An abnormal segment identification method based on continuous blood pressure information and related equipment

Abstract

The application provides an abnormal segment identification method based on continuous blood pressure information and related equipment, which can more comprehensively capture blood pressure changes at different time scales, effectively improve the detection rate of abnormal intervals, avoid abnormal missed detection caused by single segmentation, and effectively improve the reliability of the identification result. The application comprises: synchronously collecting continuous blood pressure signals; based on a preset blood pressure physiological rhythm prior rule, the continuous blood pressure signals are non-overlapping segmented in three dimensions of long time, medium time and short time to obtain effective segmentation windows; the effective segmentation windows are subjected to abnormal preliminary screening to obtain candidate abnormal intervals; multi-dimensional time sequence features of the continuous blood pressure signals in the candidate abnormal intervals are extracted, and the multi-dimensional time sequence features are spliced to obtain a fusion feature vector.

Description

Technical Field

[0001] This application relates to the field of medical testing technology, and in particular to a method and related equipment for identifying abnormal segments based on continuous blood pressure information. Background Technology

[0002] Continuous blood pressure monitoring is an important means of capturing dynamic changes in human blood pressure and reflecting the functional status of the cardiovascular system. It plays an irreplaceable role in hypertension screening and early warning, cardiovascular disease prevention and control, and the monitoring of critically ill patients. Currently, with the development of wearable devices and sensor technology, the stable acquisition of long-term continuous blood pressure signals has gradually become widespread. After acquiring the blood pressure signal, the next step is to identify and detect the blood pressure signal.

[0003] Current methods for identifying continuous blood pressure abnormalities fail to adequately consider the physiological rhythm characteristics of blood pressure, resulting in poor performance in identifying potential blood pressure abnormalities. Furthermore, they are susceptible to interference from physiological blood pressure fluctuations, leading to a high false positive rate and affecting the reliability of the identification results. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides a method and related equipment for identifying abnormal segments based on continuous blood pressure information.

[0005] The technical solution provided in this application is described below:

[0006] The first aspect of this application provides a method for identifying abnormal segments based on continuous blood pressure information, including: Simultaneous acquisition of continuous blood pressure signals; Based on the preset blood pressure physiological rhythm prior rules, the continuous blood pressure signal is segmented into three non-overlapping dimensions: long time, medium time, and short time, in order to obtain an effective segmentation window; Perform initial anomaly screening on the effective segmented window to obtain candidate anomaly intervals; Extract multi-dimensional temporal features of continuous blood pressure signals within the candidate abnormal interval, and concatenate the multi-dimensional temporal features to obtain a fused feature vector; The fused feature vector is input into a pre-trained two-branch twin contrast learning model, which is used to filter out physiological fluctuation intervals in the candidate abnormal intervals and retain pathological abnormal intervals. The pathological abnormal intervals are segmented temporally to determine the start and end boundaries of the abnormal segments. At the same time, the abnormal segments after boundary positioning are classified into pathological subtypes to obtain abnormal segments with subtype labels. Based on preset clinical evidence-based risk grading rules, the abnormal segments with subtype labels are classified into risk levels to output standardized identification results, which include full-dimensional information of the abnormal segments.

[0007] Optionally, the fused feature vector is input into a pre-trained two-branch twin contrastive learning model, which is used to filter out physiological fluctuation intervals from the candidate abnormal intervals and retain pathological abnormal intervals, including: The fused feature vector is input into the two branch networks of the two-branch Siamese contrastive learning model, and the fused feature vector is subjected to high-dimensional semantic mapping through the branch networks to generate a first embedding vector and a second embedding vector. Calculate a similarity measure between the first embedding vector and / or the second embedding vector and a preset physiological fluctuation prototype representation; the physiological fluctuation prototype representation is predetermined based on a historical physiological fluctuation sample set; The similarity metric is compared with a preset discrimination threshold to obtain a comparison result; Based on the comparison results, a comparison and discrimination result of the candidate abnormal interval is generated, and the comparison and discrimination result is used to indicate whether the candidate abnormal interval belongs to a physiological fluctuation interval or a pathological abnormal interval. Based on the comparison and discrimination results, data from the physiological fluctuation range are filtered out from the candidate abnormal range, while data from the pathological abnormal range are retained.

[0008] Optionally, based on a preset prior rule for blood pressure physiological rhythm, the continuous blood pressure signal is segmented into non-overlapping segments in three dimensions: long-term, medium-term, and short-term, to obtain an effective segmentation window, including: A prior rule library is constructed based on the preset blood pressure physiological rhythm prior rules. The prior rule library includes different physiological time periods and corresponding dynamic blood pressure baseline ranges. The prior rule base is invoked, and according to the time scale corresponding to the different physiological periods, the continuous blood pressure signal is divided into a long-term window of at least one complete circadian rhythm cycle, a medium-term window of at least one physiological event duration interval, and a short-term window of at least one instantaneous fluctuation element. The long-term window, the medium-term window, and the short-term window do not overlap with each other on the time axis. Determine whether the blood pressure data within each segmented window conforms to the dynamic baseline range of blood pressure for the corresponding time period in the prior rule base; If so, the matching segmented window will be considered a valid segmented window.

[0009] Optionally, multi-dimensional temporal features of continuous blood pressure signals within the candidate abnormal interval are extracted, and the multi-dimensional temporal features are concatenated to obtain a fused feature vector, including: Multi-domain feature extraction is performed on the continuous blood pressure signal within the candidate abnormal interval to obtain time-domain statistical features, frequency-domain energy features, and nonlinear dynamic features, respectively. Waveform morphology analysis is performed on the continuous blood pressure signal within the candidate abnormal interval to extract morphological structural features, which are used to characterize the geometric morphological changes within the blood pressure pulsation cycle. The time-domain statistical features, frequency-domain energy features, and nonlinear dynamic features are aligned and spliced ​​with the morphological structural features in the time dimension to generate a fused feature vector, which is used to characterize the multidimensional attributes of the blood pressure signal.

[0010] Optionally, the pathological abnormal intervals are temporally segmented to determine the start and end boundaries of the abnormal segments. Simultaneously, the abnormal segments after boundary localization are classified into pathological subtypes to obtain abnormal segments with subtype labels, including: The pathological abnormal regions are constructed using high-dimensional feature representation to obtain joint feature representations, which include temporal context information and morphological information. The joint feature representation is input into a preset multi-task learning model, and the multi-task learning model performs a temporal segmentation task and a pathological subtype classification task to obtain localization results and classification results; wherein, the temporal segmentation task is used to identify and locate the start point and end point of the abnormal state within the pathological abnormality interval, and the pathological subtype classification task is used to classify the abnormal state into one of a preset multiple pathological subtype categories. The location results and classification results are post-processed and fused, and the start and end boundaries after boundary calibration are associated with the corresponding pathological subtype labels to generate the abnormal segments with subtype labels.

[0011] Optionally, based on preset clinical evidence-based risk grading rules, the abnormal segments with subtype labels are classified into risk levels to output standardized identification results, including: Extract the key parameters for risk classification corresponding to the abnormal segments with subtype labels. The key parameters for risk classification include at least time-domain feature parameters, rhythm phase parameters, and waveform morphology parameters. A multi-level risk assessment matrix associated with pathological subtypes is constructed based on a pre-defined clinical evidence-based risk grading rule base. The key parameters for risk classification are input into the multi-level risk determination matrix for multi-dimensional mapping to determine the risk level to which the abnormal segment belongs. The risk level and the full-dimensional information of the abnormal segment are structurally encapsulated to generate standardized identification results and output them.

[0012] Optionally, the segmented window is subjected to initial anomaly screening to obtain candidate anomaly intervals, including: The segmented window is evaluated to obtain the evaluation results; Based on the evaluation results, invalid windows with abnormal data are removed to obtain the primary valid signal windows; The primary valid signal window is preprocessed to obtain the valid signal window.

[0013] A second aspect of this application provides an abnormal segment identification device based on continuous blood pressure information, the device comprising: The acquisition unit is used to synchronously acquire continuous blood pressure signals; The segmentation unit, based on a preset prior rule of blood pressure physiological rhythm, performs non-overlapping segmentation of the continuous blood pressure signal in three dimensions: long time, medium time, and short time, in order to obtain an effective segmentation window; The initial screening unit is used to perform initial screening of the effective segmented window to obtain candidate abnormal intervals; The acquisition unit is used to extract multi-dimensional temporal features of continuous blood pressure signals within the candidate abnormal interval, and to concatenate the multi-dimensional temporal features to obtain a fused feature vector; The input unit is used to input the fused feature vector into a pre-trained two-branch twin contrast learning model, which is used to filter out physiological fluctuation intervals in the candidate abnormal intervals and retain pathological abnormal intervals. The segmentation unit is used to perform temporal segmentation of the pathological abnormal interval, determine the start and end boundaries of the abnormal segment, and classify the abnormal segment after boundary positioning into pathological subtypes to obtain abnormal segments with subtype labels. The subdivision unit, based on preset clinical evidence-based risk grading rules, classifies the abnormal segments with subtype labels into risk levels to output standardized identification results, which include full-dimensional information of the abnormal segments.

[0014] A third aspect of this application provides an abnormal segment identification device based on continuous blood pressure information, the 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 perform the method as described in the first aspect and any one of the first aspects.

[0015] A fourth aspect of this application provides a computer-readable storage medium on which a program is stored, which, when executed on a computer, performs the method as described in the first aspect and any one of the first aspects.

[0016] As can be seen from the above technical solutions, this application has the following beneficial effects: 1. This application acquires continuous blood pressure signals synchronously and segments the signals into non-overlapping segments from three dimensions: long time, medium time, and short time. This allows for a more comprehensive capture of blood pressure changes at different time scales, effectively improving the detection rate of abnormal intervals and avoiding missed detections of abnormalities due to single segmentation.

[0017] 2. This application extracts multi-dimensional temporal features and constructs a fused feature vector, which can more completely represent the changing pattern of blood pressure signals. Then, it uses a pre-trained bi-branch twin contrast learning model to screen out physiological fluctuation ranges, significantly reducing false positive identification and improving the accuracy of pathological abnormality judgment.

[0018] 3. This application, through temporal segmentation of pathological abnormal regions, can locate the start and end boundaries of abnormal segments, achieving precise labeling of abnormal locations. Simultaneously, it completes pathological subtype classification and further combines clinical evidence-based risk grading rules to classify the risk level of abnormal segments, ultimately outputting standardized identification results containing comprehensive information. The entire process is automated, reducing manual analysis workload and thus improving the reliability of the identification results. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0020] Figure 1 This is a schematic diagram of an embodiment of the abnormal segment identification method based on continuous blood pressure information in this application; Figure 2 This is a schematic diagram of another embodiment of the abnormal segment identification method based on continuous blood pressure information in this application; Figure 3 This is a schematic diagram of another embodiment of the abnormal segment identification method based on continuous blood pressure information in this application; Figure 4 This is a schematic diagram of another embodiment of the abnormal segment identification method based on continuous blood pressure information in this application; Figure 5 This is a schematic diagram of another embodiment of the abnormal segment identification method based on continuous blood pressure information in this application; Figure 6 This is a schematic diagram of another embodiment of the abnormal segment identification method based on continuous blood pressure information in this application; Figure 7 This is a schematic diagram of an embodiment of the abnormal segment identification device based on continuous blood pressure information of this application; Figure 8This is a schematic diagram of another embodiment of the abnormal segment identification device based on continuous blood pressure information of this application; Figure 9 This is a schematic diagram of another embodiment of the abnormal segment identification device based on continuous blood pressure information in this application. Detailed Implementation

[0021] The abnormal segment identification method based on continuous blood pressure information described in this invention is not limited to any particular executing entity. It can be a terminal device with data processing and signal analysis capabilities (such as a smart blood pressure monitor, wearable physiological monitoring device, mobile terminal, dedicated medical monitoring device, etc.), or a server, cloud platform, processor, or embedded system with computing and storage capabilities. The aforementioned executing entities can either run this method independently or multiple devices / modules can cooperate and execute it in a coordinated manner. Those skilled in the art can flexibly select and deploy according to the actual application scenario, and this invention does not impose any particular limitations on this.

[0022] Current methods for identifying continuous blood pressure abnormalities fail to adequately consider the physiological rhythm characteristics of blood pressure, resulting in poor performance in identifying potential blood pressure abnormalities. Furthermore, they are susceptible to interference from physiological blood pressure fluctuations, leading to a high false positive rate and affecting the reliability of the identification results.

[0023] Based on this, this application provides an abnormal segment identification method and related equipment based on continuous blood pressure information, which can more comprehensively capture blood pressure changes at different time scales, effectively improve the detection rate of abnormal intervals, avoid missed detection of abnormalities due to single segmentation, and effectively improve the reliability of identification results.

[0024] Please see Figure 1 This application discloses a method for identifying abnormal segments based on continuous blood pressure information, the method comprising: 101. Synchronously acquire continuous blood pressure signals; 102. Based on the preset blood pressure physiological rhythm prior rules, the continuous blood pressure signal is segmented into three non-overlapping dimensions: long time, medium time, and short time, to obtain an effective segmentation window; 103. Perform initial anomaly screening on the effective segmented window to obtain candidate anomaly intervals; 104. Extract multi-dimensional temporal features of continuous blood pressure signals within the candidate abnormal interval, and concatenate the multi-dimensional temporal features to obtain a fused feature vector; 105. Input the fused feature vector into a pre-trained two-branch twin contrast learning model, wherein the two-branch twin contrast learning model is used to filter out physiological fluctuation intervals in the candidate abnormal intervals and retain pathological abnormal intervals. 106. Perform temporal segmentation on the pathological abnormal intervals to determine the start and end boundaries of the abnormal segments. At the same time, classify the abnormal segments after boundary positioning into pathological subtypes to obtain abnormal segments with subtype labels. 107. Based on preset clinical evidence-based risk grading rules, the abnormal segments with subtype labels are classified into risk levels to output standardized identification results, which include full-dimensional information of the abnormal segments.

[0025] In this embodiment, continuous blood pressure signals are first acquired synchronously. Then, based on preset blood pressure physiological rhythm prior rules, the continuous blood pressure signals are segmented into three non-overlapping dimensions: long-term, medium-term, and short-term, to obtain effective segmentation windows. Next, the effective segmentation windows are subjected to initial anomaly screening to obtain candidate abnormal intervals. Multi-dimensional temporal features of continuous blood pressure signals within the candidate abnormal intervals are extracted and concatenated to obtain a fused feature vector. Furthermore, the fused feature vector is input into a pre-trained two-branch twin contrastive learning model. The two-branch twin contrastive learning model is used to screen out physiological fluctuation intervals in the candidate abnormal intervals and retain pathological abnormal intervals. Next, the pathological abnormal intervals are segmented temporally to determine the start and end boundaries of the abnormal segments. At the same time, the abnormal segments after boundary localization are classified into pathological subtypes to obtain abnormal segments with subtype labels. Finally, based on the preset clinical evidence-based risk grading rules, the abnormal segments with subtype labels are classified into risk levels to output standardized identification results. The standardized results contain full-dimensional information of the abnormal segments.

[0026] In step 101, continuous blood pressure signals are first collected synchronously. Specifically, for example, a portable continuous blood pressure monitoring device can be used to collect blood pressure data from the subject for no less than 24 hours. During the collection process, the subject's basic physiological information, including age, gender, height, weight, basic medical history, medication, and behavioral status information during the monitoring period, is recorded synchronously to ensure the integrity and relevance of the collected data.

[0027] The continuous blood pressure signals to be collected include three parameters: systolic pressure, diastolic pressure, and mean arterial pressure. During the acquisition process, one set of blood pressure data is collected every second. The original acquired signals are initially denoised by the noise reduction module built into the device to filter out invalid noise signals caused by poor device contact, environmental electromagnetic interference, etc., and finally obtain a continuous and stable original continuous blood pressure signal dataset.

[0028] In step 102, after acquiring continuous blood pressure signals, based on preset blood pressure physiological rhythm prior rules, the continuous blood pressure signals are segmented non-overlapping into three dimensions: long-term, medium-term, and short-term, to obtain effective segmentation windows. The preset blood pressure physiological rhythm prior rules are based on the diurnal physiological rhythm characteristics of human blood pressure and are not specifically limited here. The specific segmentation process is as follows: Long-term segmentation uses 24-hour windows, corresponding to the diurnal blood pressure rhythm, to capture the overall daytime and nighttime fluctuation trend of blood pressure; medium-term segmentation uses 6-hour windows, further dividing the 24-hour long-term window into 4 non-overlapping medium-term windows, to capture the stage-wise fluctuation characteristics of blood pressure at different times; short-term segmentation uses 30-minute windows, further dividing each 6-hour medium-term window into 12 non-overlapping short-term windows, to capture the instantaneous fluctuation details of blood pressure.

[0029] During the segmentation process, the validity of each segment window is verified, and invalid windows with data missing rate or signal-to-noise ratio exceeding the preset value are removed. Finally, a set of valid segment windows covering three dimensions of long time, medium time and short time is obtained. Each valid segment window is associated with the corresponding time node, physiological state and behavioral state information.

[0030] In step 103, anomalies are initially screened within the effective segmented window to obtain candidate abnormal intervals. Specifically, the anomaly screening process employs a dual screening strategy of threshold determination and trend determination. Combining clinical blood pressure normality reference standards and physiological thresholds for blood pressure fluctuations, initial screening rules are formulated to ensure that the initial screening results can both cover potential abnormal intervals and preliminarily eliminate obvious normal fluctuation intervals.

[0031] First, for each valid segmented window, the mean, maximum, minimum, and fluctuation amplitude of systolic blood pressure, diastolic blood pressure, and mean arterial pressure within the window are calculated. The mean values ​​of each parameter are compared with the clinical normal reference range. If the mean value of any parameter exceeds the reference range, or if a single blood pressure value exceeds the threshold within the window and lasts for ≥5 minutes, the window is preliminarily determined as a suspected abnormal window. Next, the suspected abnormal window is trend-determined by calculating the linear trend slope of the blood pressure signal within the window. If the absolute value of the slope is ≥0.5 mmHg / min, indicating a rapid increase or decrease in blood pressure, the window is confirmed as an abnormal window.

[0032] Finally, adjacent abnormal windows are merged. If the time interval between two adjacent abnormal windows is ≤10 minutes and the blood pressure fluctuation trend is consistent, they are merged into a continuous candidate abnormal interval. At the same time, the time range, the long / medium / short time window to which each candidate abnormal interval belongs, and the statistical information of the blood pressure parameters within the window are recorded to complete the initial screening of abnormalities and obtain a set of candidate abnormal intervals.

[0033] In step 104, multi-dimensional temporal features of continuous blood pressure signals within the candidate abnormal interval are extracted, and these multi-dimensional temporal features are concatenated to obtain a fused feature vector. Specifically, since blood pressure signals within the candidate abnormal interval may contain both physiological fluctuations and pathological abnormalities, single-dimensional features cannot effectively distinguish between the two. Therefore, this application extracts multi-dimensional temporal features to comprehensively capture the fluctuation patterns, trend characteristics, and detailed features of blood pressure signals.

[0034] The extracted multi-dimensional time-series features fall into three categories: The first category comprises statistical features, including the mean, standard deviation, variance, maximum, and minimum values ​​of systolic blood pressure, diastolic blood pressure, and mean arterial pressure within the candidate abnormal interval; the second category comprises time-domain features, including the autocorrelation coefficient and cross-correlation coefficient of the blood pressure signal; and the third category comprises trend features, including the linear trend slope, trend goodness of fit, and fluctuation period of the blood pressure signal. During feature extraction, each candidate abnormal interval is processed by extracting these three types of features, each corresponding to a specific numerical indicator. All extracted multi-dimensional time-series features are then normalized to eliminate dimensional differences between different feature dimensions. Finally, all normalized time-series features are concatenated in a preset order to form a fixed-dimensional fusion feature vector. Each fusion feature vector uniquely corresponds to a candidate abnormal interval, thus providing standardized feature input for the subsequent input into the two-branch twin contrastive learning model.

[0035] In step 105, the fused feature vector is input into the pre-trained two-branch twin contrastive learning model. The two-branch twin contrastive learning model is used to filter out physiological fluctuation intervals in the candidate abnormal intervals and retain pathological abnormal intervals.

[0036] The purpose of the bi-branch twin contrastive learning model is to accurately distinguish between physiological fluctuations in candidate abnormal intervals, such as transient increases in blood pressure caused by emotional excitement, exercise, or eating, and pathological abnormalities, such as hypertensive crisis, hypotensive episodes, and sudden rises or falls in blood pressure, through contrastive learning, thereby solving the technical problem of high abnormality misjudgment rate in existing methods.

[0037] The pre-training process of the two-branch twin contrastive learning model is as follows: A training dataset is constructed, which contains a large number of labeled feature vectors of physiological fluctuation range and feature vectors of pathological abnormal range. The physiological fluctuation range samples are from blood pressure fluctuation data of healthy subjects under different behavioral states, and the pathological abnormal range samples are from blood pressure data of clinically diagnosed patients with hypertension, hypotension and abnormal blood pressure fluctuation. Each sample is associated with a corresponding label. The model adopts a two-branch structure, and the network structure of the two branches is completely identical, each containing 3 fully connected layers and 1 feature mapping layer. The input is a fused feature vector, and the output is a low-dimensional feature embedding.

[0038] During training, a contrastive loss function is used. By bringing the feature distances of similar samples (both physiological and pathological) closer together and widening the feature distances of dissimilar samples, the model learns to distinguish between two different types of abnormal regions.

[0039] After training, the model is validated. The fused feature vector obtained in step 104 is input into the pre-trained two-branch twin contrast learning model. The model analyzes and compares the feature vectors and outputs the category determination result for each candidate abnormal interval: physiological fluctuation interval or pathological abnormal interval. Then, all physiological fluctuation intervals are eliminated, and pathological abnormal intervals are retained. At the same time, the determination confidence of each pathological abnormal interval is recorded to provide a reference for subsequent boundary localization and subtype classification.

[0040] In step 106, the pathological abnormality intervals are segmented temporally to determine the start and end boundaries of the abnormal segments. Simultaneously, the abnormal segments after boundary localization are classified into pathological subtypes to obtain abnormal segments with subtype labels. Specifically, since the merged pathological abnormality intervals may contain multiple consecutive pathological abnormal events and the boundaries are not precise enough, temporal segmentation is needed to achieve accurate localization of the abnormal segments, while pathological subtype classification further refines the abnormality types.

[0041] Specifically, the time-series segmentation process employs a dynamic time warping algorithm combined with abrupt change detection of blood pressure signals: First, abrupt change detection is performed on continuous blood pressure signals within the pathological abnormal interval. The difference between two adjacent blood pressure data points is calculated. When the absolute value of the difference indicates a sudden rise or fall in blood pressure, it is identified as an abrupt change, and the time position of the abrupt change is recorded. Next, using the abrupt change as a segmentation node, and considering the stability of blood pressure fluctuations, the pathological abnormal interval is time-series segmented. If the blood pressure signal fluctuation amplitude between two abrupt change points is ≤10 mmHg and the duration is ≥3 minutes, it is classified as an independent abnormal segment. If there is no obvious abrupt change within a certain interval, but the blood pressure consistently exceeds the normal range and the duration is ≥10 minutes, it is also classified as an independent abnormal segment.

[0042] After segmentation, the start and end time boundaries of each abnormal segment are determined, and parameters such as duration and blood pressure fluctuation range of each abnormal segment are calculated. The pathological subtype classification process is based on the clinical blood pressure abnormality pathological classification standards, combined with the blood pressure characteristics of the abnormal segments, dividing them into 5 pathological subtypes: persistently elevated hypertension, sudden increase in hypertension, persistently decreased hypotension, sudden decrease in hypotension, and abnormal blood pressure fluctuation. During the classification process, a support vector machine classifier is used to complete the subtype classification, combining the multi-dimensional temporal features extracted in step 104. Each abnormal segment corresponds to a unique pathological subtype label, ultimately resulting in a set of abnormal segments with subtype labels. Each abnormal segment includes information such as start and end boundaries, duration, blood pressure parameters, and pathological subtype label.

[0043] In step 107, based on preset clinical evidence-based risk grading rules, the abnormal segments with subtype labels are classified into risk levels to output standardized identification results. The standardized results contain full-dimensional information of the abnormal segments.

[0044] The pre-defined clinical evidence-based risk grading rules are based on clinical guidelines for the diagnosis and treatment of hypertension and hypotension, as well as risk assessment standards for blood pressure abnormality-related complications. Combining the pathological subtype, duration, blood pressure fluctuation amplitude, and the subject's basic physiological information, abnormal segments are divided into three levels: low-risk, medium-risk, and high-risk. During the risk grading process, the core information of each abnormal segment labeled with a subtype is first extracted. Combined with the subject's basic physiological information and compared against the grading standards, the risk level of each abnormal segment is determined. Then, all dimensions of information from all abnormal segments are integrated to form a standardized identification result. This standardized result specifically includes: the start and end times of the abnormal segment, its duration, the monitoring period, the corresponding long / medium / short time window, core blood pressure parameters, pathological subtype label, risk level, confidence level, as well as the subject's basic physiological information and behavioral status information during the monitoring period. The standardized identification result is output in a structured format.

[0045] Please refer to Figure 2 According to some embodiments of the present invention, in step 105, the fused feature vector is input into a pre-trained two-branch twin contrast learning model. The two-branch twin contrast learning model is used to filter out physiological fluctuation intervals in the candidate abnormal intervals and retain pathological abnormal intervals. Specifically, it may include, but is not limited to, the following: 201. Input the fused feature vector into the two branch networks of the two-branch Siamese contrastive learning model respectively, and perform high-dimensional semantic mapping on the fused feature vector through the branch networks to generate a first embedding vector and a second embedding vector; 202. Calculate the similarity measure between the first embedding vector and / or the second embedding vector and the preset physiological fluctuation prototype representation; the physiological fluctuation prototype representation is predetermined based on a historical physiological fluctuation sample set; 203. Compare the similarity measure with a preset discrimination threshold to obtain a comparison result; 204. Generate a comparison and discrimination result of the candidate abnormal interval based on the comparison result, wherein the comparison and discrimination result is used to indicate whether the candidate abnormal interval belongs to a physiological fluctuation interval or a pathological abnormal interval; 205. Based on the comparison and discrimination results, filter out the data of the physiological fluctuation range from the candidate abnormal range, and retain the data of the pathological abnormal range.

[0046] In this embodiment, the fused feature vector is input into the two branch networks of the two-branch Siamese contrastive learning model. The branch networks perform high-dimensional semantic mapping on the fused feature vector to generate a first embedding vector and a second embedding vector. The two branch networks of the two-branch Siamese contrastive learning model are structurally symmetrical deep fully connected networks. Each branch network contains an input layer, three hidden layers, and an output layer. The two branch networks share the weight parameters obtained during the pre-training phase, ensuring consistency in the mapping rules.

[0047] Before inputting the fused feature vector, the fused feature vector obtained in step 104 is first validated to confirm that its dimension is the preset 256 dimensions. After the validation is passed, the fused feature vectors of the same candidate anomaly interval are simultaneously input into the input layers of the two branch networks. After receiving the fused feature vector, the input layer of the branch network performs nonlinear transformation through the activation function of the hidden layer in sequence, performing high-dimensional semantic fusion and mapping on the statistical features, temporal features, and trend features in the fused feature vector, compressing and mapping the original 256-dimensional fused feature vector into a 128-dimensional low-dimensional embedding vector. The low-dimensional embedding vector output by the first branch network is defined as the first embedding vector, and the low-dimensional embedding vector output by the second branch network is defined as the second embedding vector.

[0048] Based on the generated first and second embedding vectors, a similarity metric is calculated between the first and second embedding vectors and the preset physiological fluctuation prototype representation. It should be noted that the physiological fluctuation prototype representation is predetermined based on a historical physiological fluctuation sample set.

[0049] The process of constructing the physiological fluctuation prototype is as follows: A large amount of blood pressure fluctuation data from healthy subjects under physiological states such as emotional excitement, mild exercise, and post-meal conditions is collected. Following the feature extraction method in step 104, corresponding physiological fluctuation fusion feature vectors are obtained. These feature vectors are input into the branch network of a two-branch Siamese contrastive learning model to obtain the corresponding set of embedding vectors. The K-means clustering algorithm is applied to this set, and the 128-dimensional embedding vector corresponding to the cluster center is taken as the preset physiological fluctuation prototype representation. This prototype representation can characterize the feature patterns of physiological blood pressure fluctuations. The similarity measurement uses the cosine similarity algorithm, which can effectively measure the directional similarity between two vectors in high-dimensional space.

[0050] The similarity measure is then compared with a preset discrimination threshold to obtain the comparison result. Based on the comparison result, a comparison discrimination result for the candidate abnormal interval is generated, indicating whether the candidate abnormal interval belongs to a physiological fluctuation interval or a pathological abnormal interval.

[0051] Specifically, the criteria are as follows: If the comparison result is "similarity meets the threshold requirement," it means that the blood pressure fluctuation characteristics of the candidate abnormal interval are highly similar to the physiological fluctuation prototype, and therefore the generated comparison result is "the candidate abnormal interval belongs to the physiological fluctuation interval." If the comparison result is "similarity does not meet the threshold requirement," it means that the blood pressure fluctuation characteristics of the candidate abnormal interval are significantly different from the physiological fluctuation prototype and do not conform to the characteristic pattern of physiological fluctuations, and therefore the generated comparison result is "the candidate abnormal interval belongs to the pathological abnormal interval."

[0052] Finally, based on the comparison and discrimination results, data from physiological fluctuation intervals are filtered out from the candidate abnormal intervals, while data from pathological abnormal intervals are retained. The specific execution process is as follows: The comparison and discrimination results of all candidate abnormal intervals are iterated. Candidate abnormal intervals with a discrimination result of "physiological fluctuation interval" are marked as "invalid intervals" and removed from the candidate abnormal interval set, while retaining the basic information of these intervals for subsequent source tracing analysis. Candidate abnormal intervals with a discrimination result of "pathological abnormal interval" are marked as "valid intervals" and retained in a new abnormal interval set. The retained data includes the interval's fusion feature vector, similarity measure value, discrimination result, time range, core blood pressure parameters, and other full information. After the filtering process is completed, the output contains only a set of pathological abnormal intervals.

[0053] Please refer to Figure 3 According to some embodiments of the present invention, in step 102, based on a preset prior rule for blood pressure physiological rhythm, the continuous blood pressure signal is segmented into non-overlapping segments in three dimensions: long-term, medium-term, and short-term, to obtain an effective segmentation window. Specifically, this may include, but is not limited to, the following: 301. Construct a prior rule base according to the preset blood pressure physiological rhythm prior rules, wherein the prior rule base includes different physiological time periods and corresponding dynamic blood pressure baseline ranges; 302. Call the prior rule base and, according to the time scale corresponding to the different physiological periods, cut the continuous blood pressure signal into at least one long-term window of a complete circadian rhythm cycle, at least one medium-term window of a physiological event duration interval, and at least one short-term window of an instantaneous fluctuation element, wherein the long-term window, the medium-term window, and the short-term window do not overlap with each other on the time axis. 303. Determine whether the blood pressure data in each segmented window conforms to the dynamic baseline range of blood pressure for the corresponding time period in the prior rule base; 304. If yes, then the matching segmented window will be considered a valid segmented window.

[0054] In this embodiment, a prior rule base is constructed based on preset blood pressure physiological rhythm prior rules. The prior rule base includes different physiological time periods and corresponding dynamic baseline ranges of blood pressure. The preset blood pressure physiological rhythm prior rules are standardized rules formed after statistical analysis, verification and optimization based on the diurnal rhythm characteristics of human blood pressure, clinical blood pressure physiological research data and a large number of blood pressure monitoring samples from subjects.

[0055] The construction process of the prior rule base is as follows: First, different physiological time periods are divided. Combining the daily physiological activity patterns of the human body and the sleep-wake cycle, the 24 hours of a day are divided into several distinct physiological time periods, including the nighttime sleep period (23:00-06:00), the morning awakening period (06:00-08:00), the morning activity period (08:00-12:00), the afternoon rest period (12:00-14:00), the afternoon activity period (14:00-18:00), and the evening relaxation period (18:00-23:00). Each physiological time period corresponds to a specific time range and is associated with the corresponding human physiological state. Next, the dynamic baseline range of blood pressure corresponding to each physiological time period is determined. For each physiological time period, a large amount of continuous blood pressure monitoring data from healthy subjects is collected within that time period. Through statistical analysis, the mean and standard deviation of each blood pressure parameter are calculated. Combined with the clinical normal blood pressure reference standard, the dynamic baseline range of blood pressure for each physiological time period is determined, that is, the normal fluctuation range of blood pressure in healthy individuals within that time period.

[0056] It should be noted that this baseline range is not a fixed value, but rather is differentiated based on different age groups, genders, and baseline physiological states. For example, the dynamic blood pressure baseline range during nighttime sleep is lower than that during daytime activity, and the baseline range for the elderly is slightly higher than that for younger people, thus ensuring the specificity and accuracy of the baseline range. Finally, the different physiological time periods, corresponding time ranges, dynamic blood pressure baseline ranges, and associated physiological state information are integrated and constructed into a priori rule base, stored in the system database for easy retrieval in subsequent steps. A rule update interface is also reserved, allowing for dynamic optimization of the priori rule base based on new clinical data and research findings, improving the accuracy of subsequent segmentation and screening.

[0057] Next, the prior rule base is invoked, and based on the time scale corresponding to different physiological periods, the continuous blood pressure signal is segmented into a long-term window of at least one complete circadian rhythm cycle, a medium-term window of at least one physiological event duration interval, and a short-term window of at least one instantaneous fluctuation primitive. The long-term window, the medium-term window, and the short-term window do not overlap on the time axis. By matching the time scale of different physiological periods, segmented windows with three dimensions of long, medium, and short are constructed to take into account the overall trend, stage characteristics, and instantaneous fluctuation details of the blood pressure signal.

[0058] The specific segmentation process is as follows: First, the physiological time period and time scale information in the prior rule base are called to determine the segmentation criteria for the three-dimensional windows. The long-term window corresponds to the complete circadian rhythm cycle of the human body, with 24 hours as a segmentation unit. Each long-term window covers a complete 24-hour cycle, ensuring the overall trend of blood pressure fluctuations throughout the day and night is captured. Segmentation begins at the start time of data collection in step 101, and is performed sequentially at 24-hour intervals. If the collection duration exceeds 24 hours, it is segmented into multiple continuous and non-overlapping long-term windows. The medium-term window corresponds to the duration interval of a single physiological event defined in the prior rule base, with 6 hours as a segmentation unit. Each medium-term window corresponds to one or more adjacent physiological time periods. Based on the long-term window, each 24-hour long-term window is evenly segmented into four non-overlapping medium-term windows. Each medium-term window precisely corresponds to a specific physiological event duration interval, used to capture the phased fluctuation characteristics of blood pressure within that interval. Finally, the short-time window corresponds to the instantaneous fluctuation primitive of blood pressure, with 30-minute segments used to capture the details of instantaneous blood pressure fluctuations. The segmentation is based on the medium-time window, with each 6-hour medium-time window evenly divided into 12 non-overlapping short-time windows. Each short-time window corresponds to a small time unit, capable of capturing subtle blood pressure fluctuations within 30 minutes. The entire segmentation process strictly follows the timeline order, ensuring that the long-time, medium-time, and short-time windows do not overlap, and that each window can be accurately associated with the corresponding physiological period in the prior rule base, achieving multi-dimensional and refined segmentation of continuous blood pressure signals.

[0059] After segmentation, it is determined whether the blood pressure data within each segment window conforms to the dynamic baseline range of blood pressure for the corresponding time period in the prior rule base. Specifically, by comparing the blood pressure data of each segment window with the baseline range of the corresponding physiological time period in the prior rule base, abnormal windows that do not conform to physiological patterns are initially eliminated, providing reliable and effective windows for subsequent anomaly identification. The specific judgment process is as follows: First, for each segment window, all blood pressure data within the window are extracted, and the statistical indicators of each blood pressure parameter within the window are calculated, including mean, maximum, minimum, standard deviation, and fluctuation range. Among them, the fluctuation range is the difference between the maximum and minimum values, used to describe the overall fluctuation of blood pressure within the window. Then, based on the time range corresponding to the segment window, the corresponding physiological time period is matched from the prior rule base, and the dynamic baseline range of blood pressure for that physiological time period is retrieved. Then, the mean values ​​of each blood pressure parameter calculated within the window are compared with the corresponding dynamic baseline range of blood pressure for the same time period. Simultaneously, it is determined whether the fluctuation range of the blood pressure data within the window is within a reasonable range. If the mean values ​​of systolic blood pressure, diastolic blood pressure, and mean arterial pressure within the window are all within the corresponding baseline range, and the fluctuation range is within a reasonable range, then the segmented window is preliminarily determined to meet the requirements of the prior rule base. If the mean value of any blood pressure parameter exceeds the corresponding baseline range, or the fluctuation range exceeds a reasonable range, then the segmented window is determined to not meet the requirements of the prior rule base and will be removed as an invalid window.

[0060] If the criteria are met, the matching segmented windows are designated as valid segmented windows. Specifically, firstly, each segmented window judged as "meeting the requirements of the prior rule base" in step 303 is verified, checking the time range, corresponding physiological period, blood pressure statistical indicators, and judgment criteria for each window to ensure the accuracy of the screening results and avoid omissions or misscreening. Then, the verified matching windows are uniformly marked as valid segmented windows, and the core information of each valid segmented window is recorded, including window type, time range, corresponding physiological period, blood pressure parameter statistical values, and matching baseline range. Finally, the selected valid segmented windows are integrated to form a set of valid segmented windows. This set covers three dimensions: long-term, medium-term, and short-term, and the windows do not overlap on the time axis, comprehensively covering the overall trend, stage characteristics, and instantaneous fluctuation details of continuous blood pressure signals.

[0061] Please refer to Figure 4 According to some embodiments of the present invention, step 104 involves extracting multi-dimensional temporal features of continuous blood pressure signals within the candidate abnormal interval, and concatenating the multi-dimensional temporal features to obtain a fused feature vector. Specifically, this may include, but is not limited to, the following: 401. Perform multi-domain feature extraction on the continuous blood pressure signal within the candidate abnormal interval to obtain time-domain statistical features, frequency-domain energy features, and nonlinear dynamic features, respectively; 402. Perform waveform morphological analysis on the continuous blood pressure signal within the candidate abnormal interval to extract morphological structural features, which are used to characterize the geometric morphological changes within the blood pressure pulsation cycle. 403. Align and splice the time-domain statistical features, the frequency-domain energy features, and the nonlinear dynamic features with the morphological structural features in the time dimension to generate a fused feature vector, which is used to characterize the multidimensional attributes of the blood pressure signal.

[0062] In this embodiment, multi-domain feature extraction is first performed on the continuous blood pressure signal within the candidate abnormal interval to obtain time-domain statistical features, frequency-domain energy features, and nonlinear dynamic features. The multi-domain feature extraction is based on complete time-series data of systolic blood pressure, diastolic blood pressure, and mean arterial pressure within the candidate abnormal interval. Features are extracted from each of the three parameters and then aggregated to ensure comprehensive capture of the multi-dimensional intrinsic characteristics of the blood pressure signal.

[0063] Among them, the time-domain statistical features focus on the time-domain distribution and fluctuation patterns of blood pressure signals. The specific features extracted include: the mean, variance, standard deviation, maximum value, minimum value, peak value, trough value, fluctuation amplitude, and the proportion of abnormal data points of blood pressure parameters within the interval. At the same time, the time-series stability index of blood pressure signals is calculated. Through the above specific features, the overall level, local fluctuations, and abnormal distribution of blood pressure signals can be intuitively reflected.

[0064] Among them, the frequency domain energy features are obtained by performing Fast Fourier Transform (FFT) on the blood pressure signals within the candidate abnormal interval. First, the blood pressure time series data is zero-padded to an integer power of 2 length, and then transformed to the frequency domain by FFT. The extracted features include: the peak power spectral density and total energy of the low-frequency, mid-frequency and high-frequency components, the energy ratio of each frequency band and the frequency domain centroid, which are used to reveal the frequency distribution law of blood pressure signals and the correlation characteristics of physiological regulation mechanisms.

[0065] Nonlinear dynamic features are used to capture the nonlinear fluctuation characteristics of blood pressure signals, including correlation dimension, maximum Lyapunov exponent, approximate entropy, and sample entropy, which can effectively capture the differences in nonlinear characteristics between physiological fluctuations and pathological abnormalities.

[0066] After the three features were extracted, Z-score standardization was applied to convert the feature values ​​into standardized features with a mean of 0 and a variance of 1, thus eliminating the dimensional differences between features of different dimensions.

[0067] Next, waveform morphology analysis is performed on the continuous blood pressure signal within the candidate abnormal interval to extract morphological features, which are used to characterize the geometric changes within the blood pressure pulsation cycle. First, the blood pressure time series signal within the candidate abnormal interval is segmented. Taking a single blood pressure pulsation cycle as the basic unit, the inflection point detection algorithm is used to locate the start point, systolic peak, diastolic trough, and descending limb inflection point of each pulsation cycle, thus completing the segmentation of a single pulsation cycle.

[0068] For continuous blood pressure signals within the candidate abnormal interval, N complete pulsation cycles are segmented. If the waveform of a certain pulsation cycle has significant distortion, it is considered an invalid cycle and discarded. Only morphological structural features are extracted based on the valid pulsation cycles. The extracted morphological structural features specifically include: the duration of a single pulsation cycle, the duration of systole, the duration of diastole, the ratio of systolic to diastolic duration, the height of the systolic peak, the slope of the systolic peak, the slope of the descending branch, and the symmetry of the waveform. This allows for a comprehensive characterization of the geometric morphological changes and fluctuation patterns of the blood pressure pulsation cycle.

[0069] Finally, the time-domain statistical features, frequency-domain energy features, and nonlinear dynamic features are aligned and spliced ​​with the morphological and structural features in the time dimension to generate a fused feature vector, which is used to characterize the multidimensional attributes of the blood pressure signal.

[0070] Specifically, the temporal dimension alignment is first performed. Based on the timestamp of the candidate anomaly interval, the temporal statistical features, frequency energy features, and nonlinear dynamic features extracted in step 401 are temporally synchronized with the morphological and structural features extracted in step 402. This ensures that both types of features correspond to the complete temporal range of the same candidate anomaly interval, avoiding feature fusion distortion due to temporal misalignment.

[0071] After alignment, features are concatenated in the order of "time-domain statistical features, frequency-domain energy features, nonlinear dynamic features, and morphological structural features." Specifically, after each of the three blood pressure parameters is concatenated, the concatenated features corresponding to the three parameters are merged to generate a fused feature vector. Each candidate abnormal interval corresponds to one fused feature vector. This fused feature vector simultaneously encompasses the statistical characteristics, frequency-domain energy characteristics, nonlinear characteristics, and waveform morphological characteristics of the blood pressure signal, comprehensively characterizing the multidimensional attributes of the blood pressure signal.

[0072] Please refer to Figure 5 According to some embodiments of the present invention, in step 106, the pathological abnormal interval is temporally segmented to determine the start and end boundaries of the abnormal segment. Simultaneously, the abnormal segment after boundary localization is classified into pathological subtypes to obtain abnormal segments with subtype labels. Specifically, this may include, but is not limited to, the following: 501. Construct a high-dimensional feature representation for the pathological abnormal region to obtain a joint feature representation, wherein the joint feature representation includes temporal context information and morphological information; 502. Input the joint feature representation into a preset multi-task learning model, and perform temporal segmentation task and pathological subtype classification task through the multi-task learning model to obtain localization results and classification results; wherein, the temporal segmentation task is used to identify and locate the start point and end point of the abnormal state within the pathological abnormal interval, and the pathological subtype classification task is used to classify the abnormal state into one of a preset multiple pathological subtype categories. 503. Perform post-processing fusion on the positioning results and the classification results, and associate the start and end boundaries after boundary calibration with the corresponding pathological subtype labels to generate the abnormal segments with subtype labels.

[0073] In this embodiment, a high-dimensional feature representation is first constructed for the pathological abnormal region to obtain a joint feature representation, which includes temporal context information and morphological information. The high-dimensional feature representation construction is based on the generated fused feature vector, supplemented with temporal context features and optimized with morphological information representation, to achieve deep fusion of multi-dimensional features and ensure that the joint features can comprehensively reflect the temporal change patterns and waveform morphological characteristics of the pathological abnormal region.

[0074] Temporal context information is extracted through a temporal attention mechanism. Taking the pathological abnormal interval as the core, blood pressure signals 30 minutes before and after it are associated, and the feature correlation between the core interval and the extended interval is calculated. Then, attention weights are assigned to obtain abnormally related temporal features and suppress irrelevant baseline features.

[0075] It should be noted that the extracted temporal context features include: the difference in mean blood pressure between the core interval and the extended interval, the difference in fluctuation amplitude, the similarity of feature embedding vectors, as well as the blood pressure change trend before the anomaly and the recovery rate after the anomaly ends. The temporal context features obtained above are used to capture the correlation and dynamic change process before and after the anomaly occurs.

[0076] Morphological information is optimized based on the extracted morphological structural features, adding texture features and waveform distortion features of the blood pressure pulsation waveform. Subsequently, the temporal context features and optimized morphological features are aligned and concatenated with the fused feature vector generated in step 403. Layer normalization is used to eliminate differences in feature dimensions, and finally a joint feature representation is constructed. This joint feature not only retains the statistical, frequency domain, and nonlinear characteristics of the blood pressure signal, but also incorporates temporal contextual information and more comprehensive morphological information.

[0077] Based on the joint feature representation obtained in step 501, the joint feature representation is input into a pre-defined multi-task learning model. The multi-task learning model is used to perform temporal segmentation and pathological subtype classification tasks to obtain localization and classification results.

[0078] The temporal segmentation task identifies and locates the start and end points of abnormal states within the pathological abnormality intervals, while the pathological subtype classification task classifies the abnormal states into one of several preset pathological subtypes. The preset multi-task learning model employs a shared feature extraction and dual-task branching structure. The temporal segmentation task branch is constructed using fully connected layers and activation functions. The pathological subtype classification task branch is also constructed using fully connected layers and activation functions. After training, inputting the joint feature representation simultaneously outputs both the localization and classification results.

[0079] Finally, the localization and classification results are post-processed and fused, and the start and end boundaries after boundary calibration are associated with the corresponding pathological subtype labels to generate the abnormal segments with subtype labels. The post-processing fusion consists of three steps: localization result calibration, classification result verification, and feature association, to ensure the accuracy and reliability of the results.

[0080] Specifically, boundary calibration is performed first. Based on the output positioning results and the actual fluctuations of blood pressure time-series data, the boundary is optimized using a "dynamic window smoothing and clinical threshold constraint" approach: A 5-minute dynamic smoothing window is set centered on the starting and ending points obtained from the positioning. The rate of change of blood pressure values ​​within the window is calculated. If the blood pressure value within the window does not reach the clinical abnormal threshold, the boundary is adjusted to the timestamp when the blood pressure value within the window first reaches the abnormal threshold. If the starting point of the positioning is earlier than the start time of the pathological abnormal interval, it is calibrated to the start time of the pathological abnormal interval; if the ending point is later than the end time of the pathological abnormal interval, it is calibrated to the end time of the pathological abnormal interval, thus avoiding the boundary from exceeding a reasonable range.

[0081] The classification results were then validated. Combining the collected basic clinical information of the subjects with the extracted joint features, the rationality of the classification results was verified: if the classification result was secondary blood pressure abnormality, it was necessary to confirm that the subject had a corresponding underlying disease, and that the nonlinear dynamic features in the joint features matched the blood pressure fluctuation features related to the underlying disease. If the classification result was hypertensive crisis or hypertensive emergency, it was necessary to confirm that the duration of the abnormal segment and the peak blood pressure in the localization results met the clinical diagnostic criteria. If the validation failed, the classification probability distribution was recalculated, and the subtype corresponding to the second highest probability was taken as the final classification result. Finally, feature association was performed. The calibrated start and end boundaries of the abnormal segment, the validated pathological subtype label, and the joint features, core blood pressure parameters, and clinical auxiliary information corresponding to the abnormal segment were associated and bound together to generate an abnormal segment with a subtype label. Simultaneously, the boundary calibration deviation and classification confidence were recorded, providing complete basic data for subsequent risk stratification and ensuring the completeness and accuracy of the abnormal segment information.

[0082] Please refer to Figure 6 According to some embodiments of the present invention, in step 107, the abnormal segments with subtype labels are classified into risk levels based on preset clinical evidence-based risk grading rules to output standardized identification results. Specifically, this may include, but is not limited to, the following: 601. Extract the key parameters for risk classification corresponding to the abnormal segments with subtype labels. The key parameters for risk classification include at least time-domain feature parameters, rhythm phase parameters, and waveform morphology parameters. 602. Construct a multi-level risk assessment matrix that is associated with pathological subtypes based on a pre-defined clinical evidence-based risk grading rule base; 603. Input the key parameters of risk classification into the multi-level risk judgment matrix for multi-dimensional mapping to determine the risk level to which the abnormal segment belongs; 604. The risk level and the full-dimensional information of the abnormal segment are structurally encapsulated to generate standardized identification results and output them.

[0083] In this embodiment, key parameters for risk grading corresponding to abnormal segments with subtype labels are extracted. These key parameters include at least time-domain feature parameters, rhythm phase parameters, and waveform morphology parameters. The extraction process focuses on abnormal segments with subtype labels, combining the blood pressure time-series data, joint feature representations, and morphological features obtained in previous steps to extract three types of key parameters. All parameters are standardized to ensure consistent parameter dimensions, providing accurate data support for subsequent risk assessment.

[0084] Among them, the temporal characteristic parameters focus on the temporal distribution and fluctuation characteristics of blood pressure signals within the abnormal segment, directly related to the severity and lasting impact of the abnormality. The rhythm phase parameter characterizes the physiological rhythm phase of blood pressure within the abnormal segment, correlates with the diurnal blood pressure fluctuation pattern, and helps determine the impact of the physiological background of the abnormality on the risk level. The waveform morphology parameter, based on previous waveform morphology analysis results, focuses on the geometric shape and distortion characteristics of the blood pressure pulsation cycle, reflecting the impact of the abnormality on the physiological mechanism of blood pressure pulsation. These three types of parameters comprehensively cover the statistical characteristics, physiological rhythm correlation characteristics, and waveform morphology characteristics of the abnormal segment, constituting a key parameter set for risk grading. Each abnormal segment with a subtype label corresponds to a complete set of key parameters.

[0085] Next, a multi-level risk assessment matrix associated with pathological subtypes is constructed based on a pre-defined clinical evidence-based risk grading rule base. The pre-defined rule base includes risk influencing factors, key parameter thresholds, and risk level determination logic for each pathological subtype, ensuring the clinical adaptability of the rules. The multi-level risk assessment matrix adopts a "three-dimensional association" structure, with pathological subtypes as the row dimension, key risk grading parameters as the column dimension, and risk levels as the hierarchical dimension, realizing the mapping between pathological subtypes, key parameters, and risk levels.

[0086] The matrix construction process is as follows: First, the matrix is ​​divided into rows according to the five pathological subtypes, with each pathological subtype corresponding to an independent judgment logic row; second, the key parameters extracted from step 601 are used as matrix columns, with each parameter corresponding to one column, and the judgment threshold for the parameter under different risk levels is defined; finally, based on the judgment logic in the rule base, a risk level mapping relationship is assigned to each parameter column of each pathological subtype, that is, when a key parameter corresponding to a certain pathological subtype reaches a specific threshold, the corresponding risk level score is triggered, and finally, the risk level of the abnormal segment of the pathological subtype is determined by the accumulation of multiple parameter scores and threshold judgment.

[0087] For example, the maximum systolic blood pressure thresholds for hypertensive crisis are set as follows: ≥180 mmHg (triggers high-risk score), 160-179 mmHg (triggers medium-risk score), and <160 mmHg (triggers low-risk score). Conversely, the maximum systolic blood pressure thresholds for hypotension are set as follows: <70 mmHg (triggers high-risk score), 70-80 mmHg (triggers medium-risk score), and 80-90 mmHg (triggers low-risk score), ensuring the targeted and accurate nature of the matrix assessment. Simultaneously, the matrix employs a weighting mechanism: time-domain feature parameters account for 40% of the weight, rhythm and phase parameters account for 25%, and waveform morphology parameters account for 35%, prioritizing parameters with a more significant impact on risk level and enhancing the rationality of the assessment matrix.

[0088] Next, key parameters for risk grading are input into a multi-level risk assessment matrix for multi-dimensional mapping to determine the risk level of the abnormal segment. The entire process relies on the constructed multi-level risk assessment matrix to ensure the objectivity and consistency of the assessment results. First, parameter matching is performed. Key parameters of a subtype-labeled abnormal segment extracted in step 601 are matched one by one to the parameter columns in the row corresponding to the pathological subtype of that abnormal segment in the matrix. Based on the actual parameter values, the corresponding risk level score is matched. If a parameter value is within the critical range, the lowest score for the corresponding risk level is taken. Then, scores are accumulated. According to the matrix's preset weighting mechanism, the risk scores corresponding to the parameters are weighted and accumulated to calculate the total risk score for the abnormal segment. The calculation formula is: Total Risk Score = (Sum of temporal feature parameter scores × 40%) + (Sum of rhythm phase parameter scores × 25%) + (Sum of waveform morphology parameter scores × 35%). Finally, a threshold determination is performed, and the risk level is determined based on the total risk score. The preset threshold standards are: a total risk score of 1-8 points is judged as low risk; 9-18 points is judged as medium risk; and 19-30 points is judged as high risk. Finally, the risk level and the full-dimensional information of the abnormal segment are structured and encapsulated to generate standardized identification results and output them.

[0089] Please refer to Figure 7 According to some embodiments of the present invention, in step 103, the segmented window is subjected to initial anomaly screening to obtain candidate anomaly intervals, which may specifically include, but is not limited to, the following: 701. Perform signal evaluation on the segmented window to obtain evaluation results; 702. Based on the evaluation results, invalid windows with abnormal data are removed to obtain the primary valid signal windows; 703. Preprocess the primary valid signal window to obtain a valid signal window.

[0090] In this embodiment, signal evaluation is performed on segmented windows to obtain evaluation results. Then, based on the obtained evaluation results, invalid windows with abnormal data are removed to obtain preliminary valid signal windows. Specifically, the same removal criteria are used for segmented windows of long-term, medium-term, and short-term dimensions to ensure a consistent validity judgment scale across different dimensions and avoid screening bias caused by dimensional differences. After removal, all remaining provisionally valid windows constitute the preliminary valid signal windows. Each preliminary valid signal window corresponds to a complete evaluation result record, clearly defining the values ​​of its various evaluation indicators, providing a targeted basis for subsequent preprocessing steps.

[0091] Finally, the primary valid signal windows are preprocessed to obtain valid signal windows. For minor data issues present in the primary valid signal window evaluation results, targeted processing strategies are adopted to ensure that the blood pressure signals within the preprocessed windows reflect the subject's true blood pressure status. After preprocessing, each window undergoes secondary verification, recalculating data integrity rate, signal-to-noise ratio, and outlier percentage to confirm that all indicators meet the validity criteria. Windows that pass verification are the final valid signal windows.

[0092] Please see Figure 8 The second aspect of this application provides an abnormal segment identification device based on continuous blood pressure information, the device comprising: Acquisition unit 801 is used to synchronously acquire continuous blood pressure signals; Segmentation unit 802, based on preset blood pressure physiological rhythm prior rules, performs non-overlapping segmentation of the continuous blood pressure signal in three dimensions: long time, medium time, and short time, in order to obtain an effective segmentation window; The initial screening unit 803 is used to perform initial screening of the effective segmented window to obtain candidate abnormal intervals; The acquisition unit 804 is used to extract multi-dimensional temporal features of continuous blood pressure signals within the candidate abnormal interval, and to concatenate the multi-dimensional temporal features to obtain a fused feature vector. The input unit 805 is used to input the fused feature vector into a pre-trained two-branch twin contrast learning model, which is used to filter out physiological fluctuation intervals in the candidate abnormal intervals and retain pathological abnormal intervals. The segmentation unit 806 is used to perform temporal segmentation of the pathological abnormal interval, determine the start and end boundaries of the abnormal segment, and classify the abnormal segment after boundary positioning into pathological subtypes to obtain abnormal segments with subtype labels. The segmentation unit 807, based on preset clinical evidence-based risk grading rules, classifies the abnormal segments with subtype labels into risk levels to output standardized identification results, which include full-dimensional information of the abnormal segments.

[0093] Please see Figure 9 This application also provides an abnormal segment identification device based on continuous blood pressure information, the device comprising: Processor 901, memory 902, input / output unit 903, bus 904; The processor 901 is connected to the memory 902, the input / output unit 903, and the bus 904; The memory 902 stores a program, and the processor 901 calls the program to execute any of the methods described above.

[0094] 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 methods described above.

[0095] 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.

[0096] 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 through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0097] 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.

[0098] 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.

[0099] 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 of 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 identifying abnormal segments based on continuous blood pressure information, characterized in that, include: Simultaneous acquisition of continuous blood pressure signals; Based on the preset blood pressure physiological rhythm prior rules, the continuous blood pressure signal is segmented into three non-overlapping dimensions: long time, medium time, and short time, in order to obtain an effective segmentation window; Perform initial anomaly screening on the effective segmented window to obtain candidate anomaly intervals; Extract multi-dimensional temporal features of continuous blood pressure signals within the candidate abnormal interval, and concatenate the multi-dimensional temporal features to obtain a fused feature vector; The fused feature vector is input into a pre-trained two-branch twin contrast learning model, which is used to filter out physiological fluctuation intervals in the candidate abnormal intervals and retain pathological abnormal intervals. The pathological abnormal intervals are segmented temporally to determine the start and end boundaries of the abnormal segments. At the same time, the abnormal segments after boundary positioning are classified into pathological subtypes to obtain abnormal segments with subtype labels. Based on preset clinical evidence-based risk grading rules, the abnormal segments with subtype labels are classified into risk levels to output standardized identification results, which include full-dimensional information of the abnormal segments.

2. The abnormal segment identification method based on continuous blood pressure information according to claim 1, characterized in that, The fused feature vector is input into a pre-trained two-branch twin contrast learning model, which is used to filter out physiological fluctuation intervals from the candidate abnormal intervals and retain pathological abnormal intervals, including: The fused feature vector is input into the two branch networks of the two-branch Siamese contrastive learning model, and the fused feature vector is subjected to high-dimensional semantic mapping through the branch networks to generate a first embedding vector and a second embedding vector. Calculate a similarity measure between the first embedding vector and / or the second embedding vector and a preset physiological fluctuation prototype representation; the physiological fluctuation prototype representation is predetermined based on a historical physiological fluctuation sample set; The similarity metric is compared with a preset discrimination threshold to obtain a comparison result; Based on the comparison results, a comparison and discrimination result of the candidate abnormal interval is generated, and the comparison and discrimination result is used to indicate whether the candidate abnormal interval belongs to a physiological fluctuation interval or a pathological abnormal interval. Based on the comparison and discrimination results, data from the physiological fluctuation range are filtered out from the candidate abnormal range, while data from the pathological abnormal range are retained.

3. The abnormal segment identification method based on continuous blood pressure information according to claim 1, characterized in that, Based on a preset prior rule for blood pressure physiological rhythm, the continuous blood pressure signal is segmented into non-overlapping segments across three dimensions: long-term, medium-term, and short-term, to obtain an effective segmentation window, including: A prior rule library is constructed based on the preset blood pressure physiological rhythm prior rules. The prior rule library includes different physiological time periods and corresponding dynamic blood pressure baseline ranges. The prior rule base is invoked, and according to the time scale corresponding to the different physiological periods, the continuous blood pressure signal is divided into a long-term window of at least one complete circadian rhythm cycle, a medium-term window of at least one physiological event duration interval, and a short-term window of at least one instantaneous fluctuation element. The long-term window, the medium-term window, and the short-term window do not overlap with each other on the time axis. Determine whether the blood pressure data within each segmented window conforms to the dynamic baseline range of blood pressure for the corresponding time period in the prior rule base; If so, the matching segmented window will be considered a valid segmented window.

4. The abnormal segment identification method based on continuous blood pressure information according to claim 1, characterized in that, Extract multi-dimensional temporal features from continuous blood pressure signals within the candidate abnormal interval, and concatenate the multi-dimensional temporal features to obtain a fused feature vector, including: Multi-domain feature extraction is performed on the continuous blood pressure signal within the candidate abnormal interval to obtain time-domain statistical features, frequency-domain energy features, and nonlinear dynamic features, respectively. Waveform morphology analysis is performed on the continuous blood pressure signal within the candidate abnormal interval to extract morphological structural features, which are used to characterize the geometric morphological changes within the blood pressure pulsation cycle. The time-domain statistical features, frequency-domain energy features, and nonlinear dynamic features are aligned and spliced ​​with the morphological structural features in the time dimension to generate a fused feature vector, which is used to characterize the multidimensional attributes of the blood pressure signal.

5. The method for identifying abnormal segments based on continuous blood pressure information according to claim 1, characterized in that, The pathological abnormal regions are segmented temporally to determine the start and end boundaries of the abnormal segments. Simultaneously, the abnormal segments after boundary localization are classified into pathological subtypes to obtain abnormal segments with subtype labels, including: The pathological abnormal regions are constructed using high-dimensional feature representation to obtain joint feature representations, which include temporal context information and morphological information. The joint feature representation is input into a preset multi-task learning model, and the multi-task learning model performs a temporal segmentation task and a pathological subtype classification task to obtain localization results and classification results; wherein, the temporal segmentation task is used to identify and locate the start point and end point of the abnormal state within the pathological abnormality interval, and the pathological subtype classification task is used to classify the abnormal state into one of a preset multiple pathological subtype categories. The location results and classification results are post-processed and fused, and the start and end boundaries after boundary calibration are associated with the corresponding pathological subtype labels to generate the abnormal segments with subtype labels.

6. The method for identifying abnormal segments based on continuous blood pressure information according to claim 1, characterized in that, Based on preset clinical evidence-based risk grading rules, the abnormal segments with subtype labels are classified into risk levels to output standardized identification results, including: Extract the key parameters for risk classification corresponding to the abnormal segments with subtype labels. The key parameters for risk classification include at least time-domain feature parameters, rhythm phase parameters, and waveform morphology parameters. A multi-level risk assessment matrix associated with pathological subtypes is constructed based on a pre-defined clinical evidence-based risk grading rule base. The key parameters for risk classification are input into the multi-level risk determination matrix for multi-dimensional mapping to determine the risk level to which the abnormal segment belongs. The risk level and the full-dimensional information of the abnormal segment are structurally encapsulated to generate standardized identification results and output them.

7. The method for identifying abnormal segments based on continuous blood pressure information according to claim 1, characterized in that, The segmented window is initially screened for anomalies to obtain candidate anomaly intervals, including: The segmented window is evaluated to obtain the evaluation results; Based on the evaluation results, invalid windows with abnormal data are removed to obtain the primary valid signal windows; The primary valid signal window is preprocessed to obtain the valid signal window.

8. An abnormal segment identification device based on continuous blood pressure information, characterized in that, The device includes: The acquisition unit is used to synchronously acquire continuous blood pressure signals; The segmentation unit, based on a preset prior rule of blood pressure physiological rhythm, performs non-overlapping segmentation of the continuous blood pressure signal in three dimensions: long time, medium time, and short time, in order to obtain an effective segmentation window; The initial screening unit is used to perform initial screening of the effective segmented window to obtain candidate abnormal intervals; The acquisition unit is used to extract multi-dimensional temporal features of continuous blood pressure signals within the candidate abnormal interval, and to concatenate the multi-dimensional temporal features to obtain a fused feature vector; The input unit is used to input the fused feature vector into a pre-trained two-branch twin contrast learning model, which is used to filter out physiological fluctuation intervals in the candidate abnormal intervals and retain pathological abnormal intervals. The segmentation unit is used to perform temporal segmentation of the pathological abnormal interval, determine the start and end boundaries of the abnormal segment, and classify the abnormal segment after boundary positioning into pathological subtypes to obtain abnormal segments with subtype labels. The subdivision unit, based on preset clinical evidence-based risk grading rules, classifies the abnormal segments with subtype labels into risk levels to output standardized identification results, which include full-dimensional information of the abnormal segments.

9. An abnormal segment identification device based on continuous blood pressure information, characterized in that, The device includes: 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 perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains a program that, when executed on a computer, performs the method as described in any one of claims 1 to 7.

Images