Non-contact blood pressure measurement method and system based on deep learning and traditional Chinese medicine face diagnosis theory

Abstract

The invention belongs to the field of medical image processing, and provides a non-contact blood pressure measurement method based on deep learning and traditional Chinese medicine face diagnosis theories, which comprises the following steps: acquiring a face green channel vPPG signal through an RGB camera, accurately positioning a region of interest in combination with the traditional Chinese medicine face diagnosis theories, and performing efficient and accurate blood pressure prediction by using a deep learning model combining GNN and Transformer. Compared with a traditional blood pressure measuring method, the non-contact blood pressure measuring method has the remarkable advantages in the aspects of real-time performance, continuity and accuracy, meanwhile, the measuring operation process is simplified, the operation complexity and difficulty are reduced, tedious equipment connection and complex early-stage preparation work are not needed, and the method is suitable for popularization and application. Unprecedented convenient experience is brought to the field of blood pressure measurement, important breakthrough significance is achieved in the development of medical image processing and blood pressure measurement technologies, wide application prospects are shown, and the blood pressure measurement technology is promoted to advance to a new height.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, specifically a non-contact blood pressure measurement method and system based on deep learning models and traditional Chinese medicine facial diagnosis theory. Background Technology

[0002] Hypertension, a global health problem, is a major risk factor for cardiovascular disease, causing millions of deaths each year. Because hypertension often presents with no obvious symptoms in its early stages, it is known as the "silent killer." Early diagnosis, prevention, and treatment of hypertension are crucial for reducing the incidence and mortality of cardiovascular disease. However, existing blood pressure monitoring technologies have many limitations, making it difficult to meet these needs in practical applications. For example, measurements in clinical settings can be affected by the "white coat hypertension" phenomenon, where patients experience elevated blood pressure due to anxiety in front of a doctor, affecting the accuracy of the measurement results. Furthermore, while traditional cuff-type blood pressure monitors can provide relatively accurate blood pressure results, their measurement process is intermittent, unable to achieve real-time continuous monitoring, and frequent use may increase discomfort for the test subject, leading to inaccurate results.

[0003] In recent years, the continuous development of photoplethysmography (PPG) technology has provided new possibilities for non-contact blood pressure measurement. In 2008, Verkruysse et al. first proposed remote photoplethysmography (vPPG) signals to extract facial video signals under ambient light, making non-contact continuous blood pressure measurement a reality. This method uses a regular RGB camera to acquire facial video frame images, especially the green light channel signal, for processing and analysis to obtain accurate blood pressure estimates. The non-contact testing process of vPPG signals is simple and does not cause any discomfort to the tester, making it particularly suitable for special working environments such as mines and fire stations, as well as continuous monitoring in everyday environments such as homes and offices. However, existing non-contact testing technologies for vPPG signals suffer from poor signal acquisition and quality, and insufficient prediction accuracy, reliability, and continuity. Summary of the Invention

[0004] To address the problems mentioned in the background section, this invention provides a non-contact blood pressure measurement system based on deep learning and Traditional Chinese Medicine (TCM) facial diagnosis theory. TCM theory posits that changes in facial color and luster can reflect a person's health status, particularly changes in the cardiovascular system. Through deep learning technology, multimodal features related to blood pressure can be extracted from facial video signals and combined with modern medical methods for blood pressure estimation, achieving continuous, non-invasive, and highly accurate blood pressure monitoring around the clock. This method not only avoids the limitations of traditional measurement methods, such as the "white coat hypertension" phenomenon and the discomfort of frequent measurements, but also provides reliable monitoring results in various abnormal scenarios, contributing to the early diagnosis, prevention, and treatment of hypertension, and further reducing the incidence and mortality of cardiovascular diseases. Based on vPPG signals and through the acquisition of facial video, a non-contact, continuous, real-time, and accurate blood pressure measurement method is achieved. This system can be deployed on a terminal device platform, allowing testers to determine whether their blood pressure is normal based on the system's output.

[0005] The technical solution adopted in this invention is as follows: A non-contact blood pressure measurement method based on deep learning and traditional Chinese medicine facial diagnosis theory, the method comprising: S1. Deploy the blood pressure measurement system of the present invention on a terminal device, thereby logging in and running the non-contact blood pressure measurement system based on deep learning and traditional Chinese medicine facial diagnosis theory; S2. Use an RGB camera to collect video signals from the test subjects, and use the theory of traditional Chinese medicine facial diagnosis and the Dlib database to locate key points of the facial model, and initially determine 5 regions of interest for signal extraction. S3. Convert the video signal into a time-series signal. Convert the acquired video into frame images and divide the images into three channels: red, green, and blue. Process only the green channel image, calculate the average gray value of the image in this channel, and determine the three regions of interest as the optimal regions of interest for signal extraction, thereby obtaining an approximate vPPG signal. S4. Preprocess the signal. The approximate vPPG signal extracted from the green channel is preprocessed, specifically by: using a Kalman filter to remove noise and outliers from the approximate vPPG signal; and using a second-order Butterworth bandpass filter to remove baseline drift, ultimately obtaining the desired vPPG signal. S5. Feature Extraction and Feature Fusion. A graph structure (GNN) is constructed using GNNs to represent the spatiotemporal relationships in the vPPG signal. Local features are extracted by iteratively aggregating information from neighboring nodes, and graph convolution operations are used to enhance signal quality and improve the accuracy of subsequent feature extraction. Transformer processes sequence data through a self-attention mechanism to model the time-series features in the vPPG signal, capturing long-range dependencies to extract global features. Finally, the extracted features are fused. S6. Construct a blood pressure prediction regression model: Repeat steps 2-5 above, and use the fused features to construct a blood pressure prediction regression model. The input of the blood pressure prediction regression model is the fused features in S5, and the output is the blood pressure estimate. The blood pressure measurement system is constructed through this blood pressure prediction regression model.

[0006] Furthermore, the key to the non-contact blood pressure measurement method based on deep learning and traditional Chinese medicine facial diagnosis theory lies in the blood pressure prediction regression model constructed in S6. This model is trained on a pre-defined training set to obtain the regression model parameters. The model performance is then evaluated on an independent test set, using the evaluation metrics mean absolute error (MAE) and standard deviation (STD) to measure the accuracy of the prediction.

[0007] Furthermore, in step S5, feature extraction is performed on the vPPG signal extracted from the facial video based on the GNN graph neural network and Transformer, and the extracted features are fused as input to the blood pressure measurement model, specifically as follows: By creating a Graph Neural Network (GNN) model, a graph structure is constructed based on the separated green channel images. Nodes represent pixels or regions in the video, and edges represent the correlations or connections between these pixels. The feature vectors of nodes can contain variations in pixel intensity. The GNN model updates the node state by aggregating the features of each node and its neighborhood through graph convolutional layers. It enhances the signal quality through graph convolution operations and captures local patterns and relationships using a message-passing mechanism, thereby extracting local features of the vPPG signal. By creating a Transformer network model and utilizing its self-attention mechanism to process sequential data, this model is used to model the time-series features in vPPG signals. The separated green channel images are further segmented into several blocks, serving as both training and test sets. The training set is input into the Transformer model with added positional encoding, and the self-attention mechanism is used to capture long-range dependencies, resulting in a trained Transformer model. The test set is then input into the trained Transformer model for testing, thereby extracting global features. The features extracted by the GNN (Graph Neural Network) and Transformer are then fused, and the fused features are input into the Transformer model for classification and detection. The Transformer layer expands the receptive field of the GNN, and the GNN helps the Transformer capture graph structure information. The self-attention mechanism of the Transformer is used to capture more local and global features of the vPPG signal.

[0008] Furthermore, regarding the selection of the region of interest for extracting the vPPG signal, the specific details are as follows: Facial videos captured by an RGB camera are processed and converted into frame images. Using traditional Chinese medicine facial diagnosis theory and the Dlib database, key facial points are located to initially identify five regions of interest (ROIs). Within each frame, these five ROIs are separated into red, green, and blue (RGB) channels. Only the average grayscale value of the green channel ROI is calculated to determine the three optimal ROIs.

[0009] In the formula, M is the height of the image (number of rows), N is the width of the image (number of columns), and I(i,j) is the gray value of the pixel in the i-th row and j-th column of the image. The meaning of this formula is: sum the gray values ​​of each pixel in the image and then divide by the sum of all pixels in the image (i.e., the area of ​​the image, M×N).

[0010] Furthermore, in step S6, a blood pressure prediction regression model is constructed, and a blood pressure measurement system is built based on this regression model, specifically as follows: A regression model based on a graph neural network (GNN) and a transformer network is used as input, with its fused features serving as the model's input. The model outputs the measured blood pressure value. A blood pressure measurement system is built based on this regression model and deployed on a device terminal. The tester sits quietly in front of the terminal for 15 seconds, and the terminal interface displays the tester's blood pressure value, along with the normal blood pressure values ​​specified in the 2024 revised "Guidelines for the Prevention and Treatment of Hypertension in China," to indicate whether the tester's blood pressure is normal.

[0011] A non-contact blood pressure measurement system based on deep learning and traditional Chinese medicine facial diagnosis theory includes: The blood pressure system deployment platform is used to deploy blood pressure prediction models on terminal device platforms, allowing testers to log in and run a non-contact blood pressure measurement system based on deep learning and traditional Chinese medicine facial diagnosis theory. The signal extraction unit extracts signals from facial videos captured by the RGB camera after processing them accordingly, and transmits the extracted green channel signal to the signal preprocessing unit for preprocessing. The signal preprocessing unit is used to preprocess the extracted green channel signal. The extracted green channel signal is passed through a Kalman filter to remove noise and outliers. A second-order Butterworth bandpass filter is used to remove baseline drift, and finally the desired vPPG signal is obtained. The preprocessed vPPG signal is then transmitted to the feature extraction unit for feature extraction. The signal feature extraction unit is used to receive the preprocessed vPPG signal and extract features from the vPPG signal using the GNN graph neural network model and the Transformer network model. The GNN graph neural network captures its local features, and the Transformer network captures its global features.

[0012] The output unit fuses the features extracted from the signal feature extraction unit, then returns the fused features to the Transformer model for training and outputs the corresponding blood pressure value, which is then displayed on the terminal device interface.

[0013] Furthermore, the signal feature extraction unit includes: The model creation module creates GNN graph neural network models and Transformer network models based on GNN graph neural network and Transformer respectively, and combines them to extract systolic and diastolic blood pressure feature information from vPPG signals respectively; The encoder module encodes the preprocessed vPPG signal to obtain a feature vector set; The model training module is used to import the feature vector set into the GNN graph neural network model and the Transformer network model for training, and to import the test set into the trained Transformer network model for testing, and finally obtain the test results of the test set. The output module is used to output the systolic and diastolic blood pressure values ​​after passing through the GNN graph neural network model and the Transformer network model.

[0014] Compared with the methods mentioned in the background section, the present invention has the following advantages: Firstly, this invention utilizes an RGB camera to capture facial video and extracts vPPG signals through selected regions of interest (ROIs). In this step, the invention innovatively proposes a method combining traditional Chinese medicine facial diagnosis theory with observation of specific facial areas to infer the health of internal organs. Firstly, five ROIs are identified. Then, facial key points are located using the Dlib library, and combined with testing, three ROIs capable of extracting excellent signal quality are determined, thus improving the efficiency of obtaining stable and high-quality signals.

[0015] Secondly, in the process of extracting signals from images divided into red, green, and blue channels, this invention only needs to extract the signal from the green channel for preprocessing and analysis to obtain the desired vPPG signal. This is because, firstly, the wavelength of green light is closer to the absorption peak of hemoglobin, meaning the green channel can more sensitively reflect changes in light intensity caused by blood circulation; secondly, different colors of light are absorbed differently by the skin, and green light waves can better penetrate the skin surface to reach the more vascularized dermis, making the green channel more suitable for detecting changes in blood pressure.

[0016] Thirdly, this invention only requires the acquisition of a vPPG signal through facial video. Compared with many existing technologies that require the simultaneous acquisition of PPG and ECG signals, this method is simpler and faster, reducing the complexity of signal extraction and preprocessing. Attached Figure Description

[0017] Figure 1 This is a flowchart of a non-contact blood pressure measurement method based on deep learning and traditional Chinese medicine facial diagnosis theory according to the present invention.

[0018] Figure 2 This is a diagram illustrating the components of a non-contact blood pressure measurement system based on deep learning and traditional Chinese medicine facial diagnosis theory, according to the present invention.

[0019] Figure 3 shows a chart of different facial regions of interest based on the theory of face diagnosis in Traditional Chinese Medicine. Detailed Implementation

[0020] The technical solution will be clearly and completely described below with reference to the design flowchart attached to the specification. Obviously, the implementation examples described below are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, those skilled in the art, without making creative efforts to obtain all other embodiments, are all within the protection scope of the present invention.

[0021] To address the challenge of existing blood pressure measuring instruments being difficult to apply in various locations while maintaining the accuracy of real-time and continuous blood pressure measurement, this embodiment proposes the following technical solution: A non-contact blood pressure measurement method based on deep learning and traditional Chinese medicine facial diagnosis theory, the technical steps of which are as follows: S1. Deploy the blood pressure measurement system of the present invention on a terminal device, thereby logging in and running a non-contact blood pressure measurement system based on deep learning and traditional Chinese medicine facial diagnosis theory. The system also provides registration services to record the basic information of the tester, such as the tester's photo, name, gender, age, etc., so that the tester can monitor his / her own blood pressure changes over a long period of time, which is helpful for the early diagnosis, prevention and treatment of hypertension.

[0022] S2. Use an RGB camera to collect video signals from the test subject. Utilize traditional Chinese medicine facial diagnosis theory and the Dlib database to locate key points on the facial model, and preliminarily determine five regions of interest for signal extraction, as shown in Figure 3(b) of the instruction manual.

[0023] S3. Convert the video signal into a time series signal: Convert the acquired video into frame images and divide the images into three channels: red, green, and blue. Process only the green channel image, calculate the average gray value of the image in this channel, and determine the three regions of interest as the best regions of interest for signal extraction, as shown in Figure 3(c) of the manual, thereby obtaining an approximate vPPG signal. Specifically, the selection of the region of interest for extracting the vPPG signal is as follows: The facial video captured by the RGB camera is processed to convert it into frame images. Using the theory of traditional Chinese medicine facial diagnosis and the Dlib database, the key points of the face are located to initially determine 5 regions of interest. In each frame image, the 5 regions of interest are separated into red, green and blue (RGB) channels. Only the gray value of the green channel region of interest is calculated to determine 3 optimal regions of interest.

[0024] In the formula, M is the height of the image (number of rows), N is the width of the image (number of columns), and I(i,j) is the gray value of the pixel in the i-th row and j-th column of the image. The meaning of this formula is: sum the gray values ​​of each pixel in the image and then divide by the sum of all pixels in the image (i.e., the area of ​​the image, M×N).

[0025] S4. Preprocessing the signal: The approximate vPPG signal extracted from the green channel is preprocessed. Specifically, the following operations are performed: a Kalman filter is used to remove noise and outliers from the vPPG signal; a second-order Butterworth bandpass filter is used to remove baseline drift, and finally the desired vPPG signal is obtained; since some vPPG signals have the disappearance of diphtheria waves, the present invention can use the instantaneous heart rate method.

[0026] In the formula, HR represents the instantaneous heart rate, RRI represents the time between contraction peaks, and FPS is the frame rate during video capture.

[0027] S5. Feature Extraction and Feature Fusion: A graph structure is constructed using GNN to represent the spatiotemporal relationships in the vPPG signal. Local features are extracted by iteratively aggregating information from neighboring nodes. Furthermore, graph convolution operations are used to enhance signal quality and improve the accuracy of subsequent feature extraction. Transformer processes sequence data through a self-attention mechanism to model the time series features in the vPPG signal, captures long-distance dependencies, and extracts global features of the signal. Finally, the extracted features are fused. Specifically, feature extraction of vPPG signals from facial videos is performed based on GNN graph neural networks and Transformer, and the extracted features are fused as input to the blood pressure measurement model. By creating a Graph Neural Network (GNN) model, a graph structure is constructed based on the separated green channel images. Nodes represent pixels or regions in the video, and edges represent the correlations or connections between these pixels. The feature vectors of nodes can contain variations in pixel intensity. The GNN model updates the node state by aggregating the features of each node and its neighborhood through graph convolutional layers. It enhances the signal quality through graph convolution operations and captures local patterns and relationships using a message-passing mechanism, thereby extracting local features of the vPPG signal. By creating a Transformer network model and utilizing its self-attention mechanism to process sequential data, this model is used to model the time-series features in vPPG signals. The separated green channel images are further segmented into several blocks, serving as both training and test sets. The training set is input into the Transformer model with added positional encoding, and the self-attention mechanism is used to capture long-range dependencies, resulting in a trained Transformer model. The test set is then input into the trained Transformer model for testing, thereby extracting global features. The features extracted by the GNN (Graph Neural Network) and Transformer are then fused, and the fused features are input into the Transformer model for classification and detection. The Transformer layer expands the receptive field of the GNN, and the GNN helps the Transformer capture graph structure information. The self-attention mechanism of the Transformer is used to capture more local and global features of the vPPG signal.

[0028] S6. Construct a blood pressure prediction regression model: Repeat steps 2-5 above, and use the fused features to construct a blood pressure prediction regression model. The input of the blood pressure prediction regression model is the fused features in S5, and the output is the blood pressure estimate. The blood pressure measurement system is constructed through this blood pressure prediction regression model.

[0029] Furthermore, the key to the non-contact blood pressure measurement method based on deep learning and traditional Chinese medicine facial diagnosis theory lies in the blood pressure prediction regression model constructed in S6. This model is trained on a pre-defined training set to obtain the regression model parameters. The model performance is then evaluated on an independent test set, using the evaluation metrics mean absolute error (MAE) and standard deviation (STD) to measure the accuracy of the prediction.

[0030] Specifically, in step S6, a blood pressure prediction regression model is constructed, and a blood pressure measurement system is built based on this regression model. A regression model based on a graph neural network (GNN) and a transformer network is used as input, with its fused features serving as the model's input. The model outputs the measured blood pressure value. A blood pressure measurement system is built based on this regression model and deployed on a device terminal. The tester sits quietly in front of the terminal for 15 seconds, and the terminal interface displays the tester's blood pressure value, along with the normal blood pressure values ​​specified in the 2024 revised "Guidelines for the Prevention and Treatment of Hypertension in China," to indicate whether the tester's blood pressure is normal.

[0031] A non-contact blood pressure measurement system based on deep learning and traditional Chinese medicine facial diagnosis theory includes: The blood pressure system deployment platform is used to deploy blood pressure prediction models on terminal device platforms, allowing testers to log in and run a non-contact blood pressure measurement system based on deep learning and traditional Chinese medicine facial diagnosis theory. The signal extraction unit extracts signals from facial videos captured by the RGB camera after processing them accordingly, and transmits the extracted green channel signal to the signal preprocessing unit for preprocessing. The signal preprocessing unit is used to preprocess the extracted green channel signal. The extracted green channel signal is passed through a Kalman filter to remove noise and outliers. A second-order Butterworth bandpass filter is used to remove baseline drift, and finally the desired vPPG signal is obtained. The preprocessed vPPG signal is then transmitted to the feature extraction unit for feature extraction. The signal feature extraction unit is used to receive the preprocessed vPPG signal and extract features from the vPPG signal using the GNN graph neural network model and the Transformer network model. The GNN graph neural network captures its local features, and the Transformer network captures its global features.

[0032] The output unit fuses the features extracted from the signal feature extraction unit, then returns the fused features to the Transformer model for training and outputs the corresponding blood pressure value, which is then displayed on the terminal device interface. Furthermore, the signal feature extraction unit includes: The model creation module creates GNN graph neural network models and Transformer network models based on GNN graph neural network and Transformer respectively, and combines them to extract systolic and diastolic blood pressure feature information from vPPG signals respectively; The encoder module encodes the preprocessed vPPG signal to obtain a feature vector set; The model training module is used to import the feature vector set into the GNN graph neural network model and the Transformer network model for training, and to import the test set into the trained Transformer network model for testing, and finally obtain the test results of the test set. The output module is used to output the systolic and diastolic blood pressure values ​​after passing through the GNN graph neural network model and the Transformer network model.

[0033] The working principle of the above is as follows: The blood pressure system is deployed on a terminal device platform. The RGB camera equipped on the terminal device captures facial video of the tester, converts the facial video into frame images, separates them into 3-channel images, processes only the green channel to extract the vPPG signal, and then uses a GNN graph neural network and Transformer to extract features from the vPPG signal. The extracted features are fused to obtain fused features, and finally input into the Transformer network model for classification and recognition to obtain the corresponding systolic and diastolic blood pressure values. Combined with the normal blood pressure values ​​specified in the 2024 revised "Guidelines for the Prevention and Treatment of Hypertension in China", the system reminds the tester whether their blood pressure is normal.

[0034] It should be further clarified that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not imply a requirement for an actual relationship or order between these entities or operations. Furthermore, the use of terms such as "including," "containing," or similar expressions indicates an open-ended inclusion relationship. This means that referring to a series of elements, steps, items, or features does not exclude the existence of other elements, steps, items, or features not explicitly mentioned.

[0035] Secondly, without departing from the core principles and purpose of this invention, these embodiments can be adjusted, modified, substituted, and evolved in various ways. The scope of protection of this invention should be determined based on the appended patent claims and their equivalent technical solutions.

Claims

1. A non-contact blood pressure measurement method based on deep learning and traditional Chinese medicine facial diagnosis theory, characterized in that, Includes the following steps: S1. Deploy the blood pressure measurement system of the present invention on a terminal device, thereby logging in and running the non-contact blood pressure measurement system based on deep learning and traditional Chinese medicine facial diagnosis theory; S2. Use an RGB camera to collect video signals from the test subjects, and use the theory of traditional Chinese medicine facial diagnosis and the Dlib database to locate key points of the facial model, and initially determine 5 regions of interest for signal extraction. S3. Convert the video signal into a time series signal: Convert the acquired video into frame images and divide the images into three channels: red, green, and blue. Process only the green channel image, calculate the average gray value of the image in this channel, determine the three regions of interest as the best regions of interest for signal extraction, and thus obtain the approximate vPPG signal in the green channel. S4. Preprocess the signal: Preprocess the approximate vPPG signal extracted from the green channel. The specific operations include: using a Kalman filter to remove noise and outliers from the signal. A second-order Butterworth bandpass filter is used to remove baseline drift, and the vPPG signal is finally obtained. S5. Feature Extraction and Feature Fusion: A graph structure is constructed using GNN to represent the spatiotemporal relationships in the vPPG signal. Local features are extracted by iteratively aggregating information from neighboring nodes, and graph convolution operations are used to enhance signal quality and improve the accuracy of subsequent feature extraction. Transformer processes sequence data through a self-attention mechanism to model the time-series features in the vPPG signal, captures long-distance dependencies to extract global features of the signal, and finally fuses the extracted features. S6. Construct a blood pressure prediction regression model: Repeat steps 2-5 above, and use the fused features to construct a blood pressure prediction regression model. The input of the blood pressure prediction regression model is the fused features in S5, and the output is the blood pressure estimate. The blood pressure measurement system is constructed through this blood pressure prediction regression model.

2. The non-contact blood pressure measurement method based on deep learning and traditional Chinese medicine facial diagnosis theory according to claim 1, characterized in that: The blood pressure prediction regression model built in S6 is trained on a predefined training set to obtain the regression model parameters. The model performance is then evaluated on an independent test set, and the accuracy of the prediction is measured using the evaluation metrics mean absolute error (MAE) and standard deviation (STD).

3. The non-contact blood pressure measurement method based on deep learning and traditional Chinese medicine facial diagnosis theory according to claim 1, characterized in that: In S5, feature extraction is performed on the vPPG signal extracted from facial videos based on GNN graph neural networks and Transformer. The extracted features are then fused and used as input to the blood pressure measurement model. Specifically: By creating a GNN graph neural network model, a graph structure is constructed based on the separated green channel images. Nodes represent pixels or regions in the video, and edges represent the correlations or connections between these pixels. The feature vectors of nodes can contain changes in pixel intensity. The GNN graph neural network model updates the state of each node by aggregating the features of each node and its neighborhood through graph convolutional layers. The quality of the signal is enhanced through graph convolution operations, and local patterns and relationships are captured using message passing mechanisms, thereby extracting the local features of the vPPG signal. By creating a Transformer network model and utilizing its self-attention mechanism to process sequential data, this model is used to model the time-series features in vPPG signals. The separated green channel images are further segmented into several image blocks, which serve as the training set and the test set, respectively. The training set is input into the Transformer model and positional encoding is added. The self-attention mechanism is used to capture long-distance dependencies to obtain the trained Transformer model. The test set is then input into the trained Transformer model for testing, thereby extracting global features. The features extracted by the GNN graph neural network and Transformer are then fused, and the fused features are input into the Transformer model for classification and detection. The Transformer layer expands the receptive field of the GNN, and the GNN graph neural network helps Transformer capture graph structure information. The self-attention mechanism of Transformer is used to capture more local and global features of the vPPG signal.

4. The non-contact blood pressure measurement method based on deep learning and traditional Chinese medicine facial diagnosis theory according to claim 1, characterized in that: The determination of the three regions of interest as the optimal regions of interest for signal extraction, as described in S3, is as follows: Facial videos captured by an RGB camera are processed and converted into frame images. Using traditional Chinese medicine facial diagnosis theory and the Dlib database, key facial points are located to initially identify five regions of interest (ROIs). Within each frame, these five ROIs are separated into red, green, and blue (RGB) channels. Only the average grayscale value of the green channel ROI is calculated to determine the three optimal ROIs. In the formula, M is the height of the image (number of rows), N is the width of the image (number of columns), and I(i,j) is the gray value of the pixel in the i-th row and j-th column of the image. The meaning of this formula is: sum the gray values ​​of each pixel in the image and then divide by the sum of all pixels in the image (i.e., the area of ​​the image, M×N).

5. The non-contact blood pressure measurement method based on deep learning and traditional Chinese medicine facial diagnosis theory according to claim 1, characterized in that: In step S6, a blood pressure prediction regression model is constructed, and a blood pressure measurement system is built based on this regression model. Specifically: A regression model based on GNN (Graph Neural Network) and Transformer network is used as the input of the model, and the model output is the measured blood pressure value. A blood pressure measurement system is built on this regression model and deployed on the device terminal. The tester sits quietly in front of the terminal device for 15 seconds, and the terminal device interface displays the tester's blood pressure value, combined with the normal blood pressure value specified in the 2024 revised "Guidelines for the Prevention and Treatment of Hypertension in China", to remind the tester whether the blood pressure is normal.

6. A non-contact blood pressure measurement system based on deep learning and traditional Chinese medicine facial diagnosis theory, applied to the non-contact blood pressure measurement method based on deep learning and traditional Chinese medicine facial diagnosis theory as described in any one of claims 1-5. Its features include: The blood pressure system deployment platform is used to deploy blood pressure prediction models on terminal device platforms, allowing testers to log in and run a non-contact blood pressure measurement system based on deep learning and traditional Chinese medicine facial diagnosis theory. The signal extraction unit extracts signals from facial videos captured by the RGB camera after processing them accordingly, and transmits the extracted green channel signal to the signal preprocessing unit for preprocessing. The signal preprocessing unit is used to preprocess the extracted green channel signal. The extracted green channel signal is passed through a Kalman filter to remove noise and outliers. A second-order Butterworth bandpass filter is used to remove baseline drift, and finally the desired vPPG signal is obtained. The preprocessed vPPG signal is then transmitted to the feature extraction unit for feature extraction. The signal feature extraction unit receives the preprocessed vPPG signal and extracts features from it using a GNN (Graph Neural Network) model and a Transformer network model. The GNN captures local features, while the Transformer network captures global features. The output unit fuses the features extracted from the signal feature extraction unit, then returns the fused features to the Transformer model for training and outputs the corresponding blood pressure value, which is then displayed on the terminal device interface.

7. A non-contact blood pressure measurement system based on deep learning and traditional Chinese medicine facial diagnosis theory according to claim 6, characterized in that: The signal feature extraction unit includes: The model creation module creates GNN graph neural network models and Transformer network models based on GNN graph neural network and Transformer respectively, and combines them to extract systolic and diastolic blood pressure feature information from vPPG signals respectively; The encoder module encodes the preprocessed vPPG signal to obtain a feature vector set; The model training module is used to import the feature vector set into the GNN graph neural network model and the Transformer network model for training, and to import the test set into the trained Transformer network model for testing, and finally obtain the test results of the test set. The output module is used to output the systolic and diastolic blood pressure values ​​after passing through the GNN graph neural network model and the Transformer network model.

Images