Intelligent non-invasive blood pressure real-time monitoring device and method based on multimodality

Abstract

The present invention provides a multimodal intelligent non-invasive real-time blood pressure monitoring device and method, relating to the field of intelligent sensors. The method comprises: predicting based on motion signals to obtain a first error influence parameter and a contact influence parameter; performing error analysis based on skin color information and the contact influence parameter to obtain a second error influence parameter; predicting blood pressure based on motion signals to obtain a first predicted blood pressure interval; predicting the user's blood pressure based on PPG signals to obtain a predicted blood pressure; performing error compensation based on the first and second error influence parameters to obtain a second predicted blood pressure interval; and calculating a blood pressure monitoring error coefficient based on the first and second predicted blood pressure intervals, and labeling the predicted blood pressure. This method aims to address the problem that traditional blood pressure monitoring methods cannot effectively identify external interference factors, resulting in insufficient comprehensiveness, accuracy, and reliability of blood pressure monitoring results. It can provide users with more accurate, comprehensive, and reliable blood pressure monitoring data.

Description

Technical Field

[0001] The present invention relates to the field of intelligent sensors, and in particular to a multimodal intelligent non-invasive blood pressure real-time monitoring device and method. Background Art

[0002] With the rapid development of smart wearable devices, especially the popularity of smart watches, new possibilities have been proposed for blood pressure monitoring based on non-contact and non-invasive sensing technology.

[0003] As a non-invasive physiological signal measurement technology, PPG sensors have been used in blood pressure prediction. PPG sensors obtain blood pressure-related information by detecting changes in blood flow in blood vessels on the skin surface. However, their signals are greatly affected by external factors (such as exercise, sweating, skin color, etc.). Traditional blood pressure monitoring methods fail to effectively identify these influencing factors, resulting in deviations and instability in data results, affecting the accuracy of blood pressure monitoring. Summary of the Invention

[0004] The purpose of the present invention is to provide a multimodal intelligent non-invasive real-time blood pressure monitoring device and method to address the technical problem that traditional blood pressure monitoring methods cannot effectively identify external interference factors, resulting in insufficient comprehensiveness, accuracy, and reliability of blood pressure monitoring results. The present invention includes:

[0005] In a first aspect, the present invention provides an intelligent non-invasive blood pressure real-time monitoring device based on multimodality, the device including a PPG sensor and a motion sensor, and further including: a data acquisition module for collecting PPG signals and motion signals of a target user through the PPG sensor and the motion sensor to obtain PPG signals and motion signals, and collecting skin color information of the target user; an error influence analysis module for performing sensing error influence analysis and contact influence parameter prediction based on the motion signal to obtain first sensing error influence parameters and contact influence parameters, performing sensing error influence analysis based on the skin color information and contact influence parameters to obtain second sensing error influence parameters; a user blood pressure prediction module for predicting the user's blood pressure based on the motion signal to obtain a first predicted blood pressure interval, predicting the user's blood pressure based on the PPG signal to obtain a predicted blood pressure, performing error compensation based on the first sensing error influence parameter and the second sensing error influence parameter to obtain a second predicted blood pressure interval; and an error coefficient labeling module for calculating a blood pressure monitoring error coefficient based on the first predicted blood pressure interval and the second predicted blood pressure interval, and labeling the predicted blood pressure as a blood pressure monitoring result.

[0006] Preferably, the multimodal intelligent non-invasive blood pressure real-time monitoring device further includes: collecting PPG signals and motion signals of the target user within a recent preset time window through a PPG sensor and a motion sensor to obtain PPG signals and motion signals, wherein the motion sensor is a gyroscope; and obtaining skin color information of the target user collected and recorded within a historical time period.

[0007] Preferably, the intelligent non-invasive blood pressure real-time monitoring device based on multimodality also includes: collecting a set of sample motion signals based on monitoring data in historical time, and collecting the error amplitudes of different sample motion signals and blood pressure predicted based on PPG signals in a non-exercise state, marking them as sample first sensing error influence parameters, and obtaining a set of sample first sensing error influence parameters; collecting the user's skin sweating parameters under different sample motion signals, marking them as sample contact influence parameters, and obtaining a set of sample contact influence parameters; using machine learning to construct a motion sensing error influence prediction branch and a contact influence prediction branch; using the sample motion signal set as input features, and using the sample first sensing error influence parameter set and the sample contact influence parameter set as output features respectively, and performing supervised training and testing on the motion sensing error influence prediction branch and the contact influence prediction branch until convergence; inputting the motion signal into the motion sensing error influence prediction branch and the contact influence prediction branch respectively, and predicting the output to obtain the first sensing error influence parameter and the contact influence parameter.

[0008] Preferably, the intelligent non-invasive blood pressure real-time monitoring device based on multimodality also includes: collecting a sample skin color information set and a sample contact influence parameter set based on monitoring data within a historical period; collecting different sample skin color information and sample contact influence parameters compared with the preset skin color and preset contact influence parameters, predicting the error amplitude of blood pressure based on the PPG signal, marking it as the sample second sensing error influence parameter, and obtaining the sample second sensing error influence parameter set; constructing a skin color sensing error influence prediction branch based on machine learning; using the sample skin color information set, sample contact influence parameter set and sample second sensing error influence parameter set, the skin color sensing error influence prediction branch is supervised trained and tested until convergence; the skin color information and contact influence parameters are input into the skin color sensing error influence prediction branch, and the prediction output obtains the second sensing error influence parameter.

[0009] Preferably, the intelligent non-invasive blood pressure real-time monitoring device based on multimodality also includes: collecting a set of sample motion signals based on the blood pressure monitoring data of the target user in the historical time, and collecting the blood pressure intervals of the target user under different sample motion signals, and marking to obtain a set of sample first predicted blood pressure intervals; constructing an exercise blood pressure prediction branch based on machine learning; using the sample motion signal set and the sample first predicted blood pressure interval set, the exercise blood pressure prediction branch is supervised trained and tested until convergence; the motion signal is input into the exercise blood pressure prediction branch, and the prediction output obtains the first predicted blood pressure interval.

[0010] Preferably, the multimodal intelligent non-invasive blood pressure real-time monitoring device further includes: inputting the PPG signal into a pre-trained photoelectric blood pressure prediction branch, and predicting and outputting the predicted blood pressure, wherein the photoelectric blood pressure prediction branch is constructed based on machine learning, and uses a set of sample PPG signals and a set of sample blood pressure within the target user's historical time to perform supervised training until convergence; calculating and obtaining a sensing error influence parameter based on the first sensing error influence parameter and the second sensing error influence parameter; and using the sensing error influence parameter to perform error compensation on the predicted blood pressure to obtain a second predicted blood pressure interval.

[0011] Preferably, the multimodal intelligent non-invasive blood pressure real-time monitoring device also includes: obtaining the intersection of the first predicted blood pressure interval and the second predicted blood pressure interval to obtain the intersection blood pressure interval; respectively calculating the ratio of the intersection blood pressure interval to the first predicted blood pressure interval and the second predicted blood pressure interval, and calculating the first blood pressure monitoring error coefficient and the second blood pressure monitoring error coefficient; calculating the blood pressure monitoring error coefficient based on the first blood pressure monitoring error coefficient and the second blood pressure monitoring error coefficient; and using the blood pressure monitoring error coefficient to mark the predicted blood pressure to obtain a blood pressure monitoring result.

[0012] In a second aspect, the present invention also provides an intelligent non-invasive blood pressure real-time monitoring method based on multimodality, including: collecting PPG signals and motion signals of a target user through a PPG sensor and a motion sensor to obtain PPG signals and motion signals, and collecting skin color information of the target user; performing sensing error influence analysis and contact influence parameter prediction based on the motion signal to obtain first sensing error influence parameters and contact influence parameters, performing sensing error influence analysis based on the skin color information and contact influence parameters to obtain second sensing error influence parameters; predicting the user's blood pressure based on the motion signal to obtain a first predicted blood pressure interval, predicting the user's blood pressure based on the PPG signal to obtain a predicted blood pressure, performing error compensation based on the first sensing error influence parameter and the second sensing error influence parameter to obtain a second predicted blood pressure interval; calculating the blood pressure monitoring error coefficient based on the first predicted blood pressure interval and the second predicted blood pressure interval, marking the predicted blood pressure as the blood pressure monitoring result.

[0013] The embodiments of the present invention include the following advantages:

[0014] By collecting PPG signals and motion signals of the target user, the PPG signals and motion signals are obtained, and the skin color information of the target user is collected; then, a sensing error influence analysis and a contact influence parameter prediction are performed based on the motion signal to obtain a first sensing error influence parameter and a contact influence parameter, and a sensing error influence analysis is performed based on the skin color information and the contact influence parameter to obtain a second sensing error influence parameter; then, based on the motion signal, the user's blood pressure is predicted to obtain a first predicted blood pressure range; on the other hand, based on the PPG signal, the user's blood pressure is predicted to obtain a predicted blood pressure, and error compensation is performed based on the first sensing error influence parameter and the second sensing error influence parameter to obtain a second predicted blood pressure range; finally, based on the first predicted blood pressure range and the second predicted blood pressure range, the blood pressure monitoring error coefficient is calculated and the predicted blood pressure is marked as the blood pressure monitoring result. In other words, by combining multimodal sensor data for sensor error analysis, it is possible to accurately identify and quantify sensor errors caused by different factors, thereby improving the accuracy and reliability of sensor error acquisition; then, the predicted blood pressure is labeled according to the error analysis results, which can provide users with feedback information on measurement reliability, helping users to effectively identify the reliability of predicted blood pressure, thereby providing users with more accurate, comprehensive and reliable blood pressure monitoring data. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a schematic structural diagram of a multimodal intelligent non-invasive blood pressure real-time monitoring device according to the present invention;

[0016] Figure 2This is a flowchart of the steps of a multimodal intelligent non-invasive blood pressure real-time monitoring method of the present invention.

[0017] Description of reference numerals:

[0018] Data collection module 11, error impact analysis module 12, user blood pressure prediction module 13, error coefficient marking module 14. DETAILED DESCRIPTION

[0019] The present invention provides a multimodal, intelligent, non-invasive, real-time blood pressure monitoring device and method, addressing the technical problem that traditional blood pressure monitoring methods are unable to effectively identify external interference factors, resulting in insufficient comprehensiveness, accuracy, and reliability in blood pressure monitoring results. By combining multimodal sensor data for sensor error analysis, it is possible to accurately identify and quantify sensor errors caused by different factors, improving the accuracy and reliability of sensor error acquisition. Predicted blood pressure is then annotated based on the error analysis results, providing users with feedback on measurement reliability, helping them effectively identify the reliability of predicted blood pressure, and thus providing users with more accurate, comprehensive, and reliable blood pressure monitoring data.

[0020] Below, the technical solutions of the present invention will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments described herein. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should also be noted that, for the convenience of description, only the parts related to the present invention, rather than all, are shown in the accompanying drawings.

[0021] For example, see the attached Figure 1 The present invention provides a multimodal intelligent non-invasive blood pressure real-time monitoring device, the device comprising a PPG sensor and a motion sensor, and further comprising:

[0022] The data acquisition module 11 is used to collect PPG signals and motion signals of a target user through a PPG sensor and a motion sensor, obtain PPG signals and motion signals, and collect skin color information of the target user.

[0023] Furthermore, the data acquisition module 11 is also used for:

[0024] The PPG signal and motion signal of the target user are collected within a recent preset time window through a PPG sensor and a motion sensor, where the motion sensor is a gyroscope; and the skin color information of the target user collected and recorded in a historical time is obtained.

[0025] Specifically, the present invention relates to a multimodal, intelligent, non-invasive, real-time blood pressure monitoring device. Specifically, it is a blood pressure monitoring device that combines a PPG sensor and a motion sensor. This device can be installed on wearable devices such as smartwatches, enabling real-time monitoring of a user's blood pressure. By analyzing multimodal data, the accuracy, comprehensiveness, and reliability of blood pressure monitoring are improved. The PPG (photoplethysmography) sensor detects changes in blood flow in blood vessels on the surface of a user's skin. By emitting a light source to illuminate the skin surface and receiving the reflected light signal, the PPG sensor analyzes changes in blood flow and thereby obtains pulse wave information related to blood pressure. This is a core technology for non-invasive blood pressure monitoring. The motion sensor is a gyroscope, used to monitor the user's motion status in real time. Changes in motion status, such as walking, running, resting, or other activities, can affect blood pressure measurements. By collecting motion data, the impact of motion on blood pressure prediction can be analyzed, thereby improving the accuracy of blood pressure predictions. The PPG sensor and motion sensor are integrated into wearable devices such as smart watches, which ensures that the device can be worn on the user at any time for non-invasive blood pressure monitoring, avoiding the many inconveniences and discomforts of traditional measurement methods based on air bag blood pressure monitors.

[0026] First, using PPG and motion sensors, the target user's PPG and motion signals are collected within a recently preset time window (which can be set based on the actual monitoring scenario, such as the last minute). Specifically, the PPG sensor (photoplethysmography) collects photoelectric signals of blood flow on the target user's skin surface. This sensor illuminates the skin with a light source and receives reflected light signals, analyzing blood flow based on changes in the reflected light. Motion sensors (such as gyroscopes) collect the target user's movement status. Gyroscopes measure the user's angular velocity and directional changes, providing data related to their activity. This data can help analyze the impact of exercise on blood pressure prediction, especially in dynamic conditions. PPG and motion signals are obtained through data collection. The PPG signal optically monitors changes in blood flow in blood vessels, while the motion signal includes the user's exercise intensity and exercise pattern (such as walking, running, etc.). Exercise affects heart rate, which in turn affects the accuracy of blood pressure prediction.

[0027] On the other hand, the skin color information of the target user collected and recorded in the historical time (such as the last three months), that is, the most recent skin color information of the target user, such as light skin color, medium skin color and dark skin color, can be obtained. The skin image can be captured by the camera, and the RGB value of the skin area can be extracted for skin color identification. Among them, the skin color will affect the absorption rate of the PPG signal. Dark skin color may lead to stronger absorption of light, thereby affecting the transmission of the photoelectric signal and the accuracy of blood pressure prediction.

[0028] The error influence analysis module 12 is used to perform sensing error influence analysis and contact influence parameter prediction based on the motion signal to obtain first sensing error influence parameters and contact influence parameters, and to perform sensing error influence analysis based on the skin color information and contact influence parameters to obtain second sensing error influence parameters.

[0029] Furthermore, the error impact analysis module 12 is further configured to:

[0030] According to the monitoring data in the historical time, a set of sample motion signals is collected, and the error amplitudes of different sample motion signals and blood pressure predicted based on PPG signals in the non-exercise state are collected, marked as sample first sensing error influence parameters, and a set of sample first sensing error influence parameters is obtained; the user's skin sweating parameters under different sample motion signals are collected, marked as sample contact influence parameters, and a set of sample contact influence parameters is obtained; machine learning is used to construct a motion sensing error influence prediction branch and a contact influence prediction branch; the sample motion signal set is used as input features, and the sample first sensing error influence parameter set and the sample contact influence parameter set are used as output features respectively, and the motion sensing error influence prediction branch and the contact influence prediction branch are supervised for training and testing until convergence; the motion signal is input into the motion sensing error influence prediction branch and the contact influence prediction branch respectively, and the prediction output obtains the first sensing error influence parameter and the contact influence parameter.

[0031] Specifically, based on the monitoring data in the historical period (such as the past six months), multiple sample motion signals are collected to construct a sample motion signal set. These data include different types of exercise types (such as walking, running, etc.) and exercise intensities; then, the error amplitude of the blood pressure predicted based on the different sample motion signals and the PPG signal in the non-exercise state is collected, such as obtaining motion data through motion sensors (such as gyroscopes, accelerometers, etc.), distinguishing the exercise state and the non-exercise state according to the exercise intensity or state; then, the predicted blood pressure based on the PPG signal in the exercise state and the non-exercise state is obtained respectively, and the blood pressure in the non-exercise state is used as the error amplitude of the PPG signal. The predicted blood pressure based on the PPG signal under the condition of α is used as a benchmark to calculate the error margin, where the error margin is the ratio of the absolute value of the blood pressure difference between the predicted blood pressure in the exercise state and the predicted blood pressure in the non-exercise state to the predicted blood pressure in the non-exercise state. For example, assuming that the predicted blood pressure in the exercise state is 105 and the predicted blood pressure in the non-exercise state is 100, the error margin is (105-100) / 100, which is equal to 5%. The error margin is marked as the sample first sensing error influence parameter to obtain the sample first sensing error influence parameter set, where the sample motion signal and the sample first sensing error influence parameter have a one-to-one correspondence.

[0032] On the other hand, the user's skin sweat parameters under different sample motion signals are collected, and a sweat sensor is used to directly measure the flow or humidity of sweat, such as the sweat flow per unit time (such as the amount of sweat per minute, in milliliters / minute). For example, during walking, a user's sweat volume is 0.2 milliliters / minute, and during running it is 0.5 milliliters / minute. These are marked as sample contact influencing parameters, and a set of sample contact influencing parameters is obtained. Among them, sweating will change the contact quality between the skin surface and the PPG sensor, especially when sweat accumulates, which will cause changes in light scattering and absorption, thereby affecting the quality and accuracy of the PPG signal.

[0033] Next, machine learning is used to construct a motion sensing error impact prediction branch and a contact impact prediction branch. The motion sensing error impact prediction branch is used to predict the sensing error caused by motion. By training the relationship between the motion signal and the blood pressure error, this branch can predict the error amplitude of the blood pressure prediction based on the user's motion state. The contact impact prediction branch is used to predict the skin sweating parameter based on the user's motion state. For example, the motion sensing error impact prediction branch and the contact impact prediction branch are constructed based on a BP neural network, wherein the motion sensing error impact prediction branch and the contact impact prediction branch are BP neural network models that can be iteratively optimized in machine learning, including an input layer, multiple hidden layers, and an output layer. The input data of the input layer of the motion sensing error impact prediction branch is a motion signal, and the output data is a first sensing error impact parameter. The input data of the contact impact prediction branch is a motion signal, and the output data is a contact impact parameter (skin sweating parameter).

[0034] The sample motion signal set, the sample first sensing error influence parameter set and the sample contact influence parameter set are further used as training data, and divided into a training set and a test set according to a predetermined ratio, usually the training set accounts for 85% and the test set accounts for 15%; then, the sample motion signal is used as the input feature, the sample first sensing error influence parameter is used as the supervisory feature of the motion sensing error influence prediction branch, and the sample contact influence parameter is used as the supervisory feature of the contact influence prediction branch. The motion sensing error influence prediction branch and the contact influence prediction branch are supervised trained and tested using the training set and the test set. During the training process, each batch of training data will be forward propagated through the neural network to generate prediction results; then, by calculating the loss function (such as the mean square error), the error will be backpropagated to each layer of the neural network to update the weight parameters in the neural network; the model parameters will be continuously optimized through multiple iterations until the training error converges; during the testing process, the generalization ability of the model is verified by evaluation on the test set. When the training error tends to stabilize and the error on the test set also reaches the predetermined tolerance range, the training process is considered to have converged, and the training of the motion sensing error impact prediction branch and the contact impact prediction branch is completed.

[0035] Finally, the motion signal is respectively input into the motion sensing error influence prediction branch and the contact influence prediction branch for prediction, and the first sensing error influence parameter and the contact influence parameter are output; by adopting machine learning to construct the motion sensing error influence prediction branch and the contact influence prediction branch for prediction, the intelligence level of the prediction can be improved, thereby improving the accuracy and efficiency of the prediction results.

[0036] Furthermore, the error impact analysis module 12 is further configured to:

[0037] According to the monitoring data in the historical time, a sample skin color information set and a sample contact influence parameter set are collected; the error amplitude of blood pressure predicted based on the PPG signal under different sample skin color information and sample contact influence parameters compared with the preset skin color and preset contact influence parameters is recorded as the sample second sensing error influence parameter to obtain the sample second sensing error influence parameter set; based on machine learning, a skin color sensing error influence prediction branch is constructed; the sample skin color information set, the sample contact influence parameter set and the sample second sensing error influence parameter set are used to perform supervised training and testing on the skin color sensing error influence prediction branch until convergence; the skin color information and contact influence parameters are input into the skin color sensing error influence prediction branch, and the prediction output is obtained to obtain the second sensing error influence parameter.

[0038] Specifically, based on monitoring data from a historical period (e.g., the last three months), a set of sample skin color information and a set of sample contact influence parameters are collected. Skin color information can be obtained through wearable devices or manual input, and each sample's skin color information can be represented by numerical features, such as skin color RGB values. Contact influence parameters (perspiration) are measured and recorded via sensors. Next, the predicted blood pressure based on the PPG signal is collected for different sample skin color information and sample contact influence parameters, as well as the predicted blood pressure based on the PPG signal for a preset skin color (a standard skin color is selected for comparison, such as medium skin color) and a preset contact influence parameter (standard perspiration). The error margin of the blood pressure monitoring values ​​under the sample state and the standard state is calculated, and the ratio of the absolute value of the blood pressure monitoring difference to the blood pressure monitoring value under the standard state is collected and labeled as the sample second sensing error influence parameter, thereby obtaining the sample second sensing error influence parameter set.

[0039] Based on machine learning, a skin color sensor error impact prediction branch is constructed. This branch is used to predict the impact of skin color and sweat volume on PPG sensor error using a machine learning model. For example, this branch is constructed based on a BP neural network, comprising an input layer, multiple hidden layers, and an output layer. The input layer's input data is skin color information and contact influence parameters, and the output layer's output data is the second sensor error influence parameter. The sample skin color information set, sample contact influence parameter set, and sample second sensor error influence parameter set are further used as sample data and divided into a sample training set and a sample test set according to a certain ratio.

[0040] The skin color sensor error influence prediction branch is further supervised and trained using sample skin color information and sample contact influence parameters as input and sample second sensor error influence parameters as supervision, using sample training sets and sample test sets until convergence. During the training process, the neural network first performs forward propagation, and the input data (skin color information and contact influence parameters) are passed into the neural network through the input layer. The input data passes through each layer (from the input layer to the hidden layer, and then to the output layer), and after weighting and activation function calculation, the predicted value is finally output. Then, the loss function is calculated using the result of the forward propagation and the actual value (the second sensor error influence parameter). The loss function measures the difference between the predicted value and the actual value. Then, the backpropagation algorithm is used to calculate the gradient of the loss function with respect to each weight. The gradient descent algorithm is used to optimize the weights and update the weights of the neural network to minimize the loss. An optimization algorithm (such as the Adam optimizer or SGD) is used to adjust the weights and biases in the neural network so that the model output is closer to the true value at each iteration. The process of forward propagation, loss calculation, backpropagation and weight update is repeated until the network converges. The convergence standard is usually that the loss function changes very little or the set number of iterations is reached. During the testing process, an unseen test set is used to evaluate the generalization ability of the model. The data in the test set (skin color information, contact influence parameters) is input into the trained neural network, and the predicted second sensor error influence parameters are output. The predicted results of the test set are compared with the actual results, and the performance indicators of the model are calculated. If the test results meet the expected results, the trained skin color sensor error influence prediction branch is obtained.

[0041] Finally, the skin color information and contact influence parameter are input into the skin color sensing error influence prediction branch for prediction, and a second sensing error influence parameter is output. The second sensing error influence parameter represents the error size in the PPG signal due to the skin color information and contact influence, and can accurately identify and quantify the errors caused by skin color and contact factors.

[0042] The user blood pressure prediction module 13 is used to predict the user's blood pressure based on the motion signal to obtain a first predicted blood pressure interval, predict the user's blood pressure based on the PPG signal to obtain a predicted blood pressure, and perform error compensation based on the first sensor error influence parameter and the second sensor error influence parameter to obtain a second predicted blood pressure interval.

[0043] Furthermore, the user blood pressure prediction module 13 is further configured to:

[0044] Based on the blood pressure monitoring data of the target user in the historical period, a set of sample motion signals is collected, and the blood pressure intervals of the target user under different sample motion signals are collected, and the sample first predicted blood pressure interval set is marked; based on machine learning, an exercise blood pressure prediction branch is constructed; the sample motion signal set and the sample first predicted blood pressure interval set are used to supervise the exercise blood pressure prediction branch and test it until convergence; the motion signal is input into the exercise blood pressure prediction branch, and the prediction output obtains the first predicted blood pressure interval.

[0045] Specifically, first, based on the blood pressure monitoring data of the target user in the historical period (such as the last three months), a set of sample motion signals (exercise type, exercise intensity, etc.) is collected; then, the blood pressure intervals (including low pressure intervals and high pressure intervals) of the target user under different sample motion signals are collected and marked as the sample first predicted blood pressure intervals, and a set of sample first predicted blood pressure intervals is obtained, in which each sample records the range of blood pressure changes of the target user under a specific exercise state (that is, the blood pressure interval, including the low pressure interval and the high pressure interval).

[0046] Next, an exercise blood pressure prediction branch is constructed based on machine learning. This branch utilizes a machine learning model (such as a regression model or a neural network) to predict the user's blood pressure range based on their exercise status. For example, this branch is constructed based on a generative adversarial network (GAN). A GAN is a deep learning model consisting of two neural networks (a generator and a discriminator) that uses adversarial training to generate realistic data. The generator's task is to generate predicted blood pressure ranges based on the input motion signal. The generator attempts to generate realistic blood pressure range data that looks like real data; the discriminator's task is to determine whether the input blood pressure range data is real or generated. The discriminator evaluates the generator's output and provides feedback based on its judgment. The generator receives the motion signal as input and generates predicted blood pressure ranges. The generator's goal is to generate data similar to the real blood pressure range based on the motion signal. The generator model can use a multilayer perceptron or a convolutional neural network. The discriminator receives the motion signal and the generated blood pressure range (generated by the generator) as input and outputs a binary classification result, determining whether the blood pressure range is real or generated by the generator. The discriminator's task is to distinguish between real and generated data as accurately as possible.

[0047] Then, the sample motion signal set and the sample first predicted blood pressure interval set are used as training data and divided into a training set and a test set according to a certain ratio; then, the sample motion signal is used as input and the sample first predicted blood pressure interval is used as supervision, and the training set and the test set are used to perform supervised training and testing on the exercise blood pressure prediction branch until convergence; during the training process, the generator receives the sample motion signal as input and generates a predicted blood pressure interval; then, the discriminator evaluates the difference between the generated blood pressure interval and the true blood pressure interval, and according to the feedback of the discriminator, the generator updates the parameters and optimizes the generated blood pressure interval; the discriminator receives a true blood pressure interval and the blood pressure interval output by the generator as input, and the discriminator outputs a probability, indicating the possibility that the blood pressure interval is from the real data; then the discriminator updates the parameters according to the difference between the generated blood pressure interval and the true blood pressure interval, and optimizes its judgment ability between the generated data and the real data. Training a generative adversarial network is an alternating process. Each training session simultaneously updates the parameters of the generator and discriminator. In each iterative training phase, the discriminator is trained first, followed by the generator, alternating until the blood pressure intervals generated by the generator are close enough to the true blood pressure intervals that the discriminator cannot easily distinguish between the generated data and the true data. The generator's loss function is typically based on feedback from the discriminator, optimizing the generator by maximizing the probability that the discriminator determines that the generated data is true data. The discriminator's loss function is typically based on the error in its judgment results, aiming to minimize the error between the generated data and the true data, ensuring that the discriminator can accurately distinguish between the two. During testing, sample motion signals from the test set are input, and the generator uses these inputs to generate predicted blood pressure intervals. The accuracy of the generator's predicted blood pressure intervals is then evaluated. Metrics such as mean squared error can be used to measure the difference between the generated blood pressure intervals and the true blood pressure intervals. When the generative adversarial network training is complete and performance meets expectations, the trained exercise blood pressure prediction branch is obtained.

[0048] Finally, the motion signal is input into the exercise blood pressure prediction branch for prediction, and a first predicted blood pressure interval is output. The predicted interval provides the user with a blood pressure fluctuation range in an exercise state.

[0049] Furthermore, the user blood pressure prediction module 13 is further configured to:

[0050] The PPG signal is input into a pre-trained photoelectric blood pressure prediction branch, and the predicted blood pressure is obtained by prediction output, wherein the photoelectric blood pressure prediction branch is constructed based on machine learning, and supervised training is performed until convergence using a set of sample PPG signals and a set of sample blood pressure within the target user's historical time; the sensing error influence parameter is calculated based on the first sensing error influence parameter and the second sensing error influence parameter; the sensing error influence parameter is used to perform error compensation on the predicted blood pressure to obtain a second predicted blood pressure range.

[0051] Specifically, first, a photoelectric blood pressure prediction branch is constructed based on machine learning, and the photoelectric blood pressure prediction branch is used to predict blood pressure through a machine learning model. For example, a photoelectric blood pressure prediction branch is constructed based on a BP neural network, wherein the photoelectric blood pressure prediction branch includes an input layer, multiple hidden layers and an output layer, wherein the input data of the input layer is a PPG signal, and the output data of the output layer is blood pressure data; then, a set of sample PPG signals and a set of sample blood pressure within the historical time of the target user are used as training data, and the sample PPG signal is used as input and the sample blood pressure is used as supervision to perform supervised training on the photoelectric blood pressure prediction branch. During training, the PPG signal is input to the input of the neural network. The network then processes the data through the hidden layer, ultimately generating a predicted blood pressure value. The network adjusts the weights in the output layer by calculating the difference between the loss function (such as mean squared error) and the actual blood pressure value. Next, the backpropagation algorithm calculates the gradient based on the loss function and adjusts the weights and biases in the network to reduce the error. Gradient descent (or its variants, such as the Adam optimizer) is typically used to optimize the weights to minimize the loss function. During training, the neural network continuously iteratively updates parameters through forward and backward propagation until convergence. Convergence typically means that the loss function no longer decreases significantly, or that a preset number of training rounds has been reached, resulting in a trained photoelectric blood pressure prediction branch. The PPG signal is then fed into the trained photoelectric blood pressure prediction branch for prediction, outputting the predicted blood pressure.

[0052] Then, the first sensing error influence parameter (indicating the impact of motion-related errors on blood pressure prediction) and the second sensing error influence parameter (indicating the error impact of factors such as skin color and contact on blood pressure prediction) are added and summed to obtain a sensing error influence parameter, which represents the total error impact on blood pressure prediction caused by the combined action of all external factors (such as motion, skin color, contact, etc.); the sensing error influence parameter is further used to proportionally expand the predicted blood pressure to obtain a second predicted blood pressure interval. For example, assuming that the sensing error influence parameter is 10%, the predicted high pressure is 120, and the low pressure is 80, then the predicted high pressure interval is 120*(1-10%) to 120*(1+10%), that is, the predicted high pressure interval is 108 to 132, and the predicted low pressure interval is 80*0.9 to 80*1.1, that is, 72 to 88, and the predicted high pressure interval and the predicted low pressure interval are set as the second predicted blood pressure interval.

[0053] The error coefficient marking module 14 is used to calculate the blood pressure monitoring error coefficient based on the first predicted blood pressure interval and the second predicted blood pressure interval, and mark the predicted blood pressure as the blood pressure monitoring result.

[0054] Furthermore, the error coefficient marking module 14 is further configured to:

[0055] Obtain the intersection of the first predicted blood pressure interval and the second predicted blood pressure interval to obtain the intersection blood pressure interval; calculate the ratio of the intersection blood pressure interval to the first predicted blood pressure interval and the second predicted blood pressure interval respectively, and calculate the first blood pressure monitoring error coefficient and the second blood pressure monitoring error coefficient; calculate the blood pressure monitoring error coefficient based on the first blood pressure monitoring error coefficient and the second blood pressure monitoring error coefficient; use the blood pressure monitoring error coefficient to mark the predicted blood pressure to obtain the blood pressure monitoring result.

[0056] Specifically, an intersection operation is performed on the first predicted blood pressure interval and the second predicted blood pressure interval to obtain an intersection blood pressure interval. The intersection blood pressure interval represents the blood pressure range that is satisfied by both (including the high pressure intersection and the low pressure intersection). Then, the interval length ratio of the intersection blood pressure interval to the first predicted blood pressure interval is calculated, and the difference between the two is set as the first blood pressure monitoring error coefficient. The first blood pressure monitoring error coefficient reflects the degree of difference between the first predicted blood pressure interval and the intersection blood pressure interval. The smaller the error coefficient, the more accurate the predicted blood pressure interval. On the other hand, the interval length ratio of the intersection blood pressure interval to the second predicted blood pressure interval is calculated, and the difference between the two is set as the second blood pressure monitoring error coefficient. The second blood pressure monitoring error coefficient reflects the degree of difference between the second predicted blood pressure interval and the intersection blood pressure interval. A smaller error coefficient indicates that the difference between the two blood pressure intervals is smaller, and the reliability of the predicted blood pressure is higher. A larger error coefficient indicates that the difference is larger, and the predicted blood pressure may have a larger error.

[0057] A weighted calculation is further performed on the first blood pressure monitoring error coefficient and the second blood pressure monitoring error coefficient. Since the accuracy of predicting the user's blood pressure through motion signals is low, the weight of the second blood pressure monitoring error coefficient is much greater than the weight of the first blood pressure monitoring error coefficient. For example, the first weight is set to 0.1 and the second weight is set to 0.9. A weighted calculation is then performed to obtain the blood pressure monitoring error coefficient. The weighted blood pressure monitoring error coefficient obtained through the weighted calculation can be used to evaluate the accuracy and reliability of the overall blood pressure prediction. Finally, the predicted blood pressure is annotated using the blood pressure monitoring error coefficient to obtain a blood pressure monitoring result, thereby providing the user with more accurate blood pressure monitoring data.

[0058] In summary, the multimodal intelligent non-invasive blood pressure real-time monitoring device provided by the present invention has the following technical effects:

[0059] By collecting PPG signals and motion signals of the target user, the PPG signals and motion signals are obtained, and the skin color information of the target user is collected; then, a sensing error influence analysis and a contact influence parameter prediction are performed based on the motion signal to obtain a first sensing error influence parameter and a contact influence parameter, and a sensing error influence analysis is performed based on the skin color information and the contact influence parameter to obtain a second sensing error influence parameter; then, based on the motion signal, the user's blood pressure is predicted to obtain a first predicted blood pressure range; on the other hand, based on the PPG signal, the user's blood pressure is predicted to obtain a predicted blood pressure, and error compensation is performed based on the first sensing error influence parameter and the second sensing error influence parameter to obtain a second predicted blood pressure range; finally, based on the first predicted blood pressure range and the second predicted blood pressure range, the blood pressure monitoring error coefficient is calculated and the predicted blood pressure is marked as the blood pressure monitoring result. In other words, by combining multimodal sensor data for sensor error analysis, it is possible to accurately identify and quantify sensor errors caused by different factors, thereby improving the accuracy and reliability of sensor error acquisition; then, the predicted blood pressure is labeled according to the error analysis results, which can provide users with feedback information on measurement reliability, helping users to effectively identify the reliability of predicted blood pressure, thereby providing users with more accurate, comprehensive and reliable blood pressure monitoring data.

[0060] In the second embodiment, based on the same inventive concept as the multimodal intelligent non-invasive blood pressure real-time monitoring device in the above embodiment, the present invention also provides a multimodal intelligent non-invasive blood pressure real-time monitoring method, please refer to the attached Figure 2 , including: collecting PPG signals and motion signals of a target user through a PPG sensor and a motion sensor to obtain PPG signals and motion signals, and collecting skin color information of the target user; performing sensing error influence analysis and contact influence parameter prediction based on the motion signal to obtain first sensing error influence parameters and contact influence parameters, performing sensing error influence analysis based on the skin color information and contact influence parameters to obtain second sensing error influence parameters; predicting the user's blood pressure based on the motion signal to obtain a first predicted blood pressure range, predicting the user's blood pressure based on the PPG signal to obtain a predicted blood pressure, performing error compensation based on the first sensing error influence parameter and the second sensing error influence parameter to obtain a second predicted blood pressure range; calculating a blood pressure monitoring error coefficient based on the first predicted blood pressure range and the second predicted blood pressure range, marking the predicted blood pressure as a blood pressure monitoring result.

[0061] Furthermore, the multimodal intelligent non-invasive blood pressure real-time monitoring method also includes: collecting PPG signals and motion signals of the target user within a recent preset time window through a PPG sensor and a motion sensor to obtain PPG signals and motion signals, wherein the motion sensor is a gyroscope; and obtaining skin color information of the target user collected and recorded within a historical time period.

[0062] Furthermore, the multimodal intelligent non-invasive blood pressure real-time monitoring method also includes: collecting a set of sample motion signals based on monitoring data in historical time, and collecting the error amplitudes of different sample motion signals and blood pressure predicted based on PPG signals in a non-exercise state, marking them as sample first sensing error influence parameters, and obtaining a set of sample first sensing error influence parameters; collecting the user's skin sweating parameters under different sample motion signals, marking them as sample contact influence parameters, and obtaining a set of sample contact influence parameters; using machine learning to construct a motion sensing error influence prediction branch and a contact influence prediction branch; using the sample motion signal set as input features, and using the sample first sensing error influence parameter set and the sample contact influence parameter set as output features respectively, and performing supervised training and testing on the motion sensing error influence prediction branch and the contact influence prediction branch until convergence; inputting the motion signal into the motion sensing error influence prediction branch and the contact influence prediction branch respectively, and predicting the output to obtain the first sensing error influence parameter and the contact influence parameter.

[0063] Furthermore, the multimodal intelligent non-invasive blood pressure real-time monitoring method also includes: collecting a sample skin color information set and a sample contact influence parameter set based on monitoring data within a historical period; collecting different sample skin color information and sample contact influence parameters compared with the preset skin color and preset contact influence parameters, predicting the error amplitude of blood pressure based on the PPG signal, marking it as the sample second sensing error influence parameter, and obtaining the sample second sensing error influence parameter set; constructing a skin color sensing error influence prediction branch based on machine learning; using the sample skin color information set, sample contact influence parameter set and sample second sensing error influence parameter set, the skin color sensing error influence prediction branch is supervised trained and tested until convergence; the skin color information and contact influence parameters are input into the skin color sensing error influence prediction branch, and the prediction output obtains the second sensing error influence parameter.

[0064] Furthermore, the multimodal intelligent non-invasive blood pressure real-time monitoring method also includes: collecting a set of sample motion signals based on the blood pressure monitoring data of the target user in the historical time, and collecting the blood pressure intervals of the target user under different sample motion signals, and marking to obtain a set of sample first predicted blood pressure intervals; constructing an exercise blood pressure prediction branch based on machine learning; using the sample motion signal set and the sample first predicted blood pressure interval set, the exercise blood pressure prediction branch is supervised trained and tested until convergence; the motion signal is input into the exercise blood pressure prediction branch, and the prediction output obtains the first predicted blood pressure interval.

[0065] Furthermore, the multimodal intelligent non-invasive blood pressure real-time monitoring method also includes: inputting the PPG signal into a pre-trained photoelectric blood pressure prediction branch, and predicting and outputting the predicted blood pressure, wherein the photoelectric blood pressure prediction branch is constructed based on machine learning, and uses a set of sample PPG signals and a set of sample blood pressure within the target user's historical time to perform supervised training until convergence; calculating the sensing error influence parameter based on the first sensing error influence parameter and the second sensing error influence parameter; and using the sensing error influence parameter to perform error compensation on the predicted blood pressure to obtain a second predicted blood pressure range.

[0066] Furthermore, the multimodal intelligent non-invasive blood pressure real-time monitoring method also includes: obtaining the intersection of the first predicted blood pressure interval and the second predicted blood pressure interval to obtain the intersection blood pressure interval; respectively calculating the ratio of the intersection blood pressure interval to the first predicted blood pressure interval and the second predicted blood pressure interval, and calculating the first blood pressure monitoring error coefficient and the second blood pressure monitoring error coefficient; calculating the blood pressure monitoring error coefficient based on the first blood pressure monitoring error coefficient and the second blood pressure monitoring error coefficient; and using the blood pressure monitoring error coefficient to mark the predicted blood pressure to obtain a blood pressure monitoring result.

[0067] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The intelligent non-invasive blood pressure real-time monitoring device based on multimodality and the specific examples in the aforementioned embodiment 1 are also applicable to the intelligent non-invasive blood pressure real-time monitoring method based on multimodality in this embodiment. Through the aforementioned detailed description of the intelligent non-invasive blood pressure real-time monitoring device based on multimodality, those skilled in the art can clearly understand the intelligent non-invasive blood pressure real-time monitoring method based on multimodality in this embodiment, so for the sake of brevity of the specification, it will not be described in detail here. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0068] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

[0069] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the present invention and its equivalents, the present invention is intended to include these modifications and variations.

Claims

1. An intelligent non-invasive blood pressure real-time monitoring device based on multimodality, characterized in that: The device includes a PPG sensor and a motion sensor, and further includes: A data acquisition module is used to collect PPG signals and motion signals of a target user through a PPG sensor and a motion sensor, obtain the PPG signals and motion signals, and collect skin color information of the target user; An error influence analysis module is configured to perform a sensing error influence analysis and a contact influence parameter prediction based on the motion signal to obtain a first sensing error influence parameter and a contact influence parameter, and to perform a sensing error influence analysis based on the skin color information and the contact influence parameter to obtain a second sensing error influence parameter, including: Based on the monitoring data in the historical time period, a set of sample motion signals is collected, and the error ranges of different sample motion signals and blood pressure predicted based on the PPG signal in a non-exercise state are collected and marked as sample first sensor error influence parameters to obtain a set of sample first sensor error influence parameters; Collect the user's skin sweat parameters under different sample motion signals, mark them as sample contact influence parameters, and obtain a set of sample contact influence parameters; Using machine learning, we build motion sensing error impact prediction branches and contact impact prediction branches. Using the sample motion signal set as input features, and using the sample first sensing error influence parameter set and the sample contact influence parameter set as output features, respectively, to perform supervised training and testing on the motion sensing error influence prediction branch and the contact influence prediction branch until convergence; Inputting the motion signal into the motion sensing error influence prediction branch and the contact influence prediction branch respectively, and predicting and outputting a first sensing error influence parameter and a contact influence parameter; Based on the monitoring data in the historical period, a set of sample skin color information and a set of sample contact impact parameters are collected; Collect different sample skin color information and sample contact influence parameters, compare them with the preset skin color and preset contact influence parameters, and predict the error range of blood pressure based on the PPG signal, mark them as sample second sensor error influence parameters, and obtain the sample second sensor error influence parameter set; Based on machine learning, a skin color sensing error impact prediction branch is constructed; Using the sample skin color information set, the sample contact influence parameter set, and the sample second sensing error influence parameter set, supervised training and testing are performed on the skin color sensing error influence prediction branch until convergence; Inputting the skin color information and the contact influence parameter into the skin color sensing error influence prediction branch, and predicting and outputting a second sensing error influence parameter; a user blood pressure prediction module, configured to predict the user's blood pressure based on the motion signal to obtain a first predicted blood pressure interval, predict the user's blood pressure based on the PPG signal to obtain a predicted blood pressure, and perform error compensation based on the first and second sensing error influencing parameters to obtain a second predicted blood pressure interval; An error coefficient marking module is configured to calculate a blood pressure monitoring error coefficient based on the first predicted blood pressure interval and the second predicted blood pressure interval, and mark the predicted blood pressure as a blood pressure monitoring result, including: Obtaining the intersection of the first predicted blood pressure interval and the second predicted blood pressure interval to obtain an intersection blood pressure interval; Calculating the ratio of the intersection blood pressure interval to the first predicted blood pressure interval and the second predicted blood pressure interval respectively, and obtaining a first blood pressure monitoring error coefficient and a second blood pressure monitoring error coefficient by calculation; Calculating a blood pressure monitoring error coefficient based on the first blood pressure monitoring error coefficient and the second blood pressure monitoring error coefficient; The predicted blood pressure is marked using the blood pressure monitoring error coefficient to obtain a blood pressure monitoring result.

2. The multimodal intelligent non-invasive blood pressure real-time monitoring device according to claim 1, characterized in that: The PPG sensor and the motion sensor are used to collect PPG signals and motion signals of the target user, thereby obtaining the PPG signals and motion signals, and collecting skin color information of the target user, including: The PPG signal and motion signal of the target user are collected within a recent preset time window by using a PPG sensor and a motion sensor, wherein the motion sensor is a gyroscope; Get the skin color information of the target user collected and recorded in the historical time.

3. The multimodal intelligent non-invasive blood pressure real-time monitoring device according to claim 1, characterized in that: Predicting the user's blood pressure based on the motion signal to obtain a first predicted blood pressure range includes: Based on the target user's blood pressure monitoring data over a historical period, a set of sample motion signals is collected, and the target user's blood pressure intervals under different sample motion signals are collected, and the first predicted blood pressure interval set of the samples is obtained by annotation; Based on machine learning, build an exercise blood pressure prediction branch; Using the sample motion signal set and the sample first predicted blood pressure interval set, the exercise blood pressure prediction branch is supervised trained and tested until convergence; The exercise signal is input into the exercise blood pressure prediction branch, and a prediction output is obtained to obtain a first predicted blood pressure interval.

4. The multimodal intelligent non-invasive blood pressure real-time monitoring device according to claim 1, characterized in that: Predicting the user's blood pressure based on the PPG signal to obtain a predicted blood pressure, and performing error compensation based on the first sensor error influencing parameter and the second sensor error influencing parameter to obtain a second predicted blood pressure range, including: Inputting the PPG signal into a pre-trained photoelectric blood pressure prediction branch, and predicting the output to obtain the predicted blood pressure, wherein the photoelectric blood pressure prediction branch is constructed based on machine learning and uses a set of sample PPG signals and a set of sample blood pressure within the target user's historical time period for supervised training until convergence; Calculating a sensing error influencing parameter according to the first sensing error influencing parameter and the second sensing error influencing parameter; The sensor error influencing parameter is used to perform error compensation on the predicted blood pressure to obtain a second predicted blood pressure interval.

5. The multimodal intelligent non-invasive blood pressure real-time monitoring method is characterized by: The method is performed by the multimodal intelligent non-invasive blood pressure real-time monitoring device according to any one of claims 1 to 4, further comprising: Collecting PPG signals and motion signals of a target user through a PPG sensor and a motion sensor, obtaining the PPG signals and motion signals, and collecting skin color information of the target user; Performing a sensing error influence analysis and a contact influence parameter prediction based on the motion signal to obtain a first sensing error influence parameter and a contact influence parameter, and performing a sensing error influence analysis based on the skin color information and the contact influence parameter to obtain a second sensing error influence parameter; Predicting the user's blood pressure based on the motion signal to obtain a first predicted blood pressure interval, predicting the user's blood pressure based on the PPG signal to obtain a predicted blood pressure, and performing error compensation based on the first and second sensing error influencing parameters to obtain a second predicted blood pressure interval; A blood pressure monitoring error coefficient is calculated based on the first predicted blood pressure interval and the second predicted blood pressure interval, and the predicted blood pressure is marked as a blood pressure monitoring result.

Images