Blood pressure measuring method based on smart phone
By optimizing signal acquisition and preprocessing, and combining deep learning technology with signal quality assessment, a CNN model was constructed to solve the problem of unstable signal quality in smartphone blood pressure measurement methods, achieving high-precision and robust blood pressure measurement.
Patent Information
- Application Number
- CN202511357950.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-23
- Publication Date
- 2025-10-28
AI Technical Summary
Existing smartphone-based blood pressure measurement methods are sensitive to external factors such as slight user movement, ambient light fluctuations, and changes in finger pressure, resulting in poor signal quality, affecting measurement accuracy and robustness. Furthermore, it is difficult to maintain consistent measurement performance across different devices, and there is a lack of effective signal quality assessment.
By optimizing signal acquisition and preprocessing, introducing deep learning technology, a blood pressure measurement model based on convolutional neural networks (CNN) is constructed. Combined with signal quality assessment, blood pressure-related features are extracted and integrated to achieve accurate blood pressure prediction.
It significantly improves the accuracy and robustness of blood pressure measurement, enabling reliable high-precision blood pressure monitoring on different devices, thus enhancing user experience and measurement reliability.
Smart Images
Figure CN120837041A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical and health care and smartphone application technology, and relates to a technical method for achieving high-precision and high-robust blood pressure measurement by using the camera, flash and accelerometer of a smartphone and by acquiring and processing photoplethysmography (PPG) signals and cardiac oscillation (SCG) signals. Background Technology
[0002] Blood pressure, as an important physiological health indicator, plays a crucial role in cardiovascular disease screening and daily health monitoring. Currently, the number of people with hypertension is increasing year by year. Regular blood pressure monitoring allows for the timely detection of abnormalities, enabling early intervention and treatment, which is beneficial for the prevention and treatment of hypertension. Traditional blood pressure measurement mainly relies on specialized medical equipment, such as mercury or electronic sphygmomanometers. Although these devices are highly accurate, they have limitations such as high cost, inconvenience, the need for cuffs, and difficulty in integrating with mobile devices, thus limiting their application in daily health monitoring.
[0003] In recent years, with the widespread adoption of smartphones and their enhanced computing power, cuffless blood pressure measurement using mobile phones has become possible. The basic principle is as follows: blood pressure affects the elasticity of blood vessel walls. Pulse wave transit time (PTT), the time from the onset of cardiac contraction to the arrival of the pulse wave at the distal detection point, reflects the elasticity of the blood vessel walls and is one of the key parameters for assessing blood pressure, showing a negative correlation with blood pressure. The smartphone's accelerometer, placed close to the chest, acquires the proximal SCG signal reflecting the heartbeat, while the flash and camera acquire the distal PPG signal through fingertip video. The SCG signal reflects the onset of cardiac contraction, and the time difference between it and the PPG signal can effectively estimate the PTT. Therefore, smartphone-based blood pressure measurement methods can be performed using built-in sensors and are easily processed via the cloud. They offer advantages such as technological accessibility, portability, and functional integration, making measurement convenient in daily life.
[0004] Existing methods are highly sensitive to external factors such as slight user movement, ambient light fluctuations, and changes in finger pressure. These factors can generate artifacts, leading to poor SCG and PPG signal quality and ultimately affecting the accuracy of blood pressure measurements. Furthermore, differences in hardware and algorithm settings across different smartphone brands and models make it difficult for existing methods to maintain consistent and reliable measurement performance across various devices. In addition, many existing applications lack effective signal quality assessment mechanisms. Even with poor signal quality, inaccurate blood pressure data may still be output, causing unnecessary anxiety for users. Therefore, developing a more advanced smartphone-based blood pressure measurement method, improving accuracy and robustness through signal processing and model optimization, is essential. This will facilitate the wider adoption of blood pressure monitoring and increase awareness, treatment, and control rates among hypertension patients. This is of great significance for promoting the popularization of blood pressure monitoring and improving public health. Summary of the Invention
[0005] This invention addresses the problems of high cost, inconvenience, and the need for cuffs in traditional methods, as well as the low accuracy, poor robustness, and lack of reliability assessment commonly found in existing mobile terminal-based blood pressure measurement methods. It proposes a smartphone-based blood pressure measurement method. This method systematically optimizes signal acquisition, preprocessing, and quality assessment, and introduces deep learning technology to construct a convolutional neural network (CNN)-based blood pressure measurement model. This significantly improves measurement accuracy and robustness, providing a reliable and user-friendly solution for mobile health monitoring, and is of great significance for improving the awareness and control rates of hypertension patients.
[0006] The specific technical solution of this invention is as follows:
[0007] A smartphone-based method for measuring blood pressure includes the following steps:
[0008] (1) Set the camera and three-axis accelerometer parameters of the smartphone, hold the phone in your hand and place it on your chest near your heart, and collect three-axis accelerometer data containing heart movement information; at the same time, lightly cover the rear camera with your fingers and collect video images of fingertip blood flow information with the flash.
[0009] (2) Calculate the pixel mean of the RGB channels of the video frame in the video image to obtain the color component of the volume pulse wave signal PPG, process the triaxial accelerometer data to obtain the cardiac vibration signal SCG, and perform signal preprocessing; then evaluate the quality of the preprocessed color component and cardiac vibration signal; select the one with the best color signal quality score as the volume pulse wave signal used for calculation, and when the quality score of the used color signal or the quality score of the cardiac vibration signal is lower than the preset threshold, blood pressure calculation is not performed, and operation suggestions are provided to the user;
[0010] (3) Extract features from PPG and SCG signals that meet the quality score, obtain the shape and time features related to blood pressure measurement in the two signals respectively, and perform signal screening; at the same time, align the two signals and calculate the pulse wave conduction time (PTT) feature by combining the specific features of the two signals; these features will be used as the input of the model.
[0011] (4) Build a blood pressure measurement model, including a feature extraction module, a feature integration module and a prediction output module; the feature extraction module uses two sets of three-layer one-dimensional convolutional layers plus a batch normalization layer to further extract more meaningful representations from the input features; the feature integration module consists of two fully connected layers that integrate the scattered local features extracted by the convolutional layers into a global, high-level feature representation; the final prediction output module consists of a linear layer that linearly combines the high-dimensional features output by the last fully connected layer into the final blood pressure prediction value;
[0012] (5) Train the blood pressure measurement model built in step (4) using the pre-collected dataset, adjust the hyperparameters to obtain the best performance, and save the optimal model that makes the loss value converge to the minimum value.
[0013] (6) Input the features obtained by the smartphone into the trained model to output the blood pressure measurement value.
[0014] Furthermore, for the calculation of PTT, the peak AO point of the SCG signal and the peak, peak-valley and peak value of the PPG signal derivative will be used to calculate PTT in various ways, and experiments will be conducted to screen and determine one or more effective PTT values.
[0015] Furthermore, the input model features include signal shape and time features such as the time interval, amplitude, and shape characteristics (area, etc.) of PPG and SCG signal feature points, as well as physiological characteristics such as the subject's age, gender, height, weight, and BMI. Heart rate is also used as an input feature of the model. Since different features have different properties, the effective information and noise contained in the calculation are also different. In the experiment, we will screen the features to accurately determine which features and how many feature combinations to select so that the model can achieve the best results in actual operation.
[0016] The present invention has the following beneficial effects:
[0017] (1) In this invention, the built-in camera and accelerometer of a smartphone are used to conveniently measure blood pressure in everyday environments without the need to wear any external devices.
[0018] (2) The present invention optimizes signal processing and feature extraction techniques to calculate more reliable pulse conduction time and other important features; and introduces a signal quality assessment step to effectively improve the robustness of blood pressure measurement.
[0019] (3) This invention uses deep learning to simulate the relationship between signal features and blood pressure changes, which solves the problem of insufficient feature extraction and realizes accurate blood pressure measurement. Attached Figure Description
[0020] Figure 1 This is an overall flowchart of the blood pressure measurement method based on a smartphone according to the present invention;
[0021] Figure 2 This is a network structure diagram of the model used for blood pressure measurement in this invention, corresponding to... Figure 1 Machine learning models in; Detailed Implementation
[0022] To enable those skilled in the art to better understand the technical solution of the present invention, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0023] A smartphone-based method for measuring blood pressure. (For example...) Figure 1 As shown, the specific implementation process is as follows:
[0024] Step 1: Signal Acquisition
[0025] Upon entering the measurement interface, detailed measurement instructions are provided. Specifically, the user holds the phone close to their heart, holding it against their chest, and gently places their fingers over the smartphone's rear camera, activating the flash. This invention presets and locks key hardware parameters upon entering measurement mode. Specifically, it sets the camera frame rate to the device's maximum adaptive value and locks automatic exposure and automatic white balance. The camera uses a single-column interval pixel sampling method, meaning only one pixel is processed from two adjacent pixels in a column. The RGB format is converted from the original YUV pixel format. The average pixel values for the red (R), green (G), and blue (B) channels are calculated to obtain the color components of the PPG signal. The accelerometer samples according to the smartphone's maximum performance capabilities, obtaining the axis data corresponding to heart motion as the SCG signal; this invention selects the z-axis.
[0026] Step Two: Signal Processing and Quality Assessment
[0027] For the preprocessing operation, the PPG signal is first interpolated and resampled to 150Hz according to the original frame timestamp. The resampled signal is then subjected to maximum and minimum normalization to adjust the signal amplitude to a specific range. Then, a third-order bandpass filter of 0.75-3.5 Hz is used to remove low-frequency baseline drift and high-frequency noise from the signal.
[0028] SCG signal processing first applies a 4th-order bandpass filter (1-45 Hz) to the x, y, and z axes. Then, the PMS envelope is calculated based on the sampling rate and window length. A threshold is determined based on the median of the envelope. The PMS envelope and the threshold are compared to detect and remove motion artifacts. The longest artifact-free segment is selected by comparing the three-axis accelerometer signals. The total acceleration of the three axes is calculated using the Euclidean distance formula. A Hilbert transform is applied to the total acceleration signal to obtain the amplitude envelope. The amplitude envelope is then filtered with a high cutoff frequency of 3 and a low cutoff frequency of 0.5, thus smoothing the signal.
[0029] A template matching method is used for quality assessment. This method includes: segmenting the PPG and SCG signals into independent beat-per-wave patterns, normalizing the amplitude and time of each beat-per-wave pattern, calculating the correlation coefficient between each normalized waveform segment and a preset median template, and taking the median of the correlation coefficients. Quality assessment indicators , and These are the minimum and maximum thresholds for the preset correlation coefficient, respectively. When the quality score of the color signal or the quality score of the cardiac signal used is lower than the preset threshold (e.g., 60 points), blood pressure calculation is not performed, and operation suggestions are provided to the user.
[0030] Step 3: Feature Extraction
[0031] For the preprocessed SCG signal, an adaptive threshold strategy is used to periodically traverse the signal, continuously updating the signal and noise levels. Based on this, peak values that meet the conditions are detected and recorded, along with their positions and corresponding threshold information. Using the acquired local peak points, the nearest filtered z-axis SCG signal is found. Based on the sampling rate, the nearest maximum value point for each local maximum value within a set distance is calculated, ultimately yielding the accurate location of the peak points.
[0032] For the PPG signal, the initial peak values are obtained using the `find_peaks` function in Python. The extracted peak values are then sorted by size, and a height threshold is selected. Among all the peak values, those peaks that are greater than the set height threshold are identified as the final peak points. Simultaneously, the peaks and troughs of the PPG derivative are calculated to determine key points such as the maximum slope point and the starting point. Furthermore, the amplitude, time interval, ratio, and area under the line between the key points are calculated.
[0033] Four PTT values are calculated based on the SCG peak value, PPG peak and trough, and signal derivative peak. All PTT values calculated within the measurement time are combined, sorted, and the extreme data of the first and last 25% are removed. The average value is taken as the final characteristic value.
[0034] Step 4: Model Building
[0035] The constructed blood pressure measurement model includes feature extraction, feature integration, and prediction output modules, with the following structure: Figure 2 As shown, the feature extraction module uses two sets of three-layer one-dimensional convolutional layers plus a batch normalization layer. It takes a sequence of shape (26, 1) as input. The convolutional layers progressively extract features and reduce the sequence length to 2, with each layer reducing by 4 while maintaining a feature dimension of 64. The stride is 1, and the kernel size is 5. The batch normalization layer normalizes each channel of the 64 feature maps without changing the data shape, which helps stabilize and accelerate training. The feature integration module consists of two fully connected layers of 128 neurons each, integrating the scattered local features extracted by the convolutional layers into a global, high-level feature representation. The final prediction output module is a linear layer that linearly combines the high-dimensional features output from the last fully connected layer into the final blood pressure prediction value.
[0036] Step 5: Model Training and Testing
[0037] Data collection was conducted using the same-arm sequential measurement method. After subjects had rested sufficiently, data was first collected using a smartphone, followed by measurement of the same arm using a standard cuff blood pressure monitor. The data was recorded, and after a one-minute wait, the measurement was repeated three times. If poor signal quality was observed, the measurement was repeated. After collecting sufficient data, 60% of the subjects were used as the training set, 40% as the test set, and 20% of the samples were randomly selected from the training set as the validation set.
[0038] Hyperparameters were set and the model was trained. During training, Huber was used as the loss function, combining the advantages of mean squared error (MSE) and mean absolute error (MAE). The optimizer used was the Adam optimizer, which can adaptively update the learning rate. The epoch was set to 1000, and the batch size was set to 32. For systolic blood pressure measurement, a CNN model with 26 input features achieved an MAE of 6.7917 mmHg after fine-tuning the last two layers (fully connected and linear layers) using the first data from the subjects. The optimal general model parameters were saved for the final smartphone blood pressure measurement.
Claims
1. A method for measuring blood pressure based on a smartphone, characterized in that, Includes the following steps: (1) Set the camera and three-axis accelerometer parameters of the smartphone, hold the phone in your hand and place it on your chest near your heart, and collect three-axis accelerometer data containing heart movement information; at the same time, lightly cover the rear camera with your fingers and collect video images of fingertip blood flow information with the flash. (2) Calculate the pixel mean of the RGB channels of the video frame in the video image to obtain the color component of the volume pulse wave signal PPG, process the triaxial accelerometer data to obtain the cardiac vibration signal SCG, and perform signal preprocessing; then evaluate the quality of the preprocessed color component and cardiac vibration signal; select the one with the best color signal quality score as the volume pulse wave signal used for calculation, and when the quality score of the used color signal or the quality score of the cardiac vibration signal is lower than the preset threshold, blood pressure calculation is not performed, and operation suggestions are provided to the user; (3) Extract features from PPG and SCG signals that meet the quality score, obtain the shape and time features related to blood pressure measurement in the two signals respectively, and perform signal screening; at the same time, align the two signals and calculate the pulse wave conduction time (PTT) feature by combining the specific features of the two signals; (4) Build a blood pressure measurement model, including a feature extraction module, a feature integration module and a prediction output module; the feature extraction module uses two sets of three-layer one-dimensional convolutional layers plus a batch normalization layer to extract more meaningful representations from the input features; the feature integration module consists of two fully connected layers that integrate the scattered local features extracted by the convolutional layers into a global, high-level feature representation; the final prediction output module consists of a linear layer that linearly combines the high-dimensional features output by the last fully connected layer into the final blood pressure prediction value; (5) Train the blood pressure measurement model built in step (4) using the pre-collected dataset, adjust the hyperparameters to obtain the best performance, and save the optimal model that makes the loss value converge to the minimum value. (6) Input the features obtained by the smartphone into the trained model to output the blood pressure measurement value.
2. The blood pressure measurement method based on a smartphone according to claim 1, characterized in that, In step (1), the camera parameters of the smartphone are adaptive maximum sampling frame rate, and locked automatic exposure and automatic white balance.
3. The blood pressure measurement method based on a smartphone according to claim 1, characterized in that, In step (1), the camera selects a single-column interval pixel sampling method, that is, only one pixel data is processed in two adjacent pixel data in a column, and the RGB format is converted from the original pixel format YUV; for the preprocessing operation in step (2), the PPG signal is first resampled to a fixed sampling rate by interpolation according to the original frame timestamp, and then processed by a third-order bandpass filter; the SCG signal is processed by a fourth-order bandpass filter.
4. The smartphone-based blood pressure measurement method according to claim 1, characterized in that, In step (2), the quality assessment method is a template matching method, which includes: segmenting the PPG and SCG signals into independent beat waveforms, normalizing the amplitude and time of the beat waveforms, and calculating the correlation coefficient between each normalized waveform segment and a preset median template to obtain the quality assessment index; the feature extraction method in step (3) includes peak detection, key point calculation based on peak values, shape and time feature calculation based on key points, or PTT calculation combining the two waveforms.
5. The smartphone-based blood pressure measurement method according to claim 5, characterized in that, The method for calculating PTT using signal feature processing techniques is as follows: (1) Apply Hilbert transform to the SCG signal to obtain the amplitude envelope and then filter it. Use an adaptive threshold strategy to find all local maximum points in the signal. Based on this point, find the nearest filtered Z-axis SCG signal and calculate the nearest maximum point of each local maximum within a certain distance as the peak value. (2) Perform peak detection on the PPG signal to obtain the initial peak value, sort the extracted peak values to set a threshold, and find those peak values that are greater than the set threshold value as the final peak value among all peak values; at the same time, use the same method to calculate the peak valley and the peak value of the derivative of the PPG signal. (3) Calculate the time difference based on the peak values of the two signals to obtain all PTT values within the measurement duration; (4) Combine the calculated PTTs into an array, exclude extreme values, and take the average value as the final feature value.