Blood pressure measurement system and method based on oscillatory wave dynamic characteristics and regression modeling

Abstract

The invention provides a blood pressure measurement system and method based on oscillatory wave dynamic characteristics and regression modeling, and relates to the technical field of wearable medical health monitoring. A signal processing method serves as a main body, oscillatory wave key features are extracted through time domain, frequency domain, time frequency and a statistical method, traditional features are finely adjusted through a lightweight deep learning model, and finally the mapping relation between feature wave band changes and blood pressure values is established through machine learning. Not only can the complete oscillation wave waveform be predicted through the characteristic change of part of oscillation waves to realize low-pressure comfortable measurement, but also personalized blood pressure measurement can be realized by identifying the characteristic mutation point of the oscillation waves. According to the method, the interpretability and the stability of a traditional method are guaranteed, the feature recognition precision and the individual adaptive capacity are improved through lightweight deep learning fine adjustment, edge calculation can be achieved, and therefore the measurement time is shortened, and the user comfort level, the measurement accuracy and the system practicability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wearable medical health monitoring, and particularly to a blood pressure measurement system and method based on the dynamic characteristics of oscillation waves and regression modeling. Background Art

[0002] In the field of non-invasive blood pressure measurement, the traditional oscillometric method calculates blood pressure values with the help of a single airbag cuff and empirical formulas, but there are many drawbacks. The human blood circulation is complex, and hemodynamic fluctuations are easily induced during measurement, affecting the morphology of oscillation waves and resulting in poor measurement accuracy; it has poor adaptability to individual physiological differences and dynamic blood pressure fluctuations, and the measurement accuracy drops significantly in the face of abnormal pulse waves or motion interference; moreover, the hardware cost is relatively high. Although multi-airbag or multi-sensor solutions can improve accuracy, they increase the system complexity and manufacturing cost, hindering the large-scale popularization and application of the technology. These problems have become the main obstacles to the development of current non-invasive blood pressure measurement technology.

[0003] Chinese Patent CN116509354A constructs SBP and DBP prediction models using machine learning algorithms. Although it enhances the pertinence of blood pressure calculation, it relies on a large amount of labeled training data, the feature extraction and dimensionality reduction operations are complex, the computational cost is high, it is difficult to run in real time on resource-constrained wearable devices, and the model generalization ability is limited, with poor prediction accuracy for patients with hypertension or hypotension. Chinese Patent CN116570259A uses the discrete spectrum decomposition of a single-period pulse wave by the Schrödinger operator, extracts the discrete spectrum features of the reconstructed signal and combines a machine learning model to predict blood pressure, which has a certain robustness under low signal-to-noise ratio conditions. However, it is highly dependent on the pulse wave morphology, and non-typical pulse wave signals will affect the accuracy of feature extraction, resulting in an increase in blood pressure prediction error, and it has high computational resource requirements and is not suitable for low-power wearable devices. Chinese Patent CN119453960A expands the dimension of PPG signals and combines a lightweight neural network model to improve blood pressure prediction accuracy, but it relies on synchronously collected aortic pressure data to train the model, has insufficient robustness to motion artifacts, and needs to be frequently calibrated due to individual differences and dynamic blood pressure fluctuations in practical applications, with limited universality and practicality. The deficiencies of these patented technologies together reflect that the existing blood pressure measurement technologies need to be improved in terms of continuity, comfort, real-time performance, and personalized adaptability.

[0004] Generally speaking, the existing blood pressure measurement technologies have limitations in terms of continuity, comfort, real-time performance, and personalized adaptability. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a blood pressure measurement system and method based on the dynamic characteristics of oscillation waves and regression modeling to achieve accurate estimation of blood pressure in view of the above-mentioned deficiencies of the prior art.

[0006] To solve the above technical problem, the technical solution adopted by the present invention is: On the one hand, the present invention provides a blood pressure measurement system based on the dynamic characteristics of oscillatory waves and regression modeling, including an oscillatory wave database, a pressure control module, a data quality assessment module, a data preprocessing module, a feature extraction module, a deep learning fine-tuning module, and a blood pressure estimation module; The database stores oscillatory waves of at least 3000 subjects and their corresponding blood pressure results, where individual subjects need to have multiple blood pressure measurement data in a quiet state; The pressure control module controls the inflation and deflation of the cuff through a micro solenoid valve to ensure that the pressure of the cuff can be adjusted in real time as needed; the pressure control module includes an MCU, a cuff pressure sensor, a cuff pressure AC isolation and filtering circuit, and a cuff pressure DC isolation and filtering circuit; the cuff pressure sensor continuously collects the cuff pressure data on the subject's arm in real time and uses this as the basis for blood pressure estimation; the cuff pressure AC isolation and filtering circuit converts the AC signal of the cuff pressure into a DC signal and provides it to the MCU for subsequent negative feedback control and adjustment; the cuff pressure DC isolation and filtering circuit is responsible for isolating the DC signal of the cuff pressure and obtaining the AC signal of the cuff pressure to collect the real-time value of the cuff pressure oscillatory wave; The data quality assessment module is used to screen out signals with poor quality assessment and remove them; The data preprocessing module is used to filter the obtained oscillatory waves to remove common noises; The feature extraction module extracts time-domain, frequency-domain, time-frequency, and statistical features from the preprocessed oscillatory wave signals using traditional signal processing methods and divides the training set and the test set; The deep learning fine-tuning module uses a lightweight deep learning fine-tuning model to fine-tune the extracted features to achieve personalized blood pressure prediction; The blood pressure estimation module uses a machine learning model to match the signals in the database according to the fine-tuned features input in real time, predicts the complete oscillatory wave waveform, and simultaneously identifies the positions of the two single-cycle oscillatory waves with the largest changes in the characteristic bands of the systolic blood pressure, diastolic blood pressure, and mean arterial pressure oscillatory waves, and maps the changes in the oscillatory wave characteristic bands to blood pressure values.

[0007] Further, the deep learning fine-tuning module first performs normalization preprocessing on the time-domain, frequency-domain, time-frequency, and statistical features extracted by the feature extraction module, and then inputs them into the lightweight deep learning fine-tuning model; the lightweight deep learning fine-tuning model is a fully connected network, including an input layer, one or two hidden layers, a Dropout layer, and a fully connected output layer; an appropriate number of neurons are set in the hidden layer, and the ReLU activation function is used to capture the non-linear relationship between features; the Dropout layer is added between the hidden layers to prevent overfitting; the fully connected output layer uses a linear activation function to output the fine-tuned feature representation; The network of the lightweight deep learning fine-tuning model is trained using the training set data. Through forward propagation, loss calculation, backpropagation, and parameter update, the network automatically learns how to optimize the extracted features. The fine-tuned model is evaluated using the test set to detect the generalization performance of the model and perform hyperparameter tuning as needed.

[0008] Further, the blood pressure estimation module first loads the fine-tuned feature vectors and the corresponding blood pressure data of the lightweight deep learning fine-tuning model, divides the data into a training set and a test set to ensure that there are sufficient samples for each blood pressure level. The support vector machine, random forest, and extreme gradient boosting models are trained using the training set respectively, and their hyperparameters are tuned. The performance of each model on the training set is evaluated using the cross-validation method, and the best model parameter combination is selected. The trained model is applied to the test set, and the model is evaluated using metrics such as mean squared error, root mean squared error, or mean absolute error. The prediction accuracies of the models on the test set are compared, and the best-performing model is selected or the model ensemble method is used to further improve the prediction stability. The selected model is used for actual prediction. The fine-tuned features are input, and the complete oscillatory wave waveform prediction is output. At the same time, the mutation positions of the oscillatory wave features of the identified systolic blood pressure, diastolic blood pressure, and mean arterial pressure and the corresponding cuff pressures are output to achieve non-invasive blood pressure measurement. The prediction results are corrected according to the need by combining personalized calibration data.

[0009] On the other hand, the present invention also provides a blood pressure measurement method based on oscillatory wave dynamic features and regression modeling, which is implemented by the above-mentioned blood pressure measurement system based on oscillatory wave dynamic features and regression modeling, and includes the following steps: Step 1: Construct an oscillatory wave signal database, including complete oscillatory wave signals and their corresponding oscillatory wave feature bands of systolic blood pressure, diastolic blood pressure, and mean arterial pressure. The inflation and deflation of the cuff are controlled by a micro solenoid valve to ensure that the pressure of the cuff can be adjusted in real time as needed. Step 2: Evaluate the quality of the oscillatory wave signals in the database, including waveform smoothness and stationarity evaluation based on the time domain and spectrum consistency evaluation based on the frequency domain. Step 3: Preprocess the oscillatory wave signals that pass the quality evaluation, including low-pass filtering, high-pass filtering, and empirical wavelet transform, to optimize the noise and baseline drift in the oscillatory wave signals. Step 4: Perform feature band identification. Traditional signal processing methods are used to extract time domain, frequency domain, time-frequency, and statistical features from the preprocessed oscillatory wave signals, and the training set and test set are divided. Step 5: Use the lightweight deep learning fine-tuning model to fine-tune the extracted traditional features. Step 6: Train the machine learning model to automatically learn the mapping relationship between diastolic blood pressure and mean arterial pressure and the oscillometric wave signal, output the complete oscillometric waveform, and at the same time learn the change of the oscillometric characteristic bands of systolic blood pressure, diastolic blood pressure and mean arterial pressure, map the change of the characteristic bands to blood pressure values, and realize non-invasive blood pressure measurement.

[0010] Further, the specific method of step 1 is as follows: Step 1.1: Correctly place the oscillometric wave signal device on the upper arm of the subject and fix it with a cuff. Step 1.2: Gradually pressurize the cuff, record the oscillometric wave data of the subject at different pressures, and store it. Step 1.3: Measure the systolic blood pressure and diastolic blood pressure of the subject using the built-in blood pressure measurement device. Step 1.4: The cuff pressure sensor is responsible for continuously collecting the cuff pressure data on the user's arm in real time. Step 1.5: Obtain the DC and AC signals of the cuff pressure. Step 1.5.1: The cuff pressure AC isolation and filtering circuit converts the AC signal of the cuff pressure into a DC signal, which is provided to the MCU for real-time negative feedback control and regulation; the negative feedback control and regulation signal generated by the MCU is transmitted to the micro solenoid valve; the micro solenoid valve controls the inflation and deflation of the cuff to ensure that the pressure of the cuff is adjusted in real time as needed. Step 1.5.2: The cuff pressure DC isolation and filtering circuit is responsible for isolating the DC signal of the cuff pressure and obtaining the AC signal of the cuff pressure for collecting the real-time value of the cuff pressure oscillometric wave. Step 1.6: Output the processed oscillometric wave signal and cuff pressure signal. Step 1.7: Fit by identifying the waveform characteristics of diastolic blood pressure and mean arterial pressure in the oscillometric waveform. If the fitting is good, stop inflation. If the fitting is not good, continue to pressurize until a normal blood pressure measurement is completed, and upload the obtained complete oscillometric waveform to the database.

[0011] Further, the specific method of step 2 is as follows: Step 2.1: Based on the evaluation of waveform smoothness in the time domain, detect spikes or noises in the waveform to ensure smooth transition of the waveform on the time axis. Step 2.2: Based on the evaluation of spectrum consistency in the frequency domain, ensure that the obtained partial waveform is consistent with the overall waveform, and ensure that the frequency distribution of the fitted waveform is similar to that of the actual waveform.

[0012] Further, the specific method of step 3 is as follows: Step 3.1: Use a low-pass filter to remove the high-frequency noise in the original oscillometric wave signal, so as to retain the low-frequency components in the oscillometric wave signal. Step 3.2: Use a high-pass filter to remove the influence of baseline drift on the oscillatory wave, thereby retaining the high-frequency components in the oscillatory wave signal; Step 3.3: Construct a band-pass filter based on empirical wavelet transform to process each cycle of the oscillatory wave to eliminate artifacts caused by various factors, as shown in the following formula: (1); where, a and b are the scale and translation parameters, ψ is the mother wavelet function, f(t) is the original signal, which is a function of t.

[0013] Furthermore, the specific method of step 4 is as follows: Step 4.1: Perform data preparation, load the oscillatory wave signal and its corresponding characteristic bands of systolic blood pressure, diastolic blood pressure, and mean arterial pressure, and divide the training set and the test set; Step 4.2: Use traditional signal processing methods to extract features from the preprocessed oscillatory wave signal. The specific methods include: Step 4.2.1: Time-domain feature extraction; Calculate the maximum value, minimum value, peak value, number of peaks / troughs, and amplitude of the second peak of the signal; Calculate the root mean square value and average power of the signal; the root mean square value RMS of the signal is: (5); where, T is the time length of the signal, and x(t) is the time-domain representation of the signal, that is, the oscillatory wave signal changing with time; The average power P of the signal avg is: (6); Extract waveform slope, pulse width, and waveform asymmetry index; Waveform slope: Define a single-cycle signal as x(n), n is the discrete time point, n0 is the starting point of the single-cycle signal, n peak is the peak of the single-cycle waveform, n D is at the dicrotic notch, n SecondPeak is at the second peak, n valley is the trough of the single-cycle waveform, that is, the end point of the single-cycle waveform. Calculate the slope of the line connecting adjacent two points of the above characteristic points as the waveform slope; Pulse width: The pulse width refers to the duration of the signal above a certain threshold. Set the threshold as V th , and the time interval from the moment t start when the rising edge of the signal exceeds the threshold to the moment t end when the falling edge is lower than the threshold is the pulse width PW, as shown in the following formula: PW = t end -t start (7); Waveform asymmetry index: The waveform asymmetry is measured by calculating the energy ratio of the signal on the positive and negative half-axes; the energy E of the signal on the positive half-axis + is: (8); The energy E of the signal on the negative half-axis - is: (9); The waveform asymmetry index AI is: (10); Step 4.2.2: Frequency-domain feature extraction; Convert the time-domain signal to a frequency-domain signal using the Fourier transform; Extract the main frequency, power spectral density, and spectral entropy index of the signal; The specific calculation method is as follows: Fourier transform: Convert the time-domain signal x(t) to the frequency-domain signal X(f), where f is the frequency, as shown in the following equation: (11); Main frequency of the signal: In the frequency-domain signal X(f), the frequency corresponding to the frequency component with the largest amplitude is the main frequency, as shown in the following equation: (12); The power spectral density is: (13); where, X T (f) is the Fourier transform of the signal x(t) in the time interval , and PSD(f) represents the power spectral density; The spectral entropy is: (14); where, , sum the squares of the power amplitudes of all frequency points in the signal, N is the number of frequency points of the frequency-domain signal, X(f i ) is the value of the frequency-domain signal at the frequency f i , and H represents the spectral entropy.

[0014] Step 4.2.3: Time-frequency feature extraction; Use wavelet transform to extract the features of the time-domain signal at different scales; decompose the time-domain signal using the empirical mode decomposition method to extract the time-frequency features of each intrinsic mode component; Empirical mode decomposition decomposes the signal x(t) into a series of intrinsic mode functions IMF m (t). Assuming that M intrinsic mode functions are obtained by decomposition, then (15); For each intrinsic mode function IMF m (t), extract its time-frequency characteristics; Step 4.2.4: Statistical feature extraction; Calculate the mean, standard deviation, skewness and kurtosis statistics of the signal, as follows respectively: Mean μ : (16); where x i is the discrete sample value of the signal, and N is the number of samples; Standard deviation σ : (17); Skewness: (18); Kurtosis: (19).

[0015] Furthermore, the specific method of step 5 is as follows: Step 5.1: Perform normalization preprocessing on the time domain, frequency domain, time-frequency and statistical feature vectors obtained by traditional signal processing methods in step 4; Step 5.2: Input the normalized traditional feature vectors into the lightweight deep learning fine-tuning model, specifically as follows: Step 5.2.1: Construct a small fully connected network and use the normalized feature vectors as the input; Hidden layer: Construct one or two fully connected hidden layers, set an appropriate number of neurons in each layer, and use the ReLU activation function to capture the non-linear relationship between features; Step 5.2.3: Set a Dropout layer between the hidden layers to prevent overfitting; Step 5.2.4: Finally, construct a fully connected output layer and use a linear activation function to fine-tune the features and generate the final regression output; Step 5.3: Train the fine-tuning model on the training set data, and perform regression optimization using the features extracted by traditional methods and the changes in the feature bands of systolic blood pressure, diastolic blood pressure and mean arterial pressure; Step 5.4: Use the test set to evaluate the fine-tuned model, detect the generalization performance of the model, and perform hyperparameter tuning as needed.

[0016] Further, the specific method of step 6 is as follows: Step 6.1: Data preparation; Load the changes in the feature vectors and corresponding feature bands after fine-tuning the lightweight deep learning fine-tuning model in step 5, and divide them into a training set and a test set to ensure that there are sufficient samples for each blood pressure level; Step 6.2: Model selection and feature optimization; Select a set of machine learning algorithms, including support vector machine, random forest, and extreme gradient boosting algorithm; Step 6.3: Model training; Step 6.3.1: Use the training set to train the support vector machine, random forest, and extreme gradient boosting algorithm models respectively, and tune their respective hyperparameters; Step 6.3.2: Use the cross-validation method to evaluate the performance of the model on the training set and select the best model parameter combination; Step 6.4: Model evaluation and selection; Step 6.4.1: Apply the trained model to the test set, and use the mean squared error, root mean squared error, or mean absolute error as indicators to evaluate the machine learning model; Step 6.4.2: Compare the prediction accuracies of the machine learning models on the test set, select the machine learning model with the best performance, or use the model integration method to further improve the prediction stability; Step 6.5: Model integration and output; Step 6.5.1: Use the selected machine learning model for actual prediction, input the fine-tuned traditional features, and output the complete oscillatory wave waveform and the feature band mutation points of the identified systolic blood pressure, diastolic blood pressure, and mean arterial pressure; Step 6.5.2: At the same time, output the cuff pressure corresponding to the feature band mutation point of the oscillatory wave to achieve non-invasive blood pressure measurement; Step 6.5.3: According to needs, correct the measurement results by combining personalized calibration data to further improve the measurement accuracy.

[0017] The beneficial effects of adopting the above technical solutions are as follows: The blood pressure measurement system and method based on the dynamic characteristics of oscillatory waves and regression modeling provided by the present invention can effectively extract the characteristics of oscillatory waves and capture the temporal dependence of signals by combining traditional signal processing methods and machine learning techniques, predict the complete waveform of oscillatory waves, and at the same time establish the mapping relationship with blood pressure values by accurately identifying the morphological mutation points of the characteristic bands of oscillatory waves. This hybrid model not only retains the interpretability and computational efficiency of traditional methods but also improves the feature expression ability through deep learning fine-tuning, thus significantly improving the accuracy of blood pressure measurement. This technical solution is particularly suitable for home monitoring and wearable devices, reducing the discomfort and measurement time of subjects when repeatedly measuring blood pressure, and providing a reliable tool for the long-term management of hypertensive patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a structural block diagram of a blood pressure measurement system based on the dynamic characteristics of oscillatory waves and regression modeling provided by an embodiment of the present invention; Figure 2 It is an implementation block diagram of a pressure control module provided by an embodiment of the present invention; Figure 3 It is an implementation block diagram of a data quality assessment module provided by an embodiment of the present invention; Figure 4 It is an implementation block diagram of a data preprocessing module provided by an embodiment of the present invention; Figure 5 It is a structural block diagram of a deep learning fine-tuning module provided by an embodiment of the present invention; Figure 6 It is a structural block diagram of a lightweight fully connected fine-tuning network provided by an embodiment of the present invention; Figure 7 It is a structural block diagram of a blood pressure estimation module provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The following further describes in detail the specific embodiments of the present invention with reference to the drawings and embodiments. The following embodiments are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0020] In this embodiment, a blood pressure measurement system based on the dynamic characteristics of oscillatory waves and regression modeling, as Figure 1 shown, includes an oscillatory wave database, a pressure control module, a data quality assessment module, a data preprocessing module, a feature extraction module, a deep learning fine-tuning module, and a blood pressure estimation module.

[0021] The database stores the oscillatory waves and corresponding blood pressure results of at least 3,000 subjects, where individual subjects need to have multiple blood pressure measurement data in a quiet state. To construct the oscillatory wave signal database, first, place the oscillatory wave signal device correctly on the upper arm of the subject and fix it with a cuff to ensure accuracy and comfort; gradually pressurize the cuff, record the oscillatory wave data of the subject at different pressures, and store it. At the same time, use the built-in blood pressure measurement device to measure the systolic blood pressure and diastolic blood pressure of the subject.

[0022] The pressure control module controls the inflation and deflation of the cuff through a micro solenoid valve to ensure that the pressure of the cuff can be adjusted in real time as needed. As Figure 2 shown, continuously collect the cuff pressure data on the subject's arm in real time through the cuff pressure sensor and use this as the basis for blood pressure estimation. Convert the AC signal of the cuff pressure into a DC signal through the cuff pressure AC isolation and filtering circuit and provide it to the MCU for subsequent negative feedback control and adjustment; the cuff pressure DC isolation and filtering circuit is responsible for isolating the DC signal of the cuff pressure and obtaining the AC signal of the cuff pressure to collect the real-time value of the oscillatory wave of the cuff pressure. Based on the obtained real-time oscillatory wave, extract features through the feature extraction module and fit the complete oscillatory wave through the blood pressure estimation module, and evaluate the fitting curve. If the fitting is good, stop inflating and pressurizing, and use the fitting curve as the basis for the final measurement and blood pressure estimation; if the fitting is poor, continue to pressurize until a normal blood pressure measurement is completed, and upload the obtained complete oscillatory wave waveform and blood pressure measurement value to the database.

[0023] The data quality evaluation module is used to screen out signals with poor quality evaluation and remove them. As Figure 3As shown, first is the evaluation of waveform smoothness based on the time domain. First-order differentiation is performed on the obtained partial waveform data to calculate the change rate of the waveform. If the change rate fluctuates within a relatively small range, it proves that the waveform is relatively smooth. If the change rate is too large, it indicates that there are uneven spikes in the waveform. Second-order difference operation is performed on the waveform data to further detect sharp fluctuations in the waveform, and the change of the change rate is observed. If there are large fluctuations, it means that there are unreasonable spikes in the waveform. The sliding average method is used to remove noise, and it is detected whether there are still spikes after denoising. If the fluctuations of the waveform decrease and remain smooth through the above detections, the waveform can be used. Second is the evaluation of spectrum consistency based on the frequency domain. Fourier transform is performed on the obtained waveform to convert it from the time domain to the frequency domain to obtain its frequency distribution. The spectrogram is analyzed to find the main frequency of the waveform and its harmonic distribution, and at the same time, it is observed whether there are significant high-order harmonics to judge the periodicity and regularity of the waveform. The obtained spectrum is compared with the spectrum of the complete waveform in the database to check whether the amplitudes and distributions of the main frequency and harmonics are consistent, so as to judge whether the waveform can be fitted with high quality. Finally, the power spectral density is calculated to analyze the energy distribution at different frequencies of the waveform, and visually display whether there are abnormal frequency components or noise in the waveform.

[0024] The data preprocessing module is used to filter the obtained oscillatory wave to remove common noises. As Figure 4 shown, it mainly includes low-pass and high-pass filtering processing. First, a low-pass filter is used to reduce the noise of the original oscillatory wave signal. The main purpose of low-pass filtering is to remove high-frequency noises in the signal, such as power supply interference and high-frequency random noises, so as to retain the low-frequency components in the oscillatory wave signal. Then high-pass filtering is performed. The main purpose of high-pass filtering is to remove low-frequency components in the signal, such as baseline drift and respiratory artifacts and other low-frequency noises, so as to retain the high-frequency components in the oscillatory wave signal. Next, a band-pass filter is constructed based on the empirical wavelet transform to process each cycle of the oscillatory wave to eliminate motion artifacts caused by various external factors, and its calculation formula is as shown in Equation (1): (1); where a and b are the scale and translation parameters, ψ is the mother wavelet function, f(t) is the original signal, which is a function of t.

[0025] The feature extraction module uses traditional signal processing methods to extract time-domain, frequency-domain, time-frequency and statistical features from the preprocessed oscillatory wave signal, and divides the data into a training set and a test set.

[0026] The deep learning fine-tuning module uses a lightweight deep learning fine-tuning model to fine-tune the above traditional features to achieve personalized blood pressure prediction. As Figure 5As shown below, the specific method is as follows: Step 1: Feature normalization processing: Normalize the obtained traditional feature vectors to ensure that each feature is within the same numerical range, facilitating subsequent network learning.

[0027] Step 2: Construct a lightweight fully connected fine-tuning network (MLP), as Figure 6 shown.

[0028] Step 2.1: Input layer: Use the normalized feature vectors as the input.

[0029] Step 2.2: Hidden layer: Construct one or two fully connected hidden layers, set an appropriate number of neurons in each layer (such as 64 or 128), and use the ReLU activation function to capture the non-linear relationships between features.

[0030] Step 2.3: Add a Dropout layer (such as a Dropout rate of 0.5) between the hidden layers to prevent overfitting.

[0031] Step 2.4: Output layer: Construct a fully connected output layer, use a linear activation function, and output the fine-tuned feature representation.

[0032] Step 3: Training and evaluation.

[0033] Step 3.1: Use the training set data to train the lightweight fine-tuning network. Through steps such as forward propagation, loss calculation, backpropagation, and parameter update, enable the network to automatically learn how to optimize the traditionally extracted features.

[0034] Step 3.2: Use the test set to evaluate the fine-tuned model, detect the generalization performance of the model, and perform hyperparameter tuning as needed.

[0035] The blood pressure estimation module uses a machine learning model to match the signals in the database according to the real-time input fine-tuned features, predict the complete oscillatory wave waveform, and at the same time identify the positions of the two single-cycle oscillatory waves (mutation points) with the largest changes in the oscillatory wave feature bands of systolic blood pressure, diastolic blood pressure, and mean arterial pressure, and map the changes in the oscillatory wave feature bands to blood pressure values.

[0036] Train the machine learning model to automatically learn the mapping relationship between systolic blood pressure, diastolic blood pressure, and oscillatory wave signals, fit and output the complete oscillatory wave waveform. At the same time, learn the mapping relationship between the fine-tuned traditional features and the changes in the oscillatory wave feature bands of systolic blood pressure, diastolic blood pressure, and mean arterial pressure, fit the positions of the two single-cycle oscillatory waves (mutation points) with the largest changes in the feature bands, and find their corresponding cuff pressures to achieve non-invasive blood pressure measurement. The specific implementation process is as Figure 7 shown.

[0037] First, load the lightweight deep learning fine-tuning model, and the feature vectors after fine-tuning and the corresponding blood pressure data (diastolic blood pressure, mean arterial pressure, and systolic blood pressure if necessary); divide the data into a training set and a test set to ensure that there are sufficient samples for each blood pressure level. Select a group of traditional machine learning algorithms, including support vector machine, random forest, and extreme gradient boosting.

[0038] Use the training set to train the support vector machine, random forest, and extreme gradient boosting models respectively, and tune their respective hyperparameters (for example, the kernel function parameter in SVM, the number of trees in random forest, the learning rate and tree depth in extreme gradient boosting). Use the cross-validation method to evaluate the performance of the machine learning models on the training set and select the best model parameter combination.

[0039] Model evaluation and selection: Apply the trained machine learning models to the test set, and use the mean squared error (MSE), root mean squared error (RMSE), or mean absolute error (MAE) as indicators to evaluate the machine learning models. Compare the prediction accuracies of the machine learning models on the test set, and select the best-performing machine learning model or use the model ensemble method to further improve the prediction stability. The specific calculation formulas are shown in Equations (2), (3), and (4): (2); (3); (4); where n is the number of samples in the test set, y i is the true blood pressure value of the i-th sample in the test set, is the predicted blood pressure value of the i-th sample in the test set.

[0040] Model ensemble and output: Use the selected model for actual prediction, input the fine-tuned traditional features, output the complete oscillatory wave waveform prediction, and at the same time output the mutation positions of the oscillatory wave characteristic waves of the identified systolic blood pressure, diastolic blood pressure, and mean arterial pressure and the corresponding cuff pressure, so as to achieve non-invasive measurement of blood pressure; correct the prediction results according to the personalized calibration data as needed to further improve the prediction accuracy.

[0041] The blood pressure measurement method based on the dynamic characteristics of oscillatory waves and regression modeling in this embodiment is implemented through the above-mentioned blood pressure measurement system based on the dynamic characteristics of oscillatory waves and regression modeling, and includes the following steps: Step 1: Build an oscillatory wave signal database, including the complete oscillatory wave signal and its corresponding oscillatory wave characteristic bands of systolic blood pressure, diastolic blood pressure, and mean arterial pressure; control the inflation and deflation of the cuff through a micro solenoid valve to ensure that the pressure of the cuff can be adjusted in real time as needed. The specific method is as follows: Step 1.1: Correctly place the oscillometric wave signal device on the upper arm of the subject and fix it with a cuff to ensure accuracy and comfort; Step 1.2: Gradually pressurize the cuff, record the oscillometric wave data of the subject at different pressures, and store it; Step 1.3: Measure the systolic blood pressure and diastolic blood pressure of the subject using the built-in blood pressure measurement device; Step 1.4: The cuff pressure sensor is responsible for continuously collecting the cuff pressure data on the user's arm in real time; Step 1.5: Obtain the DC and AC signals of the cuff pressure; Step 1.5.1: The cuff pressure AC isolation and filtering circuit converts the AC signal of the cuff pressure into a DC signal and provides it to the MCU for real-time negative feedback control and regulation; the negative feedback control and regulation signal generated by the MCU is transmitted to the micro solenoid valve; the micro solenoid valve controls the inflation and deflation of the cuff to ensure that the pressure of the cuff is adjusted in real time as needed; Step 1.5.2: The cuff pressure DC isolation and filtering circuit is responsible for isolating the DC signal of the cuff pressure and obtaining the AC signal of the cuff pressure for collecting the real-time value of the oscillometric wave of the cuff pressure; Step 1.6: Output the processed oscillometric wave signal and cuff pressure signal; Step 1.7: Fit by identifying the waveform characteristics of the diastolic blood pressure and mean arterial pressure in the oscillometric waveform. If the fitting is good, stop inflation. If the fitting is not good, continue to pressurize until a normal blood pressure measurement is completed, and upload the obtained complete oscillometric waveform to the database.

[0042] Step 2: Evaluate the quality of the oscillometric wave signal in the database, including the evaluation of waveform smoothness and stationarity based on the time domain and the evaluation of spectrum consistency based on the frequency domain.

[0043] Step 2.1: Based on the evaluation of waveform smoothness in the time domain, detect the spikes or noises in the waveform to ensure smooth transition of the waveform on the time axis; Step 2.2: Based on the evaluation of spectrum consistency in the frequency domain, ensure that the obtained partial waveform is consistent with the overall waveform and ensure that the frequency distribution of the fitted waveform is similar to the actual waveform.

[0044] Step 3: Preprocess the oscillometric wave signal that passes the quality evaluation, including low-pass filtering, high-pass filtering, and empirical wavelet transform, to optimize the noise and baseline drift in the oscillometric wave signal. The specific method is as follows: Step 3.1: Use a low-pass filter to remove the high-frequency noise in the original oscillometric wave signal, thereby retaining the low-frequency components in the oscillometric wave signal; Step 3.2: Use a high-pass filter to remove the influence of baseline drift on the oscillometric wave, thereby retaining the high-frequency components in the oscillometric wave signal; Step 3.3: Construct a band-pass filter based on the empirical wavelet transform to process each periodic oscillation wave to eliminate the artifacts caused by various factors, as shown in the following formula: (1); Wherein, a and b are the scale and translation parameters, ψ is the mother wavelet function, f(t) is the original signal, which is a function of t.

[0045] Step 4: Perform feature band identification. Use traditional signal processing methods to extract time-domain, frequency-domain, time-frequency, and statistical features from the preprocessed oscillation wave signal, and divide the training set and the test set. The specific method is as follows: Step 4.1: Prepare data. Load the oscillation wave signal and its corresponding feature bands of systolic blood pressure, diastolic blood pressure, and mean arterial pressure, and divide the training set and the test set.

[0046] Step 4.2: Use traditional signal processing methods to extract features from the preprocessed oscillation wave signal. The specific methods include: Step 4.2.1: Time-domain feature extraction; Calculate the maximum value, minimum value, peak value, number of peaks / troughs, and amplitude of the second peak of the signal; calculate the root mean square value and average power of the signal; extract waveform slope, pulse width, and waveform asymmetry index.

[0047] The specific calculation formulas are as follows: Root mean square value RMS of the signal: (5); Wherein, T is the time length of the signal, and x(t) is the time-domain representation of the signal, that is, the oscillation wave signal changing with time.

[0048] Average power P of the signal avg : (6); Waveform slope: Define the single-cycle signal as x(n), where n is the discrete time point, n0 is the starting point of the single-cycle signal, n peak is the peak of the single-cycle waveform, n D is at the dicrotic notch, n SecondPeak is at the second peak, n valley is the trough of the single-cycle waveform, that is, the end point of the single-cycle waveform. Calculate the slope of the line connecting adjacent two points of the above feature points as the waveform slope.

[0049] Pulse width: The pulse width refers to the duration of the signal above a certain threshold. Set the threshold as V th , from the moment t when the rising edge of the signal exceeds the thresholdstart The time interval between the moment when the falling edge is lower than the threshold and the moment t is the pulse width PW, as shown in the following formula: end PW = t end - t start (7); The waveform asymmetry index can be used to measure the waveform asymmetry by calculating the energy ratio of the signal on the positive and negative half axes. The energy E of the signal on the positive half axis + is: (8); The energy E of the signal on the negative half axis - is: (9); The waveform asymmetry index AI is: (10); Step 4.2.2: Frequency domain feature extraction; Convert the time domain signal to a frequency domain signal using Fourier transform; extract the main frequency, power spectral density, and spectral entropy index of the signal.

[0050] The specific calculation method is as follows: Fourier transform: Convert the time domain signal x(t) to a frequency domain signal X(f), where f is the frequency, as shown in the following formula: (11); The main frequency of the signal: In the frequency domain signal X(f), the frequency corresponding to the frequency component with the largest amplitude is the main frequency, as shown in the following formula: (12); Power spectral density: (13); where X T (f) is the Fourier transform of the signal x(t) in the time interval , and PSD(f) represents the power spectral density; The spectral entropy is: (14); where, , sum the squares of the power amplitudes of all frequency points in the signal, N is the number of frequency points of the frequency domain signal, X(f i ) is the value of the frequency domain signal at the frequency f i , and H represents the spectral entropy.

[0051] Step 4.2.3: Time-frequency feature extraction; ​Extract the features of the time-domain signal at different scales using wavelet transform; decompose the time-domain signal using the empirical mode decomposition method and extract the time-frequency features of each intrinsic mode component.

[0052] The empirical mode decomposition decomposes the signal x(t) into a series of intrinsic mode components IMF m (t). Assuming that M intrinsic mode components are obtained by decomposition, then (15); For each intrinsic mode component IMF m (t), its time-frequency features can be extracted.

[0053] Step 4.2.4: Statistical feature extraction; Calculate the mean, standard deviation, skewness, and kurtosis statistics of the signal.

[0054] Mean μ : (16); where x i is the discrete sample value of the signal, and N is the number of samples; Standard deviation σ : (17); Skewness: (18); Kurtosis: (19).

[0055] Step 5: Fine-tune the extracted traditional features using a lightweight deep learning fine-tuning model. The specific method is as follows: Step 5.1: Perform normalization preprocessing on the time-domain, frequency-domain, time-frequency, and statistical feature vectors obtained by the traditional signal processing method in Step 4; Step 5.2: Input the normalized traditional feature vectors into the lightweight deep learning fine-tuning model, specifically as follows: Step 5.2.1: Construct a small fully connected network and use the normalized feature vectors as input; Step 5.2.2: Hidden layer: Construct one or two fully connected hidden layers, set an appropriate number of neurons in each layer, and use the ReLU activation function to capture the non-linear relationships between features; Step 5.2.3: Set a Dropout layer between the hidden layers to prevent overfitting; Step 5.2.4: Finally, construct a fully connected output layer and use a linear activation function to fine-tune the features and generate the final regression output; Step 5.3: Fine-tune the model using the training set data, and perform regression optimization on the features extracted by traditional methods and the changes in the characteristic bands of systolic blood pressure, diastolic blood pressure, and mean arterial pressure; Step 5.4: Evaluate the fine-tuned model using the test set, detect the generalization performance of the model, and perform hyperparameter tuning as needed.

[0056] Step 6: Train a machine learning model to automatically learn the mapping relationship between diastolic blood pressure, mean arterial pressure, and oscillometric wave signals, output the complete oscillometric waveform, and at the same time learn the changes in the characteristic bands of systolic blood pressure, diastolic blood pressure, and mean arterial pressure, map the changes in the characteristic bands to blood pressure values, and achieve non-invasive blood pressure measurement. The specific method is as follows: Step 6.1: Data preparation; Load the feature vectors after fine-tuning and the corresponding changes in the characteristic bands of the lightweight deep learning fine-tuning model in Step 5, and divide them into a training set and a test set to ensure that there are sufficient samples for each blood pressure level; Step 6.2: Model selection and feature optimization; Select a group of machine learning algorithms, including support vector machine, random forest, and extreme gradient boosting algorithm; Step 6.3: Model training; Step 6.3.1: Use the training set to train the support vector machine, random forest, and extreme gradient boosting algorithm models respectively, and tune their respective hyperparameters (for example, the kernel function parameter in the support vector machine, the number of trees in the random forest, the learning rate and tree depth in the extreme gradient boosting); Step 6.3.2: Use the cross-validation method to evaluate the performance of the model on the training set and select the best model parameter combination; Step 6.4: Model evaluation and selection; Step 6.4.1: Apply the trained model to the test set, and use the mean squared error (MSE), root mean squared error (RMSE), or mean absolute error (MAE) as indicators to evaluate the machine learning model; Step 6.4.2: Compare the prediction accuracies of each machine learning model on the test set, select the best-performing machine learning model, or use the model integration method to further improve the prediction stability; Step 6.5: Model integration and output; Step 6.5.1: Use the selected machine learning model for actual prediction, input the fine-tuned traditional features, and output the complete oscillometric waveform and the identified mutation points of the characteristic bands of systolic blood pressure, diastolic blood pressure, and mean arterial pressure; Step 6.5.2: At the same time, output the cuff pressure corresponding to the mutation points of the oscillometric characteristic bands to achieve non-invasive blood pressure measurement; Step 6.5.3: According to requirements, correct the measurement results in combination with personalized calibration data to further improve the measurement accuracy.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A blood pressure measurement system based on the dynamic characteristics of oscillatory waves and regression modeling, characterized in that: It includes an oscillatory wave database, a pressure control module, a data quality assessment module, a data preprocessing module, a feature extraction module, a deep learning fine-tuning module, and a blood pressure estimation module; The database stores the oscillatory waves of at least 3000 subjects and their corresponding blood pressure results, where individual subjects need to have multiple blood pressure measurement data in a quiet state; The pressure control module controls the inflation and deflation of the cuff through a micro solenoid valve to ensure that the pressure of the cuff can be adjusted in real time as needed; the pressure control module includes an MCU, a cuff pressure sensor, a cuff pressure AC isolation and filtering circuit, and a cuff pressure DC isolation and filtering circuit; the cuff pressure sensor continuously collects the cuff pressure data on the subject's arm in real time and uses this as the basis for blood pressure estimation; the cuff pressure AC isolation and filtering circuit converts the AC signal of the cuff pressure into a DC signal and provides it to the MCU for subsequent negative feedback control and adjustment; the cuff pressure DC isolation and filtering circuit is responsible for isolating the DC signal of the cuff pressure and obtaining the AC signal of the cuff pressure to collect the real-time value of the cuff pressure oscillatory wave; The data quality assessment module is used to screen out signals with poor quality assessment and remove them; The data preprocessing module is used to filter the obtained oscillatory waves to remove common noises; The feature extraction module uses traditional signal processing methods to extract time-domain, frequency-domain, time-frequency, and statistical features from the preprocessed oscillatory wave signals and divides them into a training set and a test set; The deep learning fine-tuning module uses a lightweight deep learning fine-tuning model to fine-tune the extracted features to achieve personalized blood pressure prediction; The blood pressure estimation module uses a machine learning model to match the signals in the database according to the real-time input fine-tuned features, predicts the complete oscillatory wave waveform, and at the same time identifies the positions of the two single-cycle oscillatory waves with the largest changes in the characteristic bands of the systolic blood pressure, diastolic blood pressure, and mean arterial pressure oscillatory waves, and maps the changes in the oscillatory wave characteristic bands to blood pressure values.

2. The blood pressure measurement system based on the dynamic characteristics of the oscillatory wave and regression modeling according to claim 1, characterized in that: The deep learning fine-tuning module first performs normalization preprocessing on the time-domain, frequency-domain, time-frequency, and statistical features extracted by the feature extraction module, and then inputs them into the lightweight deep learning fine-tuning model; the lightweight deep learning fine-tuning model is a fully connected network, including an input layer, one or two hidden layers, a Dropout layer, and a fully connected output layer; an appropriate number of neurons are set in the hidden layer, and the ReLU activation function is used to capture the non-linear relationship between features; the Dropout layer is added between the hidden layers to prevent overfitting; the fully connected output layer uses a linear activation function to output the fine-tuned feature representation; The network of the lightweight deep learning fine-tuning model is trained using the training set data, and through forward propagation, loss calculation, backpropagation, and parameter update, the network automatically learns how to optimize the extracted features; the fine-tuned model is evaluated using the test set to detect the generalization performance of the model and perform hyperparameter tuning as needed.

3. The blood pressure measurement system based on the dynamic characteristics of the oscillometric wave and regression modeling according to claim 2, wherein: The blood pressure estimation module first loads the feature vectors and corresponding blood pressure data after fine-tuning of the lightweight deep learning fine-tuning model, divides the data into a training set and a test set to ensure that there are sufficient samples for each blood pressure level; uses the training set to train the support vector machine, random forest, and extreme gradient boosting models respectively, and tunes their respective hyperparameters; uses the cross-validation method to evaluate the performance of each model on the training set, and selects the best model parameter combination; Applies the trained model to the test set, and evaluates the model using metrics such as mean squared error, root mean squared error, or mean absolute error; compares the prediction accuracy of each model on the test set, selects the best-performing model or uses the model ensemble method to further improve the prediction stability; Uses the selected model for actual prediction, inputs the fine-tuned features, outputs the complete oscillatory wave waveform prediction, and at the same time outputs the mutation positions of the oscillatory wave features of the recognized systolic blood pressure, diastolic blood pressure, and mean arterial pressure and the corresponding cuff pressure to achieve non-invasive blood pressure measurement; corrects the prediction results according to the need by combining personalized calibration data.

4. A blood pressure measurement method based on the dynamic characteristics of oscillatory waves and regression modeling, characterized in that: It is implemented by the blood pressure measurement system based on the dynamic features of oscillatory waves and regression modeling described in claim 1, including the following steps: Step 1: Construct an oscillatory wave signal database, including complete oscillatory wave signals and their corresponding oscillatory wave characteristic bands of systolic blood pressure, diastolic blood pressure, and mean arterial pressure; control the inflation and deflation of the cuff through a micro solenoid valve to ensure that the pressure of the cuff can be adjusted in real time as needed; Step 2: Evaluate the quality of the oscillatory wave signals in the database, including waveform smoothness and stationarity evaluation based on the time domain and spectral consistency evaluation based on the frequency domain; Step 3: Preprocess the oscillatory wave signals that pass the quality evaluation, including low-pass filtering, high-pass filtering, and empirical wavelet transform, to optimize the noise and baseline drift in the oscillatory wave signals; Step 4: Perform characteristic band identification, extract time domain, frequency domain, time-frequency, and statistical features from the preprocessed oscillatory wave signals using traditional signal processing methods, and divide them into a training set and a test set; Step 5: Use a lightweight deep learning fine-tuning model to fine-tune the extracted traditional features; Step 6: Train a machine learning model to automatically learn the mapping relationship between diastolic blood pressure, mean arterial pressure, and oscillatory wave signals, output the complete oscillatory wave waveform, and at the same time learn the changes in the oscillatory wave characteristic bands of systolic blood pressure, diastolic blood pressure, and mean arterial pressure, and map the changes in the characteristic bands to blood pressure values to achieve non-invasive blood pressure measurement.

5. The blood pressure measurement method based on the dynamic characteristics of the oscillatory wave and regression modeling according to claim 4, wherein: The specific method of step 1 is as follows: Step 1.1: Place the oscillatory wave signal device correctly on the upper arm of the subject and fix it with a cuff; Step 1.2: Gradually pressurize the cuff, record the oscillatory wave data of the subject at different pressures, and store them; Step 1.3: Measure the systolic blood pressure and diastolic blood pressure of the subject using the built-in blood pressure measurement device; Step 1.4: The cuff pressure sensor is responsible for continuously collecting the cuff pressure data on the user's arm in real time; Step 1.5: Obtain the DC and AC signals of the cuff pressure; Step 1.5.1: The AC isolation and filtering circuit of the cuff pressure converts the AC signal of the cuff pressure into a DC signal, which is provided to the MCU for real-time negative feedback control and regulation; the negative feedback control and regulation signal generated by the MCU is transmitted to the micro solenoid valve; the micro solenoid valve controls the inflation and deflation of the cuff to ensure that the pressure of the cuff is regulated in real time as required; Step 1.5.2: The DC isolation and filtering circuit of the cuff pressure is responsible for isolating the DC signal of the cuff pressure to obtain the AC signal of the cuff pressure, which is used to collect the real-time value of the cuff pressure oscillation wave; Step 1.6: Output the processed oscillation wave signal and cuff pressure signal; Step 1.7: Fit by identifying the waveform characteristics of the diastolic blood pressure and mean arterial pressure in the oscillation waveform. If the fitting is good, the inflation stops. If the fitting is not good, continue to pressurize until a normal blood pressure measurement is completed, and upload the obtained complete oscillation waveform to the database.

6. The blood pressure measurement method based on the dynamic characteristics of the oscillatory wave and regression modeling according to claim 4, wherein: The specific method of Step 2 is as follows: Step 2.1: Based on the evaluation of waveform smoothness in the time domain, detect the spikes or noises in the waveform to ensure smooth transition of the waveform on the time axis; Step 2.2: Based on the evaluation of spectrum consistency in the frequency domain, ensure that the obtained partial waveform is consistent with the overall waveform, and ensure that the frequency distribution of the fitted waveform is similar to that of the actual waveform.

7. The blood pressure measurement method based on the dynamic characteristics of the oscillatory wave and regression modeling according to claim 4, characterized in that: The specific method of Step 3 is as follows: Step 3.1: Use a low-pass filter to remove the high-frequency noise in the original oscillation wave signal, so as to retain the low-frequency components in the oscillation wave signal; Step 3.2: Use a high-pass filter to remove the influence of baseline drift on the oscillation wave, so as to retain the high-frequency components in the oscillation wave signal; Step 3.3: Construct a band-pass filter based on empirical wavelet transform to process each cycle of the oscillation wave to eliminate the artifacts caused by various factors, as shown in the following formula: （1）； wherein, a and b are scale and translation parameters, ψ is the mother wavelet function, and f(t) is the original signal, which is a function of t.

8. The blood pressure measurement method based on the dynamic characteristics of the oscillatory wave and regression modeling according to claim 4, characterized in that: The specific method of Step 4 is as follows: Step 4.1: Perform data preparation, load the oscillation wave signal and its corresponding characteristic bands of systolic blood pressure, diastolic blood pressure and mean arterial pressure, and divide the training set and test set; Step 4.2: Adopt traditional signal processing methods to extract features from the preprocessed oscillation wave signal. The specific methods include: Step 4.2.1: Time-domain feature extraction; Calculate the maximum value, minimum value, peak value, number of peaks / troughs, and amplitude of the second peak of the signal; Calculate the root mean square value and average power of the signal; the root mean square value RMS of the signal is: （5）； where T is the time length of the signal, and x(t) is the time-domain representation of the signal, that is, the oscillation wave signal changing with time; The average power P of the signal avg is as follows: （6）； Extract the waveform slope, pulse width and waveform asymmetry index; Waveform slope: Define a single-cycle signal as x(n), where n is the discrete time point, n0 is the starting point of the single-cycle signal, n peak is the peak of the single-cycle waveform, n D is at the dicrotic notch, n SecondPeak is at the second peak, n valley is the trough of the single-cycle waveform, i.e., the end point of the single-cycle waveform. Calculate the slope of the line connecting adjacent two of the above feature points as the waveform slope; Pulse width: The pulse width refers to the duration of a signal above a certain threshold. The threshold is set as V th , from the moment t start when the rising edge of the signal exceeds the threshold to the moment t end when the falling edge is below the threshold, the time interval between them is the pulse width PW, as shown in the following formula: PW = t end -t start (7); Waveform asymmetry index: Measures the waveform asymmetry by calculating the energy ratio of the signal in the positive and negative half-axes; the energy E of the signal in the positive half-axis + is as follows: （8）； The energy E of the signal in the negative half-axis - is as follows: （9）； The waveform asymmetry index AI is: （10）； Step 4.2.2: Frequency-domain feature extraction; Use Fourier transform to convert the time-domain signal into a frequency-domain signal; Extract the main frequency, power spectral density and spectral entropy index of the signal; The specific calculation methods are as follows: Fourier transform: Convert the time-domain signal x(t) into a frequency-domain signal X(f), where f is the frequency, as shown in the following formula: （11）； The main frequency of the signal: In the frequency-domain signal X(f), the frequency corresponding to the frequency component with the largest amplitude is the main frequency, as shown in the following formula: （12）； The power spectral density is: （13）； where X T (f) is the Fourier transform of the signal x(t) in the time interval and PSD(f) represents the power spectral density; The spectral entropy is: （14）； Among them, , sum the squares of the power amplitudes of all frequency points in the signal. N is the number of frequency points of the frequency-domain signal, X(f i ) is the value of the frequency-domain signal at frequency f i , and H represents spectral entropy; Step 4.2.3: Time-frequency feature extraction; Extract the features of the time-domain signal at different scales using wavelet transform; decompose the time-domain signal using empirical mode decomposition method and extract the time-frequency features of each intrinsic mode component; Empirical mode decomposition decomposes the signal x(t) into a series of intrinsic mode functions IMF m (t). Assuming that M intrinsic mode functions are obtained by decomposition, then （15）； For each intrinsic mode function IMF m (t), extract its time-frequency characteristics; Step 4.2.4: Statistical feature extraction; Calculate the mean, standard deviation, skewness and kurtosis statistics of the signal as follows: Mean value μ : （16）； where x i is the discrete sample value of the signal, and N is the number of samples; Standard deviation σ : （17）； Skewness: （18）； Kurtosis: （19）。 9. The blood pressure measurement method based on the dynamic characteristics of the oscillatory wave and regression modeling according to claim 8, characterized in that: The specific method of step 5 is as follows: Step 5.1: Perform normalization preprocessing on the time-domain, frequency-domain, time-frequency and statistical feature vectors obtained by traditional signal processing methods in step 4; Step 5.2: Input the normalized traditional feature vectors into the lightweight deep learning fine-tuning model as follows: Step 5.2.1: Construct a small fully-connected network with the normalized feature vectors as the input; Step 5.2.2: Hidden layer: Construct one or two fully-connected hidden layers, set an appropriate number of neurons in each layer, and use the ReLU activation function to capture the non-linear relationship between features; Step 5.2.3: Set a Dropout layer between the hidden layers to prevent overfitting; Step 5.2.4: Finally, construct a fully-connected output layer with a linear activation function to fine-tune the features and generate the final regression output; Step 5.3: Train the fine-tuning model on the training set data, and perform regression optimization using the features extracted by traditional methods and the changes in the feature bands of systolic blood pressure, diastolic blood pressure and mean arterial pressure; Step 5.4: Use the test set to evaluate the fine-tuned model, detect the generalization performance of the model and perform hyperparameter tuning as needed.

10. The blood pressure measurement method based on the dynamic characteristics of the oscillatory wave and regression modeling according to claim 9, characterized in that: The specific method of step 6 is as follows: Step 6.1: Data preparation; Load the fine-tuned feature vectors in step 5 and the corresponding changes in the feature bands, and divide them into a training set and a test set to ensure that there are sufficient samples for each blood pressure level; Step 6.2: Model selection and feature optimization; Select a group of machine learning algorithms, including support vector machine, random forest and extreme gradient boosting algorithm; Step 6.3: Model training; Step 6.3.1: Use the training set to train the support vector machine, random forest and extreme gradient boosting algorithm models respectively, and tune their respective hyperparameters; Step 6.3.2: Use the cross-validation method to evaluate the performance of the model on the training set and select the best model parameter combination; Step 6.4: Model evaluation and selection; Step 6.4.1: Apply the trained model to the test set and evaluate the machine learning model using the mean squared error, root mean squared error or mean absolute error as the index; Step 6.4.2: Compare the prediction accuracy of each machine learning model on the test set, select the best-performing machine learning model or use the model integration method to further improve the prediction stability; Step 6.5: Model integration and output; Step 6.5.1: Use the selected machine learning model for actual prediction, input the fine-tuned traditional features, and output the complete oscillatory wave waveform and the mutation points of the feature bands of systolic blood pressure, diastolic blood pressure and mean arterial pressure identified; Step 6.5.2: At the same time, output the cuff pressure corresponding to the mutation points of the oscillatory wave feature bands to achieve non-invasive measurement of blood pressure; Step 6.5.3: Correct the measurement results according to requirements in combination with personalized calibration data.

Images