A blood pressure accurate measurement method based on a portable pulse diagnosis instrument

Abstract

The application provides a blood pressure accurate measurement method based on a portable pulse diagnosis instrument, which realizes accurate prediction of blood pressure by collecting pulse and blood pressure data of the subject and performing systematic data processing and analysis. The method includes collecting left and right hand pulse data using a portable pulse diagnosis instrument, and synchronously collecting blood pressure data to establish the mapping relationship between pulse and blood pressure. The data processing steps include data cleaning, Shannon sampling and wavelet threshold denoising algorithm to ensure the accuracy and reliability of the pulse signal. By constructing a machine learning regression model based on the improved Hunter Prey Optimization (HPO), combined with the XGBoost algorithm for prediction, the generalization ability and stability of the model are improved. The model uses mean square error (MSE) as the loss function, and iteratively optimizes through the adaptive learning rate adjustment strategy, effectively reduces overfitting, and improves prediction accuracy. Finally, the model predicts the blood pressure of the subject through the pulse data, and performs error analysis to verify the clinical application effect of the model. The application has high efficiency, accuracy and stability, and provides a non-contact and convenient technical solution for portable blood pressure measurement.

Description

Technical Field

[0001] This invention relates to the field of computer vision application technology; and more particularly to a method for accurate blood pressure measurement based on a portable pulse diagnostic instrument. Technical Background

[0002] Hypertension, a common chronic disease, is a significant risk factor for cardiovascular and cerebrovascular diseases. Therefore, accurate and convenient blood pressure monitoring is crucial for the prevention and control of hypertension. However, traditional blood pressure measurement methods mainly rely on mercury sphygmomanometers or electronic sphygmomanometers. These devices are typically bulky, poorly portable, and require professional guidance to operate, making them difficult to use for convenient daily monitoring. To meet the growing personalized health needs of individuals, portable blood pressure monitoring devices have gradually gained attention in recent years.

[0003] Most existing portable blood pressure measuring devices are based on the cuff compression principle. Although their size has been reduced, wearing and operation are still relatively complex, and their accuracy is easily affected by factors such as posture and wearing position, resulting in unstable measurement accuracy. While some cuffless photoelectric blood pressure measuring devices have improved portability, issues with data accuracy and equipment cost limit their widespread application in home and daily health monitoring.

[0004] Pulse diagnosis is an important method in traditional Chinese medicine (TCM) for assessing a person's health. TCM practitioners observe the strength, depth, and speed of a patient's pulse to determine their blood pressure. Based on different pulse characteristics, TCM can identify the circulation of Qi and blood, as well as the health of the heart and blood vessels. Modern pulse diagnosis instruments combine traditional methods with sensor technology, using data analysis algorithms to analyze indicators such as pulse frequency, amplitude, and waveform changes, providing a digital and objective way to assess blood pressure. This method continuously monitors pulse fluctuations, quantifies pulse information, assists in assessing blood pressure, and provides scientific support for health management.

[0005] This invention proposes a blood pressure measurement method based on a portable pulse oximeter. This method combines the advantages of traditional Chinese medicine pulse diagnosis and modern sensor technology. It accurately collects pulse data using a highly sensitive inertial sensor and combines this with machine learning algorithms for blood pressure measurement, achieving a cuffless, portable design for convenient daily use. Simultaneously, the pulse oximeter can capture subtle changes in pulse fluctuations, digitizing and standardizing the data, significantly improving the objectivity and consistency of the measurement. Compared to traditional cuff-type blood pressure monitors, this portable pulse oximeter not only eliminates the need for pressure on the arm, avoiding discomfort, but also reduces external interference by directly reading pulse characteristics, providing continuous, real-time blood pressure data suitable for dynamic monitoring and trend analysis. This method improves the user experience while maintaining the accuracy and portability of blood pressure measurement, providing an efficient and innovative solution for personalized health management and disease prevention. Summary of the Invention

[0006] This invention proposes a method for accurate blood pressure measurement based on a portable pulse oximeter. The method collects pulse and blood pressure data from subjects, processes the data using techniques such as data cleaning, Shannon sampling, and wavelet thresholding denoising, and extracts pulse data features by constructing a machine learning regression model based on Hunter Prey Optimization (HPO) to achieve accurate blood pressure prediction. This method is efficient, accurate, and stable, providing reliable support for clinical blood pressure measurement.

[0007] The specific technical solution of the present invention is as follows:

[0008] The pulse data of the subjects' left and right hands were collected using a portable pulse diagnostic device, and the corresponding blood pressure data were simultaneously collected using an OMRON U30 electronic blood pressure monitor to establish a mapping relationship between pulse and blood pressure. At the same time, the blood pressure measurement time, pulse data, and the subjects' physiological state were recorded to provide background information to assist in data analysis.

[0009] The collected pulse data underwent data cleaning, with upper and lower thresholds set to remove outliers and noise interference, ensuring data accuracy. Based on Shannon's sampling theorem, an appropriate sampling frequency was set to retain pulse signals in the 0-50Hz frequency band, improving signal accuracy and reliability. A wavelet thresholding denoising algorithm was used to filter the pulse data, removing noise and obtaining a smooth and accurate pulse signal. Six-level wavelet decomposition using the Bior 3.5 basis wavelet was employed to extract signal features from different frequency bands, and a soft thresholding function was used for signal smoothing to reduce the impact of noise. Furthermore, the threshold value was dynamically adjusted based on noise levels and signal characteristics to optimize noise removal.

[0010] The final pulse dataset was divided into training and testing sets in an 8:2 ratio to ensure sample independence for model training and testing, thereby enhancing the model's generalization ability.

[0011] A machine learning regression model based on the HPO algorithm is constructed and combined with the eXtreme Gradient Boosting (XGBoost) regression algorithm to improve the model's predictive performance. The improved regression model undergoes iterative parameter optimization, and an adaptive learning rate dynamic adjustment method is applied to reduce overfitting and improve the model's stability and accuracy. Mean squared error (MSE) is used as the loss function, and the model training process is optimized by adjusting the learning rate decay strategy.

[0012] The training set is input into the improved regression model for training, thereby enhancing the model's prediction accuracy. The trained regression model is then used to predict blood pressure using pulse data from subjects in the test set. The predicted results are compared with actual blood pressure, and error analysis is performed to provide reliable predictive support for clinical applications.

[0013] This invention achieves accurate blood pressure prediction based on pulse data through the above technical solutions, with high measurement accuracy and stability. It can be widely used in portable devices, providing technical support for non-contact and convenient blood pressure measurement.

[0014] Beneficial technical effects

[0015] 1. This invention collects pulse and blood pressure data from subjects using a portable pulse diagnostic instrument. By employing Shannon sampling theorem and frequency band selection method, it preserves key frequency band information of the pulse signal (0 to 50 Hz), effectively avoiding the impact of low-frequency noise and high-frequency interference on blood pressure prediction results. This provides a good data foundation for accurate signal extraction and subsequent analysis, ensuring the stability of blood pressure prediction.

[0016] 2. This invention uses a wavelet threshold denoising algorithm to process pulse data, successfully extracting high-quality pulse signals, removing outliers and noise interference, providing accurate and reliable input data for subsequent blood pressure prediction models, and significantly improving the accuracy and precision of blood pressure measurement.

[0017] 3. This invention constructs an efficient blood pressure prediction model by combining the HPO algorithm with XGBoost regression, and adopts an adaptive learning rate optimization strategy to improve the model's generalization ability and robustness, reduce overfitting, and ensure the accuracy of blood pressure prediction under different individuals and physiological conditions.

[0018] 4. This invention is the first to apply machine learning regression models to the accurate measurement of blood pressure using a portable pulse diagnostic instrument. By automatically extracting pulse data features, it successfully eliminates the influence of environmental interference, measurement errors, and data noise on the prediction results, providing an efficient, intelligent, and stable blood pressure prediction method and offering a new technical solution for clinical and convenient blood pressure measurement.

[0019] 5. This invention breaks through the limitations of traditional blood pressure measurement methods by using a portable pulse diagnostic instrument for non-invasive blood pressure prediction. While improving measurement accuracy, it reduces costs and increases efficiency, making it particularly suitable for large-scale screening, health monitoring of the elderly, and convenient daily blood pressure management. Attached Figure Description

[0020] Figure 1 A flowchart of a method for accurate blood pressure measurement based on a portable pulse diagnostic instrument provided in an embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram of a portable pulse diagnostic instrument used in this invention for left-hand pulse detection.

[0022] Among them: 1 is an inertial sensor located at the left Cun pulse; 2 is an inertial sensor located at the left Guan pulse; 3 is an inertial sensor located at the left Chi pulse; 4 is a PCB circuit board equipped with a Bluetooth chip; 5 is a lithium battery that powers the pulse diagnosis instrument; 6 is an FPC flexible circuit board.

[0023] Figure 3 The electronic blood pressure monitor used in this invention is the OMRON U30.

[0024] Figure 4 This is a flowchart of the wavelet threshold denoising algorithm used in this invention;

[0025] Figure 5 This is a comparison chart of data before and after the application of the wavelet threshold denoising algorithm in this invention;

[0026] Figure 6 This is a diagram of the machine learning model architecture that combines the HPO algorithm with XGBoost regression used in this invention.

[0027] Figure 7 This invention presents a comparison between the actual and predicted blood pressure values ​​after regression prediction using a machine learning model. Detailed Implementation

[0028] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0029] Figure 1 The flowchart for a precise blood pressure measurement method based on a portable pulse oximeter is as follows:

[0030] Step 1: Obtain pulse and blood pressure data for the subject's left and right hands, specifically including the following steps:

[0031] Step 1.1: As Figure 2 As shown, pulse data from the left and right hands of 100 subjects were collected using a portable pulse diagnostic device. The device accurately recorded pulse waveform data via sensors, with a sampling frequency set to 100Hz to ensure the capture of subtle pulse changes. Pulse data was collected from both hands of each subject separately to ensure comparative analysis of the differences between their left and right hand pulses. Meanwhile, Figure 3 The electronic blood pressure monitor shown will simultaneously measure blood pressure by applying pressure to the subject's upper arm through an air bladder, measuring systolic and diastolic blood pressure values, and storing the results together with pulse data to establish a correspondence between pulse and blood pressure.

[0032] Step 1.2: Record the subject's blood pressure measurement time, pulse data, and corresponding physiological state to provide background information for data analysis. In addition to recording the time of each blood pressure measurement, it is also necessary to record the subject's physiological state during the measurement process in detail, such as emotional state (relaxed, tense, etc.) and pre-measurement activity (whether exercised). This information helps analyze the potential relationship between blood pressure and pulse waveforms and provides more comprehensive background information for subsequent data modeling. By combining physiological state data, the accuracy of the model can be further optimized, and the reliability of blood pressure prediction can be improved.

[0033] Step 2: Preprocess the left and right hand pulse data obtained in Step 1 to obtain the final pulse dataset. This includes the following steps:

[0034] Step 2.1: Perform data cleaning on the pulse data collected in Step 1.1, setting upper and lower threshold values ​​for the pulse, removing outliers and interference factors to ensure data accuracy. First, through time-domain analysis of the pulse signal, identify and eliminate sudden abnormal data caused by equipment malfunction or external interference, such as extremely high or low values ​​in the pulse data, which usually exceed the physiologically reasonable range. Set upper and lower threshold values ​​for the pulse signal, typically based on the normal physiological range (60 to 100 beats / minute). Through data screening, remove abnormal fluctuations and interference caused by misoperation or environmental factors to ensure data purity. Furthermore, by observing the periodicity and regularity of the pulse waveform, eliminate sudden instantaneous signal errors to ensure the data quality meets the requirements of subsequent analysis.

[0035] Step 2.2: Based on Shannon's sampling theorem, set the sampling frequency to retain the pulse signal in the 0-50Hz frequency band from the data obtained in Step 2.1, improving the accuracy and reliability of pulse detection. According to Shannon's sampling theorem, the sampling frequency needs to be at least twice the signal frequency to avoid aliasing. In the case of pulse signals, the frequency of physiological pulses is usually between 0 and 2Hz. Since the components of the pulse waveform usually include lower frequency components (such as respiration, changes in body position, etc.), the sampling rate is set to above 100Hz to ensure that all changes in the pulse signal can be completely captured. By filtering and retaining the pulse signal between 0 and 50Hz, interference signals outside the frequency band can be effectively eliminated, further improving the clarity of the pulse signal.

[0036] Step 2.3: Filter the pulse data from Step 2.2 using a wavelet thresholding denoising algorithm. The wavelet thresholding denoising algorithm is a signal denoising method based on wavelet transform. It separates noise from the valid signal through thresholding, thereby improving data quality and obtaining a smooth and accurate pulse signal. The flowchart of the wavelet thresholding denoising algorithm is shown below. Figure 4As shown. In this process, wavelet transform can decompose the pulse signal into different frequency sub-bands, and then noise suppression can be performed on each sub-band. First, the signal is decomposed into multiple layers of wavelet transform, gradually extracting different feature information from low frequency to high frequency. Next, a threshold denoising algorithm is used to denoise the decomposed signal, using wavelet functions to remove noise while preserving the useful components of the signal to the maximum extent and avoiding distortion. Specifically, the following steps are included:

[0037] Step 2.3.1: To evaluate the denoising effect of different wavelet bases at different decomposition levels, the signal-to-noise ratio (SNR) and root mean square error (RMSE) are used to evaluate the signal. SNR is the ratio of signal power to noise power, used to measure signal quality. RMSE is the square root of the mean square of the deviations between the predicted and true values, used to evaluate the model prediction accuracy. By using these two methods, the denoising performance is objectively measured, ensuring that the signal quality is effectively improved. The SNR formula is shown in Equation (1), and the RMSE formula is shown in Equation (2).

[0038]

[0039] Where x(i) is the original signal component and y(i) is the denoised signal component. For the Bior3.5 wavelet basis: when using the SNR evaluation method, the result is 31.1 for 2 layers, 30.9 for 4 layers, and 31.5 for 6 layers. When using the RMSE evaluation method, the result is 3.96654 for 2 layers, 4.57983 for 4 layers, and 2.05465 for 6 layers. The larger the SNR value, the better the denoising effect; the smaller the RMSE, the closer the reconstructed signal is to the original signal, and the better the denoising effect.

[0040] Step 2.3.2: To better extract signal features from different frequency bands, this invention employs the Bior 3.5 basis wavelet to perform a six-level wavelet decomposition on the pulse signal. The Bior 3.5 basis wavelet is a biororthogonal wavelet with excellent time-frequency localization capabilities and linear phase characteristics. Its asymmetric design can effectively extract local features of non-stationary signals such as the pulse wave. In this process, the pulse signal undergoes six levels of wavelet decomposition, and the signal is decomposed into multiple frequency bands, so that the wavelet coefficients of each level can represent the characteristics of the pulse signal at different scales. Through this decomposition method, important features and noise in the pulse signal can be more accurately identified and separated, providing an effective foundation for subsequent noise removal and signal smoothing.

[0041] Step 2.3.3: The soft threshold function is used to smoothly reduce the signal amplitude close to the threshold in step 2.3.2, thereby reducing the influence of noise and making the signal transition smoother. The soft threshold function is a commonly used wavelet denoising method. It reduces the interference of noise on the signal by smoothing the wavelet coefficients. The formula for the soft threshold function is shown in equation (3).

[0042]

[0043] Where t represents the wavelet coefficients and λ is the threshold, representing the degree of denoising. In this process, each layer of wavelet coefficients is first processed according to a preset threshold. Coefficients below the threshold are zeroed out, while coefficients above the threshold have their amplitude smoothly reduced, thus removing the influence of noise. This operation effectively suppresses high-frequency noise, making the pulse signal transition smoother while preserving the information of the valid signal.

[0044] Step 2.3.4: By calculating the noise estimate for each layer, the threshold value in Step 2.3.3 is dynamically adjusted based on the noise level and signal characteristics, allowing the high-frequency layer to remove more noise while the low-frequency layer retains more signal information. In the denoising process, the noise estimate is crucial for adjusting the threshold. Based on the distribution of wavelet coefficients at each layer, the noise is first estimated, typically represented by the standard deviation of the signal. For the high-frequency portion, where there are more noise components, a lower threshold is needed to effectively remove noise. For the low-frequency portion, where the principal components of the signal are usually located in the low-frequency band, a higher threshold is required to retain more signal information. In this way, the threshold can be dynamically adjusted according to the noise characteristics of different frequency components, achieving a more accurate denoising effect. A comparison of the data before and after the wavelet threshold denoising algorithm is shown in the figure below. Figure 5 As shown.

[0045] Step 3: Divide the final pulse dataset obtained in Step 2 into a training set and a test set in an 8:2 ratio to ensure that the samples for model training and testing are independent and to enhance the generalization ability of the model.

[0046] Step 4: Build a machine learning regression model based on HPO to improve the accuracy and stability of blood pressure prediction. The specific implementation includes the following steps:

[0047] Step 4.1: Construct a machine learning regression model based on the HPO algorithm. The HPO algorithm is a biomimetic optimization algorithm that simulates the behavioral strategies between predators and prey. It can search for optimal solutions in high-dimensional space and improves the convergence speed and global search capability through group cooperation and competition mechanisms. First, the model structure and optimization process of HPO are defined and applied to the model parameter selection in the blood pressure regression task. The HPO algorithm selects the optimal individuals through fitness evaluation and iteratively finds the best parameter combination, enabling the regression model to fully fit the blood pressure data characteristics. Through the global search capability of the HPO algorithm, the local optimum trap in traditional algorithms can be effectively avoided, ensuring better prediction results on complex datasets.

[0048] Step 4.2: Combine the HPO algorithm from Step 4.1 with the XGBoost regression model to effectively improve the model's performance. The machine learning model framework diagram is as follows. Figure 6 As shown, XGBoost is a high-performance machine learning algorithm based on the gradient boosting framework. It significantly improves the model's prediction accuracy and generalization ability through parallel computation and regularization optimization. This algorithm demonstrates excellent performance in regression tasks thanks to its efficient distributed computing architecture and built-in feature importance evaluation mechanism. By combining HPO with XGBoost, using HPO for XGBoost parameter optimization, the model's prediction performance is enhanced. The global optimization characteristics of HPO bring higher accuracy to XGBoost, while also enhancing the algorithm's convergence and stability, enabling the model to maintain high prediction accuracy under different physiological states and interference environments.

[0049] Step 4.3: Iteratively optimize the parameters of the improved machine learning regression model from Step 4.2, and apply an adaptive learning rate to dynamically adjust the learning rate, reducing overfitting and improving model stability. The adaptive learning rate uses the dynamic decay mechanism shown in formula (4):

[0050]

[0051] Where new_lr is the new learning rate, initial_lr is the initial learning rate, γ is the decay factor, epoch is the current training epoch, and step_size is the learning rate update interval. By dynamically adjusting the learning rate, the model can converge quickly in the early learning stage and gradually reduce the learning rate in the later stage to avoid parameter fluctuations, thereby reducing overfitting. At the same time, through the adaptive learning rate strategy, the model can dynamically adjust the optimization step size according to different data complexities, making the model training more stable and improving the overall prediction performance. In terms of the choice of loss function, MSE is used as the evaluation standard. MSE is the average of the squared differences between the predicted value and the true value, which is used to measure the error of the prediction model, and is expressed by formula (5):

[0052]

[0053] Where n is the number of samples, and x(i) is the true value. The mean squared error (MSE) is the model's predicted value. MSE effectively reflects how well the model fits the real data. As an optimization objective for the regression model, it helps the HPO algorithm fine-tune the parameters, enabling the model to better capture subtle differences in blood pressure data. In this way, the improved model achieves high-precision blood pressure prediction while further enhancing its robustness against interference and its performance in complex scenarios.

[0054] Step 5: Use the trained machine learning regression model to predict the subject's blood pressure. This includes the following steps:

[0055] Step 5.1: Input the preprocessed training set obtained in Step 3 into the optimized regression model in Step 4.3. Through multiple iterations of training, the model gradually enhances its ability to recognize and extract pulse signal features. A strategy of dynamically adjusting the learning rate is used to improve training efficiency, and the sample proportions in each training batch are controlled so that samples with systolic blood pressure less than 120 mmHg, 120-140 mmHg, and greater than 140 mmHg account for 30%, 50%, and 20%, respectively, to optimize data distribution. When the loss decrease rate is less than 1% after 10 consecutive iterations, the model is considered converged and training is terminated. During training, the convergence of the model is judged by monitoring the decreasing trend of the loss function, ensuring that the model stops training after reaching optimal predictive performance, thereby improving the accuracy of blood pressure prediction.

[0056] Step 5.2: Use the trained regression model to predict blood pressure using the pulse data of the subjects in the test set divided in Step 3. Post-process the prediction results, automatically removing abnormal predictions with systolic blood pressure less than 70 mmHg or greater than 200 mmHg, and retesting. Calculate the average of five consecutive predictions as the final output. Compare the processed prediction results with the actual blood pressure to evaluate the model's generalization performance and clinical applicability. The comparison results of the actual and predicted blood pressure values ​​after regression prediction using the machine learning model are shown below. Figure 7 As shown. The predicted value is compared with the subject's actual blood pressure value one by one. The accuracy and stability of the prediction results are evaluated by the mean absolute error (MAE). MAE is the average of the absolute differences between the predicted value and the actual value, which is used to measure the accuracy of the prediction. It is expressed by formula (6).

[0057]

[0058] After using machine learning models for regression prediction, the MAE (Moment of Effect) of the three algorithms was compared. The MAE of Gradient Boosting was 3.98, that of Random Forests was 3.6, and that of HPO+XGBoost was 3.07. The results show that the HPO+XGBoost algorithm has higher accuracy and stability in predicting blood pressure values.

Claims

1. A method for accurate blood pressure measurement based on a portable pulse oximeter, characterized in that: Includes the following steps: Step 1: Obtain pulse and blood pressure data for the subject's left and right hands; Step 2: Preprocess the left and right hand pulse data obtained in Step 1 to obtain the final pulse dataset; Step 3: Divide the final pulse dataset obtained in Step 2 into a training set and a test set in an 8:2 ratio to ensure that the samples for model training and testing are independent and to enhance the model's generalization ability. Step 4: Build a machine learning regression model based on Hunter Prey Optimization (HPO) to improve the accuracy and stability of blood pressure prediction; Specifically, it includes the following sub-steps: Step 4.1: Construct a machine learning regression model based on the HPO algorithm and optimize the model's weight parameters to improve the model's convergence speed and accuracy; Step 4.2: Combine the HPO algorithm from Step 4.1 with the eXtreme Gradient Boosting (XGBoost) regression model to effectively improve the model's performance; Step 4.3: Iteratively optimize the parameters of the improved machine learning regression model in Step 4.2, and apply an adaptive learning rate to dynamically adjust the learning rate, reduce overfitting and improve the stability of the model; The formula for the dynamic decay mechanism of the adaptive learning rate is: in, For the new learning rate, The initial learning rate, As the attenuation factor, For the current training round, The learning rate is set at the interval for updates. By dynamically adjusting the learning rate, the model can converge quickly in the early learning stages and gradually decrease the learning rate in the later stages to avoid parameter fluctuations, thereby reducing overfitting. Simultaneously, through an adaptive learning rate strategy, the model can dynamically adjust the optimization step size according to different data complexities, making model training more stable and improving overall prediction performance. For the loss function, mean squared error (MSE) is used as the evaluation criterion. The MSE formula is: in, For the sample size, For the true value, The model's predicted value; MSE can effectively reflect the degree of fit of the model to the real data. As the optimization target of the regression model, it helps the HPO algorithm to fine-tune the parameters, enabling the model to better capture subtle differences in blood pressure data. Step 5: Use the trained machine learning regression model to predict the subject's blood pressure.

2. The method for accurate blood pressure measurement based on a portable pulse oximeter as described in claim 1, characterized in that: The specific process of step 1 includes the following steps: Step 1.1: Collect pulse waveform data of the subject's left and right hands using a portable pulse diagnostic instrument, and simultaneously collect the corresponding blood pressure data to establish the correspondence between pulse and blood pressure; Step 1.2: Record the subject's blood pressure measurement time, pulse data, and corresponding physiological state to provide background information for data analysis.

3. The method for accurate blood pressure measurement based on a portable pulse oximeter as described in claim 1, characterized in that: The specific process of step 2 includes the following steps: Step 2.1: Clean the pulse data collected in Step 1.1, set upper and lower threshold values ​​for the pulse, remove outliers and interference factors, and ensure the accuracy of the data; Step 2.2: Based on Shannon's sampling theorem, set the sampling frequency and retain the pulse signal in the 0 to 50 Hz frequency band from the data obtained in Step 2.1 to improve the accuracy and reliability of pulse detection; Step 2.3: Use wavelet thresholding denoising algorithm to filter the pulse data in Step 2.2 to remove noise and obtain a smooth and accurate pulse signal; Step 2.3.1: The signal-to-noise ratio (SNR) and root mean square error (RMSE) are used to evaluate the denoising effect of different wavelet bases at different decomposition levels, so as to objectively measure the denoising performance and ensure that the signal quality is effectively improved. The SNR formula is shown in Equation (1), and the RMSE formula is shown in Equation (2). （1） （2） in, It is the original signal component. These are the denoised signal components; Step 2.3.2: The preprocessed pulse signal is decomposed into six levels and multiple scales using the Bior 3.5 wavelet. Its excellent frequency band segmentation capability enables effective extraction of signal features. This decomposition can accurately separate physiological features and noise components in different frequency bands, providing a reliable signal basis for subsequent denoising and feature analysis. Step 2.3.3: Use a soft threshold function to smoothly reduce the signal amplitude close to the threshold in step 2.3.2, thereby reducing the influence of noise and making the signal transition smoother; the formula for the soft threshold function is shown in equation (3): （3） in, These are wavelet coefficients. It is the threshold, representing the degree of noise reduction; Step 2.3.4: By calculating the noise estimate of each layer, the threshold value in step 2.3.3 is dynamically adjusted according to the noise level and signal characteristics, so that the high-frequency layer can remove more noise, while the low-frequency layer can retain more signal information.

4. The method for accurate blood pressure measurement based on a portable pulse oximeter as described in claim 1, characterized in that: The specific process of step 5 includes the following steps: Step 5.1: Input the preprocessed training set into the optimized regression model in Step 4.3 to further optimize the model's prediction accuracy. Through multiple iterations of training, the model gradually enhances its ability to identify and extract pulse signal features. A strategy of dynamically adjusting the learning rate is used to improve training efficiency, and the sample proportions in each training batch are controlled so that samples with systolic blood pressure less than 120 mmHg, 120-140 mmHg, and greater than 140 mmHg account for 30%, 50%, and 20% respectively, to optimize data distribution. When the loss decrease rate is less than 1% after 10 consecutive iterations, the model is considered converged and training is terminated. These measures collectively improve the prediction accuracy of blood pressure. During training, the convergence of the model is also judged by monitoring the decreasing trend of the loss function, ensuring that the model stops training after reaching optimal prediction performance. Step 5.2: Use the trained regression model to predict blood pressure using the pulse data of the subjects in the test set divided in Step 3. Post-process the prediction results, automatically remove abnormal prediction values ​​with systolic blood pressure less than 70 mmHg or greater than 200 mmHg, and retest. Calculate the average of the five consecutive prediction values ​​as the final output. Compare the processed prediction results with the actual blood pressure to evaluate the model's generalization performance and clinical applicability. Compare the predicted values ​​with the subjects' actual blood pressure values ​​one by one, and evaluate the accuracy and stability of the prediction results using the mean absolute error (MAE).

Images