Blood pressure detection method based on PPG

Abstract

The invention discloses a PPG-based blood pressure detection method, and particularly relates to the technical field of biomedical signal detection, and the method specifically comprises the following steps: 1, collecting a PPG original waveform through a PPG sensor; step 2, preprocessing the collected PPG signals; step 3, carrying out feature extraction on the preprocessed PPG signal; 4, the extracted PPG features serve as input, reference blood pressure parameters serve as output, an LSTM network model is constructed, the network weight is trained, and the mapping relation between the features and the blood pressure is established; step 5, obtaining a personalized model; and 6, collecting a large-scale multi-source heterogeneous data sample, and optimizing the model by adopting a transfer learning technology and an incremental learning technology. By processing and modeling PPG signals, accurate detection of blood pressure parameters is achieved, non-invasive continuous blood pressure monitoring can be achieved, and the long-term monitoring requirements of old people and patients with cardiovascular diseases are met.

Description

Technical Field

[0001] The present invention relates to the technical field of biomedical signal detection, and more specifically to a blood pressure detection method based on PPG. Background Art

[0002] At present, the "gold standard" for non-invasive blood pressure measurement in clinical applications is the mercury sphygmomanometer, but it requires professional medical staff for operation. The cuff-type electronic sphygmomanometer is widely used in clinics due to its simple operation. However, both of these methods belong to intermittent blood pressure detection. The cuff pressurization will block the blood flow in the blood vessels, which is not suitable for long-term blood pressure detection and management, and cannot be used in the case of wrist or arm injury or special operations. Photoplethysmography (PPG) is a method of detecting the change in blood volume in living tissues by means of optoelectronics, and can be continuously detected for a long time by wearing a wearable device. At present, there are many applications of PPG in the detection and evaluation of various hemodynamic parameters such as blood oxygen and heart rate. Research shows that PPG and arterial blood pressure are morphologically similar and can be used for blood pressure detection. The heartbeat generates a continuous pressure wave, which is transmitted through the blood vessels and slightly changes the diameter of the blood vessels. The change in the blood volume of the blood vessels can be detected by PPG, and its signal has the characteristics of non-stationarity and pseudo-periodicity.

[0003] In recent years, there have been many studies on blood pressure detection algorithms based on PPG, mainly using model algorithms such as linear regression algorithm (LR), long short-term memory network (LSTM), autoregressive moving average model (ARMA), support vector regression model (SVR), error backpropagation neural network (BP), extreme learning machine neural network (ELM), convolutional neural network (CNN), etc., and combining filtering, FFT function, least squares method, etc. to construct a detection and prediction algorithm model. It can be mainly divided into two categories: neural network algorithms and non-neural network algorithms:

[0004] (1) Neural network algorithms:

[0005] In 2017, Lo et al. proposed a method for establishing a long short-term memory (LSTM) neural network by calculating the root mean square error (RMSE) between the estimated blood pressure value and the predicted value. The experimental results show that LSTM-based blood pressure estimation has great potential for embedding into monitoring systems, with good accuracy and generality. In the future, the number of subjects between different age groups can be increased for further training of the LSTM network, which has better accuracy, generalization, and stability.

[0006] In 2018, Ertugurl et al. proposed a method for estimating blood pressure using ECG / PPG signals through an extreme learning machine neural network (ELM). By this method, systolic blood pressure, mean blood pressure, and diastolic blood pressure can be calculated simultaneously. This training algorithm has a fast training phase and high generalization ability.

[0007] In 2019, Lei Suli carried out work on suppressing the noise of photoplethysmogram, detecting characteristic parameters, and establishing a blood pressure monitoring model with the human fingertip photoplethysmogram as the research object, and proposed an algorithm for the error backpropagation neural network (BP). The output results have a good measurement experience in healthy people, but are not applicable to people with high or low blood pressure.

[0008] In the same year, Slapnicar et al. established a residual network model using PPG and its first and second derivatives as inputs. The experiments show that this method is suitable for one-time measurement, while fingertip devices are usually not suitable for mobile use, so different devices must be used for non-invasive continuous blood pressure measurement monitoring.

[0009] In 2021, Liu et al. used several machine learning algorithms to predict the continuous and cuffless estimation of diastolic and systolic blood pressure. Their results show that compared with linear regression and support vector regression methods, the artificial neural network optimized by the genetic algorithm has the advantages of efficient signal decomposition for extracting microarterial PTT, sensitivity, etc., and has better accuracy in predicting blood pressure.

[0010] (2) Frame difference method:

[0011] In 2016, Ghosh et al. adopted the least squares algorithm to predict the constant deviation of higher blood pressure values in order to improve the calculation of accurate PTT. The observation error of this measurement is within 1% of the manual PTT measurement. This research is more accurate for static measurement of objects, but not accurate enough for dynamic measurement.

[0012] In 2018, Wang et al. established the relationship between effective features and blood pressure using learning algorithms of artificial neural network, support vector regression, and linear regression respectively. Their results show that the measurement effect of the artificial neural network optimized by the genetic algorithm is better than other methods and has more potential application prospects.

[0013] In 2019, Lazazzera et al. estimated SBP and DBP from a linear regression model (LR) using the time difference dependence between positive maximum points and the time difference between the feet of the PPG waveform.

[0014] Predicting blood pressure parameters directly from the PPG waveform is a complex process, and existing PPG-based blood pressure detection algorithms still have some problems:

[0015] 1) It is unable to effectively extract the effective features in the PPG signal. The PPG signal is mixed with various information such as morphological features and frequency domain features, but existing algorithms cannot effectively distinguish prominent blood pressure-related features, resulting in inaccurate feature extraction. Identifying effective features is necessary for modeling and prediction, but the current methods cannot complete this step.

[0016] 2) The modeling method is simple and the prediction effect is poor. Existing algorithms mostly use simple methods such as linear regression for modeling, which is difficult to reflect the complex non-linear relationship between the PPG signal and blood pressure, resulting in a large deviation in the prediction results. Using only simple linear methods cannot simulate its complex correspondence.

[0017] 3) The model generalization ability is poor. Existing models are trained based on small samples, with weak generalization performance, poor effects on new samples, and unable to adapt to individual differences. The limited sample size results in poor generalization of existing models to new situations.

[0018] The above problems limit the application of PPG technology in blood pressure monitoring. There is an urgent need for a new PPG signal processing and modeling method to improve the accuracy of blood pressure prediction. The current methods cannot meet the functional requirements of continuous non-invasive monitoring. Therefore, the present invention designs a continuous long-term non-invasive blood pressure detection method based on the PCG signal that is superior to the performance of existing mainstream algorithms. Summary of the Invention

[0019] Aiming at the problems existing in the PPG blood pressure detection algorithm in the prior art, such as the inability to effectively extract signal features, simple modeling means, and poor generalization ability, the present invention provides a PPG-based blood pressure detection method, which can accurately extract features from the PPG signal and establish a personalized prediction model, thereby improving the accuracy of blood pressure detection.

[0020] To achieve the above object, the present invention provides the following technical solution: A PPG-based blood pressure detection method, the specific steps are as follows:

[0021] Step 1: PPG signal acquisition: Collect the original PPG waveform through a PPG sensor;

[0022] Step 2. PPG signal preprocessing: First, perform wavelet transform on the collected original PPG waveform signal to decompose it into multiple frequency layers. Then, introduce an attention mechanism between different frequency layers to assign different weight coefficients to different layers. Finally, reconstruct each frequency band to obtain the preprocessed PPG signal;

[0023] Step 3. Feature extraction: Use the LSTM neural network model to train the preprocessed PPG signal, automatically learn its features in the time domain and frequency domain, and obtain a PPG feature set describing blood pressure changes;

[0024] Step 4. Model establishment: Use the extracted PPG features as input and the reference blood pressure parameters as output to construct an LSTM network model, train the network weights, establish a mapping relationship between features and blood pressure, and be able to directly output blood pressure values when inputting the PPG feature vector;

[0025] Step 5. Individual adjustment: After obtaining the overall model, when introducing the PPG data of a new individual, first use this model for prediction, and then make fine adjustments to the model with blood pressure data to obtain a personalized model;

[0026] Step 6: Model optimization: Over time, collect large-scale multi-source heterogeneous data samples, and then use transfer learning technology and incremental learning technology to update the model.

[0027] Preferably, the PPG sensor in Step 1 is wearable and designed in the form of a bracelet or finger sleeve. The sensor contains an infrared light-emitting diode and a light-receiving diode;

[0028] The specific steps of PPG signal acquisition are as follows: The infrared rays emitted by the infrared light-emitting diode of the sensor penetrate the skin tissue, and the light-receiving diode detects the change in the intensity of the infrared rays transmitted through the skin and converts it into an electric current signal to obtain the original PPG waveform reflecting the change in blood volume, and collect the original waveform according to the Nyquist sampling theorem.

[0029] Preferably, the specific steps of PPG signal preprocessing in Step 2 are as follows:

[0030] A1: On the obtained original PPG waveform signal, first perform wavelet transform, and then use the Mallat fast algorithm to take 8-layer db4 wavelet transform. After the transform, it is decomposed into a low-frequency component cA8 and high-frequency components cD1-8;

[0031] A2: Introduce an attention mechanism to set different weight coefficients for different layers;

[0032] A3: According to the preset weight coefficients, use the weighted reconstruction method to linearly combine each frequency component to synthesize the preprocessed PPG signal.

[0033] Preferably, the specific steps of feature extraction in step three are as follows:

[0034] B1. Construct an LSTM neural network model. Set the number of nodes in the input layer to the number of sampling points of the PPG waveform, set the number of nodes in the output layer to the dimension of the desired PPG feature vector, and set the network hidden layer according to the task complexity;

[0035] Initialize the network parameters, and then input the collected PPG waveform sequence for training. Optimize the training network weight parameters through the BPTT algorithm. The network automatically learns to extract the effective features of the PPG waveform. After multiple rounds of training, the network output stabilizes at a 10-dimensional vector, which contains the global and local feature information of the PPG waveform;

[0036] B2. Input the preprocessed PPG waveform samples into the LSTM neural network model, configure random learning parameters, start training, and optimize the network weights using the gradient descent algorithm;

[0037] B3. Repeat multiple rounds of iterative training. The LSTM neural network model automatically learns the time-domain and frequency-domain features of the PPG waveform, with the convergence of the loss function as the termination condition;

[0038] B4. The network outputs a 10-dimensional feature vector, which is the feature expression of the PPG waveform, reflecting the overall shape of the waveform, interval statistical characteristics, and frequency band component information. This low-dimensional vector is the effective feature of the PPG signal.

[0039] Preferably, the specific steps of model establishment in step four are as follows:

[0040] C1. After obtaining the feature vector of the PPG waveform, the network input layer is the PPG feature vector. Set the number of nodes in the first LSTM hidden layer to 8 to learn low-level feature combinations, set the number of nodes in the second LSTM hidden layer to 16 to learn high-level feature abstractions, and set the number of nodes in the output layer to the dimension of the blood pressure parameter. Collect PPG training samples with blood pressure annotations as supervised data, train the network parameters, minimize the mean square error of blood pressure prediction. Through multiple rounds of training, the LSTM network learns the non-linear mapping relationship between PPG features and blood pressure parameters, and obtains an end-to-end prediction model. Input the PPG feature vector, and it can directly output the blood pressure value;

[0041] C2. Collect PPG waveform samples with reference blood pressure parameters as training data, and use the reference blood pressure parameters as the supervised signal to train the LSTM network weights;

[0042] C3. After constructing the LSTM network and initializing the parameters, it is necessary to repeat multiple rounds of iterative training to gradually optimize the model.

[0043] Preferably, the specific steps of individual adjustment in step five are as follows:

[0044] D1. Collect the PPG waveform sample data of a specific individual and the corresponding true blood pressure parameters as new training samples;

[0045] D2. After collecting enough samples, input these new individual PPG waveform data into a pre-trained blood pressure prediction model. The model will output the predicted blood pressure value, and then calculate the error between the predicted blood pressure and the true blood pressure. These errors constitute a new loss function, indicating the fitting effect of the current general model on this individual;

[0046] D3. Based on the newly added loss function, calculate the gradient contribution of each network parameter to the loss function through the backpropagation algorithm, and adjust the model parameters accordingly based on this gradient information;

[0047] D4. By repeating the gradient-guided parameter adjustment in multiple rounds, the prediction error of the personalized model will gradually decrease and finally converge to an acceptable range, indicating that a personalized model that can provide high-precision blood pressure prediction for this specific individual has been obtained. If a new individual user is encountered, only need to repeat the parameter adjustment and model fine-tuning process, so as to gradually build a blood pressure health monitoring model library with strong personalization effects for different individuals.

[0048] Preferably, in step six, the transfer learning technology includes the transfer learning technology of the joint training mode and the transfer learning technology of model cloning;

[0049] The transfer learning technology of the joint training mode enables the model to learn the common knowledge across datasets during the process of processing each dataset by simultaneously inputting different source datasets, so as to have a certain adaptability to different population distributions and greatly improve the generalization ability of the model to new examples;

[0050] The transfer learning technology of model cloning generates multiple new models based on a source model through the partial parameter replication method. The new models not only inherit the generalization knowledge of the source model but also adapt to their respective sample distributions. Finally, an integrated model is formed through model fusion, which has stronger generalization performance;

[0051] The incremental learning technology is to first train only with the samples of the elderly, and then gradually add other samples for iterative optimization to avoid the decline of the generalization ability of the model to the elderly samples and make the resulting model more suitable for hypertensive patients.

[0052] The technical effects and advantages of the present invention:

[0053] 1. The PPG signal acquisition is simple. It is acquired using a wearable device, and the acquisition method is convenient;

[0054] 2. The preprocessing of the PPG signal retains the key features, improves the quality of subsequent feature extraction, and ensures that sufficient blood pressure-related information is included;

[0055] 3. When using the LSTM neural network model to extract time-frequency domain features, it can learn the internal laws of the data and obtain more expressive features;

[0056] 4. Using the LSTM network for modeling to adapt to the non-linear relationship between the PPG waveform and blood pressure, making the prediction more accurate;

[0057] 5. Obtaining a personalized model through individual calibration, which can adapt to the differences of different individuals;

[0058] 6. The constructed model can be continuously optimized to increase the generalization range;

[0059] 7. Finally, non-invasive continuous blood pressure monitoring is achieved to meet the long-term monitoring needs of patients and the elderly;

[0060] 8. The entire system can be implemented through software and hardware integration;

[0061] 9. A new PPG signal detection idea is provided, which can promote the progress of related technologies and at the same time promote the research and development and application of related wearable devices.

[0062] In summary, the blood pressure detection method of the present invention overall optimizes the processing efficiency, extraction effect and prediction accuracy of PPG signals, realizes the function of non-invasive continuous blood pressure monitoring, meets the long-term monitoring needs of the elderly and patients with cardiovascular diseases, and has important application value. Specific Embodiments

[0063] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0064] The present invention provides a blood pressure detection method based on PPG, and the specific steps are as follows:

[0065] Step 1. PPG signal acquisition:

[0066] A wearable PPG sensor is adopted, and the sensor can be designed in the form of a bracelet or a finger ring for easy wearing by the subject on the wrist or finger;

[0067] Among them, the sensor contains an infrared light-emitting diode and a light-receiving diode, and the infrared light-emitting diode and the light-receiving diode are fixed at intervals, and the light-emitting and receiving distance is 5-10 mm; the infrared light-emitting diode emits infrared light with a wavelength of about 940 nm, which can penetrate the skin tissue by 3-5 mm and enter the subcutaneous capillaries;

[0068] Then, the subject wears the sensor on the wrist or finger, turns on the infrared light-emitting diode to emit light. After the infrared light penetrates the skin tissue and enters the subcutaneous capillaries, part of the light will be absorbed by hemoglobin. When the heart contracts to pump blood, the subcutaneous blood volume increases, and the amount of infrared light absorbed increases, so the transmitted light intensity on the skin decreases; when the heart relaxes, the subcutaneous blood volume decreases, and the amount of infrared light absorbed decreases, so the transmitted light intensity on the skin increases.

[0069] The light-receiving diode detects the change in the intensity of the infrared light transmitted through the skin and converts it into an electrical signal. Then, the electrical signal is amplified and filtered by the circuit, and the original PPG waveform reflecting the change in blood volume can be obtained. Set the sampling frequency above 100 Hz, and collect the original waveform according to the Nyquist sampling theorem.

[0070] Step 2: PPG signal preprocessing:

[0071] A1: On the obtained original PPG waveform signal, first perform wavelet transform, and then adopt the Mallat fast algorithm to take the 8-layer db4 wavelet transform. After the transform, it is decomposed into a low-frequency component cA8 and high-frequency components cD1-8. The specific steps are as follows:

[0072] First, select db4 as the wavelet basis. As the 4th wavelet in the Daubechies series, it has a good expression effect and regularity for pulse signals. Its waveform is smooth, symmetric, and has no retracement, which is suitable for analyzing physiological signals such as PPG.

[0073] Then, adopt the Mallat algorithm to perform fast wavelet transform; this algorithm is based on the theory of multi-resolution analysis, and through the quadratic folding filt-conv operation, recursively perform low-pass and high-pass filtering, so as to achieve efficient multi-level wavelet decomposition.

[0074] Next, perform 8-layer wavelet decomposition. Each layer of decomposition generates a low-frequency component and a high-frequency component; the low-frequency component reflects the signal trend, and the high-frequency component reflects the details. The 8th-layer low-frequency component is denoted as cA8, and the 1-8th layer high-frequency components are denoted as cD1-8.

[0075] Through 8-layer decomposition, the PPG signal is resolved into different frequency sub-bands such as 0.4 - 250 Hz; among them, cA8 is the trend information of 0 - 0.4 Hz, and cD1-8 contains more detailed frequency component information above 0.4 Hz, which realizes the multi-resolution analysis from low frequency to high frequency.

[0076] A2: Introduce the attention mechanism to set different weight coefficients for different layers:

[0077] In the spectral analysis of PPG signals, different frequency ranges contain different physiological information. By setting reasonable frequency weight coefficients, useless information can be suppressed, the proportion of effective information can be increased, and the signal preprocessing effect can be optimized. After the wavelet transform decomposition of the PPG signal, the specific meanings of each frequency band are as follows:

[0078] The lowest-frequency cA8 component mainly contains the trend term of the PPG signal, reflecting the overall change law of skin blood circulation; this part has little relation with the heartbeat fluctuation and has low information content.

[0079] The highest-frequency cD1-2 component corresponds to more than half of the sampling frequency of the original PPG signal, containing high-frequency random noise and some subtle ripple changes; this part of the information is not very helpful for blood pressure monitoring.

[0080] The middle-frequency cD3-5 component corresponds to the fluctuation range of 0.4 Hz to 2.5 Hz, matching the physiological frequency band of the heart rate; this part contains the main pulse wave information and is the key frequency band reflecting blood pressure changes.

[0081] The higher-frequency cD6-8 contains higher-frequency random noise and some high-frequency physiological noise, and the information content is also low.

[0082] In summary, when setting the frequency weight coefficients, emphasis is placed on the middle and low-frequency cD3-5 components, giving this part a higher weight coefficient, above 0.7; a lower weight coefficient is given to the lowest-frequency cA8 and the highest-frequency cD1-2, cD6-8, around 0.1, so as to suppress useless information and increase the proportion of effective pulse signal components.

[0083] The weight coefficients adopt a linear adjustment form, and a non-linear mapping can also be constructed to automatically learn the optimal weight distribution. A feature selection algorithm can also be introduced to automatically identify the frequency band with the largest amount of information and achieve self-adaptive data processing;

[0084] Setting reasonable frequency weight coefficients, filtering out the trend term and random noise, and retaining the pulse component can significantly improve the input effectiveness of the PPG signal in blood pressure monitoring and play an important role in improving the monitoring accuracy.

[0085] A3: According to the preset weight coefficients, using the weighted reconstruction method, linearly combine each frequency component to synthesize the preprocessed PPG signal. The specific steps of weighted reconstruction are as follows:

[0086] The band signals at each layer obtained after wavelet transform are perturbed according to the set weight coefficients. The high-weight bands remain unchanged, while the low-weight bands reduce their amplitudes. The band signals after weight modulation are subjected to simple linear superposition. The signal obtained by superposition is the preprocessed PPG signal. This signal suppresses the low-frequency trend and high-frequency random noise, and the intermediate-frequency blood-pressure-related components are retained and amplified. Nonlinear mapping can be introduced to replace linear superposition. Through neural networks or kernel methods, the complex nonlinear relationships of band combinations are learned to generate a better PPG preprocessing signal. After band filtering, resampling and interpolation can be used to reconstruct the waveform, and a more refined time-frequency method can be adopted to generate a smoother and more continuous preprocessing signal.

[0087] Using weighted reconstruction can not only suppress the ineffective components, but also strengthen the effective components, enhancing the prediction input effect of the PPG signal. Compared with the direct filtering method, this method can retain more original details and avoid over-smoothing.

[0088] Combined with wavelet transform, a feature selection method is introduced to automatically select the bands with the most blood-pressure information for reconstructing the PPG signal and optimizing the preprocessing effect. The specific steps are as follows:

[0089] Perform wavelet transform on the PPG signal to obtain multiple band component signals. Use correlation analysis, mutual information method, etc. to evaluate the correlation degree of each band with blood-pressure parameters. According to the correlation index, select the top N bands with the strongest relationship with blood pressure. These bands have the largest amount of information and the strongest prediction ability. Only retain these N bands and set the other bands to zero. Then perform signal reconstruction. Repeat the above steps to search for various different band combinations and find the optimal band set with the largest correlation. A correlation threshold can also be set to filter out all bands below the threshold to avoid introducing too many useless bands. Re-evaluate on new sample data and update the band selection to achieve adaptive optimization.

[0090] Compared with manually setting weights, the feature selection algorithm can search for the optimal bands more intelligently and automatically. At the same time, combined with the multi-resolution analysis of wavelet transform, band selections at multiple granularities can be obtained, extracting richer information.

[0091] Step 3: Feature extraction:

[0092] B1. Construct an LSTM neural network model: The LSTM network has a special memory unit structure and has the ability to learn long-sequence data, which can be used for feature learning of PPG pulse waveforms. The specific steps are as follows:

[0093] Build an LSTM network model. Set the number of nodes in the input layer to the number of sampling points of the PPG waveform. For example, if the sampling frequency is 100 Hz and 1000 sampling points are collected for a 10-second waveform, then the number of nodes in the input layer is set to 1000; set the number of nodes in the output layer to the dimension of the desired PPG feature vector, such as 10 dimensions; the hidden layer of the network is set according to the task complexity and can have 1 - 3 LSTM layers, with the number of nodes in each layer ranging from 100 to 300; initialize the network parameters, and then input the collected PPG waveform sequence for training; optimize the training network weight parameters through the BPTT algorithm. The network automatically learns to extract the effective features of the PPG waveform; after multiple rounds of training, the network output stabilizes at a 10-dimensional vector, which contains the global and local feature information of the PPG waveform; multi-task learning can also be set to simultaneously predict physiological parameters such as heart rate and blood pressure to assist in the learning of PPG features.

[0094] Compared with manually designed features, the LSTM network can autonomously learn data features, extract more implicit non-linear relationships, and can process inputs of any length, making it suitable for waveform sequences.

[0095] B2. Input the preprocessed PPG waveform samples into the LSTM model, configure random learning parameters, and start training. Use the gradient descent algorithm to optimize the network weights. The specific steps are as follows:

[0096] Network weight parameters, such as the weight matrix of the input layer and the weights of each gate, are initialized to small random numbers to avoid symmetry disappearance; determine the network optimization goal, and commonly use mean square error, cross-entropy, etc. as the loss function; use the gradient descent method to gradually optimize the network to minimize the loss function; in each training batch, calculate the network output forward and calculate the loss with the ground truth; efficiently calculate the gradient contribution value of each parameter to the loss through the BPTT algorithm; update the network weights in the gradient direction with a preset learning rate to achieve parameter optimization; repeat multiple training batches, and gradually improve the network performance through continuous gradient descent; optimization algorithms such as stochastic gradient descent, RMSprop, and Adam can be used to adjust the learning rate to accelerate training; when the loss function converges, the training is completed, and at this time the network has learned the optimal parameters.

[0097] Through training with the gradient descent algorithm, the LSTM network automatically learns the inherent temporal feature dependence of the data and realizes end-to-end PPG feature extraction without manually designing feature engineering.

[0098] B3. Repeat multiple rounds of iterative training. The LSTM model automatically learns the time-domain and frequency-domain features of the PPG waveform, with the convergence of the loss function as the termination condition.

[0099] B4. The network outputs a 10-dimensional feature vector, which is the feature expression of the PPG waveform, reflecting the overall shape of the waveform, the interval statistical characteristics, and the frequency band component information. This low-dimensional vector is the effective feature of the PPG signal.

[0100] Step 4. Establish a model:

[0101] The classical RNN consists of a series of computational modules, but they are sometimes limited by the effectiveness of the training process. Without using LSTM, when there is a long time lag between the relevant event and the target event in the backpropagation error, it is very easy to explode within a reasonable time, resulting in the RNN being unable to learn the model. Specifically, time lag tasks greater than 5 time steps have become difficult to handle within a reasonable time. The reason for using the LSTM network is to overcome this problem by forcing the error that does not decay to flow back in time. Through the constant error loop in the LSTM storage unit, time lags of more than thousands of time steps can be processed. Given the input sequence X=(x1,...xT), the cell vector sequence C=(c1,...cT), the hidden vector sequence H=(h1,...hT), and the output vector sequence Y=(y1,...yT) of the LSTM network, they are calculated by iterating from t = 1 to T according to the following equations:

[0102] h t = H(x t , c t-1 , h t-1 )

[0103] y t = σ(W hy h t + b y )

[0104] where σ is the sigmoid function, the W term represents the weight matrix, the b term represents the bias vector, and H is the hidden layer operator.

[0105] The LSTM storage unit H used in the present invention is implemented as the following composite function:

[0106] i t = σ(W xi x t + W hi h t-1 + W ci c t-1 + b i )

[0107] f t = σ(W xf x t + W hf h t-1 + W cf c t-1+b f )

[0108] c t =f t c t-1 +i t tanh(W xc x t +W hc h t-1 +b c )

[0109] o t =σ(W xo x t +W ho h t-1 +W co c t +b o )

[0110] h t =o t tanh(c t )

[0111] C1. After obtaining the feature vector of the PPG waveform, construct an LSTM network for blood pressure parameter prediction. The specific steps are as follows:

[0112] The input layer of the network is the PPG feature vector, such as 10-dimensional; the number of nodes in the first LSTM hidden layer is set to 8, which is used to learn low-level feature combinations; the number of nodes in the second LSTM hidden layer is set to 16, which learns high-level feature abstractions; more hidden layers can be added to hierarchically learn different levels of information in combination with features; the number of nodes in the output layer is the dimension of the blood pressure parameters, such as 2-dimensional, which respectively predict systolic blood pressure and diastolic blood pressure; collect PPG training samples with blood pressure annotations as supervised data; train the network parameters to minimize the mean square error of blood pressure prediction; through multiple rounds of training, the LSTM network learns the non-linear mapping relationship between PPG features and blood pressure parameters; obtain an end-to-end prediction model, input the PPG feature vector, and the blood pressure value can be directly output.

[0113] The multi-layer structure of the LSTM network can gradually learn the deep feature representation of the data. Compared with linear models, it can model the complex dynamics of physiological systems; compared with other networks such as CNN, LSTM is more suitable for processing time series data and learning long-distance dependence relationships.

[0114] C2. Collect PPG waveform samples with reference blood pressure parameters as training data, and use the reference blood pressure parameters as supervised signals to train the weights of the LSTM network.

[0115] C3. After constructing the LSTM network and initializing the parameters, it is necessary to repeat the training for multiple rounds of iteration to gradually optimize the model. The specific steps are as follows:

[0116] Configure the training set batch size and input a batch of PPG-blood pressure sample pairs; perform forward calculation of the network's blood pressure prediction output; calculate the error between the prediction output and the true blood pressure parameters, such as mean squared loss; backpropagate the error through BPTT to calculate the gradient of each network parameter; update the network weights according to the gradient, such as SGD momentum optimization; repeat inputting new sample batches and cycle through multiple rounds of training; through continuous gradient descent optimization, the loss function gradually converges and the prediction accuracy improves; finally, the network learns the complex non-linear relationship between the input PPG features and the blood pressure parameters; for new PPG samples, they can be directly input into the network for end-to-end blood pressure prediction, or the generalization performance of the model can be optimized by methods such as adjusting parameters and batch normalization.

[0117] Multiple rounds of training gradually approximate the network parameters to the global optimum and establish an accurate non-linear regression mapping.

[0118] Step Five: Individual Adjustment:

[0119] D1. The first step in personalized model adjustment is to collect the PPG waveform sample data of this specific individual and the corresponding true blood pressure parameters as new training samples. This requires the subject to cooperate in collecting PPG signals for a certain period of time and measuring blood pressure values with accurate equipment. To ensure the representativeness of the sample data, various common physiological states of the individual should be covered, such as quiet state, after activity, etc.

[0120] D2. After collecting enough samples, input these new individual PPG waveform data into the pre-trained general blood pressure prediction model. The model will output the predicted blood pressure values, and then calculate the error between the predicted blood pressure and the true blood pressure. Metrics such as mean squared error and mean absolute error can be used for quantification. These errors constitute a new loss function, indicating the fitting effect of the current general model on this individual.

[0121] D3. Based on the newly added loss function, the gradient contribution of each network parameter to the loss function can be efficiently calculated through the backpropagation algorithm. Then, adjust the model parameters accordingly based on this gradient information, such as fine-tuning the bias vector of the network layer and modifying the parameters of the activation function, etc., to improve the fitting degree for this individual's samples. Note that to prevent overfitting problems, in the initial stage of network adjustment, a small batch of samples of the individual needs to be used for training and updating, and in the later iterations, new samples are adopted to continuously optimize the model to adapt to the dynamic changes of the individual parameters.

[0122] D4. By repeatedly performing gradient-guided parameter adjustment in multiple rounds, the prediction error of the personalized model will gradually decrease and finally converge to an acceptable range, indicating that a personalized model capable of providing high-precision blood pressure prediction for this specific individual has been obtained. If a new individual user is encountered, only by repeating the parameter adjustment and model fine-tuning process can a blood pressure health monitoring model library with strong personalization effects for different individuals be gradually constructed.

[0123] Compared with directly applying a general model, this parameter adjustment method based on individual differences can significantly improve the long-term use effect and accuracy of the PPG blood pressure measurement technology on specific individuals. It enables the monitoring system to better adapt to individual physiological variations and has important practical significance for promoting the improvement of personalized medical levels.

[0124] Step Six: Optimize the model:

[0125] In order to further improve the generalization prediction ability of the PPG-based blood pressure detection method, the present invention adopts a strategy of large-scale multi-source heterogeneous data collection and model optimization using transfer learning technology and incremental learning technology.

[0126] Specifically, first, multi-center large-sample PPG blood pressure data collection is organized and carried out to cover different population distributions as much as possible, such as subjects of different age groups, patients with various diseases, and full-range blood pressure value coverage. The fusion of a large amount of multi-source heterogeneous data can significantly enrich the sample feature space and is conducive to improving the generalization performance of the model.

[0127] After collecting a large-scale labeled sample, a unified LSTM network model is constructed, and transfer learning technology in a joint training mode is used to optimize the model. This technology enables the model to learn common knowledge across datasets by simultaneously inputting different source datasets, so as to have a certain adaptability to different population distributions. Compared with independent training, joint training can significantly improve the generalization ability of the model for new examples.

[0128] In addition, transfer learning technology of model cloning is also adopted. This technology generates multiple new models based on a source model through methods such as partial parameter replication. The new models not only inherit the generalization knowledge of the source model but also adapt to their respective sample distributions. Finally, an integrated model is formed through model fusion, which has stronger generalization performance.

[0129] During the model optimization process, incremental learning technology is adopted. First, only samples of the elderly are used for training, and then other samples are gradually added for iterative optimization. This incremental learning method that prioritizes the elderly population can avoid the decline in the generalization ability of the model for elderly samples and make the resulting model more applicable to patients with hypertension and other diseases.

[0130] In summary, the application of technologies such as big data and transfer learning can continuously improve the generalization performance of the PPG blood pressure monitoring system, enabling it to adapt to a wide range of individual differences and achieve high-precision personalized monitoring.

[0131] Finally, the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A PPG-based blood pressure detection method, characterized in that, the specific steps are as follows: Step 1, PPG signal acquisition: Collect the original PPG waveform through a PPG sensor; Step 2, PPG signal preprocessing: First, perform wavelet transform on the collected original PPG waveform signal, decompose it into multiple frequency layers, then introduce an attention mechanism between different frequency layers, assign different weight coefficients to different layers, and finally reconstruct each frequency band to obtain the preprocessed PPG signal; Step 3, feature extraction: Use the LSTM neural network model to train the preprocessed PPG signal, automatically learn its features in the time domain and frequency domain, and obtain a PPG feature set describing blood pressure changes; Step 4, model establishment: Use the extracted PPG features as input and the reference blood pressure parameters as output to construct an LSTM network model, train the network weights, establish a mapping relationship between features and blood pressure, and input the PPG feature vector to directly output the blood pressure value; Step 5, individual adjustment: After obtaining the overall model, when introducing the PPG data of a new individual, first use this model for prediction, and then make fine adjustments to the model with blood pressure data to obtain a personalized model; Step 6: Model optimization: Over time, collect large-scale multi-source heterogeneous data samples, and then use transfer learning technology and incremental learning technology to update the model.

2. A PPG-based blood pressure detection method according to claim 1, characterized in that: the PPG sensor in Step 1 is wearable and designed in the form of a bracelet or finger sleeve, and the sensor contains an infrared light-emitting diode and a light-receiving diode; The specific steps of PPG signal acquisition are: The infrared rays emitted by the infrared light-emitting diode of the sensor penetrate the skin tissue, and the light-receiving diode detects the change in the intensity of the infrared rays transmitted through the skin and converts it into an electrical signal to obtain the original PPG waveform reflecting the change in blood volume, and collect the original waveform according to the Nyquist sampling theorem.

3. A PPG-based blood pressure detection method according to claim 1, characterized in that: The specific steps of PPG signal preprocessing in Step 2 are as follows: A1: On the obtained original PPG waveform signal, first perform wavelet transform, and then use the Mallat fast algorithm to take 8-layer db4 wavelet transform, and after transformation, it is decomposed into a low-frequency component cA8 and high-frequency components cD1-8; A2: Introduce an attention mechanism to set different weight coefficients for different layers; A3: According to the preset weight coefficients, use the weighted reconstruction method to linearly combine each frequency component to synthesize the preprocessed PPG signal.

4. A PPG-based blood pressure detection method according to claim 1, characterized in that: The specific steps of feature extraction in Step 3 are as follows: B1. Construct an LSTM neural network model, set the number of nodes in the input layer to the number of sampling points of the PPG waveform, set the number of nodes in the output layer to the dimension of the desired PPG feature vector, and set the network hidden layer according to the task complexity; Initialize the network parameters, then input the collected PPG waveform sequence for training. Optimize the training network weight parameters through the BPTT algorithm. The network automatically learns to extract the effective features of the PPG waveform. After multiple rounds of training, the network output stabilizes at a 10-dimensional vector, which contains the global and local feature information of the PPG waveform; B2. Input the preprocessed PPG waveform samples into the LSTM neural network model, configure random learning parameters, start training, and use the gradient descent algorithm to optimize the network weights; B3. Repeat multiple rounds of iterative training. The LSTM neural network model automatically learns the time-domain and frequency-domain features of the PPG waveform, with the convergence of the loss function as the termination condition; B4. The network outputs a 10-dimensional feature vector, which is the feature expression of the PPG waveform, reflecting the overall shape of the waveform, the interval statistical characteristics, and the frequency band component information. This low-dimensional vector is the effective feature of the PPG signal.

5. A PPG-based blood pressure detection method according to claim 1, characterized in that: The specific steps for model establishment in step four are as follows: C1. After obtaining the feature vector of the PPG waveform, the network input layer is the PPG feature vector. The number of nodes in the first LSTM hidden layer is set to 8 to learn low-level feature combinations. The number of nodes in the second LSTM hidden layer is set to 16 to learn high-level feature abstractions. The number of nodes in the output layer is the dimension of the blood pressure parameter. Collect the PPG training samples with blood pressure annotations as the supervised data, train the network parameters, minimize the mean square error of blood pressure prediction. Through multiple rounds of training, the LSTM network learns the non-linear mapping relationship between the PPG features and the blood pressure parameters, and obtains an end-to-end prediction model. Input the PPG feature vector, and the blood pressure value can be directly output; C2. Collect the PPG waveform samples with reference blood pressure parameters as the training data, and use the reference blood pressure parameters as the supervised signal to train the LSTM network weights; C3. After constructing the LSTM network and initializing the parameters, it is necessary to repeat multiple rounds of iterative training to gradually optimize the model.

6. A PPG-based blood pressure detection method according to claim 1, characterized in that: The specific steps for individual adjustment in step five are as follows: D1. Collect the PPG waveform sample data of a specific individual and the corresponding real blood pressure parameters as new training samples; D2. After collecting enough samples, input these new individual PPG waveform data into the pre-trained blood pressure prediction model. The model will output the predicted blood pressure value, and then calculate the error between the predicted blood pressure and the real blood pressure. These errors constitute a new loss function, indicating the fitting effect of the current general model on this individual; D3. Based on the newly added loss function, calculate the gradient contribution of each network parameter to the loss function through the backpropagation algorithm, and adjust the model parameters accordingly according to these gradient information; D4. By repeatedly performing gradient-guided parameter adjustment in multiple rounds, the prediction error of the personalized model will gradually decrease and finally converge to an acceptable range, indicating that a personalized model capable of providing high-precision blood pressure prediction for this specific individual has been obtained. If a new individual user is encountered, only the parameter adjustment and model fine-tuning process need to be repeated, thereby gradually constructing a blood pressure health monitoring model library with strong personalization effects for different individuals.

7. A PPG-based blood pressure detection method according to claim 1, characterized in that: in step six, the transfer learning technology includes transfer learning technology in a joint training mode and transfer learning technology of model cloning; The transfer learning technology in the joint training mode enables the model to learn common knowledge across datasets during the process of processing each dataset by simultaneously inputting different source datasets, thereby having a certain adaptability to different population distributions and greatly improving the generalization ability of the model to new examples; The transfer learning technology of model cloning generates multiple new models based on a source model through a partial parameter replication method. The new models not only inherit the generalization knowledge of the source model but also adapt to their respective sample distributions. Finally, an integrated model is formed through model fusion, which has stronger generalization performance; The incremental learning technology is to first train only using samples of the elderly, and then gradually add other samples for iterative optimization to avoid the decline of the generalization ability of the model to elderly samples and make the resulting model more suitable for hypertensive patients.