A blood pressure measurement system based on fedabn algorithm
The blood pressure measurement system based on the FedABN algorithm uses a camera to collect vPPG signals and performs signal quality evaluation and model updates, solving the problems of poor comfort, insufficient data and privacy in existing technologies, and achieving highly accurate personalized blood pressure measurement.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING SIGHTNOVO MEDICAL TECH CO LTD
- Filing Date
- 2023-04-19
- Publication Date
- 2026-07-07
AI Technical Summary
Existing continuous blood pressure monitoring methods suffer from poor comfort, poor sustainability, difficulty in miniaturizing equipment, insufficient data volume, and privacy issues, resulting in large measurement errors.
The blood pressure measurement system using the FedABN algorithm works collaboratively between the client and server, using a camera to collect vPPG signals, evaluate signal quality and update the pre-trained model, and combine BN layer data from multiple clients to perform personalized model aggregation, protecting user privacy while increasing the amount of training data.
It achieves improved accuracy and sustainability of blood pressure measurement while protecting user privacy, increases the amount of training data, and provides personalized blood pressure estimation models.
Smart Images
Figure CN116439679B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of intelligent medicine, and more specifically, to a blood pressure measurement system based on the FedABN algorithm. Background Technology
[0002] At present, continuous blood pressure detection methods are mainly divided into pressure-sensor-based methods and photoelectric-sensor-based methods. Among them, the pressure-sensor-based continuous blood pressure measurement method has high accuracy, but it has the following problems: (1) Due to the implantation or compression of the pressure sensor, it is uncomfortable and may cause trauma or congestion to the patient; (2) Due to its poor wearing comfort and the need for the user to remain relatively still, the device's measurement sustainability is poor, and it can only complete continuous measurement in a short period of time; (3) The related devices are often large and difficult to miniaturize, making it difficult to integrate them into small devices used in daily life.
[0003] The continuous blood pressure detection method based on photoelectric sensing utilizes photoplethysmographic (PPG) methods. Although it has many advantages such as miniaturization of related devices, good wearing comfort, and low power consumption, it still has some problems that cannot be ignored: (1) In the method based on pulse transit time (PTT), the premise is that the elasticity of blood vessels remains unchanged. However, the elasticity of blood vessels will change under different physiological conditions, and the estimated blood pressure value will often have a certain error; (2) In the method based on pulse wave feature extraction, due to the need for privacy, there is often a problem of data silos. The amount of data is limited and often cannot meet the needs of deep learning. Summary of the Invention
[0004] The purpose of this invention is to propose a blood pressure measurement system based on the FedABN algorithm that can guarantee user privacy and collect a large amount of data.
[0005] The technical solution of this invention is: to provide a blood pressure measurement system based on the FedABN algorithm, the system comprising: a client and a server;
[0006] The client-side module includes: a data acquisition module, a storage module, a signal quality evaluation module, and a blood pressure estimation module.
[0007] The data acquisition module uses a camera to capture videos of the user's skin and extracts vPPG signals based on the periodic changes in skin color in the videos;
[0008] The data acquisition module sends the vPPG signal to the signal quality evaluation module for quality evaluation, performs windowing processing on the vPPG signal and calculates the autocorrelation coefficient of the signal in each window;
[0009] The signal quality evaluation module outputs a signal to the blood pressure estimation module, which receives the signal and calculates the blood pressure using a built-in pre-trained model.
[0010] Multiple clients submit their pre-trained model's BN layer data to the server. The server calculates client similarity based on the BN layer data, generates a personalized model for each client based on the client similarity and the parameters of each client, and sends the parameters of the personalized model back to each client to update the pre-trained model.
[0011] In any of the above technical solutions, the signal quality evaluation module first performs windowing processing on the vPPG signal to generate vPPG window signal data, then calculates the autocorrelation coefficient ACF(lag) on the vPPG window signal data, and writes the autocorrelation coefficient ACF(lag) as a weight parameter into the vPPG signal and outputs it.
[0012] In any of the above technical solutions, the windowing process further includes dividing the vPPG signal into two signal segments of length n, and the calculation steps for the autocorrelation coefficient ACF(lag) are as follows:
[0013] ,
[0014] ,
[0015] ,
[0016] ,
[0017] in, This is the mean value of the signal segment. This represents the i-th signal value of the signal segment, where lag is the delay of the vPPG signal, A is the lag segment before the current signal window, and B is the lag segment after the current signal window.
[0018] In any of the above technical solutions, further, during the first N uses, where N is a positive integer greater than 2, the blood pressure estimation module uses a pre-trained model built with a ResNet model to calculate blood pressure, and after the Nth blood pressure measurement, it uses all the measurement data from the first N measurements to update the parameters of the pre-trained model.
[0019] In any of the above technical solutions, further, when the blood pressure estimation module is used for the N+1th time, it submits the BN layer data of the pre-trained model to the server. The server aggregates the BN layer data sent by each client and calculates the client similarity. The specific calculation steps are as follows:
[0020] Let z i,lLet represent the input of the first batch normalization layer for the i-th client. Its Gaussian distribution is obtained by calculating the channel statistics. (µ i,l , σ i,l If ), then the BN layer statistics for the i-th client are:
[0021] (µ i , σ i ) = [(µ i, , σ i, ), (µ i, , σ i, ), · · · , (µ i,M , σ i,M )],
[0022] Where, σ i,l = Diag(r i,l M is the total number of channels. Therefore, the approximate value for calculating the Wasserstein distance is:
[0023] ,
[0024] Let there be a coefficient K. K is infinite when the value of ACF(lag) is less than 0.6; K is 5 when the value of ACF(lag) is greater than 0.6 and less than 0.7; K is 2 when the value of ACF(lag) is greater than 0.7 and less than 0.8; and K is 1 when the value of ACF(lag) is greater than 0.8. Therefore, the distance between two clients i and j is calculated as follows:
[0025] ,
[0026] Where L is the total number of clients, then the coefficient is calculated. Value:
[0027] ,
[0028] in Represents the amount of data from the i-th client; then ω * i,j Standardized to ω i,j ,Right now:
[0029] ,
[0030] We obtain an L*L client similarity matrix, where each matrix element is... .
[0031] In any of the above technical solutions, further, after calculating the client similarity, the server receives the parameters of other layers besides the Batch Normalization (BN) layer in the pre-trained model of each client and performs an aggregation operation. The aggregation operation involves the server updating the pre-trained model parameters of each client, and the update strategy is as follows:
[0032] ,
[0033] in Let be the BN layer parameters for the i-th client at time t. For the parameters of other layers at time t of the i-th client, This will be the BN layer parameter for the i-th client in the next use. These are the parameters for the other layers of the i-th client when used next.
[0034] In any of the above technical solutions, the client further includes: a storage module;
[0035] The data acquisition module sends vPPG signals to the storage module to store records, and the blood pressure estimation module connects to the storage module to send blood pressure measurement values to it to save as historical records.
[0036] In any of the above technical solutions, the client further includes: a signal preprocessing module; the signal preprocessing module receives the signal output by the signal quality evaluation module, performs high-frequency noise filtering, moving average filtering and median filtering on it in sequence, converts it into usable data for the blood pressure estimation module and sends it to it.
[0037] In any of the above technical solutions, further, when the data acquisition module uses the camera to capture a video of the user's skin, it simultaneously turns on a flashlight to increase the illumination.
[0038] A blood pressure measurement method based on the FedABN algorithm, the method comprising:
[0039] S1. The user places the camera connected to the client close to the skin, the client controls the camera to shoot skin video, and the client extracts the vPPG signal from the skin video;
[0040] S2. The client evaluates the quality of the vPPG signal extracted in step S1, calculates the autocorrelation coefficient ACF (lag) of the signal and writes it into the vPPG signal as a weight parameter, and then performs high-frequency noise filtering, moving average filtering and median filtering on the vPPG signal in sequence.
[0041] S3. In the first N measurements on the client, the vPPG signal processed in step S2 is used to calculate the user's blood pressure through the built-in pre-trained model. The client saves the measurement data of the first N measurements and updates the parameters of the pre-trained model using all the measurement data of the first N measurements after the Nth measurement is completed. The client then uploads the updated pre-trained model parameters to the server.
[0042] S4. After receiving the pre-trained model parameters uploaded by all clients, the server aggregates the data and calculates the client similarity. The server updates the received pre-trained model parameters of each client based on the client similarity and sends the updated parameters back to each client. Each client updates its pre-trained model into a personalized model based on the received parameters.
[0043] S5. Starting from the N+1th measurement by the client, the client uses the personalized model obtained in step S4 to calculate the user's blood pressure.
[0044] The beneficial effects of this invention are:
[0045] The technical solution in this invention uses a pre-trained model to measure blood pressure. Based on updating the parameters of the pre-trained model according to each measurement data, the FedABN algorithm enables interaction between the client and the server. It uses the BN layer parameters of multiple client pre-trained models to construct client similarity, aggregates the parameters, and updates the personalized model for each client. The local model is trained using the training data obtained by each client, and the model parameters are aggregated on the server. This is equivalent to using enough training data to update the model and obtain a personalized model for each client without obtaining the user's specific measurement data and protecting user privacy, thereby achieving more accurate blood pressure estimation. Attached Figure Description
[0046] The advantages of the above and additional aspects of the present invention will become apparent and readily understood in the description of the embodiments in conjunction with the following drawings, wherein:
[0047] Figure 1 This is a schematic diagram of a blood pressure measurement system based on the FedABN algorithm according to an embodiment of the present invention;
[0048] Figure 2 This is a schematic diagram of the FedABN algorithm process of a blood pressure measurement system based on the FedABN algorithm according to an embodiment of the present invention.
[0049] Among them, 101-data acquisition module, 102-storage module, 103-signal quality evaluation module, 104-signal preprocessing module, and 105-blood pressure estimation module. Detailed Implementation
[0050] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.
[0051] In the following description, many specific details are set forth in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.
[0052] vPPG (video photoplethysmography) is a video-based photoplethysmography method that uses a mobile phone camera or a regular consumer camera to record minute changes in skin color caused by changes in blood volume in facial capillaries. It has the advantages of low cost, non-contact, and simplicity.
[0053] The peak value of the vPPG signal is caused by vasoconstriction. Therefore, we can obtain the transmission time of blood from the heart to the measurement site, which is the pulse transit time (PTT). The speed of pulse transit is directly related to blood pressure. When blood pressure is high, the pulse transit is fast, and vice versa. This enables non-invasive continuous blood pressure measurement.
[0054] like Figure 1 As shown, this embodiment provides a blood pressure measurement system based on the FedABN algorithm. The system includes: a data acquisition module 101, a storage module 102, a signal quality evaluation module 103, a signal preprocessing module 104, and a blood pressure estimation module 105.
[0055] In this embodiment, the data acquisition module 101 includes a mobile phone camera and a mobile phone flashlight. After the user runs the mobile phone's blood pressure measurement software system, he / she turns on the mobile phone camera and places the mobile phone flashlight on his / her finger. The mobile phone camera captures a video of the user's finger skin, and the vPPG signal is extracted based on the periodic changes in skin color in the captured video.
[0056] After the data acquisition module 101 obtains the vPPG signal, it sends the signal to the storage module 102 and the signal quality evaluation module 103 respectively. The storage module 102 stores and records the vPPG signal, and the signal quality evaluation module 103 evaluates the quality of the vPPG signal and adds weight parameters to it.
[0057] Specifically, the signal quality evaluation module 103 first performs windowing processing on the vPPG signal to generate vPPG window signal data, then calculates the autocorrelation coefficient on the generated vPPG window signal data, and writes the autocorrelation coefficient ACF(lag) as a weight parameter into the vPPG signal and sends it to the signal preprocessing module 104.
[0058] The specific process of windowing the vPPG signal involves dividing the signal into two segments of length n, thus:
[0059] ,
[0060] This is the mean value of the signal segment. Let represent the i-th signal value of this signal segment, and then calculate the autocorrelation coefficient (ACF) of this signal segment:
[0061] ,
[0062] ,
[0063] ,
[0064] Where lag is the delay of the vPPG signal, A is the lag before the current signal window, and B is the lag after the current signal window.
[0065] The signal preprocessing module 104 receives the signal output by the signal quality evaluation module 103, preprocesses it, and converts it into model-usable data for the subsequent blood pressure estimation module 105.
[0066] To preserve the original information to the greatest extent, the signal preprocessing module 104 uses a low-pass filter with a cutoff frequency of 8Hz to filter out high-frequency noise from the signal. In addition, since the vPPG signal is often affected by the light environment during acquisition, resulting in signal jitter, the signal preprocessing module 104 uses a moving average filter to eliminate the jitter noise after filtering the high-frequency noise, and finally uses a median filter to remove the baseline drift of the signal.
[0067] The blood pressure estimation module 105 is pre-trained using a traditional ResNet model built in-house. The blood pressure estimation module 105 is connected to the signal preprocessing module 104 and receives signals from the signal preprocessing module 104. The blood pressure estimation module 105 is also connected to the storage module 102. After calculating the measured blood pressure data, the estimated blood pressure data is sent to the storage module 102 for storage as historical data.
[0068] During the first 10 uses, the blood pressure estimation module 105 uses a pre-trained model to calculate blood pressure. After the 10th blood pressure measurement, the parameters of the pre-trained model are updated using all the measurement data from the previous 10 measurements. Then, the local model parameters are uploaded to the server, and the server returns personalized model parameters to update the local model.
[0069] Starting from the 11th use, the blood pressure estimation module 105 uses a personalized model of the user created through the FedABN algorithm, which increases the amount of training data for the model while ensuring user privacy, thereby improving the accuracy of blood pressure measurement.
[0070] like Figure 2 As shown, the client submits the BN layer and other layers of the pre-trained model to the server. The server aggregates the BN layer data sent by each client and calculates the client similarity. The specific calculation steps are as follows:
[0071] Let z i,l Let represent the input of the first batch normalization layer for the i-th client. Its Gaussian distribution is obtained by calculating the channel statistics. (µ i,l , σ i,l If ), then the BN layer statistics for the i-th client are:
[0072] (µ i , σ i ) = [(µ i,1 , σ i,1 ), (µ i,2 , σ i,2 ), · · · , (µ i,M , σ i,M )],
[0073] Where, σ i,l = Diag(r i,l M is the total number of channels. Therefore, the approximate value for calculating the Wasserstein distance is:
[0074] ,
[0075] Let there be a coefficient K. K is infinite when the value of ACF(lag) is less than 0.6; K is 5 when the value of ACF(lag) is greater than 0.6 and less than 0.7; K is 2 when the value of ACF(lag) is greater than 0.7 and less than 0.8; and K is 1 when the value of ACF(lag) is greater than 0.8. Therefore, the distance between two clients i and j is calculated as follows:
[0076] ,
[0077] Where L is the total number of clients, then the coefficient is calculated. Value:
[0078] ,
[0079] in Represents the amount of data from the i-th client; then ω * i,j Standardized to ω i,j ,Right now:
[0080] ,
[0081] This yields an L*L client similarity matrix, where each matrix element is... .
[0082] After calculating client similarity, the server loads the parameters of other layers in the pre-trained models of each client, performs aggregation operations on the server, and obtains the personalized model parameters of each client, which are then sent back to the corresponding client.
[0083] set up Let be the BN layer parameters for the i-th client at time t. The parameters of other layers for the i-th client at time t will change to the BN layer parameters for the i-th client in the next use. The parameters of the other layers of the i-th client become For aggregation operations on the server, the parameter update strategy is as follows:
[0084] ,
[0085] The parameters obtained from the server are sent back to the corresponding client. The client uses these parameters to adjust the pre-trained model to obtain a personalized model for that client. The personalized model is then used to provide the user with subsequent blood pressure monitoring services.
[0086] In summary, this invention proposes a blood pressure measurement system based on the FedABN algorithm, comprising a client and a server, wherein the client includes a data acquisition module 101, a storage module 102, a signal quality evaluation module 103, a signal preprocessing module 104, and a blood pressure estimation module 105.
[0087] The data acquisition module 101 captures a video of the user's finger skin using the mobile phone camera, extracts the vPPG signal, and sends it to the storage module 102 and the signal quality evaluation module 103. The signal quality evaluation module 103 evaluates the signal quality and then sends it to the signal preprocessing module 104. The signal preprocessing module 104 filters out noise and then sends it to the blood pressure estimation module 105. The blood pressure estimation module 105 uses a pre-trained model to measure blood pressure for the first 10 times. After the 10th use, it updates the model parameters with all the measurement data from the first 10 times and sends the model's BN layer and other layer data to the server. The server calculates the parameters for the personalized blood pressure measurement model for the client and sends them to the client. The client starts using the personalized blood pressure measurement model to measure blood pressure from the 11th time.
[0088] The steps in this invention can be adjusted, combined, or deleted according to actual needs.
[0089] The units in the device of the present invention can be merged, divided, or reduced according to actual needs.
[0090] Although the invention has been disclosed in detail with reference to the accompanying drawings, it should be understood that these descriptions are merely exemplary and not intended to limit the application of the invention. The scope of protection of the invention is defined by the appended claims and may include various modifications, alterations, and equivalents made to the invention without departing from the scope and spirit of the invention.
Claims
1. A blood pressure measurement system based on the FedABN algorithm, characterized in that, The system includes: a client and a server; The client includes: a data acquisition module (101), a signal quality evaluation module (103), and a blood pressure estimation module (105). The data acquisition module (101) uses a camera to capture a video of the user's skin and extracts the vPPG signal based on the periodically changing skin color in the video. The data acquisition module (101) sends the vPPG signal to the signal quality evaluation module (103) for quality evaluation, performs windowing processing on the vPPG signal and calculates the autocorrelation coefficient of the signal in each window; The signal quality evaluation module (103) first performs windowing processing on the vPPG signal to generate vPPG window signal data, then calculates the autocorrelation coefficient ACF(lag) on the vPPG window signal data, and writes the autocorrelation coefficient ACF(lag) as a weight parameter into the vPPG signal and outputs it. The signal quality evaluation module (103) outputs a signal to the blood pressure estimation module (105), and the blood pressure estimation module (105) receives the signal and calculates the blood pressure using a built-in pre-trained model; Multiple clients submit Batch Normalization (BN) layer data of their respective pre-trained models to the server. The server calculates client similarity based on the BN layer data, and generates a personalized model for each client based on the client similarity and the parameters of other layers besides the BN layer in each client's pre-trained model. The server then sends the parameters of the personalized model back to each client to update the pre-trained model.
2. The blood pressure measurement system based on the FedABN algorithm as described in claim 1, characterized in that, The windowing process includes dividing the vPPG signal into two signal segments of length n, and the calculation steps for the autocorrelation coefficient ACF(lag) are as follows: ， ， ， ， in, This is the mean value of the signal segment. This represents the i-th signal value of the signal segment, where lag is the delay of the vPPG signal, A is the lag before the current signal window, and B is the lag after the current signal window.
3. The blood pressure measurement system based on the FedABN algorithm as described in claim 1, characterized in that, In the first N uses, where N is a positive integer greater than 2, the blood pressure estimation module (105) uses the pre-trained model built with the ResNet model to calculate blood pressure. After the Nth blood pressure measurement, the parameters of the pre-trained model are updated using all the measurement data from the previous N measurements.
4. The blood pressure measurement system based on the FedABN algorithm as described in claim 3, characterized in that, When the blood pressure estimation module (105) is used for the N+1th time, it submits the BN layer data of the pre-trained model to the server. The server summarizes the BN layer data sent by each client and calculates the client similarity. The specific calculation steps are as follows: Let z i,l Let represent the input of the first batch normalization layer for the i-th client. Its Gaussian distribution is obtained by calculating the channel statistics. (µ i,l , σ i,l If ), then the BN layer statistics for the i-th client are: (µ i , s i ) = [(µ i,1 , s i,1 ), (µ i,2 , s i,2 ), · · · , (µ i,M , s i,M )], Where, σ i,l = Diag(r i,l M is the total number of channels. Therefore, the approximate value for calculating the Wasserstein distance is: ， Let there be a coefficient K. K is infinite when the value of ACF(lag) is less than 0.6; K is 5 when the value of ACF(lag) is greater than 0.6 and less than 0.7; K is 2 when the value of ACF(lag) is greater than 0.7 and less than 0.8; and K is 1 when the value of ACF(lag) is greater than 0.
8. Therefore, the distance between two clients i and j is calculated as follows: ， Where L is the total number of clients, then the coefficient is calculated. Value: ， in Represents the amount of data from the i-th client; then ω * i,j Standardized to ω i,j ,Right now: ， We obtain an L*L client similarity matrix, where each matrix element is... .
5. The blood pressure measurement system based on the FedABN algorithm as described in claim 4, characterized in that, After calculating the client similarity, the server receives the parameters of all layers other than the Batch Normalization (BN) layer in the pre-trained model of each client and performs an aggregation operation. This aggregation operation updates the pre-trained model parameters of each client on the server, and the update strategy is as follows: ， in Let be the BN layer parameters for the i-th client at time t. For the parameters of other layers at time t of the i-th client, This will be the BN layer parameter for the i-th client in the next use. These are the parameters for the other layers of the i-th client when used next.
6. The blood pressure measurement system based on the FedABN algorithm as described in claim 1, characterized in that, The client also includes: a storage module (102); The data acquisition module (101) sends the vPPG signal to the storage module (102) to store the record, and the blood pressure estimation module (105) connects to the storage module (102) to send the blood pressure measurement value to it for storage as a historical record.
7. The blood pressure measurement system based on the FedABN algorithm as described in claim 1, characterized in that, The client also includes a signal preprocessing module (104); the signal preprocessing module (104) receives the signal output by the signal quality evaluation module (103), performs high-frequency noise filtering, moving average filtering and median filtering on it in sequence, converts it into usable data for the blood pressure estimation module (105) and sends it to it.
8. The blood pressure measurement system based on the FedABN algorithm as described in claim 1, characterized in that, When the data acquisition module (101) uses the camera to capture a video of the user's skin, it simultaneously turns on the flashlight to increase the illumination.