Method for classifying and detecting blood pressure by inductive transfer learning of dual-channel non-invasive blood pressure
Patent Information
- Application Number
- CN202310888715.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-19
- Publication Date
- 2026-11-03
- Estimated Expiration
- 2043-07-19
AI Technical Summary
Existing non-invasive blood pressure detection methods suffer from low accuracy, cumbersome manual feature extraction, reliance on specific individuals, and limited training data, making it difficult to achieve high-precision blood pressure detection.
A dual-channel method of inductive transfer learning is adopted. By preprocessing the ECG signal and photoplethysmography pulse wave signal, a deep learning network is built for feature extraction. A five-class classification model of arrhythmia and a two-class classification model of blood pressure are constructed. A one-dimensional convolutional neural network is then used for regression to achieve non-invasive blood pressure detection.
It improves the accuracy and generalization ability of blood pressure detection, can automatically select models for regression, and output specific systolic and diastolic blood pressure values, meeting the detection standards of mean absolute error and root mean square error.
Smart Images
Figure CN117084648B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of blood pressure classification and detection, and deep learning technology, and in particular to a dual-channel non-invasive blood pressure classification and detection method based on inductive transfer learning. Background Technology
[0002] Over the years, with the continuous improvement of living standards, many people's dietary structure has changed, with an increase in fast food high in sugar, salt, and fat. Coupled with the high stress levels of modern life, and the increasing prevalence of staying up late and lack of exercise, many people are experiencing high blood pressure. Blood pressure is a crucial physiological parameter for the human body. If left unattended in its early stages, it can develop into hypertension, causing damage to the cardiovascular system. Hypertension is considered a major risk factor for cardiovascular disease. Controlling hypertension can prevent the occurrence of cardiovascular disease and reduce patient mortality. Furthermore, the age of onset and the number of people with hypertension are showing a gradual downward trend, while the trend is upward. Therefore, blood pressure monitoring is particularly important. Currently, common blood pressure measurement methods can be divided into invasive and non-invasive methods. Invasive blood pressure measurement involves inserting a catheter directly into a central or peripheral artery to obtain blood pressure values using a pressure sensor. This method provides relatively accurate blood pressure readings, but it is prone to complications such as thrombosis, ischemic injury, and arteritis. Non-invasive blood pressure measurement methods, including arterial tension method and volume compensation method, overcome the drawbacks of invasive methods to some extent, but still have the disadvantage of low accuracy.
[0003] Current research on blood pressure monitoring primarily utilizes deep learning methods for non-invasive blood pressure measurement. Normalized photoplethysmograph (PPG) and electrocardiogram (ECG) signals contain rich physiological information related to blood pressure. Blood pressure information is obtained by building detection models based on PPG and ECG signals, thus enabling blood pressure measurement. However, these methods currently suffer from limitations such as cumbersome manual feature extraction, reliance on specific individuals, and limited training data, resulting in low accuracy of blood pressure detection models. In the context of today's medical-engineering integration, how to leverage the powerful data processing capabilities of computers and deep learning technology in artificial intelligence to design a more convenient non-invasive blood pressure detection algorithm, and build a blood pressure detection model with strong generalization ability and high accuracy, allowing the public to have a general understanding of their blood pressure and thus improving the ability to detect the risk of hypertension, is an urgent problem to be solved. Summary of the Invention
[0004] This invention proposes a dual-channel non-invasive blood pressure classification and detection method based on inductive transfer learning, which can improve non-invasive blood pressure classification through inductive transfer learning, thereby improving detection accuracy.
[0005] The present invention adopts the following technical solution.
[0006] A dual-channel non-invasive blood pressure classification and measurement method based on inductive transfer learning includes the following steps;
[0007] Step S1: Perform data preprocessing on the electrocardiogram (ECG), photoplethysmograph (PPG), and arterial blood pressure (ABP) signals from the sleeveless blood pressure estimation database;
[0008] Step S2: Extract electrocardiogram waveform features from the ECG signal processed in Step S1 using a deep learning network. The features include RR interval, R wave amplitude, ST segment interval, etc., and build a five-class classification model of arrhythmia, specifically N, S, V, F and Q classes, as the source domain model for inductive transfer learning.
[0009] Step S3: Perform inductive transfer on the source domain model built in step S2 to construct a blood pressure binary classification model. Build a binary classification model Model1 for normal blood pressure + prehypertension vs. hypertension, and a binary classification model Model2 for normal blood pressure vs. prehypertension.
[0010] Step S4: The ECG and PPG signals processed in S1 are re-labeled, and features are extracted using a one-dimensional convolutional neural network to construct a regression model. Normal blood pressure data is input into the model, and regressions are performed on systolic and diastolic blood pressure respectively to construct two regression models, Model3 and Model4. Similarly, two models, Model5 and Model6, are obtained by inputting prehypertension data, and two models, Model7 and Model8, are obtained by inputting hypertension data. A total of 6 regression models are trained.
[0011] Step S5: Input the PPG and ECG dual-channel test data into the secondary binary classification model built in S3. After the model automatically determines the category, it automatically selects the corresponding regression model in S4 for regression to obtain the specific systolic blood pressure (SBP) and diastolic blood pressure (DBP) values, and obtains the detection results of the regression model on the mean absolute error (MAE), standard deviation (SD), and root mean square error (RMSE) indicators.
[0012] Step S1 specifically includes the following steps;
[0013] Step S11: PPG, ECG and ABP signals from the sleeveless blood pressure estimation database are collected simultaneously to synchronize them in the time dimension. The three signals corresponding in time are cut into data of the same length without overlap. The cut PPG and ECG signals are used as input data for model training, and the corresponding ABP signals are used to determine the data label for that time period.
[0014] Step S12: Record the maximum and minimum values of each ABP data segment, which are the systolic blood pressure (SBP) and diastolic blood pressure (DBP) of that data segment, respectively. According to the 7th report of the Joint National Committee on Prevention, Detection, Evaluation and Treatment of Hypertension (JNC7), blood pressure is divided into three categories: systolic blood pressure less than 120 mmHg, the data segment is marked as normal blood pressure; systolic blood pressure in the range of 120-139 mmHg, the data segment is marked as prehypertension; systolic blood pressure greater than or equal to 140 mmHg, the data segment is marked as hypertension.
[0015] Step S13: Process the ECG signal using discrete wavelet transform and Mallat algorithm. A fixed threshold function, sqtwolog, is used to obtain multi-level ECG decomposed signals of different frequency bands. The threshold (Thr) of the sqtwolog function is... , It is the noise standard deviation of the layer detail factor. , in It is the median of the absolute values of the high-frequency coefficients of each wavelet layer, arranged in ascending order. for:
[0016] ;
[0017] Step S14: Use the threshold and threshold function sqtwolog to perform threshold processing on the decomposed signal to filter out ECG noise signals;
[0018] Step S15: Reconstruct the multi-layer decomposition signal after filtering out noise to obtain the filtered ECG signal;
[0019] Step S16: Process the PPG signal using a median filter to obtain the baseline drift noise of the PPG signal;
[0020] Step S17: Subtract the extracted baseline drift noise from the original PPG signal to filter out the PPG noise signal.
[0021] In step S11, the data storage method of the sleeveless blood pressure estimation database is MATLAB's .mat format, the cutting length is selected to be within 10 seconds, and the data are not cut overlappingly.
[0022] In step S13, after data segmentation, the segmented ECG and PPG data are denoised separately. Noise in the ECG signal includes baseline drift, power line interference, and electromyography (EMG) interference. The baseline drift frequency is less than 1 Hz, the power line interference frequency is 50 Hz, and the EMG interference frequency range is between 10-3000 Hz. The main frequency of the ECG signal is between 0.05-100 Hz, with most of the energy concentrated between 0.5-45 Hz. Discrete wavelet transform is used to decompose the ECG signal and remove noise. The calculation formula for discretized wavelet transform is as follows:
[0023]
[0024] f(t) is the function that needs to be analyzed by wavelet analysis. Describe the wavelet basis functions. It is a scaling factor, and its value is greater than 0. It is a displacement, which can be positive or negative;
[0025] The wavelet thresholding method is used to remove noise by performing a 7-level wavelet decomposition on the ECG signal using the wavelet basis function "db8" and setting the frequency of noise to zero.
[0026] Step S2 is as follows:
[0027] Step S21: Divide the ECG data of lead II and lead V1 into heart rate categories according to the American Association for the Advancement of Medical Devices (AAMI) standards, ensuring that the training set and test set data do not overlap at all;
[0028] Step S22: Construct a deep learning network consisting of M convolutional layers and residual network layers with cross-layer connections;
[0029] Step S23: Extract waveform features from ECG data using the deep learning network built in S22, and build a five-class classification model for heart rate arrhythmia with high accuracy;
[0030] Step S24: Freeze the parameters of the first N convolutional layers of the heart rate arrhythmia five-classification model built in S23, and unfreeze the parameters of the last (MN) convolutional layers to serve as the source domain model for inductive transfer learning.
[0031] Step S3 is as follows:
[0032] Step S31: Classify the processed PPG signal dataset and ECG signal dataset into normal data, prehypertension data, and hypertension data;
[0033] Step S32: Merge the normal data and prehypertension data into a single category dataset, and use it together with the hypertension dataset as the dual-channel input data for the first binary classification transfer model, Model 1.
[0034] Step S33: Use normal data and prehypertension data as dual-channel input data for the second binary classification transfer model, Model 2;
[0035] Step S34: Change the output layer of the source domain model from five-class classification to two-class classification, and connect it with a fully connected layer to build the target domain model;
[0036] Step S35: Train the transfer model built in S34 using the Tensorflow deep learning framework. The input data uses dual-channel data classified in S32 and S33 respectively. During training, the cross-entropy loss function and stochastic gradient descent optimizer SGD are used. The evaluation metric is categorical accuracy. Build a binary classification model Model1 for normal blood pressure + prehypertension vs. hypertension, and a binary classification model Model2 for normal blood pressure vs. prehypertension.
[0037] In step S3, the dataset is first classified into normal data, prehypertension data and hypertension data. Let the dataset consisting of normal blood pressure data be N_dataset, the dataset consisting of prehypertension data be Pre_dataset, and the dataset consisting of hypertension data be H_dataset.
[0038] When training the first binary classification model, Model1, N_dataset and Pre_dataset are merged into a single class dataset, which is then input into the model along with the H_dataset for training. When training the second binary classification model, Model2, N_dataset and Pre_dataset are input into the model for training. During training, the training and validation sets are divided in an 8:2 ratio.
[0039] The source domain model is designed to classify arrhythmias into five categories, while the target domain for transfer learning is to classify blood pressure into two categories. Therefore, the number of neurons in the output layer is changed from 5 to 2. The network parameters of the first 15 layers are frozen, the parameters of the last 2 convolutional layers are unfrozen, and a fully connected layer is added to obtain the target domain model.
[0040] The transfer learning model was trained using the Tensorflow deep learning framework. During training, the cross-entropy loss function and stochastic gradient descent optimizer (SGD) were used. The evaluation metric was categorical accuracy. Both models were trained for 60 epochs, and the best-performing epoch was saved. A binary classification model Model1 was built to classify normal blood pressure + prehypertension vs. hypertension, and a binary classification model Model2 was built to classify normal blood pressure vs. prehypertension.
[0041] Step S4 is as follows:
[0042] Step S41: Relabel the dataset using the specific values of systolic and diastolic blood pressure. Each data sample has two labels: systolic blood pressure value and diastolic blood pressure value.
[0043] Step S42: Randomly shuffle the normal blood pressure data, prehypertension data, and hypertension data respectively, and divide them into training set and validation set in a 7:3 ratio;
[0044] Step S43: Input the ECG and PPG signals into a 1D-CNN for feature extraction and then perform blood pressure regression. Input normal blood pressure data into the model and perform regression on systolic and diastolic blood pressure respectively to obtain two models, Model3 and Model4. Similarly, input prehypertension data to obtain two models, Model5 and Model6, and input hypertension data to obtain two models, Model7 and Model8, for a total of 6 regression models trained.
[0045] In step S43, a 1D-CNN network is constructed, which contains four convolutional layers using the ReLU activation function, followed by a Batch Normalization layer to batch normalize the data output by the activation layers, thereby reducing the network's sensitivity to data and accelerating the network convergence process.
[0046] ECG and PPG signals were input into a pre-built 1D-CNN for feature extraction and then for blood pressure regression. During training, the Adam optimizer was used, with the loss function set to mean squared error (MSE) and the evaluation metric being mean absolute error (MSE). A step-decreasing learning rate was employed. Normal blood pressure data was input into the model, and regressions were performed on systolic and diastolic blood pressure, resulting in two models, Model 3 and Model 4. Similarly, prehypertension data yielded two models, Model 5 and Model 6, and hypertension data yielded two models, Model 7 and Model 8, resulting in a total of six regression models trained.
[0047] Step S5 is as follows:
[0048] Step S51: Input the dual-channel test data into the binary classification model Model1 for normal blood pressure + prehypertension vs. hypertension. The model will automatically output the category.
[0049] Step S52: If the output of S51 is hypertension, it means that the data belongs to hypertension; if the output of S51 is normal blood pressure + prehypertension, then input this type of data into the second-level binary classification model Model2 of normal blood pressure vs. prehypertension to determine whether the data belongs to the normal blood pressure category or the prehypertension category.
[0050] Step S53: After the model outputs the classification results, the data will be automatically input into the corresponding regression model, and the regression model will output the specific values of SBP and DBP respectively.
[0051] Step S54: Based on the detection results in S53, obtain the regression model's performance in terms of Mean Absolute Error (MAE), Standard Deviation (SD), and Root Mean Square Error (RMSE), thereby evaluating the accuracy and effectiveness of the method. The formula for calculating the Mean Absolute Error (MAE) is as follows:
[0052]
[0053] The formula for calculating the standard deviation (SD) is:
[0054]
[0055] in,
[0056] The formula for the root mean square error (RMSE) is:
[0057] .
[0058] The mean absolute error (MAE) is the average of the absolute values of the error. The mean absolute error (MAE) should meet the requirements of the AAMI standard for the overall performance of electronic or automatic blood pressure monitors. Specifically, when systolic and diastolic blood pressure are processed separately, the average difference between the paired measurements of the test system and the comparison system should be ±5 mmHg or less.
[0059] The root mean square error index (RMSE) is the arithmetic square root of the root mean square error. The root mean square error is the expected value of the square of the difference between the measured value of the test system and the true value of the comparison system. It is used to evaluate the degree of variation of the data. The smaller the value, the higher the accuracy of the experimental data measured by the test system.
[0060] This invention proposes a method for implementing a dual-channel non-invasive blood pressure classification and detection model using inductive transfer learning. The method includes: preprocessing ECG signals, photoplethysmography (PPG) pulse wave signals, and arterial blood pressure signals from a cuffless blood pressure estimation database; constructing a five-class source domain model for arrhythmias using a deep learning network; transferring the constructed source domain model to build three binary blood pressure classification models; constructing a regression model using a one-dimensional convolutional neural network, inputting normal blood pressure, prehypertension, and hypertension data into the model, and performing regression on systolic and diastolic blood pressure respectively, constructing six regression models; inputting PPG and ECG dual-channel test data into the constructed binary classification model, the model automatically determines the category and automatically selects the corresponding regression model for regression, and outputs the specific systolic and diastolic blood pressure values; achieving the effect of non-invasive blood pressure classification and detection, this invention can improve the accuracy of blood pressure detection.
[0061] Compared with existing technologies, this invention adopts dual-channel data input, making full use of the blood pressure information contained in electrocardiogram signals and photoplethysmography (PPG) pulse wave signals. It builds a non-invasive blood pressure classification model and a high-precision blood pressure detection regression model through inductive transfer learning, realizes automatic model selection, and significantly improves classification performance. It provides a methodology for exploring non-invasive blood pressure classification and detection, and makes up for the neglect of potential hypertension risks due to the limitations of existing technologies. Attached Figure Description
[0062] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:
[0063] Appendix Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;
[0064] Appendix Figure 2 This is a schematic diagram of the transfer model training process according to an embodiment of the present invention;
[0065] Appendix Figure 3 This is a schematic diagram of the regression model training process according to an embodiment of the present invention. Detailed Implementation
[0066] The specific implementation details of this patented method will now be described in more detail with reference to the accompanying drawings.
[0067] The accompanying drawings illustrate exemplary embodiments of this disclosure; however, it should be understood that this disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. These embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.
[0068] As shown in the figure, the method of dual-channel non-invasive blood pressure classification and blood pressure detection based on inductive transfer learning includes the following steps;
[0069] Step S1: Perform data preprocessing on the electrocardiogram (ECG), photoplethysmograph (PPG), and arterial blood pressure (ABP) signals from the sleeveless blood pressure estimation database;
[0070] Step S2: Extract electrocardiogram waveform features from the ECG signal processed in Step S1 using a deep learning network. The features include RR interval, R wave amplitude, ST segment interval, etc., and build a five-class classification model of arrhythmia, specifically N, S, V, F and Q classes, as the source domain model for inductive transfer learning.
[0071] Step S3: Perform inductive transfer on the source domain model built in step S2 to construct a blood pressure binary classification model. Build a binary classification model Model1 for normal blood pressure + prehypertension vs. hypertension, and a binary classification model Model2 for normal blood pressure vs. prehypertension.
[0072] Step S4: The ECG and PPG signals processed in S1 are re-labeled, and features are extracted using a one-dimensional convolutional neural network to construct a regression model. Normal blood pressure data is input into the model, and regressions are performed on systolic and diastolic blood pressure respectively to construct two regression models, Model3 and Model4. Similarly, two models, Model5 and Model6, are obtained by inputting prehypertension data, and two models, Model7 and Model8, are obtained by inputting hypertension data. A total of 6 regression models are trained.
[0073] Step S5: Input the PPG and ECG dual-channel test data into the secondary binary classification model built in S3. After the model automatically determines the category, it automatically selects the corresponding regression model in S4 for regression to obtain the specific systolic blood pressure (SBP) and diastolic blood pressure (DBP) values, and obtains the detection results of the regression model on the mean absolute error (MAE), standard deviation (SD), and root mean square error (RMSE) indicators.
[0074] Step S1 specifically includes the following steps;
[0075] Step S11: PPG, ECG and ABP signals from the sleeveless blood pressure estimation database are collected simultaneously to synchronize them in the time dimension. The three signals corresponding in time are cut into data of the same length without overlap. The cut PPG and ECG signals are used as input data for model training, and the corresponding ABP signals are used to determine the data label for that time period.
[0076] Step S12: Record the maximum and minimum values of each ABP data segment, which are the systolic blood pressure (SBP) and diastolic blood pressure (DBP) of that data segment, respectively. According to the 7th report of the Joint National Committee on Prevention, Detection, Evaluation and Treatment of Hypertension (JNC7), blood pressure is divided into three categories: systolic blood pressure less than 120 mmHg, the data segment is marked as normal blood pressure; systolic blood pressure in the range of 120-139 mmHg, the data segment is marked as prehypertension; systolic blood pressure greater than or equal to 140 mmHg, the data segment is marked as hypertension.
[0077] Step S13: Process the ECG signal using discrete wavelet transform and Mallat algorithm. A fixed threshold function, sqtwolog, is used to obtain multi-level ECG decomposed signals of different frequency bands. The threshold (Thr) of the sqtwolog function is... , It is the noise standard deviation of the layer detail factor. , in It is the median of the absolute values of the high-frequency coefficients of each wavelet layer, arranged in ascending order. for:
[0078] ;
[0079] Step S14: Use the threshold and threshold function sqtwolog to perform threshold processing on the decomposed signal to filter out ECG noise signals;
[0080] Step S15: Reconstruct the multi-layer decomposition signal after filtering out noise to obtain the filtered ECG signal;
[0081] Step S16: Process the PPG signal using a median filter to obtain the baseline drift noise of the PPG signal;
[0082] Step S17: Subtract the extracted baseline drift noise from the original PPG signal to filter out the PPG noise signal.
[0083] In step S11, the data storage method of the sleeveless blood pressure estimation database is MATLAB's .mat format, the cutting length is selected to be within 10 seconds, and the data are not cut overlappingly.
[0084] In step S13, after data segmentation, the segmented ECG and PPG data are denoised separately. Noise in the ECG signal includes baseline drift, power line interference, and electromyography (EMG) interference. The baseline drift frequency is less than 1 Hz, the power line interference frequency is 50 Hz, and the EMG interference frequency range is between 10-3000 Hz. The main frequency of the ECG signal is between 0.05-100 Hz, with most of the energy concentrated between 0.5-45 Hz. Discrete wavelet transform is used to decompose the ECG signal and remove noise. The calculation formula for discretized wavelet transform is as follows:
[0085]
[0086] f(t) is the function that needs to be analyzed by wavelet analysis. Describe the wavelet basis functions. It is a scaling factor, and its value is greater than 0. It is a displacement, which can be positive or negative;
[0087] The wavelet thresholding method is used to remove noise by performing a 7-level wavelet decomposition on the ECG signal using the wavelet basis function "db8" and setting the frequency of noise to zero.
[0088] Step S2 is as follows:
[0089] Step S21: Divide the ECG data of lead II and lead V1 into heart rate categories according to the American Association for the Advancement of Medical Devices (AAMI) standards, ensuring that the training set and test set data do not overlap at all;
[0090] Step S22: Construct a deep learning network consisting of M convolutional layers and residual network layers with cross-layer connections;
[0091] Step S23: Extract waveform features from ECG data using the deep learning network built in S22, and build a five-class classification model for heart rate arrhythmia with high accuracy;
[0092] Step S24: Freeze the parameters of the first N convolutional layers of the heart rate arrhythmia five-classification model built in S23, and unfreeze the parameters of the last (MN) convolutional layers to serve as the source domain model for inductive transfer learning.
[0093] Step S3 is as follows:
[0094] Step S31: Classify the processed PPG signal dataset and ECG signal dataset into normal data, prehypertension data, and hypertension data;
[0095] Step S32: Merge the normal data and prehypertension data into a single category dataset, and use it together with the hypertension dataset as the dual-channel input data for the first binary classification transfer model, Model 1.
[0096] Step S33: Use normal data and prehypertension data as dual-channel input data for the second binary classification transfer model, Model 2;
[0097] Step S34: Change the output layer of the source domain model from five-class classification to two-class classification, and connect it with a fully connected layer to build the target domain model;
[0098] Step S35: Train the transfer model built in S34 using the Tensorflow deep learning framework. The input data uses dual-channel data classified in S32 and S33 respectively. During training, the cross-entropy loss function and stochastic gradient descent optimizer SGD are used. The evaluation metric is categorical accuracy. Build a binary classification model Model1 for normal blood pressure + prehypertension vs. hypertension, and a binary classification model Model2 for normal blood pressure vs. prehypertension.
[0099] In step S3, the dataset is first classified into normal data, prehypertension data and hypertension data. Let the dataset consisting of normal blood pressure data be N_dataset, the dataset consisting of prehypertension data be Pre_dataset, and the dataset consisting of hypertension data be H_dataset.
[0100] When training the first binary classification model, Model1, N_dataset and Pre_dataset are merged into a single class dataset, which is then input into the model along with the H_dataset for training. When training the second binary classification model, Model2, N_dataset and Pre_dataset are input into the model for training. During training, the training and validation sets are divided in an 8:2 ratio.
[0101] The source domain model is designed to classify arrhythmias into five categories, while the target domain for transfer learning is to classify blood pressure into two categories. Therefore, the number of neurons in the output layer is changed from 5 to 2. The network parameters of the first 15 layers are frozen, the parameters of the last 2 convolutional layers are unfrozen, and a fully connected layer is added to obtain the target domain model.
[0102] The transfer learning model was trained using the Tensorflow deep learning framework. During training, the cross-entropy loss function and stochastic gradient descent optimizer (SGD) were used. The evaluation metric was categorical accuracy. Both models were trained for 60 epochs, and the best-performing epoch was saved. A binary classification model Model1 was built to classify normal blood pressure + prehypertension vs. hypertension, and a binary classification model Model2 was built to classify normal blood pressure vs. prehypertension.
[0103] Step S4 is as follows:
[0104] Step S41: Relabel the dataset using the specific values of systolic and diastolic blood pressure. Each data sample has two labels: systolic blood pressure value and diastolic blood pressure value.
[0105] Step S42: Randomly shuffle the normal blood pressure data, prehypertension data, and hypertension data respectively, and divide them into training set and validation set in a 7:3 ratio;
[0106] Step S43: Input the ECG and PPG signals into a 1D-CNN for feature extraction and then perform blood pressure regression. Input normal blood pressure data into the model and perform regression on systolic and diastolic blood pressure respectively to obtain two models, Model3 and Model4. Similarly, input prehypertension data to obtain two models, Model5 and Model6, and input hypertension data to obtain two models, Model7 and Model8, for a total of 6 regression models trained.
[0107] In step S43, a 1D-CNN network is constructed, which contains four convolutional layers using the ReLU activation function, followed by a Batch Normalization layer to batch normalize the data output by the activation layers, thereby reducing the network's sensitivity to data and accelerating the network convergence process.
[0108] ECG and PPG signals were input into a pre-built 1D-CNN for feature extraction and then for blood pressure regression. During training, the Adam optimizer was used, with the loss function set to mean squared error (MSE) and the evaluation metric being mean absolute error (MSE). A step-decreasing learning rate was employed. Normal blood pressure data was input into the model, and regressions were performed on systolic and diastolic blood pressure, resulting in two models, Model 3 and Model 4. Similarly, prehypertension data yielded two models, Model 5 and Model 6, and hypertension data yielded two models, Model 7 and Model 8, resulting in a total of six regression models trained.
[0109] Step S5 is as follows:
[0110] Step S51: Input the dual-channel test data into the binary classification model Model1 for normal blood pressure + prehypertension vs. hypertension. The model will automatically output the category.
[0111] Step S52: If the output of S51 is hypertension, it means that the data belongs to hypertension; if the output of S51 is normal blood pressure + prehypertension, then input this type of data into the second-level binary classification model Model2 of normal blood pressure vs. prehypertension to determine whether the data belongs to the normal blood pressure category or the prehypertension category.
[0112] Step S53: After the model outputs the classification results, the data will be automatically input into the corresponding regression model, and the regression model will output the specific values of SBP and DBP respectively.
[0113] Step S54: Based on the detection results in S53, obtain the regression model's performance in terms of Mean Absolute Error (MAE), Standard Deviation (SD), and Root Mean Square Error (RMSE), thereby evaluating the accuracy and effectiveness of the method. The formula for calculating the Mean Absolute Error (MAE) is as follows:
[0114]
[0115] The formula for calculating the standard deviation (SD) is:
[0116]
[0117] in,
[0118] The formula for the root mean square error (RMSE) is:
[0119] .
[0120] The mean absolute error (MAE) is the average of the absolute values of the error. The mean absolute error (MAE) should meet the requirements of the AAMI standard for the overall performance of electronic or automatic blood pressure monitors. Specifically, when systolic and diastolic blood pressure are processed separately, the average difference between the paired measurements of the test system and the comparison system should be ±5 mmHg or less.
[0121] The root mean square error index (RMSE) is the arithmetic square root of the root mean square error. The root mean square error is the expected value of the square of the difference between the measured value of the test system and the true value of the comparison system. It is used to evaluate the degree of variation of the data. The smaller the value, the higher the accuracy of the experimental data measured by the test system.
[0122] Example:
[0123] Please refer to Figure 1 This invention provides a method for implementing a dual-channel non-invasive blood pressure classification and blood pressure detection model using inductive transfer learning, comprising the following steps:
[0124] Step S1: Perform data preprocessing on the electrocardiogram (ECG), photoplethysmograph (PPG), and arterial blood pressure (ABP) signals from the sleeveless blood pressure estimation database;
[0125] Step S2: Extract electrocardiogram waveform features from the ECG signal processed in Step S1 using a deep learning network. The features include RR interval, R wave amplitude, ST segment interval, etc., and build a five-class classification model of arrhythmia, specifically N, S, V, F and Q classes, as the source domain model for inductive transfer learning.
[0126] Step S3: Perform inductive transfer on the source domain model built in step S2 to construct a blood pressure binary classification model. Build a binary classification model Model1 for normal blood pressure + prehypertension vs. hypertension, and a binary classification model Model2 for normal blood pressure vs. prehypertension.
[0127] Step S4: The ECG and PPG signals processed in S1 are re-labeled, and features are extracted using a one-dimensional convolutional neural network to construct a regression model. Normal blood pressure data is input into the model, and regressions are performed on systolic and diastolic blood pressure respectively to construct two regression models, Model3 and Model4. Similarly, two models, Model5 and Model6, are obtained by inputting prehypertension data, and two models, Model7 and Model8, are obtained by inputting hypertension data. A total of 6 regression models are trained.
[0128] Step S5: Input the PPG and ECG dual-channel test data into the secondary binary classification model built in S3. After the model automatically determines the category, it automatically selects the corresponding regression model in S4 for regression to obtain the specific systolic blood pressure (SBP) and diastolic blood pressure (DBP) values. Obtain the detection results of the regression model in terms of mean absolute error (MAE), standard deviation (SD), and root mean square error (RMSE).
[0129] In this example, step S1 specifically involves:
[0130] The sleeveless blood pressure estimation database stores data in MATLAB's .mat format. The PPG, ECG, and ABP signals are acquired simultaneously, ensuring temporal synchronization. The three time-series signals are then segmented into data of equal length. The segmentation length is chosen to be within 10 seconds, with no overlap between data segments. If the segmentation length is too long, because blood pressure changes in real time during that period, using the detected blood pressure to represent the blood pressure situation within that timeframe will introduce significant errors, affecting the assessment of hypertension patients. The segmented PPG and ECG signals are used as input data for model training, while the corresponding ABP signal is used to determine the data label for that time segment.
[0131] Record the maximum and minimum values of each ABP data segment, labeled as SBP and DBP for that data point, respectively. Based on the blood pressure classification criteria of the 7th Joint National Committee on Prevention, Detection, Evaluation and Treatment of Hypertension (JNC7), blood pressure is divided into three categories: systolic blood pressure less than 120 mmHg, labeled as normal blood pressure data, with a label of 0; systolic blood pressure between 120-139 mmHg, labeled as prehypertension data, with a label of 1; and systolic blood pressure greater than or equal to 140 mmHg, labeled as hypertension data, with a label of 2.
[0132] After data segmentation, the segmented ECG and PPG data will be denoised separately. Noise in ECG signals mainly includes baseline drift, power line interference, and electromyography (EMG) interference. Baseline drift typically occurs at frequencies less than 1 Hz, power line interference at 50 Hz, and EMG interference across a wider frequency range of 10-3000 Hz. Since the main frequencies of ECG signals are between 0.05-100 Hz, with most energy concentrated between 0.5-45 Hz, discrete wavelet transform is used to decompose the ECG signal and remove noise. The formula for calculating the discretized wavelet transform is as follows:
[0133]
[0134] f(t) is the function that needs to be analyzed by wavelet analysis. Describe the wavelet basis functions. It is a scaling factor, and its value is greater than 0. It is a displacement, which can be positive or negative.
[0135] A wavelet thresholding denoising method is employed, using the "db8" wavelet basis function to perform a 7-level wavelet decomposition on the ECG signal. Noise is removed by setting the frequencies containing the noise to zero. However, directly setting these frequencies to zero would remove useful ECG signals within that frequency band, resulting in significant waveform distortion. Therefore, thresholding is necessary for the coefficients. The Mallat algorithm is used to process the ECG signal, employing a fixed threshold sqtwolog function to obtain multi-level ECG decomposed signals at different frequency bands. The threshold value (Thr) for the sqtwolog function is... ,That It is the noise standard deviation of the layer detail factor. , in It is the median of the absolute values of the high-frequency coefficients of each wavelet layer, arranged in ascending order. for:
[0136]
[0137] The decomposed signal is thresholded using a threshold and the threshold function sqtwolog to filter out ECG noise signals. The multi-layer decomposed signal with the noise removed is then reconstructed to obtain the filtered ECG signal.
[0138] To denoise the PPG signal, the pulse wave frequency is between 0-20Hz, with most of the energy concentrated in the 0-10Hz range and a dominant frequency less than or equal to 3Hz. Therefore, a median filter is used to process the PPG signal. To ensure minimal distortion at the beginning and end of the final result, a mirror extension method is employed. The length W of the sliding window must be at least greater than one cycle; considering the human heart rate range (50-80 beats / minute), W is set to 1.5s. Finally, the baseline drift noise of the PPG signal is obtained. The extracted baseline drift noise is subtracted from the original PPG signal to filter out the PPG noise signal.
[0139] In this example, step S2 specifically involves:
[0140] ECG signal data in leads II and V1 are classified according to the American Association for the Advancement of Medical Devices (AAMI) standards, into N (normal), V (abnormal ventricular beats), S (abnormal supraventricular beats), F (fusion beats), and Q (unclassified beats), respectively. The classification of heart beats and patient data strictly follows the AAMI standards, and the training and test sets are completely independent.
[0141] A deep learning network consisting of M convolutional layers and cross-layer connected residual network layers is constructed. For example, a 17-layer convolutional network is constructed, with L2 regularization added to the weights of each layer. The optimizer uses stochastic gradient descent and momentum optimization algorithms. The cross-layer connected residual network structure can completely extract the waveform features of ECG signals while preventing gradient explosion during training. The constructed deep learning network is used to extract waveform features from ECG data, and a five-class classification model of arrhythmia with an accuracy of 94.4% is constructed. The parameters of the first 15 convolutional layers are frozen, and the parameters of the last 2 convolutional layers are unfrozen and used as the source domain model for inductive transfer learning.
[0142] In this example, step S3 specifically involves:
[0143] The transfer learning model needs to perform two binary classifications: normal blood pressure + prehypertension vs. hypertension, and normal blood pressure vs. prehypertension. Therefore, the dataset needs to be first classified into normal data, prehypertension data, and hypertension data. Let the dataset consisting of normal blood pressure data be N_dataset, the dataset consisting of prehypertension data be Pre_dataset, and the dataset consisting of hypertension data be H_dataset.
[0144] When training the first binary classification model, Model1, N_dataset and Pre_dataset are merged into a single class dataset, which is then input into the model along with the H_dataset for training. When training the second binary classification model, Model2, N_dataset and Pre_dataset are input into the model for training. During training, the training and validation sets are divided in an 8:2 ratio.
[0145] The source domain model is designed to classify cardiac arrhythmias into five categories, while the target domain for transfer learning is to classify blood pressure into two categories. Therefore, the number of neurons in the output layer is reduced from 5 to 2. The parameters of the first 15 layers are frozen, the parameters of the last two convolutional layers are unfrozen, and a fully connected layer is added to obtain the target domain model.
[0146] The transfer learning model was trained using the Tensorflow deep learning framework. During training, the cross-entropy loss function and stochastic gradient descent optimizer (SGD) were used. The evaluation metric was categorical accuracy. Both models were trained for 60 epochs, and the best-performing epoch was saved. A binary classification model Model1 was built to classify normal blood pressure + prehypertension vs. hypertension, and a binary classification model Model2 was built to classify normal blood pressure vs. prehypertension.
[0147] In this example, step S4 specifically involves:
[0148] The dataset was relabeled using specific systolic and diastolic blood pressure values, with each data sample having two labels: systolic and diastolic blood pressure values. The normal blood pressure data (N_dataset), prehypertension data (Pre_dataset), and hypertension data (H_dataset) were randomly shuffled and divided into training and validation sets in a 7:3 ratio.
[0149] A 1D-CNN network was constructed, consisting of four convolutional layers using the ReLU activation function, followed by a Batch Normalization layer. This batch normalization reduces the network's sensitivity to data changes and accelerates convergence. ECG and PPG signals were input into the 1D-CNN for feature extraction and blood pressure regression. The Adam optimizer was used during training, with the mean squared error (MSE) loss function and the mean absolute error (MAE) as the evaluation metric. A step-decreasing learning rate was employed. Normal blood pressure data was input into the model to regress systolic and diastolic blood pressure, resulting in two models, Model 3 and Model 4. Similarly, prehypertension data yielded two models, Model 5 and Model 6, and hypertension data yielded two models, Model 7 and Model 8, resulting in a total of six regression models trained.
[0150] In this example, step S5 specifically involves:
[0151] Inputting dual-channel test data into the binary classification model Model1 (normal blood pressure + prehypertension vs. hypertension), the model will automatically output the category. If the output is hypertension, the data belongs to hypertension. If the output is normal blood pressure + prehypertension, this type of data is input into the second-level binary classification model Model2 (normal blood pressure vs. prehypertension) to determine whether the data belongs to the normal blood pressure category or the prehypertension category. After the model outputs the classification results, the data will be automatically input into the corresponding regression model, which will output the specific values of SBP and DBP, respectively.
[0152] The overall performance of the regression model is evaluated using mean absolute error, standard deviation, and root mean square error, where:
[0153] Mean Absolute Error (MAE): This is calculated by averaging the absolute values of the errors, providing a good reflection of the actual error in the predicted values. The AAMI standard sets strict performance requirements for electronic or automatic blood pressure monitors: when systolic and diastolic blood pressure are processed separately, the average difference between paired measurements from the testing and comparison systems should be ±5 mmHg or less. The calculation formula is as follows:
[0154]
[0155] Standard Deviation (SD): In blood pressure testing, the standard deviation of the difference between the measured value of the testing system and the paired value of the comparison system should be 8 mmHg or less. SD reflects the degree of dispersion between the differences, and its specific calculation formula is as follows:
[0156]
[0157] in,
[0158] Root Mean Square Error (RMSE): RMSE is the arithmetic square root of the mean square error. It represents the expected value of the square of the difference between the measured value of the test system and the true value of the comparison system. It assesses the degree of data variability; the smaller the value, the higher the accuracy of the experimental data obtained by the test system. The formula for RMSE is:
[0159]
[0160] The test results showed that the overall errors of the regression model were: mean absolute error of diastolic blood pressure of 3.5834 mmHg and standard deviation of 5.7242 mmHg, and mean absolute error of systolic blood pressure of 4.5760 mmHg and standard deviation of 7.6024 mmHg. Both met the AAMI standard requirements for blood pressure monitoring instruments, namely, mean absolute error less than 5 mmHg and standard deviation less than 8 mmHg. This demonstrates the effectiveness and reliability of the non-invasive blood pressure classification model trained using transfer learning and the blood pressure monitoring model built using a one-dimensional convolutional neural network in this example.
[0161] Those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, this invention also intends to include such modifications and variations if they fall within the scope of the claims of this invention and their equivalents.
Claims
1. A method for dual-channel non-invasive blood pressure classification and detection based on inductive transfer learning, characterized by: Includes the following steps; Step S1: Perform data preprocessing on the ECG, PPG, and ABP signals from the sleeveless blood pressure estimation database; Step S2: Extract electrocardiogram waveform features from the ECG signal processed in Step S1 using a deep learning network. The features include RR interval, R wave amplitude, and ST segment interval. Build a five-class arrhythmia model with specific categories N, S, V, F, and Q, which serve as the source domain model for inductive transfer learning. Step S3: Perform inductive transfer on the source domain model built in step S2 to construct a blood pressure binary classification model. Build binary classification models Model1 for normal blood pressure and prehypertension, and Model2 for normal blood pressure and prehypertension. Step S4: The ECG and PPG signals processed in S1 are re-labeled, and features are extracted using a one-dimensional convolutional neural network to construct a regression model. Normal blood pressure data is input into the model, and regressions are performed on systolic and diastolic blood pressure respectively to construct two regression models, Model3 and Model4. Similarly, two models, Model5 and Model6, are obtained by inputting prehypertension data, and two models, Model7 and Model8, are obtained by inputting hypertension data. A total of 6 regression models are trained. Step S5: Input the PPG and ECG dual-channel test data into the secondary binary classification model built in S3. After the model automatically determines the category, it automatically selects the corresponding regression model in S4 for regression to obtain the specific systolic blood pressure SBP and diastolic blood pressure DBP values, and obtains the detection results of the regression model on mean absolute error (MAE), standard deviation (SD), and root mean square error (RMSE). Step S1 specifically includes the following steps; Step S11: PPG, ECG and ABP signals from the sleeveless blood pressure estimation database are collected simultaneously to synchronize them in the time dimension. The three signals corresponding in time are cut into data of the same length without overlap. The cut PPG and ECG signals are used as input data for model training, and the corresponding ABP signals are used to determine the data label for that time period. Step S3 is as follows: Step S31: Classify the processed PPG signal dataset and ECG signal dataset into normal data, prehypertension data, and hypertension data; Step S32: Merge the normal data and prehypertension data into a single category dataset, and use it together with the hypertension dataset as the dual-channel input data for the first binary classification transfer model, Model 1. Step S33: Use normal data and prehypertension data as dual-channel input data for the second binary classification transfer model, Model 2; Step S34: Change the output layer of the source domain model from five-class classification to two-class classification, and connect it with a fully connected layer to build the target domain model; Step S35: Train the transfer model built in S34 using the Tensorflow deep learning framework. The input data uses dual-channel data classified in S32 and S33 respectively. During training, the cross-entropy loss function and stochastic gradient descent optimizer SGD are used. The evaluation metric is categorical_accuracy. Build a binary classification model Model1 for normal blood pressure and prehypertension, and a binary classification model Model2 for normal blood pressure and prehypertension.
2. The method for dual-channel non-invasive blood pressure classification and detection using inductive transfer learning according to claim 1, characterized in that: Step S1 also Includes the following steps; Step S12: Record the maximum and minimum values of each ABP data segment, which are the systolic blood pressure (SBP) and diastolic blood pressure (DBP) for that segment, respectively. According to the blood pressure classification criteria in the 7th report of the Joint National Committee for Prevention, Detection, Evaluation and Treatment of Hypertension in the United States, blood pressure is divided into three categories: systolic blood pressure less than 120 mmHg, which is marked as normal blood pressure; systolic blood pressure between 120-139 mmHg, which is marked as prehypertension; and systolic blood pressure greater than or equal to 140 mmHg, which is marked as hypertension. Step S13: Process the ECG signal using discrete wavelet transform and Mallat algorithm. A fixed threshold function, sqtwolog, is used to obtain multi-level ECG decomposed signals of different frequency bands. The threshold Thr of the sqtwolog function is... , It is the noise standard deviation of the layer detail factor. , in It is the median of the absolute values of the high-frequency coefficients of each wavelet layer, arranged in ascending order. for: ； Step S14: Use the threshold and threshold function sqtwolog to perform threshold processing on the decomposed signal to filter out ECG noise signals; Step S15: Reconstruct the multi-layer decomposition signal after filtering out noise to obtain the filtered ECG signal; Step S16: Process the PPG signal using a median filter to obtain the baseline drift noise of the PPG signal; Step S17: Subtract the extracted baseline drift noise from the original PPG signal to filter out the PPG noise signal; Step S2 is as follows: Step S21: Divide the ECG data of lead II and lead V1 into heart rate categories according to the American Association for the Advancement of Medical Devices (AAMI) standard, ensuring that the training set and test set data do not overlap at all; Step S22: Construct a deep learning network consisting of M convolutional layers and residual network layers with cross-layer connections; Step S23: Extract waveform features from ECG data using the deep learning network built in S22, and build a five-class classification model for arrhythmias with high accuracy; Step S24: Freeze the parameters of the first N convolutional layers of the arrhythmia five-classification model built in S23, and unfreeze the parameters of the last MN convolutional layers to serve as the source domain model for inductive transfer learning.
3. The method for dual-channel non-invasive blood pressure classification and detection using inductive transfer learning according to claim 1, characterized in that: In step S11, the data storage method of the sleeveless blood pressure estimation database is MATLAB's .mat format, the cutting length is selected to be within 10 seconds, and the data are not cut overlappingly. In step S13, after data segmentation, the segmented ECG and PPG data are denoised separately. Noise in the ECG signal includes baseline drift, power line interference, and electromyography (EMG) interference. The baseline drift frequency is less than 1 Hz, the power line interference frequency is 50 Hz, and the EMG interference frequency range is between 10-3000 Hz. The main frequency of the ECG signal is between 0.05-100 Hz, with most of the energy concentrated between 0.5-45 Hz. Discrete wavelet transform is used to decompose the ECG signal and remove noise. The calculation formula for discretized wavelet transform is as follows: f(t) is the function that needs to be analyzed by wavelet analysis. Describe the wavelet basis functions. It is a scaling factor, and its value is greater than 0. It is a displacement, which can be positive or negative; A wavelet thresholding method is used to remove noise. The wavelet basis function is db8. The ECG signal is decomposed into 7 layers of wavelets, and the frequency of noise is set to zero to remove the noise.
4. The method for dual-channel non-invasive blood pressure classification and detection using inductive transfer learning according to claim 1, characterized in that: Step S5 is as follows: Step S51: Input the dual-channel test data into the binary classification model Model1 for normal blood pressure, prehypertension, and hypertension. The model will automatically output the category. Step S52: If the output of S51 is hypertension, it means that the data belongs to hypertension; if the output of S51 is normal blood pressure and prehypertension, then input this type of data into the second-level binary classification model Model2 for normal blood pressure and prehypertension to determine whether the data belongs to the normal blood pressure category or the prehypertension category. Step S53: After the model outputs the classification results, the data will be automatically input into the corresponding regression model, and the regression model will output the specific values of SBP and DBP respectively. Step S54: Based on the detection results in S53, obtain the regression model's performance in terms of Mean Absolute Error (MAE), Standard Deviation (SD), and Root Mean Square Error (RMSE), thereby evaluating the accuracy and effectiveness of the method. The formula for calculating the Mean Absolute Error (MAE) is as follows: The formula for calculating the standard deviation (SD) is: in, The formula for the root mean square error (RMSE) is: 。 5. The method for dual-channel non-invasive blood pressure classification and detection using inductive transfer learning according to claim 1, characterized in that: In step S3, the dataset is first classified into normal data, prehypertension data and hypertension data. Let the dataset consisting of normal blood pressure data be N_dataset, the dataset consisting of prehypertension data be Pre_dataset, and the dataset consisting of hypertension data be H_dataset. When training the first binary classification model, Model1, N_dataset and Pre_dataset are merged into a single class dataset, which is then input into the model along with H_dataset for training. When training the second binary classification model, Model2, N_dataset and Pre_dataset are input into the model for training. During training, the training set and validation set are divided in an 8:2 ratio. The source domain model is designed to classify arrhythmias into five categories, while the target domain for transfer learning is to classify blood pressure into two categories. Therefore, the number of neurons in the output layer is changed from 5 to 2. The network parameters of the first 15 layers are frozen, the parameters of the last 2 convolutional layers are unfrozen, and a fully connected layer is added to obtain the target domain model. The transfer learning model was trained using the Tensorflow deep learning framework. During training, the cross-entropy loss function and stochastic gradient descent optimizer (SGD) were used. The evaluation metric was categorical accuracy. Both models were trained for 60 epochs, and the best-performing epoch was saved. A binary classification model Model1 was built to classify normal blood pressure, prehypertension, and hypertension, and a binary classification model Model2 was built to classify normal blood pressure and prehypertension.
6. The method for dual-channel non-invasive blood pressure classification and detection using inductive transfer learning according to claim 1, characterized in that: Step S4 is as follows: Step S41: Relabel the dataset using the specific values of systolic and diastolic blood pressure. Each data sample has two labels: systolic blood pressure value and diastolic blood pressure value. Step S42: Randomly shuffle the normal blood pressure data, prehypertension data, and hypertension data respectively, and divide them into training set and validation set in a 7:3 ratio; Step S43: Input the ECG and PPG signals into the 1D-CNN for feature extraction and then perform blood pressure regression; input normal blood pressure data into the model and perform regression of systolic and diastolic blood pressure respectively to obtain two models, Model3 and Model4. Similarly, input prehypertension data to obtain two models, Model5 and Model6, and input hypertension data to obtain two models, Model7 and Model8, for a total of 6 regression models trained.
7. The method for dual-channel non-invasive blood pressure classification and detection using inductive transfer learning according to claim 6, characterized in that: In step S43, a 1D-CNN network is constructed, which contains four convolutional layers using the ReLU activation function, followed by a Batch Normalization layer to batch normalize the data output by the activation layers, thereby reducing the network's sensitivity to data and accelerating the network convergence process. ECG and PPG signals were input into a pre-built 1D-CNN for feature extraction and blood pressure regression. The Adam optimizer was used during training, with the loss function set to mean squared error (MSE) and the evaluation metric being mean absolute error. The learning rate was decreased in stages. Normal blood pressure data was input into the model, and regressions were performed on systolic and diastolic blood pressure, resulting in two models, Model 3 and Model 4. Similarly, prehypertension data was input to obtain two models, Model 5 and Model 6, and hypertension data was input to obtain two models, Model 7 and Model 8, for a total of six regression models trained.
8. The method for dual-channel non-invasive blood pressure classification and detection using inductive transfer learning according to claim 1, characterized in that: The mean absolute error (MAE) is the average of the absolute values of the error. The mean absolute error (MAE) should meet the requirements of the AAMI standard for the overall performance of electronic or automatic blood pressure monitors. Specifically, when systolic and diastolic blood pressure are processed separately, the average difference between the paired measurements of the test system and the comparison system should be -5 to 5 mmHg. The root mean square error index (RMSE) is the arithmetic square root of the root mean square error. The root mean square error is the expected value of the square of the difference between the measured value of the test system and the true value of the comparison system. It is used to evaluate the degree of variation of the data. The smaller the value, the higher the accuracy of the experimental data measured by the test system.