Cuffless-based blood pressure estimation modeling method and system

Abstract

Provided is a cuffless-based blood pressure estimation modeling method comprising the steps of: extracting, from biological signal data, features of biological data including waveform information necessary for blood pressure measurement; clustering the extracted features of the biological data to group same into a preset number of models; and applying a blood pressure estimation algorithm to the grouped models.

Description

Cuffless-based blood pressure estimation modeling method and system

[0001] The present invention relates to a cuffless-based blood pressure estimation modeling method and system.

[0002] Hypertension and hypotension are medical conditions associated with abnormally high and low blood pressure (BP), respectively, and are significant causes of global mortality. Long-term, uncontrolled hypertension can lead to serious complications, including stroke, kidney damage, and heart attack. Conversely, prolonged hypotension can lead to organ hypoperfusion, which can lead to life-threatening conditions such as septic shock. Therefore, early detection of hypertension and hypotension is crucial for effective prevention and treatment, and continuous BP monitoring is necessary for optimal cardiovascular health management.

[0003] Among existing continuous BP monitoring methods, arterial catheterization, which directly measures arterial BP using a pressure sensor, has been the standard diagnostic method in clinical settings. However, due to its invasive nature, this method is limited to hospital settings and has limited applicability in ambulatory settings. Cuff-based measurement, a band worn on the back of the hand, utilizes an inflatable cuff with a pressure sensor attached to the upper arm and is widely used in clinical settings. Furthermore, cuff-based oscillometric methods provide automatic, noninvasive BP measurements with high accuracy, making them suitable for ambulatory BP monitoring in outpatients.

[0004] However, despite these advantages, this method has drawbacks: it requires the use of a cuff, which can be uncomfortable for patients, and it only provides discrete BP measurements. Furthermore, it cannot rapidly detect hypertension or hypotension, and its ability to capture dynamic BP changes in response to daily physical and mental activities is limited. Therefore, efforts are being made to develop novel, patient-friendly, so-called cuffless BP monitoring methods that enable continuous monitoring.

[0005] In recent decades, rapid advances in noninvasive sensing technologies have led to significant efforts to develop reliable cuffless continuous BP measurement methods, resulting in the emergence of commercially available cuffless BP measurement devices.

[0006] Most conventional cuffless BP monitoring methods utilize the correlation between BP and pulse-wave velocity (PWV), estimated from the use of noninvasive sensing technologies such as electrocardiography (ECG) and photoplethysmography (PPG) sensors, while PWV-based methods refer to the Bramwell-Hill equation, which describes the hemodynamic relationship between BP and PWV.

[0007] At this time, because direct measurement of PWV is difficult, pulse transit time (PTT) and pulse arrival time (PAT), calculated noninvasively from physiological signals such as ECG and PPG, are generally adopted as practical surrogate markers of PWV. PTT represents the pulse wave travel time between two arterial segments, while PAT is the sum of PTT and cardiac pre-ejection time.

[0008] Although these prior art techniques have made remarkable progress in this field, most PTT-based and PAT-based methods require two or more sensors at two locations for accuracy.

[0009] In this regard, another promising technique based on pulse wave analysis (PWA) has emerged for BP estimation, utilizing various features extracted from the waveform morphology of ECG and PPG. Various PWA-based studies have attempted to estimate BP using various machine learning techniques, and for validation, their algorithms have been applied to signals stored in large biosignal databases.

[0010] Furthermore, personalized methods for cuffless BP estimation models that consider individual characteristics have been studied to improve accuracy. Zheng et al. introduced a personalized cuffless BP estimation method using a nonlinear autoregressive model with additional inputs in non-stationary situations such as water drinking (see Y. Zheng et al., "Unobtrusive blood pressure estimation using personalized autoregressive models," 42nd Annu. Int. Conf. IEEE Eng. Med. Biol. Soc.(EMBC). IEEE, 2020, DOI: 10.1109 / EMBC44109.2020.9175635).

[0011] In this prior art, the model was first trained on baseline data from all subjects to determine the basic model structure, and then further trained on individual data to obtain model parameters for personalized purposes. While this prior art reported increased systolic BP (SBP) estimation accuracy in the range of 9.3 to 9.6 mmHg, their study was limited to SBP and only 25 subjects participated in validation.

[0012] In another study, Lin et al. reported a multiple regression model for BP estimation using a personalized optimal feature subset of 13 features (out of a total feature set of 65) based on a feature selection method. The main limitation of this investigation is that the proposed method requires a pretraining process that requires the use of some individual measurements for training before estimation, and the results do not show a significant improvement in accuracy in cuffless BP estimation (see Lin, Wan-Hua, et al., "Towards accurate estimation of cuffless and continuous blood pressure using multi-order derivative and multivariate photoplethysmogram features." Biomedical Signal Processing and Control 63 (2021): 102198).

[0013] Therefore, the problem to be solved by the present invention is to provide a new cuffless blood pressure measurement method and system with high accuracy.

[0014] In order to solve the above problem, the present invention provides a cuffless-based blood pressure estimation modeling method, comprising the steps of: extracting characteristics of bio-data including waveform information necessary for blood pressure measurement among bio-signal data; clustering the extracted bio-data characteristics into a predetermined number of groups; and applying a blood pressure estimation algorithm to the clustered model.

[0015] In one embodiment of the present invention, the biometric data includes waveform data measured from an electrocardiogram (ECG) sensor and a photoplethysmography (PPG) sensor, and the clustering clusters the extracted biometric data using a K-means clustering algorithm.

[0016] In one embodiment of the present invention, the number of the predetermined groups is determined using the elbow method based on calculations within the cluster sum of squares (WSS) for each candidate K.

[0017] In one embodiment of the present invention, the biometric data includes the ratio (d / a) of the d wave to the a wave among five characteristic points (a, b, c, d, e) of the accelerated thymus (APG) waveform of the photoplethysmography sensor (PPG); the amplitude (AID); and the stability index (SI) of the following formula.

[0018]

[0019] (where h is the subject's height, is the time difference between the systolic peak and the diastolic peak / inflection point of the PPG, which is equal to the time it takes for the pulse wave to travel from the heart to the periphery.

[0020] In one embodiment of the present invention, the blood pressure estimation algorithm separates blood pressure into a low frequency component (LFC) and a high frequency component (HFC), and applies a separate estimation algorithm to the low frequency component (LFC) and the high frequency component (HFC).

[0021] In one embodiment of the present invention, the blood pressure estimation algorithm is a combination of an artificial neural network (ANN) and a recurrent neural network (RNN).

[0022] In one embodiment of the present invention, the artificial neural network (ANN) and recurrent neural network (RNN) algorithms are applied to the low frequency component (LFC) and high frequency component (HFC), respectively.

[0023] The present invention also provides a system for implementing the above-described cuffless-based blood pressure estimation modeling method, the system comprising: a biodata extraction unit (100) for extracting characteristics of biodata including waveform information necessary for blood pressure measurement among biosignal data; a cluster unit (200) for clustering biodata characteristics extracted from the biodata extraction unit (100) and grouping them into a preset number of models; and a blood pressure estimation unit (300) for applying a blood pressure estimation algorithm to the grouped models.

[0024] The present invention analyzes characteristic features for clustering and utilizes a clustering-enhanced capless BP estimation method with the PWA technique using data from a comprehensive biosignal database. Specifically, for optimal clustering, data is grouped based on key features that significantly influence BP estimation, using K-means clustering with a focus on d / a and DBW66. This meets the SBP and DBP requirements specified in the BP standard and improves the accuracy of the BP estimation model by combining ANN and RNN.

[0025] FIG. 1 is a step diagram of a cuffless-based blood pressure estimation modeling method according to one embodiment of the present invention.

[0026] Figure 2 is a block diagram of a system for implementing the above-described caprice-based blood pressure estimation modeling method.

[0027] FIG. 3 is a step diagram illustrating an extraction method based on data exclusion according to one embodiment of the present invention.

[0028] Figure 4 shows the statistics of subjects by age group.

[0029] Figure 5 is a diagram showing the elbow method used to determine the number of clusters.

[0030] Figure 6 is a step diagram of a BP estimation algorithm according to one embodiment of the present invention.

[0031] Figure 7 (left) is a diagram showing the distribution of the absolute SHAP values ​​of all features sorted in descending order of importance for clustering, and Figure 7 (right) is a diagram showing the absolute SHAP values ​​for all features.

[0032] Figure 8 is a structural diagram of the optimized model of cluster 2, showing an artificial neural network (a) and a recurrent neural network (b).

[0033] Figure 9 shows the performance comparison results obtained by two types of algorithms (with / without clustering).

[0034] Figure 10 is a diagram showing the accuracy when all test data are included in the BP estimation model trained with clusters.

[0035] Figure 11 shows the results of analysis using Bland-Altman and regression plots for SBP and DBP of the overall BP estimation after clustering, and Figure 12 lists the comparison of BP estimation performance between estimation models (after clustering) in terms of three international BP standards.

[0036] Figure 13 is an external verification result according to an embodiment of the present invention.

[0037] Hereinafter, a preferred embodiment of the present invention will be described in detail with reference to the attached drawings.

[0038] Before describing the present invention in detail, it should be noted that the terms or words used in this specification should not be interpreted as being unconditionally limited to their usual or dictionary meanings, and the inventor of the present invention may appropriately define and use the concepts of various terms in order to describe his or her invention in the best possible manner.

[0039] Furthermore, it should be noted that these terms and words should be interpreted with meanings and concepts that are consistent with the technical idea of ​​the present invention.

[0040] That is, the terms used in this specification are only used to describe preferred embodiments of the present invention, and are not intended to specifically limit the contents of the present invention.

[0041] It should be noted that these terms are defined taking into account the various possibilities of the present invention.

[0042] Additionally, in this specification, a singular expression may include a plural expression unless the context clearly indicates a different meaning.

[0043] Also, it should be noted that even if similarly expressed in plural, it can contain singular meaning.

[0044] Throughout this specification, whenever a component is described as "including" another component, it may mean that the component may further include any other component, rather than excluding any other component, unless specifically stated otherwise.

[0045] Furthermore, if a component is described as being "internal to, connected to, or installed within" another component, it is understood that the component may be directly connected to, or installed in contact with, the other component.

[0046] The present invention provides a modeling method and system for clustering-based caprice BP estimation to solve the above-described problem.

[0047] FIG. 1 is a step diagram of a cuffless-based blood pressure estimation modeling method according to one embodiment of the present invention.

[0048] Referring to FIG. 1, a cuffless-based blood pressure estimation modeling method according to one embodiment of the present invention includes a step of extracting characteristics of biodata including waveform information necessary for blood pressure measurement among biosignal data.

[0049] In one embodiment of the present invention, data including waveform information required for PWA-based blood pressure estimation is extracted from biosignal data, and for this purpose, a data exclusion process described below is performed.

[0050] Thereafter, the extracted biometric data characteristics are clustered and grouped into a predetermined number of models. In one embodiment of the present invention, the data is clustered into six clusters, taking into account 29 waveform characteristic types.

[0051] Afterwards, a blood pressure estimation algorithm is applied to the above grouped model.

[0052] In one embodiment of the present invention, the biometric data includes waveform data measured from an electrocardiogram (ECG) sensor and a photoplethysmography (PPG) sensor, and the present invention constructs a personalized PWA-based cuffless BP estimation model using a clustering approach for each subject instead of a pre-training process, thereby improving the estimation accuracy.

[0053] A clustering method according to one embodiment of the present invention is based on features extracted from the PWA process, and individual features are captured in the BP estimation model of all clusters, and the applicability of the method according to the present invention to a generalized population was confirmed using data from a wide range of biosignal databases.

[0054] The method according to the present invention can be summarized in the following flow.

[0055] Step 1) Waveform data for the study were selected from the comprehensive biosignal database VitalDB, and specific exclusion criteria are described in more detail below.

[0056] Step 2) Features were extracted from the selected waveform data and organized into a predetermined number of clusters.

[0057] Step 3) Finally, the BP estimation model according to the present invention was applied to all clusters to derive BP values. Therefore, the clustering-based BP measurement method and system according to the present invention can omit the pre-training process for individual subjects and also achieve high estimation accuracy. Therefore, the clustering-based approach according to the present invention can be applied practically and clinically for personalized BP estimation.

[0058] Figure 2 is a block diagram of a system for implementing the above-described caprice-based blood pressure estimation modeling method.

[0059] Referring to FIG. 2, a system according to one embodiment of the present invention includes a biodata extraction unit (100) that is provided on a server equipped with a process and extracts characteristics of biodata including waveform information necessary for blood pressure measurement from biosignal data stored externally or internally; a cluster unit (200) that clusters biodata characteristics extracted from the biodata extraction unit (100) and groups them into a preset number of models; and a blood pressure estimation unit (300) that applies a blood pressure estimation algorithm to the grouped models.

[0060] The present invention groups data, taking into account the characteristics of the waveform information required for blood pressure measurement, and applies a blood pressure estimation algorithm to each group. This allows for personalized blood pressure estimation, virtually without the need for a prior training process, by matching individual data characteristics to the clustered groups and applying the blood pressure estimation algorithm.

[0061]

[0062] Example

[0063] A. Data Exclusion

[0064] In one embodiment of the present invention, VitalDB, an open multi-vital signal database provided by the Department of Anesthesiology, Seoul National University Hospital, was used as a data source.

[0065] In one embodiment of the present invention, the Seoul National University Hospital Institutional Review Board (E-2301-064-1394) approved exemption from review by using a de-identified open database. In one embodiment of the present invention, the VitalDB was collected from 6,388 patients and consisted of time-synchronized intraoperative vital signs and other clinical information. In one embodiment of the present invention, Lead II ECG, finger PPG (Photoplethysmography), arterial blood pressure (ABP), and some clinical information from the VitalDB were used, and all vital signs were downsampled to 125 Hz.

[0066] Figure 3 is a step-by-step diagram illustrating an extraction method based on data exclusion according to one embodiment of the present invention. In the present invention, the exclusion criteria for data sources to which the method of the present invention is not applied are applied to screen abnormal or severely distorted data during the feature extraction process for BP estimation.

[0067] Referring to Figure 3, data that did not include ECG, PPG, and ABP sets were first excluded during the data selection process. If the signal data contained recordings exceeding 10 minutes, 10-minute signal data were extracted based on manual inspection. After data selection, the ECG and PPG signals were band-pass filtered at 0.3-25 Hz for ECG and 0.3-10 Hz for PPG.

[0068] To extract biosignal characteristics, 10 adjacent cardiac cycles were ensemble averaged through a sliding window. The ensemble averaged signal was removed if the appropriate cardiac cycle ratio was less than 70%. Subsequently, signal data with 1) SBP ≥ 50 mmHg but < 250 mmHg, 2) diastolic BP (DBP) ≥ 30 mmHg but < 160 mmHg, and 3) BP change of < 30 mmHg over 5 seconds were excluded. After the above exclusion process, data for 1,675 subjects (1,051,462 cardiac cycles) remained.

[0069]

[0070] B. Characteristics of biosignals used for BP estimation

[0071] For the cuffless BP estimation model based on PWA according to the present invention, various features extracted from ECG and PPG signals were utilized. Previously, Yang et al. investigated the contribution of features to cuffless BP estimation using a feature importance ranking method, and therefore, all features of the waveforms used in the study can be found in the reference (SM Yang; et al., "Estimation and validation of arterial blood pressure using photoplethysmogram morphology features in conjunction with pulse arrival time in large open databases" IEEE J. Biomed. Health Inform., vol. 25, no. 4, pp. 1018-1030, 2021, DOI: 10.1109 / JBHI.2020.3009658.).

[0072] Of the 41 features used in this study, 28 were selected based on their high correlation with BP estimation. Additionally, three features were added and two were removed from PPG, resulting in a total of 29 features extracted from ECG and PPG and used as input variables for the BP estimation model. The three additional features are described as follows.

[0073] In one embodiment of the present invention, d / a has a high contribution to determining clusters, which is explained in more detail in the following experimental examples.

[0074]

[0075] 1) PPG second derivative

[0076] The present invention utilized the properties of the second derivative of the PPG, also known as accelerated pulse gait (APG), for PWA. The APG waveform contains five characteristic points: the a-, b-, c-, d-, and e-waves. Since the shape of the APG varies with age and arterial stiffness, various ratios between the amplitudes of the characteristic points and the amplitudes of the a-waves can provide useful hemodynamic information related to blood pressure (see S. R. Alty et al., "Predicting arterial stiffness from the digital volume pulse waveform," IEEE Trans. Bio Med. Eng., vol. 54, no. 12, pp. 2268-2275, 2007, DOI: 10.1109 / tbme.2007.897805.).

[0077] In one embodiment of the present invention, the ratio of the d-wave to the a-wave (d / a) and the amplitude of the d-wave (AID) were adopted as characteristics considered for BP estimation. The ratio d / a tends to decrease with age, as it is reported to reflect decreased arterial stiffness.

[0078]

[0079] 2) Stiffness Index (SI)

[0080] The present invention uses an additional feature called stiffness index (SI), which represents the stiffness of the artery and is defined as follows:

[0081]

[0082] Here, h is the subject's height, is the time difference between the systolic peak and the diastolic peak / inflection point of the PPG, which is equal to the time it takes for the pulse wave to travel from the heart to the periphery. If no clear diastolic peak is observed in the PPG, this point is defined as the point where the first derivative of the PPG is closest to zero. tends to decrease with age as a result of increased aortic stiffness and increased PWV.

[0083]

[0084] C. Clustering method based on waveform characteristics

[0085] The present invention uses a clustering method that divides data into groups based on characteristic similarity and assigns labels to the group data to improve accuracy.

[0086] Clustering is an unsupervised learning technique that divides a dataset into distinct groups based on similarity. The individual feature dependencies used in BP estimation assume that hemodynamic characteristics differ across groups. Therefore, a BP estimation model used for a specific group based on a clustered dataset is closer to a personalized model that optimizes BP estimation, leading to the expectation of improved accuracy.

[0087] Some clustering methods, such as DBSCAN (Density-Based Spatial Clustering with Noise), are not suitable for data with a large number of features and samples.

[0088] Therefore, in this case, it is desirable that the number of clusters be determined in advance, and in one embodiment of the present invention, the K-means algorithm is adopted to cluster data into a predetermined number of groups.

[0089] The K-means clustering used in the present invention, especially the widely used K-means algorithm, is an algorithm that relies on the Euclidean distance between centroids and samples (AK Jain and RC Dubes, Algorithms for Clustering Data. 655 Upper Saddle River, NJ, USA: Prentice-Hall, 1988.)

[0090] A K-means clustering method according to one embodiment of the present invention includes the steps of selecting an arbitrary number as K (the desired number of clusters), arranging K center points in a feature space, assigning samples to the closest center point, updating the center value, and repeating the assignment and update until convergence.

[0091] In one embodiment of the present invention, the K-means++ algorithm was utilized for better initial value setting in the step of arranging K center points in the feature space (D. Arthur and S. Vassilvitskii, "K-means++: The advantages of careful seeding" in Soda '07, Proc. Eighteenth Ann. ACM-SIAM Symp. Discrete Algorithms, Location, Country, Data, 2007, pp. 1027-1035.)

[0092] To ensure algorithm efficiency, the preprocessing step normalizes the features to ensure isotropic behavior in the feature space. The optimal number of clusters, K, was determined using the elbow method, which calculates the cluster sum of squares (WSS) for each candidate K. The optimal K, which represents the point at which the WSS minimization level is released, was selected using this heuristic approach. In the present invention, six clusters were selected using the elbow method (Figure 5).

[0093] Figure 5 is a diagram showing the elbow method used to determine the number of clusters.

[0094] Referring to Fig. 5, the sum of squared distance (WSS) is the Y-axis and the number of clusters is the X-axis. When executing the K-means algorithm, clustering is performed while gradually increasing the number of clusters (K), and the sum of squared distance (WSS) according to the result is calculated and represented in a graph to select the optimal K value. Here, the sum of squared distance (Sum of squared distance) on the Y-axis is the sum of the squared distances between each data point and the center point of the corresponding cluster. The smaller this value is, the better the clustering performance can be determined. However, since an infinite increase in the number of clusters is also a problem, the present invention selects the K value at the point where the WSS minimization level is released.

[0095] Furthermore, Shapley values ​​and Shapley additive explanations (SHAPs) were used to analyze the contribution of each feature to clustering (SM Lundberg and L. Su-In, "A unified approach to interpreting model predictions," Adv. Neural Inform. Process Syst., vol. 30, 2017.).

[0096] Here, the Shapley value is the average marginal contribution of a feature across all possible simultaneous outcomes. This is based on simultaneous game theory, which provides a solution to the problem of how to fairly distribute each player's payoff. The Shapley value is defined as follows:

[0097]

[0098] Here, S is a subset of features (excluding the feature of interest), val(S) is the evaluation of the feature subset S, xj is the feature of interest, and F is the total number of features. After calculating the expectation by excluding a specific feature, the Shapley value analyzes how the specific feature contributes to the given expectation model. The SHAP method described above is a type of explainable artificial intelligence method based on Shapley value, and was used in the present invention to explain the importance of features for clustering. However, since the output of K-means clustering is only a regular label value, in one embodiment of the present invention, a random forest classifier that classifies samples in the same way as the given K-means clustering model used in SHAP analysis was used to apply SHAP.

[0099]

[0100] D. BP estimation algorithm

[0101] In one embodiment of the present invention, frequency separation was performed by separating the low-frequency component (LFC) and high-frequency component (HFC) of the BP and dividing them into the LFC and HFC for separate estimation. This method demonstrated better performance than prior art, and was therefore adopted for BP estimation. To separate the bit-acquired features and BP into the LFC and HFC, 1 Hz cubic spline interpolation was performed. Therefore, the overall procedure for cullless BP estimation is as follows.

[0102] Once the patient's signal data is received, an ensemble average is extracted from the 10-bit ECG and the PPG is extracted for noise reduction. After extracting 29 features from each ensemble average, the patient is classified into clusters ranging from 1 to 6 based on a predetermined clustering criterion. The patient's feature values ​​are then input as variables into an optimized BP estimation model corresponding to the classified cluster. This process is fully automated, and the results are shown in Figure 6.

[0103] Figure 6 is a step diagram of a BP estimation algorithm according to one embodiment of the present invention.

[0104] Referring to Figure 6, 1,051,462 cardiac cycles (features) obtained from 1,675 subjects were randomly divided into two groups (training and testing groups) at a 4:1 ratio within each cluster. The accuracy of all estimated models was tested using 5-fold cross-validation.

[0105]

[0106] 1) Multivariable Linear Regression (LR)

[0107] The multivariate LR model was trained and tested as a post-clustering feature for BP estimation using least squares based on a MATLAB (version R2020a, MathWorks, Natick, MA, USA) script as a simple model for comparison with other algorithms.

[0108]

[0109] 2) Extra Tree Regression (ETR)

[0110] The bagging method used for regression analysis is the Extra Tree method, also known as the Extreme Tree method. ETR uses an algorithm similar to that used in random forests, but is less sensitive to specific characteristics. This helps prevent model overfitting, and because it randomly splits nodes without using bootstrap samples, computation time is relatively low. In one embodiment of the present invention, grid search was used to identify optimal hyperparameters. The model used a minimum of 100 samples to determine leaf nodes, a maximum of 10,000 leaf nodes, and 600 trees, and was trained in a Python environment using Scikit-Learn.

[0111]

[0112] 3) 1D Squeeze U-Net

[0113] As described in the prior art, 1D Squeeze U-Net was used for real-time cuffless BP estimation to facilitate accurate comparison with other methods (Athaya, T., & Choi, S. (2022). Real-Time Cuffless Continuous Blood Pressure Estimation Using 1D Squeeze U-Net Model: A Progress toward mHealth. Biosensors, 12(8), 655.).

[0114] BP estimation based on the 1D Squeeze U-Net relies on the Squeeze U-Net model and utilizes a modified 1D Squeeze U-Net architecture. Initially, cross-correlation is used to remove the phase lag between the 1-channel raw PPG input and the ABP, and phase segments consisting of 256 artifact-free samples are used for training. The architectural components, including the contraction and expansion pathways, are consistent with those described in detail below. (See

[0031] Athaya, T., & Choi, S. (2022). Real-Time Cuffless Continuous Blood Pressure Estimation Using 1D Squeeze U-Net Model: A Progress toward mHealth. Biosensors, 12(8), 655.)

[0115]

[0116] 4) Combination of artificial neural networks (ANN) and recurrent neural networks (RNN)

[0117] Separating the estimation of LFC and HFC of BP using ANN and RNN at each frequency component can achieve significant performance improvement compared to regression analysis (LR) (see SM Yang; et al, “and validation of arterial blood pressure using photoplethysmogram morphology features in conjunction with pulse arrival time in large open databases” IEEE J. Biomed. Health Inform., vol. 25, no. 4, pp. 1018-1030, 2021, DOI: 10.1109 / JBHI.2020.3009658.).

[0118] To optimize the model architecture for each cluster, we performed hyperparameter tuning on ANNs and RNNs for LFC and HFC of the BP function and tested them in terms of BP estimation accuracy. The architecture for testing was set as follows.

[0119] For ANNs with a fixed single input / output layer, the number of hidden layers was tested from 1 to 5, the step dropout rate was tested from 0.2 to 0.5 (in 0.1-step increments), the number of nodes was tested from 100 to 1,000 (in 100 steps), and the learning rate was tested from 0.001 to 0.01 (in 0.001 steps). The batch size was 512, and He initialization with batch normalization was used. Adam and rectified linear units were selected as the optimizer and activation functions, respectively.

[0120] For RNNs, a many-to-many approach using bidirectional LSTM (Long-Term Memory) was adopted. For a single input-output layer, hidden layers were tested from 1 to 5 (consistency steps). The dropout rate was tested from 0.2 to 0.5 (0.1 steps), the number of nodes was tested from 10 to 50 (10 steps), and the learning rate was tested from 0.01 to 0.1 (0.1 steps). The batch size and time steps were fixed at 512 and 250, respectively, and He initialization and layer normalization were performed. Adam and tanh were adopted as the optimizer and activation function, respectively. Training and testing were performed in Python and PyTorch environments.

[0121] In one embodiment of the present invention, the final blood pressure values ​​(SBP, DBP) were determined by applying ANN to the LFC of each feature and RNN to the HFC and adding the blood pressure values ​​estimated by each model.

[0122]

[0123] result

[0124] A. Contribution of features for clustering

[0125] Samples from 29 feature sets representing 1,051,462 cardiac cycles extracted from VitalDB were clustered into six clusters using K-means clustering, as described above. The proportions of samples included in the six clusters were 22.6%, 18.3%, 16.5%, 15.4%, 14.2%, and 13%, respectively. The contributions of individual features to the clustering are shown in Figure 7.

[0126] Figure 7 (left) is a diagram showing the distribution of the absolute SHAP values ​​of all features sorted in descending order of importance for clustering, and Figure 7 (right) is a diagram showing the absolute SHAP values ​​for all features.

[0127] Referring to Figure 7, d / a, DBW66, AS, Pip, and SI are the top five features that contributed most to clustering among the 28 features. In particular, the average absolute SHAP values ​​of d / a and DBW66 were 3.45 and 2.94, respectively, while AS significantly decreased to 1.83. This suggests that the two features, d / a and DBW66, contributed significantly to clustering compared to the other features.

[0128] Here, DBW stands for Diastolic BandWidth, and refers to the time gap between two points in the PPG signal. That is, it refers to the difference between the time point that falls to the right of the peak among the two points where the amplitude is 60% of the peak point of each pulse and the time point corresponding to the peak.

[0129]

[0130] B. Optimization of the BP estimation model using a combination of ANN and RNN

[0131] To optimize the combination of ANN and RNN-based BP estimation model architectures, hyperparameter tuning was performed for all six clusters. The hyperparameters that yielded the best BP estimation performance were selected. For the LFC ANN model, the number of hidden layers varied from 1 to 2, the dropout rate varied from 0.2 to 0.3, and the number of nodes in each cluster varied from 200 to 300. The learning rate was fixed at 0.001. For the HFC RNN model, the number of nodes varied from 40 to 50, and the dropout rate, time steps, and learning rate were fixed at 0.3, 250, and 0.01, respectively. An example of the BP estimation architecture for the ANN and RNN in cluster 2 is shown in Figure 8.

[0132] Figure 8 is a structural diagram of the optimized model of cluster 2, showing an artificial neural network (a) and a recurrent neural network (b).

[0133]

[0134] C. Accuracy of BP estimation model after clustering

[0135] The performance results of each BP estimation model with and without clustering in terms of ME, SDE, and MAD were analyzed and are shown in Figure 9.

[0136] In this experimental example, the same clustering and training data set were used across all methods. Figure 9 compares the performance of two algorithms (with and without clustering) for all estimation models. All clustering algorithms yielded better performance for BP estimation in ME, SDE, and MAD compared to the non-clustering algorithms.

[0137] To analyze the performance of all clusters, we examined the results of the ANN + RNN model, which produced the best performance among the three BP estimation models. The model performance for all clusters and the total error with clustering were superior to those of the model without clustering in SBP. The BP estimation model after clustering also performed better for DBP in the SDE and MAD cases, but the ME results were slightly higher.

[0138] Figure 10 presents the accuracy of a cluster-trained BP estimation model when all test data are included. Comparing the results listed in Figures 9 and 10 , it can be seen that the accuracy of BP estimation using input data clustering outperforms estimation without input data clustering in terms of SDE and MAD.

[0139] Figure 11 shows the results of analysis using Bland-Altman and regression plots for SBP and DBP of the overall BP estimation after clustering, and Figure 12 lists the comparison of BP estimation performance between estimation models (after clustering) in terms of three international BP standards.

[0140] Referring to Figures 11 and 12, the performance of the proposed model satisfied all three standards' criteria, while the BP estimation model without clustering did not meet the IEEE standard criteria in terms of MAD (5.51 > 5 mmHg). The SDE of the BP estimation model with clustering yielded the same improvement (compared to the estimation without clustering) of 0.85 mmHg for SBP and 0.67 mmHg for DBP, while the MAD yielded a better improvement of 0.79 mmHg for SBP and 0.35 mmHg for DBP.

[0141] Additionally, we show the estimation error for the regression and Bland-Altman plots for the results of the BP estimation model combining ANN and RNN, and this model showed the best performance among the implemented estimation methods.

[0142]

[0143] D. External validation of the BP estimation model

[0144] The BP estimation model performed best when using a combination of clustering, ANN, and RNN, as shown in Table 3. To evaluate the performance of this optimized model and ensure that it did not overfit, we used an external database, MIMIC III, to evaluate it. A subset of MIMIC III data was selected based on the exclusion criteria specifically described in the Methods section, providing clinical information. The data were preprocessed and feature extracted using the same methodology described above. The dataset consisted of 513,623 bit pairs across 236 records.

[0145] The external validation results according to this example are summarized in Figure 13. It can be seen that when clustering is applied, the estimation accuracy significantly improves, exceeding the standards set by three international BP organizations.

[0146] As described above, the present invention analyzes the characteristic features for clustering and utilizes data from a comprehensive biosignal database, utilizing the PWA technique and a clustering-enhanced capless BP estimation method. As a result, the accuracy of BP estimation is significantly improved. Specifically, data is grouped based on key features that significantly influence BP estimation, using K-means clustering with a focus on d / a and DBW66 for optimal clustering. This satisfies the SBP and DBP requirements specified in the BP standard, and by combining ANN and RNN, the accuracy of the BP estimation model is improved.

Claims

1. A capless-based blood pressure estimation modeling method, A step of extracting characteristics of bio-data including waveform information required for blood pressure measurement among bio-signal data; A step of clustering the extracted biometric data characteristics into a preset number of groups; and A cuffless-based blood pressure estimation modeling method comprising a step of applying a blood pressure estimation algorithm to the above clustered model.

2. In paragraph 1, A cuffless-based blood pressure estimation modeling method, characterized in that the above biometric data includes waveform data measured from an electrocardiogram (ECG) sensor and a photoplethysmography (PPG) sensor.

3. In paragraph 1, A cuffless-based blood pressure estimation modeling method characterized in that the above clustering clusters the extracted biodata using a K-means clustering algorithm.

4. In paragraph 1, A cuffless-based blood pressure estimation modeling method, characterized in that the number of the above-described groups is determined using the elbow method based on calculations within the cluster sum of squares (WSS) for each candidate K.

5. In paragraph 1, The above biometric data is, A cuff-based blood pressure estimation modeling method characterized by including a ratio (d / a) of the d wave to the a wave among five characteristic points (a, b, c, d, e) of the accelerated thymus (APG) waveform of the photoplethysmography sensor (PPG); an amplitude (AID); and a stiffness index (SI) of the following formula. (where h is the subject's height, is the time difference between the systolic peak and the diastolic peak / inflection point of the PPG, which is equal to the time it takes for the pulse wave to travel from the heart to the periphery.

6. In paragraph 1, The above blood pressure estimation algorithm is a capless-based blood pressure estimation modeling method characterized in that it separates blood pressure into a low-frequency component (LFC) and a high-frequency component (HFC), and applies a separate estimation algorithm to the low-frequency component (LFC) and the high-frequency component (HFC).

7. In paragraph 6, The above blood pressure estimation algorithm is a cuffless-based blood pressure estimation modeling method characterized by a combination of an artificial neural network (ANN) and a recurrent neural network (RNN).

8. In paragraph 6, A cuffless-based blood pressure estimation modeling method, characterized in that the above artificial neural network (ANN) and recurrent neural network (RNN) algorithms are applied to the low frequency component (LFC) and high frequency component (HFC), respectively.

9. A system for implementing a caprice-based blood pressure estimation modeling method according to any one of clauses 1 to 8.

10. In paragraph 9, the system, A biodata extraction unit (100) that extracts the characteristics of biodata including waveform information required for blood pressure measurement among biosignal data; A cluster unit (200) that clusters the biodata characteristics extracted from the biodata extraction unit (100) and groups them into a preset number of models; and A system for implementing a cuffless-based blood pressure estimation modeling method, characterized in that it includes a blood pressure estimation unit (300) that applies a blood pressure estimation algorithm to the above grouped model.

Images