Blood pressure monitoring method, device and equipment based on wearable equipment and medium
By extracting the multidimensional features of PPG signals in wearable devices and combining them with hierarchical calibration and transfer learning models, the problems of single features and noise interference in blood pressure monitoring are solved, and highly accurate and adaptive blood pressure prediction and alarm are achieved.
Patent Information
- Application Number
- CN202511158261.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-18
- Publication Date
- 2025-09-23
AI Technical Summary
Existing wearable devices have a single feature dimension in blood pressure monitoring, which makes it difficult to fully express the blood pressure information in physiological signals. In addition, there is a lack of effective calibration-free blood pressure regression models, resulting in insufficient prediction accuracy.
By collecting PPG signals, the target feature vectors of time-frequency domain features, inter-pulse correlation matrix features and time-domain morphological features are extracted. The hierarchical calibration model and transfer learning algorithm are used to predict blood pressure, and the graded alarm threshold is determined in combination with HIS information.
It improves the accuracy of wearable devices in monitoring blood pressure of hospitalized patients, realizes adaptive graded alarm based on symptom characteristics, reduces prediction deviation caused by noise interference of physiological signals, and improves the reliability of monitoring.
Smart Images
Figure CN120678400A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of blood pressure monitoring, and in particular to a blood pressure monitoring method, apparatus, device and medium based on wearable devices. Background Art
[0002] Existing research on wearable device-based blood pressure prediction uses relatively limited feature dimensions, failing to fully represent the blood pressure information contained in physiological signals. Furthermore, due to noise in physiological signals, it is difficult to extract physiological information that effectively represents the overall blood pressure status from a single segment of a signal. Furthermore, blood pressure monitoring methods have stringent usage restrictions, and most wearable device-based research has failed to provide more effective calibration-free blood pressure regression models. Summary of the Invention
[0003] In view of this, the purpose of the present invention is to provide a wearable device-based blood pressure monitoring method, apparatus, device, and medium, which can improve the prediction accuracy of wearable devices for blood pressure monitoring of hospitalized patients and realize adaptive graded alarms based on symptom characteristics. The specific scheme is as follows:
[0004] In a first aspect, the present application discloses a blood pressure monitoring method based on a wearable device, comprising:
[0005] The PPG signal of the target object is collected through the wearable device, and blood pressure-related features are extracted from the PPG signal to obtain a target feature vector including time-frequency domain features, PPG signal pulse-to-pulse correlation matrix features, and time-domain morphological features;
[0006] Inputting the target feature vector into a target blood pressure stratification calibration model to obtain an initial blood pressure prediction value; the target blood pressure stratification calibration model includes a hypertension classification sub-model for distinguishing between hypertensive and normotensive populations, and target systolic blood pressure regression sub-models trained for the hypertensive and normotensive populations, respectively;
[0007] Inputting the target feature vector and the initial blood pressure prediction value into a target blood pressure individual calibration model to obtain a target blood pressure prediction value; the target blood pressure individual calibration model is a regression model trained for a target subject based on a transfer learning algorithm;
[0008] Based on the target HIS information data of the target subject in the hospital, determining a blood pressure graded alarm threshold that matches the target HIS information data, the blood pressure graded alarm threshold including at least a first abnormal threshold range and a second abnormal threshold range;
[0009] The target blood pressure prediction value is matched with the graded alarm threshold, and a graded alarm of a corresponding level and a target blood pressure prediction value are output.
[0010] Optionally, the extracting blood pressure-related features from the PPG signal to obtain a target feature vector including time-frequency domain features, PPG signal pulse-to-pulse correlation matrix features, and time-domain morphological features includes:
[0011] Performing a multi-order wavelet transform on the PPG signal using a preset wavelet filter to obtain a low-frequency component of the signal representing the profile and baseline trend of the PPG signal and a high-frequency component of the signal representing the rapidly changing details of the PPG signal;
[0012] Extracting corresponding component information from the low-frequency component of the signal and the high-frequency component of the signal to obtain time-frequency domain features;
[0013] Dividing the PPG signal into signal segments according to a preset time window to obtain a plurality of pulse cycles;
[0014] Calculating the Pearson correlation coefficient between different pulse periods to construct a PPG signal pulse-to-pulse correlation matrix based on the Pearson correlation coefficient, and then calculating the mean, variance, and minimum value of all elements in the PPG signal pulse-to-pulse correlation matrix to obtain PPG signal pulse-to-pulse correlation matrix characteristics;
[0015] performing derivative processing on the PPG signal to obtain a corresponding derivative waveform;
[0016] Waveform amplitude information, time information of different waveform periods, and area information of different waveform periods are extracted from the PPG signal and the derivative waveform to obtain time domain morphological features.
[0017] Optionally, before inputting the target feature vector into the target blood pressure stratified calibration model to obtain the initial blood pressure prediction value, the method further includes:
[0018] Collect a training dataset containing historical PPG signals and corresponding reference blood pressure values of different subjects;
[0019] Performing feature vector screening on the historical feature vectors extracted from the historical PPG signals to screen out a classification feature subset for distinguishing between a hypertensive population and a normal blood pressure population;
[0020] Using the classification feature subset to train an initial classification sub-model, so that the initial classification sub-model learns the mapping relationship between the feature vectors in the classification feature subset and the hypertension label / normal blood pressure label, thereby obtaining a hypertension classification sub-model;
[0021] Filtering out systolic blood pressure regression feature subsets corresponding to the hypertensive population and the normal blood pressure population from the historical feature vectors;
[0022] The systolic pressure regression feature subset is used to train an initial systolic pressure regression sub-model so that the initial systolic pressure regression sub-model learns the mapping relationship between the feature vectors in the systolic pressure regression feature subset and the systolic pressure of people with different blood pressures, so as to obtain a target systolic pressure regression sub-model for hypertensive people and a target systolic pressure regression sub-model for normal blood pressure people.
[0023] Optionally, the performing feature vector screening on the historical feature vectors extracted from the historical PPG signals to screen out a classification feature subset for distinguishing between hypertensive people and normal blood pressure people includes:
[0024] Extracting a historical feature vector from the historical PPG signal;
[0025] The historical feature vectors are screened using a recursive feature elimination method to screen out a classification feature subset for distinguishing between hypertensive people and normal blood pressure people.
[0026] Optionally, the performing feature vector screening on the historical feature vectors extracted from the historical PPG signals to screen out a classification feature subset for distinguishing between hypertensive people and normal blood pressure people includes:
[0027] Performing feature vector screening on historical feature vectors extracted from the historical PPG signals to obtain an initial classification feature subset;
[0028] The leave-one-out cross-validation method is used to evaluate the prediction stability of the initial classification feature subset to obtain the corresponding stability prediction results;
[0029] With the goal of minimizing the standard deviation of the stability prediction result, a target classification feature subset is determined from the feature quantity interval of the initial classification feature subset.
[0030] Optionally, before inputting the target feature vector and the initial blood pressure prediction value into the target blood pressure individual calibration model to obtain the target blood pressure prediction value, the method further includes:
[0031] The feature vector of the effective PPG signal collected by the target subject within a preset time period and the corresponding reference blood pressure value are used as the target domain training set;
[0032] The feature vectors of PPG signals of other subjects and the corresponding reference blood pressure values are used as the source domain dataset;
[0033] The linear regression base model is iteratively trained based on the TrAdaBoost transfer learning algorithm, the initial target domain sample weights, the initial source domain sample weights, the target domain training set, and the source domain data set to obtain the target blood pressure individual calibration model based on the weighted combination of the linear regression base model in each iterative process.
[0034] Optionally, before determining the blood pressure graded alarm threshold that matches the target HIS information data of the target subject in the hospital, the method further includes:
[0035] Collect historical in-hospital HIS information data of different target subjects, corresponding abnormal blood pressure sample labels and disease characteristic information;
[0036] extracting abnormal HIS features related to abnormal blood pressure from the historical in-hospital HIS information data;
[0037] The historical HIS features and the abnormal blood pressure sample labels are used as training data for training a random forest classification model, so that the random forest classification model can learn the abnormal blood pressure range rules corresponding to different disease feature combinations to determine the first abnormal threshold range and the second abnormal threshold range.
[0038] In a second aspect, the present application discloses a blood pressure monitoring device based on a wearable device, comprising:
[0039] A feature extraction module is used to collect the PPG signal of the target object through the wearable device and extract blood pressure-related features from the PPG signal to obtain a target feature vector including time-frequency domain features, PPG signal pulse-to-pulse correlation matrix features, and time-domain morphological features;
[0040] a first monitoring module, configured to input the target feature vector into a target blood pressure stratification calibration model to obtain an initial blood pressure prediction value; the target blood pressure stratification calibration model comprising a hypertension classification sub-model for distinguishing between hypertensive and normotensive populations, and a target systolic blood pressure regression sub-model trained for the hypertensive and normotensive populations, respectively;
[0041] a second monitoring module, configured to input the target feature vector and the initial blood pressure prediction value into a target blood pressure individual calibration model to obtain a target blood pressure prediction value; the target blood pressure individual calibration model is a regression model trained for a target subject based on a transfer learning algorithm;
[0042] an alarm output module, configured to determine, based on the target HIS information data of the target subject in the hospital, a blood pressure graded alarm threshold that matches the target HIS information data, the blood pressure graded alarm threshold including at least a first abnormal threshold range and a second abnormal threshold range;
[0043] The result matching module is used to match the target blood pressure prediction value with the blood pressure graded alarm threshold, and output the corresponding graded alarm and target blood pressure prediction value.
[0044] In a third aspect, the present application discloses an electronic device, comprising:
[0045] Memory, used to store computer programs;
[0046] A processor is used to execute the computer program to implement the steps of the aforementioned blood pressure monitoring method based on a wearable device.
[0047] In a fourth aspect, the present application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the steps of the aforementioned disclosed blood pressure monitoring method based on a wearable device are implemented.
[0048] As can be seen, the present application discloses a blood pressure monitoring method based on a wearable device, comprising: collecting a PPG signal of a target subject through a wearable device, and extracting blood pressure-related features from the PPG signal to obtain a target feature vector comprising time-frequency domain features, PPG signal pulse-to-pulse correlation matrix features, and time-domain morphological features; inputting the target feature vector into a target blood pressure stratification calibration model to obtain an initial blood pressure prediction value; the target blood pressure stratification calibration model comprises a hypertension classification sub-model for distinguishing between hypertensive and normotensive subjects, and a target systolic pressure regression sub-model trained for the hypertensive and normotensive subjects, respectively; inputting the target feature vector and the initial blood pressure prediction value into a target blood pressure individual calibration model to obtain a target blood pressure prediction value; the target blood pressure individual calibration model is a regression model personalized for the target subject based on a transfer learning algorithm; based on the target HIS information data of the target subject in the hospital, determining a blood pressure graded alarm threshold that matches the target HIS information data, the blood pressure graded alarm threshold comprising at least a first abnormal threshold range and a second abnormal threshold range; matching the target blood pressure prediction value with the graded alarm threshold, and outputting a graded alarm and target blood pressure prediction value of the corresponding level. As can be seen, by classifying groups (hypertensive vs. normal), model fitting bias is reduced, targeted regression is used to improve group prediction accuracy, and transfer learning is then used to integrate individual patient data to dynamically optimize the predicted value. This addresses the problem of large prediction bias caused by physiological signal noise interference and improves the reliability of blood pressure monitoring for in-hospital patients. Based on HIS information, the alarm range corresponding to the current blood pressure prediction value of the target subject is determined, and the corresponding level of alarm is then issued, binding the alarm level to the patient's actual disease risk. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.
[0050] Figure 1 This is a flow chart of a blood pressure monitoring method based on a wearable device disclosed in this application;
[0051] Figure 2 This is a flow chart of a target blood pressure stratified calibration model training and usage method disclosed in this application;
[0052] Figure 3 This is a flow chart of a method for training and using a target blood pressure individual calibration model disclosed in this application;
[0053] Figure 4 This is a schematic structural diagram of a blood pressure monitoring device based on a wearable device disclosed in this application;
[0054] Figure 5 This is a structural diagram of an electronic device disclosed in this application. DETAILED DESCRIPTION
[0055] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0056] Existing research on wearable device-based blood pressure prediction uses relatively limited feature dimensions, failing to fully represent the blood pressure information contained in physiological signals. Furthermore, due to noise in physiological signals, it is difficult to extract physiological information that effectively represents the overall blood pressure status from a single segment of a signal. Furthermore, blood pressure monitoring methods have stringent usage restrictions, and most wearable device-based research has failed to provide more effective calibration-free blood pressure regression models.
[0057] To this end, the present invention provides a blood pressure monitoring solution based on wearable devices, which can improve the prediction accuracy of wearable devices in blood pressure monitoring of hospitalized patients and realize adaptive graded alarms based on symptom characteristics.
[0058] Reference Figure 1 As shown, the present invention discloses a blood pressure monitoring method based on a wearable device, comprising:
[0059] Step S11: Collect the PPG signal of the target object through the wearable device, and extract blood pressure related features of the PPG signal to obtain a target feature vector including time-frequency domain features, PPG signal pulse inter-correlation matrix features, and time domain morphological features.
[0060] In this embodiment, a multi-order wavelet transform is performed on the PPG signal using a preset wavelet filter to obtain a low-frequency signal component that represents the profile and baseline trend of the PPG signal, and a high-frequency signal component that represents the rapidly changing details of the PPG signal; corresponding component information is extracted from the low-frequency signal component and the high-frequency signal component to obtain time-frequency domain features; the PPG signal is segmented according to a preset time window to obtain a plurality of pulse periods; the Pearson correlation coefficient between different pulse periods is calculated to construct a PPG signal pulse-to-pulse correlation matrix based on the Pearson correlation coefficient; then, the mean, variance, and minimum value of all elements in the PPG signal pulse-to-pulse correlation matrix are calculated to obtain PPG signal pulse-to-pulse correlation matrix features; the PPG signal is differentiated to obtain a corresponding derivative waveform; and waveform amplitude information, time information of different waveform periods, and area information of different waveform periods are extracted from the PPG signal and the derivative waveform to obtain time-domain morphological features. It can be understood that the wearable device collects the target subject's PPG (Photo PlephysmoGraphy) signal, and then uses the db2 wavelet to perform first-, second-, and third-order wavelet transforms on the PPG signal to obtain the detail signal PPG_cD and the approximate signal PPG_cA. The maximum value, minimum value, mean value, variance, sample entropy, frequency band energy, and frequency band energy ratio are extracted from PPG_cD and PPG_cA to obtain time-frequency domain features. The sample entropy extraction process is as follows:
[0061] Set the time series of the PPG signal to Then, using m as the window, the signal time series is divided into Q=n-m+1 sequences, that is, a new time series is reconstructed. , calculate the distance between each sequence and all Q sequences, and find the maximum absolute value of the corresponding difference between the two, that is, calculate the maximum absolute value of the difference between the elements of the corresponding positions of the two subsequences, which is used to measure the distance between the subsequences:
[0062] ;
[0063] in, 、 They represent the current values of the elements at the corresponding positions in the subsequence.
[0064] An expression that defines a threshold , where r is a coefficient in the range of 0.1-0.25 and SD is the standard deviation of the sequence. Then all the distance values are counted The ratio of the number of numbers greater than F to the total number nm excluding itself is recorded as , then according to K , calculate their average , increase the window m to m+1 and repeat the above steps. Finally, the sample entropy is calculated as:
[0065] .
[0066] The band energy is the sum of the squared amplitudes of each data point.
[0067] The frequency band energy ratio is the ratio of the sub-frequency band energy to the total frequency band energy.
[0068] Furthermore, the process of obtaining the inter-pulse correlation matrix characteristics of the PPG signal is as follows: by calculating the Pearson correlation coefficient between the PPG signal pulses over a period of time, the inter-pulse correlation matrix is obtained, which can reflect the degree of change of the PPG signal pulses in a short period of time. First, n pulse cycles are extracted from the 5s PPG signal, namely S1, S2, S3, ..., S n The calculation formula of the pulse correlation coefficient is:
[0069] .
[0070] in, and It means that the 5sPPG signal segment is divided into two different pulse signal cycles. express and The covariance between Respectively and The standard deviation of . The correlation matrix between pulses Calculate the mean, variance, and minimum value of all elements in the correlation matrix. The mean value reflects the average correlation level between PPG pulses within 5 seconds. The variance reflects the degree of variation in the correlation between PPG pulses within 5 seconds, and the minimum value reflects the maximum degree of variation between PPG pulses within 5 seconds.
[0071] The process for acquiring PPG time-domain morphological features is as follows: Morphological features are extracted from the PPG waveform, effectively reflecting physiological factors related to blood pressure, such as vascular wall resistance and vascular elasticity. First- and second-order derivatives of the PPG are taken to extract maximum amplitude, minimum amplitude, minimum amplitude / maximum amplitude, rise time, fall time, rise time / fall time, rise time + fall time, rise area, fall area, rise area / fall area, and rise area + fall area.
[0072] Step S12: Input the target feature vector into a target blood pressure stratification calibration model to obtain an initial blood pressure prediction value; the target blood pressure stratification calibration model includes a hypertension classification sub-model for distinguishing between hypertensive people and normal blood pressure people, and a target systolic blood pressure regression sub-model trained for the hypertensive people and the normal blood pressure people respectively.
[0073] In this embodiment, before inputting the target feature vector into the target blood pressure stratified calibration model to obtain the initial blood pressure prediction value, the method further includes: collecting a training data set containing historical PPG signals and corresponding reference blood pressure values of different subjects; performing feature vector screening on the historical feature vectors extracted from the historical PPG signals to screen out a classification feature subset for distinguishing between hypertensive people and normal blood pressure people; using the classification feature subset to train an initial classification sub-model, so that the initial classification sub-model learns the mapping relationship between the feature vectors in the classification feature subset and the hypertension label / normal blood pressure label, and obtains a hypertension classification sub-model; screening out the systolic blood pressure regression feature subsets corresponding to the hypertensive people and the normal blood pressure people respectively from the historical feature vectors; and using the systolic blood pressure regression feature subset to train the initial systolic blood pressure regression sub-model, so that the initial systolic blood pressure regression sub-model learns the mapping relationship between the feature vectors in the systolic blood pressure regression feature subset and the systolic blood pressure of different blood pressure people, so as to obtain a target systolic blood pressure regression sub-model for the hypertensive people and a target systolic blood pressure regression sub-model for the normal blood pressure people. Here, historical feature vectors are extracted from the historical PPG signals; recursive feature elimination is used to perform feature vector screening on the historical feature vectors to screen out a classification feature subset for distinguishing between hypertensive and normotensive populations. Specifically, feature vector screening is performed on the historical feature vectors extracted from the historical PPG signals to obtain an initial classification feature subset; the predictive stability of the initial classification feature subset is evaluated using leave-one-out cross-validation to obtain a corresponding stability prediction result; and with the goal of minimizing the standard deviation of the stability prediction result, a target classification feature subset is determined from the feature quantity interval of the initial classification feature subset.
[0074] like Figure 2As shown in the stratified cascade target blood pressure stratified calibration model, there are two stages in total. The first stage is to select the optimal classification feature subsets of normotensive and hypertensive subjects through RFE (Recursive feature elimination) to establish classification submodels for normotensive and hypertensive populations. In the second stage, the hypertension SBP (Sub-Bottom Profiling) regression feature subset and the normal blood pressure SBP regression feature subset are used to construct SBP regression models for hypertensive and normotensive populations for blood pressure prediction. The RFE algorithm recursively eliminates the features with the smallest absolute value of the feature weight (according to the feature weight provided by the model) until the set number of features is reached. The algorithm sorts the features according to the order in which the features are eliminated. Its main idea is divided into the following three steps:
[0075] First, all features are used to construct a feature set {x1, x2, ..., xn}. Based on this feature set and combined with the learner, a model is constructed, and all features are trained to obtain the weights of all features. The feature with the smallest weight is deleted from the feature set, and a new feature set is reconstructed for model training to obtain new feature weights. Among them, this model uses the linear regression algorithm: the linear regression formula is , x is a feature, and θ is the parameter of each feature. The θ in the linear regression formula is solved by minimizing the mean square error between the predicted value and the true value, that is, the loss function formula is: , where m is the number of training samples, is the predicted value of the i-th sample, is the true value of the i-th sample.
[0076] Training steps:
[0077] 1. Data preprocessing: Collect a dataset containing independent variables (features) and dependent variables, handle missing values (e.g., deletion, mean filling, regression filling, etc.), and remove outliers (using statistical methods or visual detection). Use the Pearson correlation coefficient to retain features with high correlation with the dependent variable. Transform features with nonlinear relationships and normalize features to avoid multicollinearity between features. Split the data into training and test sets in a 7:3 ratio to ensure randomness and representativeness of the data partition.
[0078] 2. Model construction and parameter solution: Randomly initialize the parameter vector θ and solve θ through the loss function until the weights of all features in the feature set are greater than the set threshold or the number of features in the feature set is equal to the set target number of features.
[0079] To reduce the computational complexity of the model, features with a strong correlation with blood pressure were selected and ranked using a recursive feature elimination method. Furthermore, to simplify the regression task and reduce the learning difficulty of a single model, RFE was used to generate multiple feature subsets: a classification feature subset for normotensive and hypertensive subjects, a feature subset for hypertensive SBP regression, and a feature subset for normotensive SBP regression. Different feature subsets corresponded to different task models. A leave-one-out approach was used for training and validation. During training, the number of feature sets was set between 5 and 50, generating 45 feature sets. Each set was validated using the leave-one-out approach, and the feature set with the smallest mean standard deviation was selected as the final feature set.
[0080] After the model training is completed, the target feature vector is input into the target blood pressure stratification calibration model so that the model can predict the initial blood pressure prediction value corresponding to the blood pressure type of the target object (hypertensive patient type or normal blood pressure patient type).
[0081] Step S13: Input the target feature vector and the initial blood pressure prediction value into a target blood pressure individual calibration model to obtain a target blood pressure prediction value; the target blood pressure individual calibration model is a regression model trained individually for the target object based on a transfer learning algorithm.
[0082] In this embodiment, the target feature vector and the initial blood pressure prediction value are input into the target blood pressure individual calibration model, and before the target blood pressure prediction value is obtained, the method further includes: using the feature vector and corresponding reference blood pressure value of the effective PPG signal collected by the target subject within a preset time period as the target domain training set; using the feature vector and corresponding reference blood pressure value of the PPG signal of other subjects as the source domain data set; iteratively training the linear regression base model based on the TrAdaBoost transfer learning algorithm, the initial target domain sample weight, the initial source domain sample weight, the target domain training set, and the source domain data set to obtain the target blood pressure individual calibration model based on the weighted combination of the linear regression base model in each iterative process. Figure 3As shown in the figure, the TrAdaBoost algorithm is used to train a personalized calibration model for each subject in the TrAdaBoost-based blood pressure individual calibration model. The method used in this model is the instance transfer method. A weight adjustment mechanism is established to adjust the weights of the source domain and the target domain at the sample level, so that the source domain data is more matched with the target domain data. The model training method is to use the valid physiological data and blood pressure data collected by the subject for the first 1 to 10 times as the training set of the target domain, the remaining valid data of the subject as the test set of the target domain, and all the data of other subjects as the source domain, making full use of the effective information of the source domain data and the target domain training set data. The model is combined with the linear regression algorithm to continuously reduce the weight of the source domain sample data that is inconsistent with the target domain data during the training process, and predict the blood pressure of the remaining data of the subject. The leave-one-out method is used for 9 rounds of cyclic training, and then the blood pressure prediction accuracy of each subject is evaluated as a whole. The training steps of this algorithm are as follows:
[0083] Initialize the weight of a single sample , where n represents the number of samples in the source domain and m represents the number of samples in the target domain. Then, all sample weight vectors .
[0084] The blood pressure prediction value of each physiological signal sample is obtained through the regression algorithm, and the error is calculated using the blood pressure reference value of the physiological signal sample. The error calculation value is as follows:
[0085] ;
[0086] in, represents the reference blood pressure of subject i, represents the model-predicted blood pressure for subject i.
[0087] Update the sample weights. The weights are calculated as follows:
[0088] ;
[0089] in, is the normalization constant, N is the number of iterations, Represents the key coefficient for adjusting sample weights. Its core function is to dynamically change its influence in subsequent iterations based on whether the sample is classified correctly.
[0090] Set the number of iterations to 10, repeat the previous iterative process, and continuously update the sample weights. Predictions closer to the reference value are assigned larger weights to obtain the target blood pressure individual calibration model.
[0091] The target blood pressure individual calibration model after training is used to re-predict the input target feature vector and the initial blood pressure prediction value. In this way, the currently predicted target blood pressure prediction value is a personalized blood pressure prediction value for the current target object.
[0092] Step S14: Based on the target HIS information data of the target subject in the hospital, determine a blood pressure graded alarm threshold that matches the target HIS information data, wherein the blood pressure graded alarm threshold includes at least a first abnormal threshold range and a second abnormal threshold range.
[0093] In this embodiment, before determining the blood pressure grading alarm threshold that matches the target HIS information data based on the target object in the hospital, the method further includes: collecting historical in-hospital HIS information data containing different target objects, corresponding abnormal blood pressure sample labels and disease characteristic information; extracting abnormal HIS features related to abnormal blood pressure from the historical in-hospital HIS information data; using the historical HIS features and the abnormal blood pressure sample labels as training data for training a random forest classification model, so that the random forest classification model learns the abnormal blood pressure range patterns corresponding to different disease feature combinations to determine the first abnormal threshold range and the second abnormal threshold range. It can be understood that a blood pressure alarm grading model is constructed, and the training process of the model is as follows: Collecting hospitalization information of inpatients: including electronic medical records, test and examination information, medication records, surgical records and adverse reaction records. Collecting patient information: including patient basic information, medication history, family history, etc. Feature extraction is performed based on the hospitalization information, patient information, discharge details, and drug information of the discharge prescription of the inpatient to form feature information; based on the extracted features and abnormal blood pressure samples, 70% of the samples are used as training samples and 30% of the samples are used as prediction samples for decision tree model training. 10 groups of training samples are randomly divided for decision tree model training to form 10 decision trees, and the prediction samples are validated and optimized using the random forest algorithm. This decision tree model training uses the random forest algorithm: the final classification is determined by voting through multiple random trees. Its formula is , where T is the number of trees, is the predicted category of the i-th tree for sample x, is an indicator function, which is 1 when the condition is met and 0 otherwise. Random forest does not require an additional validation set. The error is calculated by out-of-bag data. Each sample has a 36.8% probability of not being sampled (out-of-bag data). The formula is , where N is the number of out-of-bag samples, To use only the predicted value of sample j by the tree that does not contain j, c represents the category label in the classification task.
[0094] The training steps are as follows:
[0095] 1. Data preparation and sampling: The dataset is D. For each tree i, N samples are extracted from D with replacement to form the training set Di (repeated sampling is allowed, and about 36.8% of the samples are not selected and are used as out-of-band data).
[0096] 2. Construction method for each of the 10 decision trees: For the current node, randomly select m feature subsets from the M features. Select the optimal splitting feature and split point among the m features. Recursively split until the stopping condition is met.
[0097] 3. Integrated prediction and result aggregation: Each tree outputs the probability or category of a sample belonging to each category. The final result is determined by majority vote.
[0098] 4. The current blood pressure value is predicted using the random forest algorithm and combined with the eigenvalues of the HIS data to derive blood pressure alarm criteria based on the combination of eigenvalues from different HIS (Hospital Information System) data. Level 1: Abnormal blood pressure falls within the range of the current medical record data. Level 2: Abnormal blood pressure falls outside the range of the current medical record data.
[0099] Therefore, after the blood pressure alarm grading model training is completed, the model can output a matching blood pressure grading threshold range according to the target HIS information data of the target subject.
[0100] Step S15: Matching the target blood pressure prediction value with the graded alarm threshold, and outputting a graded alarm and target blood pressure prediction value of corresponding levels.
[0101] In this embodiment, hierarchical data of blood pressure alarm is obtained according to the HIS data characteristics of the current patient, and the blood pressure value predicted by the dual model is matched with the hierarchical data to output a first or second level alarm.
[0102] In this way, the use of a stratified blood pressure calibration model and an individual calibration model improves blood pressure prediction accuracy from two dimensions: group patterns and individual characteristics. The former reduces model bias through group segmentation, while the latter captures dynamic changes through personalized modeling. The combination of the two creates a two-tiered blood pressure prediction system: a group baseline and individual fine-tuning. Furthermore, based on the patient hospitalization scenario, the relationship between HIS information and abnormal blood pressure values for all in-hospital patients is collected. A blood pressure alarm grading model is used to measure the different abnormal blood pressure ranges reflected by different disease characteristics, allowing for more accurate blood pressure grading alerts for people with the same disease characteristics.
[0103] As can be seen, the present application discloses a blood pressure monitoring method based on a wearable device, comprising: collecting a PPG signal of a target subject through a wearable device, and extracting blood pressure-related features from the PPG signal to obtain a target feature vector comprising time-frequency domain features, PPG signal pulse-to-pulse correlation matrix features, and time-domain morphological features; inputting the target feature vector into a target blood pressure stratification calibration model to obtain an initial blood pressure prediction value; the target blood pressure stratification calibration model comprises a hypertension classification sub-model for distinguishing between hypertensive and normotensive subjects, and a target systolic pressure regression sub-model trained for the hypertensive and normotensive subjects, respectively; inputting the target feature vector and the initial blood pressure prediction value into a target blood pressure individual calibration model to obtain a target blood pressure prediction value; the target blood pressure individual calibration model is a regression model personalized for the target subject based on a transfer learning algorithm; based on the target HIS information data of the target subject in the hospital, determining a blood pressure graded alarm threshold that matches the target HIS information data, the blood pressure graded alarm threshold comprising at least a first abnormal threshold range and a second abnormal threshold range; matching the target blood pressure prediction value with the graded alarm threshold, and outputting a graded alarm and target blood pressure prediction value of the corresponding level. As can be seen, by classifying groups (hypertensive vs. normal), model fitting bias is reduced, targeted regression is used to improve group prediction accuracy, and transfer learning is then used to integrate individual patient data to dynamically optimize the predicted value. This addresses the problem of large prediction bias caused by physiological signal noise interference and improves the reliability of blood pressure monitoring for in-hospital patients. Based on HIS information, the alarm range corresponding to the current blood pressure prediction value of the target subject is determined, and the corresponding level of alarm is then issued, binding the alarm level to the patient's actual disease risk.
[0104] Reference Figure 4 As shown, the present invention also discloses a blood pressure monitoring device based on a wearable device, comprising:
[0105] A feature extraction module 11 is configured to collect a PPG signal from a target subject through a wearable device and extract blood pressure-related features from the PPG signal to obtain a target feature vector comprising time-frequency domain features, PPG signal pulse-to-pulse correlation matrix features, and time-domain morphological features;
[0106] a first monitoring module 12 configured to input the target feature vector into a target blood pressure stratification calibration model to obtain an initial blood pressure prediction value; the target blood pressure stratification calibration model comprising a hypertension classification sub-model for distinguishing between hypertensive and normotensive subjects, and a target systolic blood pressure regression sub-model trained for the hypertensive and normotensive subjects, respectively;
[0107] A second monitoring module 13 is configured to input the target feature vector and the initial blood pressure prediction value into a target blood pressure individual calibration model to obtain a target blood pressure prediction value; the target blood pressure individual calibration model is a regression model trained for a target subject based on a transfer learning algorithm;
[0108] The alarm output module 14 is configured to determine, based on the target HIS information data of the target subject in the hospital, a blood pressure graded alarm threshold that matches the target HIS information data, wherein the blood pressure graded alarm threshold includes at least a first abnormal threshold range and a second abnormal threshold range;
[0109] The result matching module 15 is used to match the target blood pressure prediction value with the blood pressure graded alarm threshold, and output the corresponding graded alarm and target blood pressure prediction value.
[0110] As can be seen, the present application discloses collecting a PPG signal of a target subject through a wearable device, and extracting blood pressure-related features from the PPG signal to obtain a target feature vector comprising time-frequency domain features, PPG signal pulse-to-pulse correlation matrix features, and time-domain morphological features; inputting the target feature vector into a target blood pressure stratification calibration model to obtain an initial blood pressure prediction value; the target blood pressure stratification calibration model includes a hypertension classification sub-model for distinguishing between hypertensive and normotensive subjects, and a target systolic pressure regression sub-model trained for the hypertensive and normotensive subjects, respectively; inputting the target feature vector and the initial blood pressure prediction value into a target blood pressure individual calibration model to obtain a target blood pressure prediction value; the target blood pressure individual calibration model is a regression model personalized for the target subject based on a transfer learning algorithm; based on the target HIS information data of the target subject in the hospital, determining a blood pressure graded alarm threshold that matches the target HIS information data, the blood pressure graded alarm threshold comprising at least a first abnormal threshold range and a second abnormal threshold range; matching the target blood pressure prediction value with the graded alarm threshold, and outputting a graded alarm and target blood pressure prediction value of the corresponding level. As can be seen, by classifying groups (hypertensive vs. normal), model fitting bias is reduced, targeted regression is used to improve group prediction accuracy, and transfer learning is then used to integrate individual patient data to dynamically optimize the predicted value. This addresses the problem of large prediction bias caused by physiological signal noise interference and improves the reliability of blood pressure monitoring for in-hospital patients. Based on HIS information, the alarm range corresponding to the current blood pressure prediction value of the target subject is determined, and the corresponding level of alarm is then issued, binding the alarm level to the patient's actual disease risk.
[0111] Furthermore, the embodiment of the present application also discloses an electronic device, Figure 5This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content in the diagram should not be considered as any limitation to the scope of application of the present application.
[0112] Figure 5 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of the present application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 is used to store a computer program, which is loaded and executed by the processor 21 to implement the relevant steps of the blood pressure monitoring method based on a wearable device disclosed in any of the aforementioned embodiments. In addition, the electronic device 20 in this embodiment may specifically be an electronic computer.
[0113] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device. The communication protocol it follows is any communication protocol that can be applied to the technical solution of this application and is not specifically limited here; the input and output interface 25 is used to obtain external input data or output data to the outside world. Its specific interface type can be selected according to specific application needs and is not specifically limited here.
[0114] Among them, the processor 21 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 21 can be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 21 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 21 may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.
[0115] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or CD, etc. The resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage method can be temporary storage or permanent storage.
[0116] The operating system 221 is used to manage and control the hardware devices and computer program 222 on the electronic device 20, so as to enable the processor 21 to calculate and process the massive amount of data 223 in the memory 22. The operating system 221 can be Windows Server, Netware, Unix, Linux, etc. In addition to including computer programs capable of performing the wearable device-based blood pressure monitoring method performed by the electronic device 20 as disclosed in any of the aforementioned embodiments, the computer program 222 can further include computer programs capable of performing other specific tasks. The data 223 can include not only data transmitted by the electronic device from an external device but also data collected by its own input and output interface 25.
[0117] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when executed by a processor, the computer program implements the aforementioned wearable device-based blood pressure monitoring method. The specific steps of this method can be referred to the corresponding contents disclosed in the aforementioned embodiments and will not be repeated here.
[0118] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from the other embodiments. Reference can be made to the descriptions of the identical or similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and the relevant parts can be referred to the descriptions of the methods.
[0119] Professionals may further appreciate that the units and algorithmic steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application. The steps of the method or algorithm described in conjunction with the embodiments disclosed herein can be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module can be placed in a random access memory RAM (Random Access Memory), memory, read-only memory ROM (Read Only Memory), electrically programmable EPROM (Electrically Programmable Read Only Memory), electrically erasable programmable EEPROM (Electric Erasable Programmable Read Only Memory), registers, hard disk, removable disk, CD-ROM (Compact Disc-Read Only Memory), or any other form of storage medium known in the technical field.
[0120] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.
[0121] The above is a detailed introduction to the solution provided by the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only applicable to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.
Claims
1. A blood pressure monitoring method based on a wearable device, characterized in that: include: The PPG signal of the target object is collected through the wearable device, and blood pressure-related features are extracted from the PPG signal to obtain a target feature vector including time-frequency domain features, PPG signal pulse-to-pulse correlation matrix features, and time-domain morphological features; Inputting the target feature vector into a target blood pressure stratification calibration model to obtain an initial blood pressure prediction value; the target blood pressure stratification calibration model includes a hypertension classification sub-model for distinguishing between hypertensive and normotensive populations, and target systolic blood pressure regression sub-models trained for the hypertensive and normotensive populations, respectively; Inputting the target eigenvector and the initial blood pressure prediction value into a target blood pressure individual calibration model to obtain a target blood pressure prediction value; The target blood pressure individual calibration model is a regression model trained individually for the target subject based on a transfer learning algorithm; Based on the target HIS information data of the target subject in the hospital, determining a blood pressure graded alarm threshold that matches the target HIS information data, the blood pressure graded alarm threshold including at least a first abnormal threshold range and a second abnormal threshold range; The target blood pressure prediction value is matched with the graded alarm threshold, and a graded alarm of a corresponding level and a target blood pressure prediction value are output.
2. The blood pressure monitoring method based on a wearable device according to claim 1, characterized in that: The blood pressure related feature extraction of the PPG signal to obtain a target feature vector including time-frequency domain features, PPG signal pulse inter-pulse correlation matrix features, and time domain morphological features includes: Performing a multi-order wavelet transform on the PPG signal using a preset wavelet filter to obtain a low-frequency component of the signal representing the profile and baseline trend of the PPG signal and a high-frequency component of the signal representing the rapidly changing details of the PPG signal; Extracting corresponding component information from the low-frequency component of the signal and the high-frequency component of the signal to obtain time-frequency domain features; Dividing the PPG signal into signal segments according to a preset time window to obtain a plurality of pulse cycles; Calculating the Pearson correlation coefficient between different pulse periods to construct a PPG signal pulse-to-pulse correlation matrix based on the Pearson correlation coefficient, and then calculating the mean, variance, and minimum value of all elements in the PPG signal pulse-to-pulse correlation matrix to obtain PPG signal pulse-to-pulse correlation matrix characteristics; performing derivative processing on the PPG signal to obtain a corresponding derivative waveform; Waveform amplitude information, time information of different waveform periods, and area information of different waveform periods are extracted from the PPG signal and the derivative waveform to obtain time domain morphological features.
3. The blood pressure monitoring method based on a wearable device according to claim 1, characterized in that: Before inputting the target feature vector into the target blood pressure stratification calibration model to obtain the initial blood pressure prediction value, the method further includes: Collect a training dataset containing historical PPG signals and corresponding reference blood pressure values of different subjects; Performing feature vector screening on historical feature vectors extracted from the historical PPG signals to screen out a classification feature subset for distinguishing between hypertensive people and normal blood pressure people; Using the classification feature subset to train an initial classification sub-model, so that the initial classification sub-model learns the mapping relationship between the feature vectors in the classification feature subset and the hypertension label / normal blood pressure label, thereby obtaining a hypertension classification sub-model; Filtering out systolic blood pressure regression feature subsets corresponding to the hypertensive population and the normal blood pressure population respectively from the historical feature vectors; The systolic pressure regression feature subset is used to train an initial systolic pressure regression sub-model so that the initial systolic pressure regression sub-model learns the mapping relationship between the feature vectors in the systolic pressure regression feature subset and the systolic pressure of people with different blood pressures, so as to obtain a target systolic pressure regression sub-model for hypertensive people and a target systolic pressure regression sub-model for normal blood pressure people.
4. The blood pressure monitoring method based on a wearable device according to claim 3, characterized in that: The performing feature vector screening on the historical feature vectors extracted from the historical PPG signals to screen out a classification feature subset for distinguishing between hypertensive people and normal blood pressure people includes: Extracting a historical feature vector from the historical PPG signal; The historical feature vectors are screened using a recursive feature elimination method to screen out a classification feature subset for distinguishing between hypertensive people and normal blood pressure people.
5. The blood pressure monitoring method based on a wearable device according to claim 3, characterized in that: The performing feature vector screening on the historical feature vectors extracted from the historical PPG signals to screen out a classification feature subset for distinguishing between hypertensive people and normal blood pressure people includes: Performing feature vector screening on historical feature vectors extracted from the historical PPG signals to obtain an initial classification feature subset; The leave-one-out cross-validation method is used to evaluate the prediction stability of the initial classification feature subset to obtain the corresponding stability prediction results; With the goal of minimizing the standard deviation of the stability prediction result, a target classification feature subset is determined from the feature quantity interval of the initial classification feature subset.
6. The blood pressure monitoring method based on a wearable device according to claim 1, characterized in that: Before inputting the target characteristic vector and the initial blood pressure prediction value into the target blood pressure individual calibration model to obtain the target blood pressure prediction value, the method further includes: The feature vector of the effective PPG signal collected by the target subject within a preset time period and the corresponding reference blood pressure value are used as the target domain training set; The feature vectors of PPG signals of other subjects and the corresponding reference blood pressure values are used as the source domain dataset; The linear regression base model is iteratively trained based on the TrAdaBoost transfer learning algorithm, the initial target domain sample weights, the initial source domain sample weights, the target domain training set, and the source domain data set to obtain the target blood pressure individual calibration model based on the weighted combination of the linear regression base model in each iterative process.
7. The blood pressure monitoring method based on a wearable device according to claim 1, characterized in that: Before determining the blood pressure graded alarm threshold value matching the target HIS information data of the target subject in the hospital, the method further includes: Collect historical in-hospital HIS information data of different target subjects, corresponding abnormal blood pressure sample labels and disease characteristic information; extracting abnormal HIS features related to abnormal blood pressure from the historical in-hospital HIS information data; The historical HIS features and the abnormal blood pressure sample labels are used as training data for training a random forest classification model, so that the random forest classification model can learn the abnormal blood pressure range rules corresponding to different disease feature combinations to determine the first abnormal threshold range and the second abnormal threshold range.
8. A blood pressure monitoring device based on a wearable device, characterized in that: include: A feature extraction module is used to collect the PPG signal of the target object through the wearable device and extract blood pressure-related features from the PPG signal to obtain a target feature vector including time-frequency domain features, PPG signal pulse-to-pulse correlation matrix features, and time-domain morphological features; a first monitoring module, configured to input the target feature vector into a target blood pressure stratification calibration model to obtain an initial blood pressure prediction value; the target blood pressure stratification calibration model comprising a hypertension classification sub-model for distinguishing between hypertensive and normotensive populations, and a target systolic blood pressure regression sub-model trained for the hypertensive and normotensive populations, respectively; a second monitoring module, configured to input the target feature vector and the initial blood pressure prediction value into a target blood pressure individual calibration model to obtain a target blood pressure prediction value; The target blood pressure individual calibration model is a regression model trained individually for the target subject based on a transfer learning algorithm; an alarm output module, configured to determine, based on the target HIS information data of the target subject in the hospital, a blood pressure graded alarm threshold that matches the target HIS information data, the blood pressure graded alarm threshold including at least a first abnormal threshold range and a second abnormal threshold range; The result matching module is used to match the target blood pressure prediction value with the blood pressure graded alarm threshold, and output the corresponding graded alarm and target blood pressure prediction value.
9. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the blood pressure monitoring method based on a wearable device as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that Used to store a computer program; wherein, when the computer program is executed by a processor, the steps of the blood pressure monitoring method based on a wearable device are implemented as described in any one of claims 1 to 7.