CN115486823B - A cuffless continuous blood pressure estimation system based on online learning - Google Patents

Abstract

The invention discloses a cuffless continuous blood pressure estimation system based on online learning, which belongs to the field of data processing and deep learning. The invention trains the network based on online learning, which can effectively solve the problem of poor universality of the blood pressure estimation model trained under large amounts of data, thereby realizing continuous real-time monitoring of individual blood pressure, reflecting the patient's true continuous blood pressure changes, effectively avoiding the randomness of blood pressure measurement in clinics, excluding white coat hypertension, and timely detecting blood pressure changes, etc., and can comprehensively analyze the changes in physiological activities or biochemical indicators in the body that increase or decrease with blood pressure fluctuations.

Description

A cuff-free continuous blood pressure estimation system based on online learning

Technical Field

The present invention relates to the fields of deep learning, biomedical engineering, and more particularly to estimating cuffless continuous arterial blood pressure using a deep learning online learning method.

Background Art

Since humans discovered arterial blood pressure in 1733, observed arterial pulsation, and successfully used mercury manometers to measure blood pressure intermittently, and realized continuous blood pressure measurement through arterial catheterization in the last century, how to achieve non-invasive and non-invasive continuous blood pressure measurement has become a major problem in the field of blood pressure measurement. More traditional methods such as arterial tension measurement and arterial volume clamp method can achieve non-invasive indirect continuous blood pressure measurement, but due to the limitations of the method itself and various other factors, such as flattening degree, sensor position setting, and motion artifacts, the accuracy that can be achieved is compared with standard invasive continuous blood pressure measurement. In addition, the measurement systems based on these methods are large and expensive, and require professional operation, which limits their application to a certain extent.

At present, the main blood pressure measurement method is the upper arm medical electronic blood pressure monitor that meets international standards (ESH, BHS and AAMI), such as "A Review of Clinical Accuracy Evaluation Standards for Electronic Blood Pressure Monitors_Li Qin". This blood pressure monitor is widely used in clinical and home blood pressure measurement, but its main problems are: 1) The inflatable cuff required for operation will cause discomfort to the user, and if monitored at night, it will also interfere with sleep; 2) It is impossible to measure blood pressure continuously for a long time; 3) The measurement accuracy of blood pressure is easily affected, resulting in the inability to correctly reflect the real blood pressure information. The reasons are as follows: a) The tightness and height of the cuff will affect the blood pressure value; b) The cuff will squeeze the muscles and blood vessels during the inflation process. After the air bag is deflated, it takes a while for the muscles and arteries to calm down. It usually takes more than 5 minutes to measure again; c) White coat hypertension values may be measured (multiple blood pressure measurements will cause neuroendocrine and emotional changes, which will also have a certain impact on short-term blood pressure measurements).

With the development of new sensor technologies, various cuffless continuous monitoring methods can be implemented in wearable and some inconspicuous devices, such as integrating sensors and estimation algorithms into daily objects such as watches, glasses, sleeping mats and smart watches to achieve non-intrusive, continuous and real-time measurement. When used by users, these devices can automatically obtain physiological signals from the human body surface, such as electrocardiogram, pulse wave and other cardiovascular-related physiological signals, and then give estimation results in time according to various blood pressure measurement models. However, wearable cuffless blood pressure measurement is not widely used in clinical practice. The main reason is that the measurement accuracy of the device application cannot meet the current clinical needs, especially when monitoring the dynamic changes of blood pressure. It is difficult to obtain accurate blood pressure estimates without calibrating the patient's blood pressure. Therefore, in order to solve the calibration and accuracy deviation problems, further research is needed and new estimation models should be explored, including more comprehensive physiological information or new potential mechanism theories of cuffless continuous pressure measurement, so as to directly use automatic calibration procedures or blood pressure estimation without any calibration.

The current cuffless continuous blood pressure measurement models are mainly divided into two categories, namely, based on physiological mechanisms or based on data-driven methods. For example, based on the pulse wave transmission time PTT, DBP and SBP are calculated as follows:

Among them, PTT _w is obtained by PTT weighting, A is a subject-dependent coefficient, and the parameters with the subscript "o" are obtained by the corresponding parameters through a calibration procedure, such as SBP ₀ and DBP ₀ are calibrated by systolic blood pressure SBP and diastolic blood pressure DBP, respectively. This PTT-based method can estimate SBP and DBP within the baseline values of 0.6±9.8mmHg and 0.9±5.6mmHg, respectively. In addition, physiological characteristics methods such as PPG intensity ratio, Worthley number, radial electrical bioimpedance, etc. are also used to estimate continuous blood pressure;

With the development of theories and applications related to machine learning technology, many data-driven deep learning methods have been introduced, such as methods based on artificial neural networks and the fusion of multiple neural networks to achieve non-invasive continuous time blood pressure prediction, such as fully connected networks, ResNet, WaveNet, LSTM and other network structures and the fusion of these networks. Among them, A. Paviglianiti et al. proposed the ResNet+LSTM network architecture, and used the MIMIC database and PPG and ECG data to obtain the minimum mean square error of systolic and diastolic blood pressure estimation of 6.414 mmHg and 3.101 mmHg. This type of method is simple and easy to implement, but there are still some problems, that is, when the amount of data for training the model is large, these network models obtained based on data drive are fixed, that is, the parameters of the network model obtained by training based on the specified data set will not change, and the trained network model cannot achieve self-optimization and update of the model with the continuous input of data. Therefore, this method is usually only applicable to groups with similar input physiological signal distribution, and has poor universality. When applied to different scenarios, it often needs to be retrained and the network model reconstructed. When the amount of data is large, the time cost is often high, and because this method cannot achieve real-time updating of the model, it cannot reflect the continuous changes in blood pressure in real time during estimation, and has certain stage limitations in predicting scenarios with large blood pressure change trends.

Summary of the invention

Based on the above problems, the present invention proposes a continuous blood pressure estimation system based on online learning, which combines the online learning method with various data-driven models (LSTM, RNN, etc.) or mechanism models, and mainly solves the problem of poor universality of the blood pressure estimation model trained under large amounts of data based on deep learning and other methods, so as to realize an individual adaptive non-invasive continuous time blood pressure estimation method, thereby providing another feasible method for high-precision continuous blood pressure ubiquitous measurement.

The technical solution of the present invention is: a cuffless continuous blood pressure estimation system based on online learning, the system comprises: a data acquisition and preprocessing module, a data input control module, an online learning model construction module, a model update optimization module, and a blood pressure prediction display module;

The data acquisition and preprocessing module realizes the acquisition and preprocessing of input signals, that is, using the wearable device to collect time-series physiological signals on the user's body surface, and simultaneously using the blood pressure measurement device to collect synchronous time-series blood pressure data, and transmitting all the collected signals to the data preprocessing module. The data preprocessing method includes data conversion, removal of non-stationary data, differentiation, and normalization;

The data input control module realizes the real-time orderly transmission of data and maintains the synchronization of multiple signal inputs. It is implemented in the distributed publish-subscribe messaging platform Kafka, that is, the pre-processed time series data is written to the Kafka message queue of the cloud server, so that the events are organized and stored persistently in the Topic, where the events can be read at any time as needed, and the events will not be deleted after use. The time that each Topic in Kafka should retain events is defined through configuration, and old events will be discarded after this time.

The online learning model construction module realizes real-time supervised learning. The continuous real-time output of the data input control module is used as the input of the online learning model construction module. In this process, the data will be divided into two parts: training set and verification set. Different networks are selected according to specific needs to establish the initial online learning model, and the time series physiological signals collected by the data acquisition and preprocessing module are further used for supervised learning and subsequent calculation and analysis.

The implementation method in the model updating and optimization module is as follows: for the streaming data transmitted by the data input control module, a time series data segment STData _Ti with a time length of T is continuously intercepted, and the STData _Ti includes multimodal signals including the reference blood pressure. After each STData _Ti is read, it is used to optimize the network model. The model parameter is W _i , and its corresponding continuous blood pressure prediction accuracy is Acc _i . When the STData _Ti data with a time length of T are continuously read, after a time length NT, N=1, 2, 3..., the optimal model W _NO is obtained. The optimization method is as follows:

A1. Set the initial optimization threshold th and the training time threshold NT;

A2. For each time series data segment STData _Ti , under the same network model, the prediction accuracy of the network model after optimization learning is Acc _i , and whether to retain the model parameters is determined according to the established optimization threshold th:

If Acc _i ≥ th, that is, the threshold condition is met, the model parameters are retained, the model parameters are recorded and saved as _Wi , and the model is set as the optimal model W _NO = _Wi , and the optimization threshold is reset to th = Acc _i .

If Acc _i ＜th, that is, it does not meet the threshold condition, then the model parameters are discarded;

A3. Input reaches the time threshold, and after optimization is completed, output the optimization model W _NO =W _i , Acc _N0 =max{Acc _i } obtained by training and learning under the collected data STData _NT ;

Finally, after N times of training and learning, the optimal model W _NO is obtained, and the model W _NO is output as the optimal prediction model of the current system, and the model is used to predict the continuous blood pressure values in the future.

Furthermore, in the model updating and optimization module, a higher prediction effect is achieved by partially overlapping STData _T(i+1) and STData _Ti data.

Beneficial effects: The present invention trains the network based on an online learning method, which can effectively solve the problem of poor universality of the blood pressure estimation model obtained by training under large amounts of data, thereby realizing continuous real-time monitoring of individual blood pressure, reflecting the patient's actual continuous blood pressure changes, effectively avoiding the arbitrariness of blood pressure measurement in the clinic, excluding white coat hypertension, and timely detecting blood pressure changes, etc., and can comprehensively analyze the changes in physiological activities or biochemical indicators in the body that increase or decrease with blood pressure fluctuations.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a block diagram of a cloud-based sequential data control module according to the present invention.

FIG2 is a training block diagram of the online learning model of the present invention.

FIG3 is a block diagram of the model optimization and updating of the present invention.

FIG4 is a flowchart of an online learning model estimation system according to the present invention.

DETAILED DESCRIPTION

To realize the above system, the present invention can collect and measure the heart and arterial pulsation related information of the human body over a period of time based on wearable sensor equipment, so as to obtain timing signals such as electrocardiogram, photoelectric volume pulse wave, and surface arterial pressure, and use these collected real-time signals and adopt online learning method to realize real-time continuous blood pressure prediction.

A cuffless continuous blood pressure estimation system based on online learning, the system comprises: a data acquisition and preprocessing module, a data input control module, an online learning model building module, a model update optimization module, and a blood pressure prediction display module;

The data acquisition and preprocessing module realizes the acquisition and preprocessing of input signals, that is, using the wearable device to collect time-series physiological signals on the user's body surface, and simultaneously using the blood pressure measurement device to collect synchronous time-series blood pressure data, and transmitting all the collected signals to the data preprocessing module. The data preprocessing method includes data conversion, removal of non-stationary data, differentiation, and normalization;

The data input control module realizes the real-time orderly transmission of data and maintains the synchronization of multi-signal input. It is implemented in the distributed publish-subscribe message platform Kafka, that is, the pre-processed time series data is written to the Kafka message queue of the cloud server, so that the events are organized and permanently stored in the Topic, where the events can be read at any time as needed, and the events will not be deleted after use. The time that each Topic in Kafka should retain the event is defined by configuration, and the old events will be discarded after the time exceeds this time. Since the Topic in Kafka has a high degree of sequentiality and reliability, the written and read messages can be completely consistent without any processing, and a buffer is provided to prevent blocking, ensuring the consistency of reading different messages. As shown in Figure 1, the message queues of different topics input to the Kafka server are independent of each other and consistent in time. They can be obtained by the wearable device through the sensor and written after preprocessing. The input data includes inputs for supervised learning, which can be time series signals such as PPG and ECG signals, and the blood pressure signal is read at the same time as a reference for supervised learning. The timing of the entire reading and writing process must be consistent to achieve streaming non-blocking output of input data in the input control module.

The online learning model construction module realizes real-time supervised learning. The STData _Ti data within the time length T will be divided into two parts: training set and verification set, which are used for supervised training of the network model; as shown in Figure 2, specifically, if the time length of STData _T is 120s, a time length of 100s can be selected for training and learning, and a time length of 20s can be selected for verifying the prediction accuracy. If the LSTM model is selected as the training model, the prediction accuracy parameter threshold selects the mean absolute error MAE: 5.000mmHg (SBP)/4.000mmHg (DBP) or the root mean square error RMSE: 6.000mmHg (SBP)/4.000mmHg (DBP), that is, the LSTM model is trained with the collected data STData _Ti . If the model conclusion is within the allowable range of the threshold parameter, the model is updated to the current optimal model, and the threshold condition is updated at the same time. In this way, the optimal model is continuously updated to realize the model optimization function. In addition, the optimal hyperparameter (such as learning rate, etc.) needs to be found during the training process. Generally speaking, the optimal hyperparameter can be found by continuous polling and traversal.

The model update optimization module implements the optimal model for STData _T for training and learning. Specifically, a time threshold such as NT (N=4, T=120s) is set, and the prediction accuracy th is the threshold condition. The optimization steps are as follows:

1) Data STData1, the model parameter obtained by training is W ₁ , and its prediction accuracy is Acc ₁ ≥ th, then W _NO ＝W ₁ , th＝Acc ₁ ;

2) Data STData2, the model parameter obtained by training is W ₂ , and its prediction accuracy is Acc ₂ ＜th, then W _NO ＝W ₁ , th＝Acc ₁ ;

3) Data STData3, the model parameter obtained by training is W ₃ , and its prediction accuracy is Acc ₃ <th, then W _NO =W ₁ , th = Acc ₁ ;

4) Data STData4, the model parameter obtained by training is W ₄ , and its prediction accuracy is Acc ₄ ≥ th, then W _NO ＝W ₄ , th＝Acc ₄ ;

The implementation process of the online learning model estimation system is shown in Figure 4. The specific steps are as follows:

1. Collect multiple synchronous time-series physiological signals of subjects including blood pressure signals from wearable devices, such as PPG, ECG, etc.

2. Input all collected signals into the preprocessing module for preprocessing.

3. Upload the preprocessed timing synchronization signal to the cloud Kafka message queue for queue synchronization buffering, and read the information in the message queue from the Kafka server as input for model training.

4. Divide the timing signals read from the Kafka message queue into two parts:

a) Time series signal division: the data stream is divided into STData _Ti data segments of 10 minutes or 15 minutes for each training.

b) STData _Ti partitioning: Each STData _Ti is divided into a training set and a validation set, such as a training set (80%), a validation set (20%), etc.

5. Model Training

a) Set the number of training times N, and let N = 0 to indicate that the model starts training. After the model is trained once, N = N + 1 is reset.

b) Select a training model. Taking LSTM as an example, LSTM(0) represents the initial state of the model. After one training, the state of the model is LSTM(1).

c) Determine the results of each training. If the model conclusion meets the threshold condition, record and save the model (parameters) and prediction results of this training. For example, if only the prediction accuracy is considered, if the prediction accuracy of the first training is 0.86, which is greater than the threshold of 0.80, save the results of this training (model parameters and prediction results). At this time, the model state is LSTM (1), and the update threshold is 0.86. At the same time, LSTM (1) is used as the initial model for the second training. The second training is carried out and the results of the second training are determined. If the threshold condition is met, LSTM (1) is further updated to LSTM (2) and the threshold is reset; if the threshold condition is not met, LSTM (1) is continued to be used as the initial model for the third training, and this process is repeated until N trainings are completed.

The degree of optimization of the system model is also related to the network type and network capacity. When selecting a model, different types of network models have different prediction performance under the same network capacity. When the network model is the same, different network capacities will also affect the prediction performance of the model. Too large a network capacity may cause the model to overfit the data, and too small a network capacity may cause the model to have insufficient learning ability for the data, both of which will lead to a decrease in the prediction performance of the final model. The selection and adjustment of the model needs to be adjusted according to the specific sample size to make it close to the boundary between underfitting and overfitting.

Implementation:

The specific method of using wearable devices to achieve continuous time blood pressure estimation based on online learning is as follows:

Step 1: Use wearable devices to measure the blood pressure, PPG and other physiological signal values of the subject within a period of time (such as 30 minutes).

Step 2: Preprocess the real-time time series signal obtained in step 1, including but not limited to denoising, removing non-stationarity, converting to a signal suitable for supervised learning, and other operations.

Step 3: Load the preprocessed signal input into the Kafka server. It should be noted that signals of different modes need to be loaded into different Kafka topics through different APIs to ensure that there is no crosstalk between different signals, and the same signal forms a reliable message queue in one Topic.

Step 4: Consume messages in the Topic to obtain synchronized streaming data of each signal. When consuming messages in the queue, you need to check whether the different signals in the Topic message queue are synchronized in time. If not, you need to delete some of the data that is ahead of time so that the data output by each signal is synchronized in a physiological sense.

Step 5: Divide the streaming data into batches (e.g., divide 1 minute of data into one batch), so that 30 batches of sampling points can be obtained continuously within 30 minutes; divide each batch into a training set and a test set (e.g., 70% for training, 10% for parameter adjustment, and 20% for verification).

Step 6: Set the update threshold. For example, setting the prediction accuracy η≥0.9 means that the model is allowed to update.

Step 7: Select an initial network model, such as LSTM, and define the initial state of the model as LSTM(1).

Step 8: Model training. Use the data from the first minute to train and verify LSTM (1). When the prediction result meets the threshold condition, update the threshold condition and record and update the model as the current optimal model. Otherwise, do not update and continue to train the current model.

Step 9: Repeat step 8 until 30 trainings are completed, and finally obtain the 30 optimized models LSTM(i)

Step 10: Use the final model for real-time predictions in the future, such as 5 minutes.

Claims (2)

1. A cuffless continuous blood pressure estimation system based on online learning, the system comprising: a data acquisition and preprocessing module, a data input control module, an online learning model building module, a model update optimization module, and a blood pressure prediction and display module; The data acquisition and preprocessing module realizes the acquisition and preprocessing of input signals, that is, using the wearable device to collect time-series physiological signals on the user's body surface, and simultaneously using the blood pressure measurement device to collect synchronous time-series blood pressure data, and transmitting all the collected signals to the data preprocessing module. The data preprocessing method includes data conversion, removal of non-stationary data, differentiation, and normalization; The data input control module realizes the real-time orderly transmission of data and maintains the synchronization of multi-signal input. It is implemented in the cloud server Kafka, that is, the pre-processed time series data is written to the cloud server Kafka message queue, so that the events are organized and persistently stored in the Topic, where the events can be read at any time as needed, and the events will not be deleted after use. The time that each Topic in the cloud server Kafka should retain the event is defined through configuration, and the old events will be discarded after the time exceeds this time; The online learning model construction module realizes real-time supervised learning. The continuous real-time output of the data input control module is used as the input of the online learning model construction module. In this process, the data will be divided into two parts: training set and verification set. Different networks are selected according to specific needs to establish the initial online learning model, and the time series physiological signals collected by the data acquisition and preprocessing module are further used for supervised learning and subsequent calculation and analysis. The implementation method in the model updating optimization module is: for the streaming data transmitted by the data input control module, a time series data segment STData _Ti with a time length of T is continuously intercepted, and the STData _Ti includes multimodal signals including reference blood pressure. After each STData _Ti is read, it is used to optimize the network model. The model parameter is W _i , and its corresponding continuous blood pressure prediction accuracy is Acc _i . When the STData _Ti data with a time length of T are continuously read, after a time length NT, N=1, 2, 3..., the optimal model W _NO is obtained. The optimization method is as follows: A1. Set the initial optimization threshold th and the training time threshold NT; A2. For each time series data segment STData _Ti , under the same network model, the prediction accuracy of the network model after optimization learning is Acc _i , and whether to retain the model parameters is determined according to the established optimization threshold th: If Acc _i ≥ th, that is, the threshold condition is met, the model parameters are retained, the model parameters are recorded and saved as _Wi , and the model is set as the optimal model W _NO = _Wi , and the optimization threshold is reset to th = Acc _i . If Acc _i ＜th, that is, it does not meet the threshold condition, then the model parameters are discarded; A3. Input reaches the time threshold, and after optimization is completed, output the optimization model W _NO =W _i , Acc _N0 =max{Acc _i } obtained by training and learning under the collected data STData _NT ; Finally, after N times of training and learning, the optimal model W _NO is obtained, and the model W _NO is output as the optimal prediction model of the current system, and the model is used to predict the continuous blood pressure values in the future. 2. A cuffless continuous blood pressure estimation system based on online learning as described in claim 1, characterized in that in the model update optimization module, a higher prediction effect is achieved by partially overlapping STData _T(i+1) and STData _Ti data.