A blood pressure monitoring method and wearable device based on PPG and cuff pressure synchronization
By employing a split-structure design and time alignment of the neural network model, along with incremental updates in the cloud, the problems of unstable accuracy and poor comfort in existing wearable blood pressure monitoring devices have been solved, achieving lightweight, second-level continuous monitoring and adaptive optimization for high-precision blood pressure management.
Patent Information
- Application Number
- CN202611475108.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-21
- Publication Date
- 2026-11-10
AI Technical Summary
Existing wearable blood pressure monitoring devices suffer from problems such as unstable accuracy, poor comfort, large device size, and inability to achieve continuous monitoring at the second level. Furthermore, existing fusion solutions fail to achieve deep coupling between the data layer, algorithm layer, and hardware layer.
It adopts a split structure design, and achieves time alignment between the airbag pressure value and the PPG signal by synchronously triggering the PPG sensor and airbag assembly, using a unified clock source and linear interpolation. It builds a user-specific neural network model, and optimizes the model through a cloud incremental update mechanism, realizing automatic switching between detachable airbag assembly and pure PPG monitoring mode.
It achieves lightweight wearability, second-level continuous blood pressure monitoring, adaptive optimization, and high-precision blood pressure estimation, and can adapt to changes in individual user physiological characteristics, providing 24-hour uninterrupted blood pressure management.
Smart Images

Figure CN123058692A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wearable medical health monitoring technology, and in particular to a blood pressure monitoring method and wearable device based on PPG and airbag pressure synchronization. Background Technology
[0002] Existing wearable blood pressure monitoring technologies fall into two categories: cuff oscillometric and pure PPG estimation. Both have inherent limitations, and their fusion solutions fail to achieve deep coupling between the data, algorithm, and hardware layers. Cuff oscillometric methods offer high accuracy and meet medical standards, but they are intermittent measurements, unable to achieve continuous monitoring down to the second level. Furthermore, the repeated inflation of the cuff compresses the skin, leading to poor comfort, and the overall airway is bulky, making it difficult to lightweight the device. Pure PPG solutions rely on photoplethysmography (PPG) to acquire signals and fit blood pressure through machine learning, enabling non-invasive, continuous, real-time monitoring with compact devices. However, their model training is highly dependent on external medical cuff calibration. External calibration and local PPG acquisition have time, position, and equipment differences, resulting in a spatiotemporal mismatch between training labels and feature signals. The model has poor versatility, with significant accuracy drift across different users and physical conditions, requiring frequent manual calibration. It also lacks self-iterative capabilities, leading to continuous accuracy degradation with long-term use.
[0003] Existing fusion products are simply hardware stack-ups, with the two data streams operating independently without synchronous acquisition, time-series alignment, or joint modeling mechanisms; the devices are mostly integrated structures, with airbags that cannot be disassembled, making it impossible to switch between high-precision calibration and lightweight monitoring modes as needed; the models only support single-cycle training, cannot continuously and incrementally optimize based on historical data, and have weak personalized adaptation capabilities.
[0004] Existing patent number EP4344621A1 discloses a wearable device that integrates airbags and PPG. In this solution, PPG acquisition and airbag inflation are deliberately staggered in time to avoid interference of airbag pressure on the PPG signal. Therefore, it is impossible to achieve synchronous supervised learning with airbag pressure value as the true value label of PPG signal. Summary of the Invention
[0005] To overcome the shortcomings of existing technologies, this invention provides a blood pressure monitoring method and wearable device based on PPG and airbag pressure synchronization.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] The first aspect of this invention provides a blood pressure monitoring method based on PPG and cuff pressure synchronization, comprising the following steps:
[0008] S1. Provide a wearable device that integrates a PPG sensor and an airbag assembly to obtain user personal information and build an initial neural network model;
[0009] S2. Inflate and pressurize the airbag assembly to obtain the airbag internal pressure value P(t), and simultaneously trigger the PPG sensor to collect the PPG signal S(t). Record and align the airbag internal pressure value P(t) and the corresponding PPG signal S(t) at the same time according to the time to obtain the paired dataset {S(t), P(t)}.
[0010] S3. Using the systolic blood pressure and / or diastolic blood pressure obtained by oscillometric processing of the airbag pressure value P(t) as supervision labels, and using the synchronously acquired PPG signal S(t) and user personal information as input features, a neural network model is trained using the paired dataset {S(t), P(t)}, and a user-specific blood pressure mapping model is constructed.
[0011] S4. Deploy the trained neural network model on the wearable device. In the non-airbag component pressurized state, only the PPG signal S(t) collected by the PPG sensor is input into the neural network model, and the blood pressure estimate is output.
[0012] Further, in step S2, the airbag pressure value P(t) and the corresponding PPG signal S(t) at the same time are aligned using linear interpolation to form a paired dataset {S(t), P(t)} with completely corresponding timestamps. The paired dataset {S(t), P(t)} is uploaded to the cloud server. The cloud server performs online incremental updates to the neural network model based on the paired dataset {S(t), P(t)} and historical paired datasets, and transmits the updated neural network model weights to the wearable device.
[0013] Furthermore, in step S2, the preset time delay for the synchronization trigger does not exceed 50ms, and the acquisition clocks of the PPG sensor and the pressure sensor are provided by the same clock source.
[0014] Furthermore, in step S3, the neural network model includes an input layer, a convolutional feature extraction layer, a bidirectional long short-term memory network layer, a fully connected layer, and an output layer; mean squared error, mean absolute error, or Huber loss are used as loss functions, and the Adam optimizer is used to complete parameter iterative convergence.
[0015] Furthermore, it also includes the following steps: S5, detecting the installation status of the airbag assembly; when the airbag assembly is detected to be not installed, automatically switching to the pure PPG sensor monitoring mode, and continuously outputting blood pressure estimates using the most recently updated neural network model.
[0016] A second aspect of the present invention provides a wearable device for blood pressure monitoring, used to perform the blood pressure monitoring method according to any one of claims 1-5, comprising:
[0017] The meter head has a built-in main control chip, a power supply battery, an air pump, and a pressure sensor. The power supply battery is used to power the device.
[0018] The watch strap is connected to the watch head, and the watch strap is provided with a slot for installing the airbag assembly;
[0019] An airbag assembly is disposed inside the watch strap. The airbag assembly is detachably connected to the air pump and pressure sensor inside the watch head and is used to collect the pressure value P(t) inside the airbag under inflation and pressurization conditions.
[0020] A PPG sensor is located at the bottom of the instrument head and / or on the skin-facing surface of the airbag assembly for acquiring PPG signal S(t);
[0021] The main control chip integrates a synchronous acquisition unit and a model inference unit.
[0022] The synchronous acquisition unit is used to align and interpolate the pressure value P(t) inside the airbag and the PPG signal S(t) according to time to obtain a paired dataset {S(t), P(t)}; the model inference unit stores a neural network model and outputs a blood pressure estimate based on the PPG signal S(t) acquired by the PPG sensor.
[0023] Furthermore, the airbag assembly and the watch strap are separate structures. The airbag assembly is connected to the air pump and pressure sensor inside the watch head through a detachable air passage interface, and a silicone sealing ring is provided at the air passage interface. The airbag assembly is provided with a metal adsorption ring, and the watch head is provided with a ring permanent magnet adapted to the ferromagnetic metal ring.
[0024] Furthermore, the PPG sensor employs a dual-wavelength LED, which includes green light with a wavelength of 530nm and infrared light with a wavelength of 940nm, to alternately emit light and be used in conjunction with a photodiode receiver to acquire the PPG signal S(t); the sampling rate of the PPG sensor can be configured to 125Hz, 250Hz or 500Hz, and the sampling rate of the pressure sensor is not less than 200Hz.
[0025] Furthermore, the main control chip also includes a detection unit, which is used to detect the installation status of the airbag assembly. When the airbag assembly is not installed, the main control chip is triggered to switch to pure PPG monitoring mode and disable the air pump and pressure sensor to save power. The most recently updated neural network model stored in the model inference unit is called to continuously output blood pressure estimates.
[0026] Furthermore, the device also includes a wireless communication module, which is electrically connected to the main control chip and is used to interact with the cloud server to pair data and perform weight difference analysis of the neural network model, thereby completing the remote iterative update of the user's exclusive blood pressure mapping model.
[0027] The present invention provides a blood pressure monitoring method and wearable device based on PPG and cuff pressure synchronization, which has the following advantages:
[0028] 1) Split-type structure design: The airbag component can be removed at any time, making it easy to wear; when the airbag component is present, it serves as a "mobile calibrator" to provide the gold standard label for oscillometric method, and when the airbag component is absent, it serves as a "continuous monitor" to provide real-time blood pressure estimation, making it a dual-purpose device.
[0029] 2) Time synchronization: The synchronization trigger delay is controlled within 50ms, and the signal is strictly aligned through a unified clock source and linear interpolation, which ensures the spatiotemporal consistency between training labels and input features, thereby improving model accuracy.
[0030] 3) Adaptive optimization: Through the cloud-based incremental update mechanism, the model can continuously learn about changes in the physiological characteristics of individual users (such as seasonal changes and drug effects), becoming more and more accurate with use.
[0031] 4) Mode switching: After the airbag assembly is removed, the latest model is automatically activated for pure PPG monitoring, which is imperceptible to the user and enables 24-hour uninterrupted blood pressure management.
[0032] 5) Deep learning architecture: CNN extracts local waveform features and captures long-term temporal dependencies, which is particularly suitable for physiological signals of PPG that have both high-frequency details and low-frequency trends. Attached Figure Description
[0033] Figure 1 A schematic flowchart of the blood pressure monitoring method provided in an embodiment of the present invention;
[0034] Figure 2 This is an exploded structural diagram of a wearable device provided in an embodiment of the present invention.
[0035] Figure 3 This is a three-dimensional structural diagram of a wearable device provided in an embodiment of the present invention;
[0036] Figure 4 A schematic diagram of the training layer structure of the neural network model provided in the embodiments of the present invention.
[0037] Figure 5 A schematic block diagram of the neural network model deployment structure provided in an embodiment of the present invention;
[0038] Figure 6 This is a schematic block diagram illustrating the module connections in an embodiment of the present invention.
[0039] Reference numerals: 1. Meter head; 11. Main control chip; 111. Synchronous acquisition unit; 112. Model inference unit; 113. Detection unit; 12. Power supply battery; 13. Air pump; 14. Pressure sensor; 15. Ring permanent magnet; 16. Wireless communication module; 2. Watch strap; 21. Card slot; 3. Airbag assembly; 31. Airway interface; 32. Metal adsorption ring; 33. Silicone sealing ring; 34. Electrical contact; 4. PPG sensor; 5. Cloud server. Detailed Implementation
[0040] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.
[0041] like Figures 1-6 As shown, the wearable device of the present invention is in the form of a smartwatch, which is composed of a watch head 1, a watch strap 2, and an airbag assembly 3. A slot 21 is provided on the inner side of the watch strap 2 for fitting and fixing the airbag assembly 3 to prevent displacement during measurement.
[0042] In this embodiment, the meter head 1 is made of medical-grade PC material, and the meter head 1 integrates a main control chip 11, a power supply battery 12, an air pump 13, and a pressure sensor 14. The main control chip 11 is a low-power Bluetooth chip, used for output command control, data acquisition scheduling, model inference calculation, and wireless communication management of the wearable device. Specifically, as shown... Figure 6 As shown, the main control chip 11 integrates a synchronous acquisition unit 111, a model inference unit 112, and a detection unit 113. The synchronous acquisition unit 111 is implemented by a timer inside the main control chip 11, used to align and interpolate the airbag pressure value and PPG signal according to time, to obtain a paired dataset {S(t), P(t)}. When the main control chip 11 starts the airbag assembly 3 pressurization measurement, the timer simultaneously triggers the air pump 13 to start inflating at t=0, and the pressure sensor 14 and PPG sensor 4 start acquiring data. The synchronous trigger delay time does not exceed 50ms. Both signals are timestamped based on the same clock source and stored in the wearable device. For data where the time points do not completely overlap due to the different sampling rates of the pressure sensor 14 and PPG sensor 4, the synchronous acquisition unit 111 calls a linear interpolation algorithm for alignment: setting the airbag pressure value as P(t) and the PPG signal as S(t), forming a paired dataset {S(t), P(t)} with completely corresponding timestamps. The paired dataset {S(t), P(t)} is uploaded to cloud server 5. Cloud server 5 performs online incremental updates to the neural network model based on the paired dataset {S(t), P(t)} and historical paired datasets, and transmits the updated neural network model weights to the wearable device.
[0043] The model inference unit 112 stores a neural network model, generates blood pressure estimates based on the signals collected by the PPG sensor 4, and performs offline high-speed inference. The model inference unit 112 is implemented based on the neural network acceleration engine inside the main control chip 11, loads the pre-trained neural network model, receives real-time PPG signals S(t) and user personal information, and outputs predicted values of systolic and diastolic blood pressure after forward propagation calculation.
[0044] In this embodiment, the main control chip 11 further includes a detection unit 113. The detection unit 113 includes a Hall sensor located next to the air passage interface 31 of the meter head 1 and a miniature magnet embedded at the corresponding position of the airbag assembly 3, used to detect the closed state of the magnetic connector. When the airbag assembly 3 is installed in place, the magnetic connector is closed, the miniature magnet approaches the Hall sensor, and the Hall sensor outputs a high-level signal; when the airbag assembly 3 is removed, the magnetic field disappears, and the Hall sensor outputs a low-level signal. The main control chip 11 reads this level state to determine the installation state of the airbag assembly 3. In another optional embodiment, the detection unit 113 may also use a mechanical micro switch, a resistance detection circuit, or detect the on / off state of the electrical contact 34 and the ejector pin connector to detect the installation state.
[0045] like Figure 2 , Figure 3 As shown, the airbag assembly 3 and the watch strap 2 are separate structures. The airbag assembly 3 is not fixed to the watch strap 2. The watch strap 2 is made of skin-friendly fluororubber and its length is adjustable to accommodate users with different wrist sizes. Specifically, the airbag assembly 3 is embedded in the slot 21 of the watch strap. At the same time, the end of the airbag assembly 3 near the watch head 1 is provided with an air passage interface 31 and an electrical contact 34. The airbag assembly 3 is connected to the air pump 13 and pressure sensor 14 inside the watch head 1 through the detachable air passage interface 31. A silicone sealing ring 33 (Shore A hardness 40°) is provided at the air passage interface 31 for compression sealing to ensure no air leakage or pressure drop, effectively improving the pressure acquisition accuracy. The electrical contact 34 contacts the pin connector inside the watch head 1 to realize the signal transmission of the detection unit 113. The airbag assembly 3 is provided with a metal adsorption ring 32, and the watch head 1 is provided with a ring permanent magnet 15 adapted to the ferromagnetic metal ring, thereby realizing magnetic attraction for rapid positioning.
[0046] The power supply battery 12 is a lithium polymer battery with a capacity ≥200mAh. The power supply battery 12 is electrically connected to the main control chip 11 via a ribbon cable and supports magnetic wireless charging for charging the device. The air pump 13 is a miniature diaphragm air pump 13 with an operating voltage of 3.7V, a maximum output pressure ≥300mmHg, and a flow rate ≥0.5L / min. The air pump 13's air inlet is open to the atmosphere, and its air outlet is connected to the air path interface 31 at the bottom of the meter 1 via an internal air path. The pressure sensor 14 is a piezoresistive pressure sensor 14 with a sampling rate configured to be no less than 200Hz, a measurement range of 0-300mmHg, and an accuracy of ±1mmHg. The air path input terminal of the pressure sensor 14 is connected to the air path between the air pump 13's air outlet and the airbag assembly 3 for real-time detection of the pressure value inside the airbag assembly 3.
[0047] In this embodiment, the PPG sensor 4 is positioned at the bottom of the watch head 1 close to the skin of the wrist, and can also be positioned on the surface of the airbag assembly 3 facing the skin. Specifically, the PPG sensor uses a dual-wavelength LED, which includes green light with a wavelength of 530nm and infrared light with a wavelength of 940nm, to alternately emit light and, together with a photodiode receiver, collect the PPG signal S(t). The sampling rate of the PPG sensor 4 can be configured to 125Hz, 250Hz, or 500Hz, and DMA transmission is triggered by the timer of the main control chip 11 to ensure uniform sampling intervals. In a preferred embodiment, the sampling rate of the PPG sensor 4 is configured to 250Hz, at which point the sampling window of the PPG sensor 4 collecting the PPG signal S(t) contains exactly 3000 sampling points.
[0048] In a preferred embodiment, the device further includes a wireless communication module 16, which is electrically connected to the main control chip 11 and establishes a wireless connection with the cloud server 5 for uploading paired datasets {S(t), P(t)}. The wireless communication module 16 supports Bluetooth and WIFI encrypted transmission. In this process, only input features and supervision labels are uploaded, and the original waveforms are not uploaded to protect user privacy. The wireless communication module 16 also receives the neural network model weight difference file sent by the cloud server 5 and completes the remote iterative update of the user's exclusive blood pressure mapping model by superimposing it with the weights of the local application model.
[0049] This embodiment employs the aforementioned watch-shaped split-type wearable device, equipped with a main control chip 11, a pressure sensor 14, a PPG sensor 4, and a detachable airbag assembly 3, to fully execute the monitoring method of this invention. Figure 1 As shown, the method steps include:
[0050] S1. Device provision and initialization: Provide a wearable device that integrates a PPG sensor 4 and a detachable airbag assembly 3. Record user personal information on the wearable device, including age, gender, height, weight and wrist circumference, as auxiliary features to participate in subsequent model training; initialize the wearable device and build an initial neural network model.
[0051] S2. Synchronous Acquisition, Interpolation Alignment, and Cloud Incremental Update: When the user activates the precise calibration mode, the device starts the micro air pump 13 to uniformly inflate and pressurize the airbag assembly. The pressure sensor 14 acquires the pressure value inside the airbag assembly at a frequency of ≥200Hz. The main control chip 11 synchronously triggers the PPG sensor 4 to acquire the PPG signal S(t) through the same clock source, with the synchronization delay time controlled within 50ms. The sampling rate of the PPG sensor 4 is configured to 250Hz. Due to the different sampling frequencies of the pressure sensor 14 and the PPG sensor 4, the original paired dataset {S(t), P(t)} has discrete time misalignment. The synchronous acquisition unit 111 uses the device timestamp as a reference and uses a linear interpolation algorithm to fill in the discrete data points, realizing frame-by-frame temporal alignment between the pressure value data inside the airbag assembly and the waveform data of the PPG signal S(t), thus obtaining the paired dataset {S(t), P(t)}. Subsequently, the wearable device uploads the paired dataset {S(t), P(t)} to the cloud server 5 via the wireless communication module 16. The cloud server 5 retrieves all historical paired datasets for the user and uses an online incremental learning strategy to fine-tune and iterate the user-specific blood pressure mapping model, then transmits the updated neural network model weights to the wearable device. This invention differs from single-cycle training schemes by jointly training with the real-time paired dataset {S(t), P(t)} and historical paired datasets. This preserves long-term user characteristics while adapting to the latest changes in physical condition. After iteration, only a small-volume weight difference file is pushed to the wearable device, quickly completing model upgrades and achieving long-term adaptive and accurate optimization.
[0052] S3. Model Training: The device's built-in oscilloscope algorithm analyzes the pressure oscillation waveform and extracts the systolic and diastolic blood pressure measurements as supervision labels for the initial neural network model training. The time-aligned complete PPG waveform sequence is concatenated with the user's personal information to form multi-dimensional input features, which are then input into the initial neural network model for supervised training. During the training process, the loss function can be adaptively selected according to the scenario (e.g., mean squared error (MSE) loss is used for static sedentary measurements, mean absolute error (MAE) loss is used for daily activity monitoring, and Huber loss is used for complex motion states). In conjunction with the Adam optimizer, the learning rate is adaptively adjusted to quickly converge the optimal parameters and establish a user-specific blood pressure mapping model.
[0053] Specifically, the paired dataset {S(t), P(t)} is input into the initial neural network model to learn the mapping relationship between input features and supervision labels, thus constructing a blood pressure mapping model. The historical database of the paired dataset {S(t), P(t)} is divided into a 70% mapping training set, a 15% mapping validation set, and a 15% mapping test set. The mapping training set is input into the blood pressure mapping model for training, and the mapping validation set is input into the trained blood pressure mapping model for iterative hyperparameter optimization. The mapping test set is input into the iteratively optimized blood pressure mapping model to calculate the mean absolute error (MAE). The MAE threshold is set to MAE0 (e.g., 5 mmHg), where:
[0054] When MAE≤MAE0, the blood pressure mapping model is deemed to have met the training criteria and is used as the output of the neural network model.
[0055] When MAE > MAE0, the blood pressure mapping model is deemed unqualified during training. The paired dataset {S(t), P(t)} is updated online and retrained until the blood pressure mapping model meets the training requirements. This updated model is then output as the neural network model, and its weights are transmitted to the wearable device. The mean absolute error (MAE) regression metric is used for evaluation to ensure the scientific validity and accuracy of the model assessment.
[0056] like Figure 4 As shown, the neural network model in this embodiment uses a hybrid architecture combining a one-dimensional convolutional neural network and a bidirectional long short-term memory network as the blood pressure mapping model. The hierarchical structure is as follows:
[0057] 1) Input layer: Receives a fixed 12s duration, 250Hz sampling rate PPG raw signal with a total of 3000 sampling points to ensure complete coverage of the pulse cycle waveform characteristics;
[0058] 2) Convolutional Feature Extraction Layer: Contains two one-dimensional convolutional layers. The first layer has a kernel size of 5, a stride of 2, and 32 output channels. The second layer has a kernel size of 3, a stride of 2, and 64 output channels. Each layer is followed by batch normalization and ReLU activation to extract shallow pulse waveform peaks and valleys, as well as deep subtle temporal features. The user information vector is mapped through an independent fully connected layer and then concatenated with the CNN output features.
[0059] 3) Bidirectional Long Short-Term Memory Network Layer: Bidirectional LSTM, 128 hidden units, bidirectionally captures the temporal correlation features of pulse signals, adapting to the dynamic fluctuation pattern of human blood pressure.
[0060] 4) Fully connected layer: 64 output units, coupled with 0.3 Dropout regularization, effectively prevents model overfitting and improves generalization ability;
[0061] 5) Output layer: dual output neurons, corresponding to the predicted systolic blood pressure and the predicted diastolic blood pressure respectively.
[0062] S4. Model Deployment and Pure PPG Continuous Blood Pressure Monitoring: After deploying the trained neural network model on the wearable device, the device enters the normal monitoring mode. The PPG sensor 4 collects the user's wrist pulse signal in real time, inputs it into the latest iterative model, and outputs a second-level continuous blood pressure estimate in real time, realizing 24-hour imperceptible dynamic monitoring.
[0063] S5. Adaptive Mode Switching: The device monitors the magnetic connection status of the airbag assembly 3 in real time. When the airbag assembly 3 is detected as not installed, the device automatically disables the air pump 13 drive and pressure sensor 14 data acquisition to save power, switches to low-power pure PPG monitoring mode, and continuously outputs blood pressure estimates using the most recently updated neural network model stored in the model inference unit 112. After reinstalling the airbag assembly 3, the synchronous calibration and model training functions are automatically restored.
[0064] Experimental testing revealed that as the synchronization trigger delay gradually increased from 0ms to 100ms, the time alignment error between the PPG signal S(t) and the airbag component pressure value in the paired dataset {S(t), P(t)} increased linearly. When the delay exceeded 50ms, the phase shift of a complete pulse cycle (approximately 0.8 seconds) in the PPG signal S(t) exceeded 6%, causing a misalignment between the systolic / diastolic blood pressure labels extracted by the oscillometric method and the PPG waveform features, resulting in a rise in the mean square error (MSE) of blood pressure estimation after model training exceeding 15%. Therefore, setting the preset time delay to no more than 50ms ensures the temporal consistency of the paired data, providing high-quality training labels for the model.
[0065] The above is a description of the present invention to help understand it; however, the implementation of the present invention is not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the principle of the present invention should be considered equivalent substitutions and are included within the protection scope of the present invention.
Claims
1. A blood pressure monitoring method based on PPG and cuff pressure synchronization, characterized in that, Includes the following steps: S1. Provide a wearable device that integrates a PPG sensor and an airbag assembly to obtain user personal information and build an initial neural network model; S2. Inflate and pressurize the airbag assembly to obtain the airbag internal pressure value P(t), and simultaneously trigger the PPG sensor to collect the PPG signal S(t). Record and align the airbag internal pressure value P(t) and the corresponding PPG signal S(t) at the same time according to the time to obtain the paired dataset {S(t), P(t)}. S3. Using the systolic blood pressure and / or diastolic blood pressure obtained by oscillometric processing of the airbag pressure value P(t) as supervision labels, and using the synchronously acquired PPG signal S(t) and user personal information as input features, a neural network model is trained using the paired dataset {S(t), P(t)}, and a user-specific blood pressure mapping model is constructed. S4. Deploy the trained neural network model on the wearable device. In the non-airbag component pressurized state, only the PPG signal S(t) collected by the PPG sensor is input into the neural network model, and the blood pressure estimate is output.
2. The blood pressure monitoring method according to claim 1, characterized in that, In step S2, the pressure value P(t) inside the airbag and the corresponding PPG signal S(t) at the same time are aligned using linear interpolation to form a paired dataset {S(t), P(t)} with completely corresponding timestamps. The paired dataset {S(t), P(t)} is uploaded to the cloud server. The cloud server performs online incremental updates to the neural network model based on the paired dataset {S(t), P(t)} and historical paired datasets, and transmits the updated neural network model weights to the wearable device.
3. The blood pressure monitoring method according to claim 1, characterized in that, In step S2, the preset time delay for the synchronization trigger does not exceed 50ms, and the acquisition clocks of the PPG sensor and the pressure sensor are provided by the same clock source.
4. The blood pressure monitoring method according to claim 1, characterized in that, In step S3, the neural network model includes an input layer, a convolutional feature extraction layer, a bidirectional long short-term memory network layer, a fully connected layer, and an output layer; mean squared error, mean absolute error, or Huber loss is used as the loss function, and the Adam optimizer is used to complete the parameter iterative convergence.
5. The blood pressure monitoring method according to claim 1, characterized in that, It also includes the following steps: S5, detecting the installation status of the airbag assembly; when the airbag assembly is not installed, automatically switching to the pure PPG sensor monitoring mode, and continuously outputting blood pressure estimates using the most recently updated neural network model.
6. A wearable device for blood pressure monitoring, used to perform the blood pressure monitoring method according to any one of claims 1-5, characterized in that, include: The meter head has a built-in main control chip, a power supply battery, an air pump, and a pressure sensor. The power supply battery is used to power the device. The watch strap is connected to the watch head, and the watch strap is provided with a slot for installing the airbag assembly; An airbag assembly is disposed inside the watch strap. The airbag assembly is detachably connected to the air pump and pressure sensor inside the watch head and is used to collect the pressure value P(t) inside the airbag under inflation and pressurization conditions. A PPG sensor is located at the bottom of the instrument head and / or on the skin-facing surface of the airbag assembly for acquiring PPG signal S(t); The main control chip integrates a synchronous acquisition unit and a model inference unit. The synchronous acquisition unit is used to align and interpolate the pressure value P(t) inside the airbag and the PPG signal S(t) according to time to obtain a paired dataset {S(t), P(t)}; the model inference unit stores a neural network model and outputs a blood pressure estimate based on the PPG signal S(t) acquired by the PPG sensor.
7. The wearable device according to claim 6, characterized in that, The airbag assembly and the watch strap are separate structures. The airbag assembly is connected to the air pump and pressure sensor inside the watch head through a detachable air passage interface. A silicone sealing ring is provided at the air passage interface. The airbag assembly is provided with a metal adsorption ring, and the watch head is provided with a ring permanent magnet adapted to the ferromagnetic metal ring.
8. The wearable device according to claim 6, characterized in that, The PPG sensor uses a dual-wavelength LED, which includes green light with a wavelength of 530nm and infrared light with a wavelength of 940nm, to emit light alternately and is used in conjunction with a photodiode receiver to collect the PPG signal S(t); the sampling rate of the PPG sensor can be configured to 125Hz, 250Hz or 500Hz, and the sampling rate of the pressure sensor is not less than 200Hz.
9. The wearable device according to claim 6, characterized in that, The main control chip also includes a detection unit, which is used to detect the installation status of the airbag assembly. When the airbag assembly is not installed, the main control chip is triggered to switch to pure PPG monitoring mode and disable the air pump and pressure sensor to save power. The most recently updated neural network model stored in the model inference unit is called to continuously output blood pressure estimates.
10. The wearable device according to claim 6, characterized in that, The device also includes a wireless communication module, which is electrically connected to the main control chip and is used to interact with the cloud server to perform pairing data and neural network model weight difference, thereby completing the remote iterative update of the user's exclusive blood pressure mapping model.
Citation Information
Patent Citations
Apparatus and method for measuring blood pressure
EP4344621A1
