorithm, and the time segment is reconstructed based on the optimal path to obtain a continuous blood pressure sequence. In this embodiment, a preset search algorithm is used to reconstruct a continuous blood pressure sequence, completing the temporal correlation of fragmented features. The preset search algorithm can select the optimal time segment connection method in a complex directed dynamic weight graph, ensuring that the reconstructed sequence retains the characteristics of each time segment and conforms to the temporal logic of blood pressure changes. For example, if the optimal path of "morning-afternoon-evening" is found in the directed dynamic weight graph, the complete continuous blood pressure sequence can be reconstructed according to this order. This preset search algorithm can be a shortest path algorithm, which is not limited here.

[0080] Specifically, a pre-defined search algorithm is used, targeting the start and end nodes of a directed dynamic weight graph, to search for the path with the optimal total weight. The optimal path represents the time sequence combination with the strongest correlation between time segments and best reflects the actual variation pattern of blood pressure. Then, according to the node order on the optimal path, the corresponding time segments are sequentially connected to complete the temporal correlation between the segments and reconstruct a continuous blood pressure sequence.

[0081] See Figure 5 The following provides a specific implementation method after searching for the optimal path in step S105, which includes: S1051. Based on the temporal sequence of each node in the directed dynamic weight graph, and combined with the preset path continuity condition, calculate the temporal continuity index of the optimal path. In this embodiment, it is necessary to first retrieve all nodes on the optimal path in the directed dynamic weight graph, sort out the original collection time sequence of blood pressure data corresponding to each node, clearly mark the time connection relationship and time interval distribution between nodes, and then, based on the preset path continuity conditions, comprehensively quantify and evaluate the optimal path from dimensions such as node time connection gap, time sequence consistency, and time coverage integrity, and finally obtain the time continuity index that can characterize the time coherence of the path.

[0082] Furthermore, the preset path continuity conditions include requirements such as no gaps or repetitions in the time intervals of nodes, and strict arrangement according to the order of data collection. These are the basic standards for evaluating temporal continuity. The time continuity index transforms the temporal logical rationality of the path into a quantitative value, making up for the limitations of determining the optimal path solely by weights. It avoids paths with optimal weights but chaotic timing, such as optimal path nodes exhibiting a reversal of the time sequence from morning to night. The time continuity index can intuitively reflect this problem.

[0083] S1052. Determine whether the time continuity index is greater than or equal to the preset index threshold. In this embodiment, the temporal continuity of the optimal path is determined to meet the reconstruction requirements by comparing quantitative thresholds. A preset exponential threshold sets a clear benchmark for temporal continuity. Direct comparison of quantitative values ​​enables an objective determination of path effectiveness, avoiding biases from subjective human judgment.

[0084] Specifically, the calculated optimal path time continuity index is retrieved and compared with a preset index threshold. The comparison result is recorded. The preset index threshold is set according to the accuracy requirements of blood pressure trend analysis and is the critical standard for determining whether time continuity meets the standard. In clinical high-precision analysis scenarios, a higher threshold can be set, while in routine screening scenarios, the threshold can be appropriately lowered. For example, when the preset index threshold is set to 85, a time continuity index of 88 indicates that the time continuity index is greater than the preset index threshold, and a time continuity index of 80 indicates that the time continuity index is less than the preset index threshold.

[0085] When the time continuity index is greater than or equal to the preset index threshold, step S1053 is executed; when the time continuity index is less than the preset index threshold, step S1054 is executed.

[0086] S1053. Perform the step of reconstructing the time segment based on the optimal path; In this embodiment, step S1053 is similar to step S105 in the previous embodiment, and will not be described again here.

[0087] S1054. Adjust the constraint parameters of the preset search algorithm, and re-execute the step of searching for the optimal path in the directed dynamic weight graph using the adjusted preset search algorithm.

[0088] When the time continuity index is less than the preset index threshold, the constraint parameters of the preset search algorithm are adjusted accordingly, with a focus on strengthening the constraints related to time continuity. For example, the penalty weight of the preset search algorithm on the time connection gap of nodes is increased, the hard constraint that the node time sequence is strictly arranged according to the collection order is added, and the filtering preference based solely on weight is reduced.

[0089] After the parameters are adjusted, the updated constraint parameters are substituted into the original preset search algorithm, and the optimal path search is re-executed from the starting node to the ending node in the directed dynamic weight graph. The new search process will take into account both the weight correlation between nodes and the time continuity.

[0090] S106. Perform linear fitting on the continuous blood pressure sequence within analysis windows of different time lengths, and generate blood pressure trend indicators for different time lengths based on the trend slope obtained from the linear fitting.

[0091] After obtaining continuous blood pressure sequences, quantitative blood pressure trend indicators are generated through multi-window linear fitting to meet the analysis needs of different scenarios.

[0092] Specifically, the reconstructed continuous blood pressure sequence is placed into analysis windows of different time lengths. Linear fitting is performed on the data in each window. During the fitting process, the trend slope of the fitted line is calculated based on the blood pressure value and the corresponding timestamp. The trend slopes under different windows are then organized into corresponding blood pressure trend indicators. A positive trend slope indicates that the blood pressure is rising during that period, while a negative trend slope indicates a falling trend. The larger the absolute value of the trend slope, the more obvious the trend.

[0093] This embodiment, after acquiring blood pressure time-series data, removes artifacts to ensure data validity. It then uses a pre-set multi-scale sliding window to segment the data, obtaining time segments with different time resolutions, overcoming the limitations of a single fixed time dimension and providing a foundation for multi-dimensional analysis. Each time segment is input into a pre-set graph node embedding generation model, which extracts segment feature vectors. This accurately captures subtle changes and potential patterns in blood pressure at different time scales and achieves structured feature representation. A pre-set dynamic time warping algorithm calculates the temporal similarity between feature vectors, and density clustering is used to divide pattern clusters. A directed dynamic weight graph is then constructed using time segments as nodes and temporal similarity and the similarity between pattern clusters as weighting factors. This achieves ordered association and pattern classification of multi-scale features, solving the problem of feature confusion across different time dimensions. A pre-set search algorithm reconstructs continuous blood pressure sequences, completing the temporal associations between fragmented features. Finally, linear fitting using sliding windows of different lengths yields blood pressure trend indicators at different time lengths, achieving quantitative output of blood pressure trends across multiple time dimensions and meeting the differentiated and high-precision analysis needs of different scenarios.

[0094] See Figure 6 The following provides a specific implementation method for step S106, which includes: S1061. Calculate the trend slope of the continuous blood pressure series corresponding to analysis windows of different time lengths. In this embodiment, the trend slope of the reconstructed continuous blood pressure sequence is quantitatively calculated using analysis windows of different time lengths, achieving preliminary quantification of multi-dimensional blood pressure trends. The trend slope obtained from linear fitting can transform the temporal variation pattern of the continuous blood pressure sequence into a quantitative value. Analysis windows of different lengths can capture blood pressure trends in different time dimensions, and sliding traversal allows the trend of each interval within the monitoring period to be accurately quantified.

[0095] Specifically, based on the differentiated needs of blood pressure trend analysis, analysis windows of different time lengths can be set, which can be short-term, medium-term, and long-term dimensions, such as a 1-hour short-term window, a 4-hour medium-term window, and a 12-hour long-term window. Each analysis window independently corresponds to different time resolution analysis needs.

[0096] Furthermore, the reconstructed continuous blood pressure sequence is matched to analysis windows of various time lengths. Each analysis window slides along the time axis of the continuous blood pressure sequence to ensure that each analysis window completely covers the entire blood pressure monitoring period without missing any time periods. Simultaneously, for the blood pressure data covered during the sliding process of each analysis window, a linear fit is performed with monitoring time as the horizontal axis and blood pressure value as the vertical axis. The fitting results yield a trend slope that characterizes the direction and magnitude of blood pressure changes within that window. A positive trend slope indicates an upward trend in blood pressure, while a negative slope indicates a downward trend. The larger the absolute value, the more significant the trend in blood pressure change.

[0097] S1062. Sort the trend slopes in chronological order and classify each trend slope according to a preset trend threshold to obtain a classified trend slope sequence. After calculating the trend slope, the blood pressure monitoring times corresponding to each trend slope are arranged in an orderly manner, strictly preserving the temporal logic of the trend slope to ensure that the time trend of the slope sequence is consistent with the continuous blood pressure sequence. Subsequently, based on clinical blood pressure analysis standards and actual needs, a preset trend threshold is determined. This preset trend threshold defines the critical range of different trend states, which can be divided into three basic states: rising, stable, and falling. It can also be further refined into sub-states such as significantly rising, slightly rising, stable, slightly falling, and significantly falling.

[0098] Furthermore, each sorted trend slope is compared one by one with a preset trend threshold. The corresponding trend state is matched and marked according to the numerical range of the slope. The trend slopes with state labels are then integrated in their original chronological order to obtain a classified trend slope sequence. This time-ordered sorting method ensures the temporal continuity of the trend slopes, and the preset trend threshold provides a unified quantitative standard for state classification, making the classification results objective and standardized, and avoiding the bias of subjective human judgment.

[0099] S1063. Based on the classified trend slope sequence, statistically analyze the blood pressure pattern, duration, and frequency of occurrence for each time length, and generate blood pressure trend indicators based on the blood pressure pattern, duration, and frequency of occurrence.

[0100] In this embodiment, fragmented trend states are transformed into structured feature information by statistical patterns, duration, and frequency of occurrence. The multi-dimensional statistical results can comprehensively reflect the changing patterns of blood pressure, and the index-based presentation makes the analysis results more intuitive and easier to interpret.

[0101] Specifically, based on the trend slope sequences of each time length after classification, they are sorted and organized according to the time length of the analysis window. For each time length of the trend slope sequence, the slope segments with the same continuous trend state are identified and defined as the corresponding blood pressure patterns. For example, continuous "significantly rising" slope segments correspond to the "sudden rise in blood pressure in the morning" pattern, and continuous "stable" slope segments correspond to the "stable blood pressure in the afternoon" pattern.

[0102] Simultaneously, the duration and frequency of each identified blood pressure pattern were statistically analyzed. Duration represents the actual blood pressure monitoring time span covered by the pattern, and frequency represents the number of times the pattern occurred during the entire monitoring period. After statistical analysis, the blood pressure patterns, corresponding durations, and frequencies for each time period were integrated and analyzed. Following standardized formats for clinical application and health management, this information was consolidated into multi-dimensional blood pressure trend indicators. These indicators clearly define typical blood pressure change characteristics, durations, and patterns of occurrence across different time dimensions.

[0103] The blood pressure trend analysis system based on time segment reconstruction provided in this application is described in detail below. Please refer to [link / reference]. Figure 7 , Figure 7 An embodiment of the blood pressure trend analysis system provided in this application includes: The removal unit 701 is used to acquire blood pressure time-series data including systolic blood pressure, diastolic blood pressure and corresponding timestamps, and to remove artifacts from the blood pressure time-series data. Extraction unit 702 is used to cut the blood pressure time series data after artifact removal using a preset multi-scale sliding window to obtain multiple time segments with different time resolutions, and extract the segment feature vectors in each time segment through a preset graph node embedding generation model. The partitioning unit 703 is used to calculate the temporal similarity between the feature vectors of each segment using a preset dynamic time warping algorithm, and to divide the feature vectors of the segments into several pattern clusters based on temporal similarity using a preset density clustering algorithm. Construction unit 704 is used to construct a directed dynamic weighted graph with time segments as nodes and temporal similarity and the similarity between each pattern cluster as weighting factors. The reconstruction unit 705 is used to search for the optimal path in the directed dynamic weight graph using a preset search algorithm, and to reconstruct the time segment based on the optimal path to obtain a continuous blood pressure sequence. The generation unit 706 is used to perform linear fitting on the continuous blood pressure sequence within analysis windows of different time lengths, and generate blood pressure trend indicators for different time lengths based on the trend slope obtained from the linear fitting.

[0104] Optionally, the preset graph node embedding generation model includes a feature extraction layer, a pooling layer, and a fully connected layer; The feature vectors of each time segment are extracted by a pre-defined graph node embedding generation model, including: The time segment is input into the feature extraction layer, which extracts local temporal feature maps from the time segment and then feeds the local temporal feature maps into the pooling layer. The local saliency screening mechanism of the pooling layer is used to select features from the local temporal feature map, and the salient features obtained after feature selection are sent to the fully connected layer. By using a fully connected layer, salient features are mapped from a high-dimensional space to a low-dimensional vector space to obtain fragment feature vectors.

[0105] Optionally, multiple time segments with different time resolutions include instantaneous time segments, short time segments, medium time segments, and long time segments; Before inputting time segments into the feature extraction layer, the blood pressure trend analysis method also includes: Perform time-series alignment preprocessing on instantaneous time segments, short-time time segments, medium-time time segments, and long-time time segments; Timing alignment preprocessing is performed on instantaneous time segments, short-time time segments, medium-time time segments, and long-time time segments, including: The neighborhood data completion method is used to handle missing values ​​in instantaneous time segments, and the linear time series interpolation method is used to handle missing values ​​in short-time, medium-time, and long-time time segments. Using the time axis of the processed long time segment as the reference axis, the time axes of short time segments, medium time segments, and instantaneous time segments are time-series scaled and aligned based on the reference axis.

[0106] Optionally, the extraction unit 702 is specifically used for: Time segments are input into the temporal convolutional attention module of the feature extraction layer; Local neighborhood features of time segments are extracted using the built-in temporal convolution kernels in the temporal convolutional attention module; The local neighborhood features are weighted according to the temporal attention mechanism in the temporal convolutional attention module to obtain a local temporal feature map.

[0107] Optionally, the extraction unit 702 is specifically used for: The temporal saliency score of each feature point in the local temporal feature map is calculated based on the local saliency screening mechanism of the pooling layer; Feature points with temporal significance scores higher than a preset score threshold are selected as the set of highly significant feature points; Feature fusion is performed on the set of highly significant feature points, and the resulting significant features are then fed into the fully connected layer.

[0108] Optionally, the preset multi-scale sliding window includes a preset instantaneous scale sliding window, a preset short-time scale sliding window, a preset medium-time scale sliding window, and a preset long-time scale sliding window; A preset multi-scale sliding window was used to segment the artifact-removed blood pressure time-series data, resulting in multiple time segments with different temporal resolutions, including: The blood pressure time series data after artifact removal is traversed along the time axis using preset instantaneous scale sliding windows, preset short-time scale sliding windows, preset medium-time scale sliding windows, and preset long-time scale sliding windows, respectively. Based on the window size and sliding step size of each scale sliding window, the blood pressure time series data in the traversal is truncated to obtain the blood pressure data subsequence corresponding to each scale. The blood pressure data subsequences corresponding to each scale are timestamped and then encapsulated according to the traversal order of the sliding window at each scale to obtain multiple time segments with different time resolutions.

[0109] Optionally, the generating unit 706 is specifically used for: Calculate the trend slope of the continuous blood pressure series corresponding to analysis windows of different time lengths; The trend slopes are sorted in chronological order, and each trend slope is classified according to a preset trend threshold to obtain a classified trend slope sequence. Based on the statistical analysis of the trend slope sequence after classification, the blood pressure pattern, duration, and frequency of occurrence for each time period are obtained, and a blood pressure trend index is generated based on the blood pressure pattern, duration, and frequency of occurrence.

[0110] Optionally, building block 704 is specifically used for: Each time segment is initialized as a node, and the temporal similarity between any two nodes is obtained. At the same time, the similarity between any two pattern clusters is calculated based on the pattern cluster to which each node belongs. The weighting factor between each node is calculated based on temporal similarity and the similarity between pattern clusters using a preset weighted fusion function. The weighting factor is the dynamic weight of the directed edge connecting two nodes. A directed dynamic weighted graph is constructed based on nodes and their corresponding weighting factors.

[0111] Optionally, a judgment unit 707 is also included, specifically used for: Based on the temporal sequence of each node in the directed dynamic weight graph, the temporal continuity index of the optimal path is calculated in combination with the preset path continuity condition. Determine whether the time continuity index is greater than or equal to a preset index threshold; If so, then proceed with the step of reconstructing the time segment based on the optimal path; If not, adjust the constraint parameters of the preset search algorithm and re-execute the step of searching for the optimal path in the directed dynamic weight graph using the adjusted preset search algorithm.

[0112] For details on the implementation method, please refer to [link / reference]. Figures 1-6 Examples will not be described in detail here.

[0113] This application also provides a blood pressure trend analysis device based on time segment reconstruction; please refer to [link to relevant documentation]. Figure 8 , Figure 8 One embodiment of the blood pressure trend analysis device provided in this application includes: Processor 801, memory 802, input / output unit 803, bus 804; The processor 801 is connected to the memory 802, the input / output unit 803, and the bus 804; The memory 802 stores a program, and the processor 801 calls the program to execute any of the blood pressure trend analysis methods described above.

[0114] This application also relates to a computer-readable storage medium on which a program is stored, which, when run on a computer, causes the computer to perform any of the blood pressure trend analysis methods described above.

[0115] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0116] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0117] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0118] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0119] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for blood pressure trend analysis based on time segment reconstruction, characterized in that, include: Acquire time-series blood pressure data including systolic blood pressure, diastolic blood pressure and corresponding timestamps, and remove artifacts from the blood pressure time-series data; The blood pressure time series data after artifact removal is segmented using a preset multi-scale sliding window to obtain multiple time segments with different time resolutions, and the segment feature vectors in each time segment are extracted by a preset graph node embedding generation model. The temporal similarity between each segment feature vector is calculated using a preset dynamic time warping algorithm, and the segment feature vectors are divided into several pattern clusters based on the temporal similarity using a preset density clustering algorithm. Using the time segments as nodes and the temporal similarity and the similarity between each pattern cluster as weighting factors for the nodes, a directed dynamic weight graph is constructed. A preset search algorithm is used to search for the optimal path in the directed dynamic weight graph, and the time segment is reconstructed based on the optimal path to obtain a continuous blood pressure sequence; The continuous blood pressure sequence is linearly fitted within analysis windows of different time lengths, and blood pressure trend indicators for different time lengths are generated based on the trend slope obtained from the linear fitting.

2. The blood pressure trend analysis method according to claim 1, characterized in that, The preset graph node embedding generation model includes a feature extraction layer, a pooling layer, and a fully connected layer. The step of extracting segment feature vectors from each time segment using a preset graph node embedding generation model includes: The time segment is input into the feature extraction layer, and local temporal feature maps are extracted from the time segment through the feature extraction layer. The local temporal feature maps are then sent to the pooling layer. The local temporal feature map is selected according to the local saliency screening mechanism of the pooling layer, and the salient features obtained after feature selection are sent to the fully connected layer. The fully connected layer maps the salient features from the high-dimensional space to the low-dimensional vector space to obtain the fragment feature vector.

3. The blood pressure trend analysis method according to claim 2, characterized in that, The multiple time segments with different time resolutions include instantaneous time segments, short-time time segments, medium-time time segments, and long-time time segments; Before inputting the time segment into the feature extraction layer, the blood pressure trend analysis method further includes: The instantaneous time segment, the short-time time segment, the medium-time time segment, and the long-time time segment are subjected to time sequence alignment preprocessing. The time-series alignment preprocessing for the instantaneous time segment, the short-time time segment, the medium-time time segment, and the long-time time segment includes: The missing values ​​of the instantaneous time segment are processed by the neighborhood data completion method, and the missing values ​​of the short time segment, the medium time segment and the long time segment are processed by the linear time series interpolation method. Using the processed long-duration time segment's time axis as a reference axis, the time axes of the short-duration time segment, the medium-duration time segment, and the instantaneous time segment are time-series scaled and aligned based on the reference axis.

4. The blood pressure trend analysis method according to claim 3, characterized in that, The step of inputting the time segment into the feature extraction layer and extracting local temporal feature maps from the time segment through the feature extraction layer includes: The time segment is input into the temporal convolutional attention module of the feature extraction layer; The local neighborhood features of the time segment are extracted using the temporal convolution kernel built into the temporal convolutional attention module; The local neighborhood features are weighted according to the temporal attention mechanism in the temporal convolutional attention module to obtain a local temporal feature map.

5. The blood pressure trend analysis method according to claim 4, characterized in that, The step of performing feature selection on the local temporal feature map according to the local saliency screening mechanism of the pooling layer, and then sending the salient features obtained after feature selection to the fully connected layer, includes: The temporal saliency score of each feature point in the local temporal feature map is calculated based on the local saliency screening mechanism of the pooling layer. Feature points whose temporal significance scores are higher than a preset score threshold are selected as a set of highly significant feature points; The set of highly salient feature points is fused, and the resulting salient features are then fed to the fully connected layer.

6. The blood pressure trend analysis method according to claim 1, characterized in that, The preset multi-scale sliding window includes a preset instantaneous scale sliding window, a preset short-time scale sliding window, a preset medium-time scale sliding window, and a preset long-time scale sliding window. The blood pressure time-series data after artifact removal is segmented using a preset multi-scale sliding window to obtain multiple time segments with different time resolutions, including: The blood pressure time series data after artifact removal is traversed along the time axis using preset instantaneous scale sliding windows, preset short-time scale sliding windows, preset medium-time scale sliding windows, and preset long-time scale sliding windows, respectively. The blood pressure time series data in the traversal is truncated according to the window size and sliding step size of the sliding window at each scale to obtain the blood pressure data subsequence corresponding to each scale. The blood pressure data subsequences corresponding to each scale are timestamped and then encapsulated according to the traversal order of the sliding window at each scale to obtain multiple time segments with different time resolutions.

7. The blood pressure trend analysis method according to claim 1, characterized in that, The process of performing linear fitting on the continuous blood pressure sequence within analysis windows of different time lengths, and generating blood pressure trend indicators for different time lengths based on the trend slope obtained from the linear fitting, includes: Calculate the trend slope of the continuous blood pressure series corresponding to analysis windows of different time lengths; The trend slopes are sorted in chronological order, and each trend slope is classified according to a preset trend threshold to obtain a classified trend slope sequence. Based on the classified trend slope sequence, the blood pressure pattern, duration, and frequency of occurrence are statistically analyzed for each time length, and a blood pressure trend index is generated based on the blood pressure pattern, duration, and frequency of occurrence.

8. The blood pressure trend analysis method according to claim 1, characterized in that, The construction of a directed dynamic weighted graph, using the time segment as a node and the temporal similarity and the similarity between each pattern cluster as weighting factors, includes: Each time segment is initialized as a node, and the temporal similarity between any two nodes is obtained. At the same time, the similarity between any two pattern clusters is calculated according to the pattern cluster to which each node belongs. The weighting factor between each node is calculated based on the temporal similarity and the similarity between the pattern clusters using a preset weighted fusion function. The weighting factor is the dynamic weight of the directed edge connecting two nodes. A directed dynamic weight graph is constructed based on the nodes and the corresponding weighting factors.

9. The blood pressure trend analysis method according to any one of claims 1 to 8, characterized in that, After searching for the optimal path in the directed dynamic weight graph using a preset search algorithm, the blood pressure trend analysis method further includes: Based on the temporal sequence of each node in the directed dynamic weight graph, the temporal continuity index of the optimal path is calculated in combination with the preset path continuity condition. Determine whether the time continuity index is greater than or equal to a preset index threshold; If so, then the step of reconstructing the time segment based on the optimal path is performed; If not, the constraint parameters of the preset search algorithm are adjusted, and the step of searching for the optimal path in the directed dynamic weight graph is re-executed using the adjusted preset search algorithm.

10. A blood pressure trend analysis system based on time segment reconstruction, characterized in that, include: The removal unit is used to acquire blood pressure time-series data including systolic blood pressure, diastolic blood pressure and corresponding timestamps, and to remove artifacts from the blood pressure time-series data. The extraction unit is used to segment the blood pressure time series data after artifact removal using a preset multi-scale sliding window to obtain multiple time segments with different time resolutions, and to extract the segment feature vectors in each time segment through a preset graph node embedding generation model. The partitioning unit is used to calculate the temporal similarity between each segment feature vector using a preset dynamic time warping algorithm, and to partition the segment feature vectors into several pattern clusters based on the temporal similarity using a preset density clustering algorithm. A construction unit is used to construct a directed dynamic weighted graph with the time segment as a node and the temporal similarity and the similarity between each pattern cluster as weighting factors for the node. The reconstruction unit is used to search for the optimal path in the directed dynamic weight graph using a preset search algorithm, and to reconstruct the time segment based on the optimal path to obtain a continuous blood pressure sequence. The generation unit is used to perform linear fitting on the continuous blood pressure sequence within analysis windows of different time lengths, and generate blood pressure trend indicators for different time lengths based on the trend slope obtained from the linear fitting.

11. A blood pressure trend analysis device based on time segment reconstruction, characterized in that, include: Processor, memory, input / output units, and bus; The processor is connected to the memory, the input / output unit, and the bus; The memory stores a program, which the processor calls to execute the blood pressure trend analysis method as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains a program that, when executed on a computer, performs the blood pressure trend analysis method as described in any one of claims 1 to 9.

Images