A PPG blood pressure monitoring method and system

Abstract

The application discloses a PPG blood pressure monitoring method and system, which comprises the following steps: collecting a user PPG signal by using a narrow-spectrum light source and a narrow-band response organic light detector, and performing sliding window segmentation on the user PPG signal; performing pretreatment on the expanded data set to obtain 20-dimensional features of each window; training an initial blood pressure prediction model by using a plurality of machine learning algorithms and the 20-dimensional features; optimizing the 20-dimensional features to obtain an optimal feature subset combination of each model; retraining the initial blood pressure prediction model by using the optimal feature subset combination, comparing the model performance with the training result obtained in the last step, and finally obtaining a personalized blood pressure prediction model; and the method effectively reduces environmental light interference, optimizes PPG signal quality, and realizes all-weather high-precision blood pressure monitoring by synergistically using a narrow-spectrum red light source and a narrow-band organic photoelectric detector and combining an innovative intelligent waveform processing algorithm.

Description

Technical Field

[0001] This invention relates to the field of physiological parameter monitoring, and in particular to a PPG blood pressure monitoring method and system. Background Technology

[0002] With an aging population and an increase in chronic diseases, blood pressure monitoring has become an important component of health management and is receiving increasing attention. Traditional blood pressure measurement methods mainly rely on cuff-type blood pressure monitors, which are semi-invasive and can only provide blood pressure data at a single point in time, failing to achieve continuous monitoring. Furthermore, the frequent inflation process can cause discomfort for the user. In recent years, non-invasive blood pressure monitoring technology based on photoplethysmography (PPG) has gained popularity due to its portability and continuous monitoring capabilities.

[0003] PPG technology is a method that uses photoelectric sensors to detect changes in blood vessel volume, and can be used to monitor physiological parameters such as heart rate, blood oxygen saturation, and blood pressure. Currently, there are various blood pressure monitoring devices based on PPG technology on the market, such as a blood pressure measurement system based on a flexible organic photodetector and deep learning algorithm, which is published in Chinese Patent Publication No. CN116548935A. This system achieves continuous monitoring of blood pressure values ​​by wearing a flexible organic photodetector and ECG electrodes on the wrist.

[0004] Existing PPG blood pressure monitoring technology still has the following problems:

[0005] First, PPG signals are low-frequency signals, making them susceptible to interference from ambient light noise, power frequency noise, and circuit disturbances during acquisition. Existing technologies primarily use broadband light sources with broad emission spectra, which easily alias with ambient light, leading to signal quality degradation. Furthermore, conventional photodetectors respond to light across all wavelengths, failing to effectively filter out interfering light signals in non-characteristic wavelength bands.

[0006] Secondly, low-quality PPG signals can negatively impact the training performance of algorithms, leading to a significant reduction in accuracy. Existing blood pressure prediction models are mostly based on traditional machine learning algorithms or more complex neural network architectures, which have high requirements for signal quality. To achieve optimal prediction results, a significant amount of time and effort is typically required to fine-tune the model architecture and parameters. Furthermore, it is often impossible to achieve a simultaneous balance between local computing resources and model performance. Complex models prioritizing performance encounter numerous difficulties in local deployment in real-world application scenarios, especially when the input signal is interfered with, resulting in a significant decrease in prediction accuracy.

[0007] Finally, current technologies cannot achieve high-precision cuffless blood pressure monitoring. Due to the complex and nonlinear relationship between PPG signals and blood pressure, existing PPG-based blood pressure monitoring methods still have significant room for improvement in terms of accuracy and stability, especially in continuous monitoring scenarios under varying ambient light conditions throughout the day.

[0008] Therefore, there is an urgent need to develop a technical solution that can acquire high-quality PPG signals and achieve accurate blood pressure monitoring under various environmental conditions. Summary of the Invention

[0009] In order to overcome the above-mentioned shortcomings and deficiencies of the prior art, the purpose of this invention is to provide a PPG blood pressure monitoring method and system.

[0010] The objective of this invention is achieved through the following technical solution:

[0011] A PPG blood pressure monitoring method, comprising:

[0012] User PPG signals were collected using a narrow-spectrum light source and a narrow-band response organic photodetector to form a dataset.

[0013] The user's PPG signal is divided into sliding window segments to expand the dataset;

[0014] The expanded dataset is preprocessed by using an intelligent waveform segmentation and feature extraction algorithm to extract 20-dimensional waveform features for each window, discarding incomplete waveforms, and then extracting the mean of the complete waveform feature data within each window to obtain 20-dimensional features for each window.

[0015] The initial blood pressure prediction model was trained using a variety of machine learning algorithms and 20-dimensional features.

[0016] Based on the recursive feature elimination cross-validation algorithm, the 20-dimensional features are optimized to obtain the optimal feature subset combination for each model;

[0017] The initial blood pressure prediction model is retrained using the optimal feature subset combination, and its performance is compared with the training results obtained in the previous step to finally obtain a personalized blood pressure prediction model.

[0018] The real-time acquired PPG signal is processed by intelligent waveform segmentation and feature extraction algorithms to extract features. After constructing a feature subset combination according to the requirements of the personalized blood pressure prediction model, the feature is finally input into the personalized blood pressure prediction model to obtain the blood pressure prediction value.

[0019] Furthermore, the narrow-spectrum light source is specifically a narrow-spectrum red light source with a human-safe power of 627nm emission peak and 3.2nm half-width, prepared from Mn4+-doped fluorinated gallate single-crystal red light material.

[0020] Furthermore, the narrowband-response photodetector is fabricated by combining an optical depletion layer with a specific absorption spectrum and an organic photoactive layer, resulting in a narrowband-response organic photodetector with a response peak of 655 nm. The fabrication conditions for the optical depletion layer and the organic photoactive layer are as follows:

[0021] Optical depletion layer: P3HT:J71:IEICO-4F (1:1:1), dissolved in chloroform (CF), spin-coated to form a film of about 400-500 nm, and heat-annealed at 100℃ / 10 min;

[0022] Organic photoactive layer: PM6:L8-BO (1:1.2), dissolved in chloroform (CF), spin-coated to form a film of about 150-200 nm, and heat-annealed at 100℃ for 10 min.

[0023] Furthermore, the process of performing sliding window segmentation on the user's PPG signal to expand the dataset specifically involves:

[0024] An 875-point sliding window was used to segment the 5000-point PPG signal acquired each time;

[0025] Each window inherits the SBP / DBP tags from the original file;

[0026] Valid windows are automatically filtered through waveform quality assessment;

[0027] Achieve a 5-10x scaling of the dataset from file level to window level.

[0028] Furthermore, the intelligent waveform segmentation and feature extraction algorithm includes:

[0029] Intelligent segmentation of PPG waveforms based on valley detection;

[0030] Tidal peak value is identified by first derivative analysis;

[0031] Automatically detect and discard incomplete waveforms;

[0032] Extract 20-dimensional waveform features;

[0033] The mean value of the complete waveform features is calculated to generate a 20-dimensional feature window.

[0034] Furthermore, the PPG signal includes the main wave, tidal wave, descending isthmus, and diabetic wave.

[0035] Furthermore, the 20-dimensional features of each window include peak features, slope features, ratio features, time features, and time ratio features.

[0036] Furthermore, the various machine learning algorithms include SVR, GBR, RF, LR, and XGB.

[0037] Furthermore, the 20-dimensional waveform features include:

[0038] Peak characteristics: systolic blood pressure peak, tidal wave peak, diastolic blood pressure peak, trough value;

[0039] Slope characteristics: the rising / falling slope between each peak;

[0040] Ratio characteristics: the relative relationship between peaks and troughs;

[0041] Temporal characteristics: pulse width, diastolic time interval;

[0042] Time ratio characteristic: the relative relationship of time intervals.

[0043] A system for implementing the PPG blood pressure monitoring method, comprising:

[0044] The signal acquisition module uses a fully integrated analog front-end module, which connects a narrow-spectrum light source and a narrow-band response organic photodetector to acquire user PPG signals.

[0045] The PPG signal processing module is used to control and drive the narrow-spectrum light source and narrow-band response organic photodetector to read signals and features in a time sequence, and to transmit the collected PPG signals and blood oxygen data to the host computer via the UART protocol.

[0046] The host computer preprocesses and extracts features from the PPG signal, then inputs it into the personalized blood pressure prediction model to obtain the blood pressure value.

[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0048] (1) This invention uses a narrow-spectrum red light source and a narrow-spectrum red light detector to collect PPG signals. Compared with commercial blood oxygen probes, the PPG signals collected by this strategy have more significant details. After using intelligent waveform segmentation and feature extraction algorithms to extract and analyze the morphological features of pulse waves, higher blood pressure monitoring accuracy can be achieved.

[0049] (2) The narrow-band organic photodetector described in this invention can adjust the response band and responsivity of the organic photodetector by flexibly controlling the thickness and material of the optical depletion layer. The device is prepared by solution processing and vapor deposition, which does not have the problem of lattice mismatch in processes such as lattice growth and vapor deposition. It has the feasibility and simplicity of operation and is expected to be applied in the fields of flexible wearable devices, mobile sensing and medical health monitoring.

[0050] (3) The narrowband response organic photodetector described in this invention has a strong response to light in only a single range of wavelengths. It can filter ambient light in the presence of ambient light interference and still collect high-quality PPG signals during the PPG signal acquisition process, thus realizing all-weather blood pressure monitoring.

[0051] (4) The complete algorithm process described in this invention has clear logic and innovation. First, it solves the problem of poor model generalization ability caused by insufficient user-personalized sample size in traditional methods by using small sample dataset expansion technology. It achieves a 5-10 times expansion of the dataset, providing sufficient training data for subsequent algorithms and ensuring that the best personalized model that best suits the user can be built under a limited dataset.

[0052] (5) The intelligent waveform segmentation and feature extraction algorithm described in this invention achieves accurate extraction of key physiological features in PPG signals through adaptive segmentation based on valley detection and intelligent identification of tidal wave features. The feature point positioning accuracy is significantly improved, laying a solid foundation for high-precision blood pressure prediction. It is perfectly adapted to the high-quality PPG signals with richer physiological health information collected by this system, and solves the problems of inaccurate and incomplete feature extraction in traditional methods.

[0053] (6) The feature subset selection and optimization algorithm described in this invention, based on RFECV multi-model feature selection, can customize the optimal feature combination for different users, realize personalized modeling, significantly improve the accuracy and stability of blood pressure prediction, eliminate feature redundancy and overfitting problems, and improve the generalization ability of the model.

[0054] (7) The personalized model building technology described in this invention achieves high-precision blood pressure prediction customized by users through a feature subset retraining strategy, which solves the problems of insufficient model generalization ability and low personalization in traditional methods, and can obtain the best personalized model that best suits the user under limited datasets.

[0055] (8) The all-weather monitoring technology described in this invention, through the collaborative work of a narrow-spectrum light source and a narrow-band detector, combined with an innovative intelligent signal processing algorithm, can maintain stable monitoring performance under various environmental conditions, meet the needs of continuous clinical monitoring, and realize high-precision cuffless blood pressure monitoring.

[0056] (9) The present invention describes an all-weather PPG blood pressure monitoring system based on a narrow-spectrum light source and a narrow-band response organic photodetector. Its intelligent algorithm supports customized feature selection and personalized modeling. The system supports cyclic acquisition, and the model algorithm can perform real-time calculation and update. It can realize personalized high-precision blood pressure parameter prediction and meet the needs of clinical applications.

[0057] (10) The all-weather PPG blood pressure monitoring system described in this invention, based on a narrow-spectrum light source and a narrow-band response organic photodetector, is simple to operate through system optimization and programming. Through the collaborative work of the computer, microcontroller, and analog acquisition front-end chip, real-time blood pressure monitoring and feedback can be achieved simply by placing a finger into the probe, resulting in a good user experience and clinical applicability. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the workflow of the present invention;

[0059] Figure 2 This is the excitation spectrum of the narrow-spectrum light source used in this invention;

[0060] Figure 3 This is the emission spectrum of the narrow-spectrum light source used in this invention;

[0061] Figure 4 This is a graph showing the external quantum efficiency data of the narrowband-response organic photodetector used in this invention;

[0062] Figure 5 This is a graph showing the specific detectivity data of the narrowband-response organic photodetector used in this invention;

[0063] Figure 6 This is a reference diagram of the high-quality pulse wave morphology of the present invention;

[0064] Figure 7 This invention is based on a PPG signal image acquired by a commercially available pulse oximeter.

[0065] Figure 8 This is a PPG signal image acquired by the present invention based on a narrow-spectrum light source and a narrow-band organic photodetector;

[0066] Figure 9 This is a diagram showing the effect of expanding the small sample dataset of this invention;

[0067] Figure 10 This is a schematic diagram of the intelligent waveform segmentation and feature extraction algorithm of the present invention;

[0068] Figure 11 This is a graph showing the performance optimization of the model in this invention as a feature subset is selected;

[0069] Figure 12 This is a comparison chart of model performance before and after personalized modeling in this invention. Detailed Implementation

[0070] The present invention will be further described in detail below with reference to the embodiments, but the implementation of the present invention is not limited thereto.

[0071] Example

[0072] like Figure 1 As shown, a PPG blood pressure monitoring method includes the following steps:

[0073] S1 is used to prepare a narrow-spectrum light source with safe power for human use, concentrated emission spectrum, and small half-peak width.

[0074] Specifically, a narrow-spectrum red light source with a peak emission of 627 nm and a full width at half maximum (FWHM) of 3.2 nm, meeting human safety power requirements, was fabricated using Mn4+-doped fluorinated gallate single-crystal red light material. The fabrication process included: firstly, preparing fluoride single-crystal red light material using a saturated solution crystallization method; and then combining and packaging the Mn4+-doped fluorinated gallate single-crystal red light material with a blue InGaN chip to obtain the final narrow-spectrum red light source. The luminous intensity of this narrow-spectrum light source depends entirely on the commercially available blue LED used in the packaging. Unlike lasers, this narrow-spectrum light source operates within a human safety power range and can be used for long-term human health monitoring. Furthermore, the emission peak of this light source remains stable at 627 nm and supports stable operation over extended periods. Compared to existing commercial light sources, the emission spectrum exhibits minimal thermal drift and dispersion with increasing temperature, while maintaining a constant FWHM of 3.2 nm. The excitation spectrum of the fabricated narrow-spectrum light source is shown below. Figure 2 As shown, the emission spectrum of the narrow-spectrum light source is as follows: Figure 3 As shown.

[0075] S2: Fabricate an organic photodetector with a corresponding narrowband response.

[0076] Specifically, a narrowband response organic photodetector with a response peak of 655 nm and a half-maximum width of 85 nm is fabricated, comprising, in sequence, an optical depletion layer, a transparent substrate, a conductive anode, a hole transport layer, an organic photoactive layer, an electron transport layer, and a conductive cathode.

[0077] The optical depletion layer is selected based on the target response band of the device, using donor materials, acceptor materials, or a blend of donor and acceptor materials. The final target response band corresponds to the weak absorption region of the optical depletion layer's absorption spectrum. In this embodiment, P3HT:J71:IEICO-4F (1:1:1) is used as the optical depletion layer material, and a thin film of 450–500 nm is prepared and thermally annealed at 100°C for 10 min. The substrate is glass; the conductive anode is ITO; the hole transport layer is PEDOT:PSS with a thickness of 5–10 nm; the organic photoactive layer is a blend of donor material PBDB-T-2F (PM6) and acceptor material L8-BO, preparing a thin film of 150–200 nm, and thermally annealed at 100°C for 10 min; the electron transport layer is PFN-Br with a thickness of 5–10 nm; and the conductive cathode is aluminum with a thickness of 80 nm. Figure 4 and Figure 5As shown, the detector features a narrow-band response spectrum with a peak value of 655 nm and a half-maximum width of 85 nm, low dark current, high specific detectivity, and high sensitivity, making it particularly suitable for the precise detection of biomedical signals.

[0078] S3 uses a narrow-spectrum light source and a narrow-band response organic photodetector to obtain PPG signals with richer physiological health information and higher quality, forming a dataset. The waveform characteristics of the PPG signal include the main wave, tidal wave, descending isthmus, and dicrotic wave (also referred to as: systolic peak, tidal wave, descending isthmus, and diastolic peak, respectively). In this embodiment, the user places their finger between the narrow-spectrum light source and the narrow-band response organic photodetector, loads the blood pressure model on the computer interface, and clicks "Start Acquisition" to display the user's real-time PPG waveform and corresponding blood pressure value.

[0079] Specifically, the narrow-spectrum red light source prepared in step one and the narrow-band response organic photodetector prepared in step two are respectively connected to the programmable integrated LED driver and low-noise detector receiving channel of the analog front-end chip AFE4490. The 627nm red light emitted by the narrow-spectrum light source penetrates the skin and interacts with hemoglobin in the blood. Part of the light is absorbed, and part is scattered, ultimately being received by the narrow-band response organic photodetector. Because the concentration of hemoglobin in the blood changes periodically with heartbeats, the received light signal also exhibits periodic changes, forming a high-quality PPG signal with obvious tidal characteristics, significantly different from commercial pulse oximeters.

[0080] By synergistically utilizing a narrow-spectrum red light source and a narrow-band organic photodetector with a customized wavelength response, interference from ambient light and motion artifacts can be effectively reduced. Because the emission spectrum of the narrow-spectrum light source is concentrated around 627 nm with a full width at half maximum (FWHM) of only 3.2 nm, compared to broadband light sources, the narrow-spectrum red light source ensures more consistent tissue penetration and reduces wavelength-dependent scattering, thereby minimizing signal broadening caused by tissue heterogeneity. The organic photodetector's response peak is at 655 nm with a FWHM of only 85 nm, thus enabling the formation of a physical-level feature acquisition correspondence of "characteristic emission + characteristic absorption." The system therefore possesses a natural suppression effect on other wavelength components of ambient light. Simultaneously, the narrow-band response characteristics also reduce interference from scattered light from the skin surface, improving the signal-to-noise ratio.

[0081] By comparing the PPG signal waveforms obtained from commercial pulse oximeters, such as Figure 7 and Figure 8 As shown, combined with high-quality pulse wave morphology reference Figure 6The results show that the method of the present invention can significantly improve the quality of PPG signals and reduce interference from non-characteristic band noise signals, thereby enabling PPG signals to display richer physiological health information. Specifically, the signal waveform is smoother, the original signal noise is lower, and detailed features such as tidal waves are more obvious. The morphology and amplitude of these features are closely related to vascular elasticity, peripheral resistance, and the state of cardiovascular coupling.

[0082] S4 performs sliding window segmentation on the small sample PPG signal dataset, expanding the dataset from file level to window level by 5-10 times, in order to solve the problem of poor model generalization ability caused by insufficient sample size in traditional PPG blood pressure monitoring.

[0083] Specifically, small sample dataset expansion technology solves the key problem of insufficient personalized user sample size in traditional PPG blood pressure monitoring, such as... Figure 9 As shown, an 875-point sliding window is used to segment 5000 PPG signals, with a window step size of 250 points, achieving a 5-10 times expansion of the dataset. Each window inherits the SBP / DBP labels from the original file to ensure data consistency. In conjunction with the intelligent waveform segmentation and feature extraction algorithms in S5, waveform features are extracted from each window to obtain 20-dimensional waveform features of each high-quality PPG signal within the window. After excluding abnormal or incomplete waveforms, the waveform feature data within the window is averaged to eliminate weight interference from accidental waveform errors, ultimately forming the expanded dataset. This technique allows for the generation of sufficient training samples even with only a small number of user-personalized PPG files, reducing the number of repeated user data collections and maximizing the utilization of PPG data from a single collection, effectively solving the problem of small user-personalized datasets. This expansion method not only increases the number of samples but, more importantly, maintains the physiological continuity and label consistency of the data, providing a reliable data foundation for subsequent model training.

[0084] Further explanation: The waveform quality of each window is automatically evaluated through intelligent algorithms, including signal integrity checks: ensuring that the window contains the complete pulse cycle and avoiding the impact of truncated waveforms on feature extraction; noise level assessment: calculating the signal-to-noise ratio, screening high-quality signals, and ensuring the purity of training data; and baseline stability analysis: detecting baseline drift, ensuring signal quality, and improving the accuracy of feature extraction.

[0085] S5 employs an intelligent waveform segmentation and feature extraction algorithm to extract 20-dimensional waveform features from each window, discarding incomplete waveforms. Then, it extracts the mean of the complete waveform feature data within each window to obtain 20-dimensional features for each window, ensuring the consistency and reliability of feature quality.

[0086] Specifically, the intelligent waveform segmentation and feature extraction algorithm achieves accurate extraction of key physiological features from PPG signals, such as... Figure 10 As shown, the algorithm can automatically identify key physiological features in PPG signals and intelligently discard incomplete waveforms to ensure feature quality.

[0087] To further explain, the intelligent waveform segmentation and feature extraction algorithm comprises the following specific steps:

[0088] S5.1 Intelligent waveform segmentation based on valley detection;

[0089] The algorithm first performs trough detection on the preprocessed PPG signal and then uses the zero-crossing point identification method based on the first derivative of the signal to achieve accurate waveform segmentation. Unlike traditional fixed-length segmentation methods, this algorithm can adaptively adjust the segmentation points according to the physiological characteristics of the PPG signal, ensuring that each segment contains a complete pulse cycle. The specific implementation method is as follows:

[0090] (1) Calculate the first derivative of the PPG signal and identify the zero-crossing points of the derivative;

[0091] (2) Find the local minimum near the zero crossover point to determine the location of the trough;

[0092] (3) Divide the waveform according to the trough position to ensure that each segment contains a complete pulse cycle;

[0093] (4) Verify the physiological rationality of the segmented fragments and discard abnormal fragments.

[0094] S5.2 Intelligent Recognition Technology for Tidal Wave Characteristics;

[0095] Tidal waves are a key physiological characteristic that distinguishes high-quality PPG signals acquired by this system from ordinary PPG signals, such as... Figure 7 and Figure 8 As shown, combined with high-quality pulse wave morphology reference Figure 6 It is closely related to vascular elasticity and peripheral resistance, such as Figure 10 As shown. This algorithm intelligently identifies the peak position of tidal waves by analyzing the negative slope segment of the first derivative of the signal. The specific implementation method is as follows:

[0096] (1) Find the negative slope segment of the first derivative after the peak of systolic blood pressure;

[0097] (2) Sort by duration and select the three longest negative slope segments;

[0098] (3) Determine the peak tidal wave value and peak diastolic pressure value by their positional relationship;

[0099] (4) Automatically verify the physiological rationality of feature points.

[0100] The extracted 20-dimensional waveform features, as shown in Table 1, include:

[0101] Peak characteristics (4 dimensions): peak systolic blood pressure, peak tidal wave, peak diastolic blood pressure, and bottom value, reflecting vascular elasticity and peripheral resistance;

[0102] Slope characteristics (6 dimensions): The rising / falling slope between each peak reflects the rate of hemodynamic change;

[0103] Ratio feature (3D): The relative relationship between peak and trough, eliminating the influence of individual differences and enhancing feature stability;

[0104] Temporal characteristics (4 dimensions): pulse width and diastolic interval (DEIT), reflecting cardiac cycle and blood flow time;

[0105] Time ratio feature (3D): The relative relationship of time intervals, which enhances feature stability and improves the model's generalization ability.

[0106] Table 1

[0107]

[0108] S5.3 Intelligent Waveform Quality Assessment and Screening

[0109] The algorithm automatically detects the integrity of each segmented waveform and evaluates waveform quality using the following criteria:

[0110] (1) Feature point integrity: Check whether all key feature points have been successfully identified to ensure the integrity of feature extraction.

[0111] (2) Physiological rationality: Verify whether the feature values ​​are within the physiological range and avoid the impact of outliers on model training.

[0112] (3) Signal quality: Evaluate signal-to-noise ratio and baseline stability to ensure the reliability of features.

[0113] The algorithm automatically discards incomplete or low-quality waveforms to ensure the reliability of the final feature vectors and provide high-quality feature inputs for subsequent model training.

[0114] S6 uses 20-dimensional features to train the initial blood pressure prediction model based on the expanded dataset. In this embodiment, the initial blood pressure prediction model includes a diastolic blood pressure model and a systolic blood pressure model.

[0115] Specifically, the initial model building technique employs multiple machine learning algorithms to provide a foundation for subsequent feature subset selection. It supports various algorithms such as SVR, GBR, RF, LR, and XGB, each trained using complete 20-dimensional waveform features to ensure the comprehensiveness and robustness of model selection. The model performance of each algorithm is evaluated on a validation set segmented from the extended dataset, providing a performance benchmark for subsequent feature subset selection.

[0116] S7 is based on the Recursive Feature Elimination Cross-Validation (RFECV) algorithm, which performs feature subset selection and optimization on 20-dimensional features. It customizes the optimal feature subset combination for different users, eliminates redundant features, and improves model accuracy and generalization ability. Figure 11 As shown.

[0117] Multi-model feature selection based on RFECV

[0118] The Recursive Feature Elimination Cross-Validation (RFECV) algorithm is employed, performing feature selection independently for each machine learning algorithm to ensure the comprehensiveness and accuracy of the selection. By recursively eliminating the least important features and using cross-validation to evaluate model performance, the RFECV algorithm can find the optimal feature subset.

[0119] Characteristic stability analysis

[0120] Consistency in feature selection was ensured through multiple validations. A stability threshold mechanism was employed, selecting waveform features with an occurrence frequency exceeding 50% as stable features to avoid overfitting. Feature stability analysis ensured the reliability of the feature selection results, preventing instability caused by differences in data partitioning.

[0121] Goal-oriented optimization

[0122] For SBP and DBP, optimal feature combinations are selected respectively, taking into account the different feature requirements of different blood pressure indicators to achieve precise feature customization. SBP and DBP have different physiological mechanisms and influencing factors. By optimizing feature combinations separately, their respective physiological characteristics can be better captured.

[0123] User Adaptability Analysis

[0124] By adjusting the feature selection strategy based on the characteristics of user PPG signals, personalized feature subset configuration can be achieved, laying the foundation for personalized modeling. Different users' PPG signals have different feature patterns; through user adaptability analysis, the most suitable feature combination can be selected for each user.

[0125] S8 retrains the optimized feature subset into a new blood pressure prediction model and compares its performance with the model trained in S6. The superior model is selected as the high-precision personalized blood pressure prediction model, achieving accurate modeling of user physiological characteristics and realizing customized high-precision blood pressure prediction. Figure 12 As shown:

[0126] Feature subset retraining

[0127] The optimized feature subset is then retrained into the model to ensure an optimal match between the model and the features. By retraining the model using the optimized feature subset, the influence of redundant features can be eliminated, improving the model's accuracy and generalization ability.

[0128] Feature adaptation optimization

[0129] Adjusting the model structure based on a subset of user features ensures an optimal match between the model and the features. Through feature adaptation optimization, specific physiological characteristics of users can be fully utilized to improve the model's prediction accuracy.

[0130] The S9 extracts features from the real-time PPG signal using intelligent waveform segmentation and feature extraction algorithms. Based on the feature subset requirements of the final high-precision personalized blood pressure prediction model, it constructs a personalized model feature combination and finally inputs it into the high-precision personalized blood pressure prediction model to obtain an accurate blood pressure prediction value.

[0131] Specifically, by inputting high-quality PPG data collected from users into a trained model, more accurate systolic blood pressure (SBP) and diastolic blood pressure (DBP) can be obtained. Compared with traditional methods, the method of this invention significantly improves the accuracy of blood pressure prediction.

[0132] Example 2

[0133] An all-weather PPG blood pressure monitoring system based on a narrow-spectrum light source and a narrow-band response organic photodetector, such as Figure 1 As shown, it includes:

[0134] The signal acquisition module, through a fully integrated analog front-end chip, coordinates the timing of the narrow-spectrum light source and the narrow-band response organic photodetector to acquire high-quality PPG signals from the user.

[0135] The PPG signal processing module obtains the high-quality PPG signal from the user through communication between the microcontroller and the fully integrated analog front-end chip based on the SPI communication protocol, and then transmits the PPG signal to the host computer via the UART communication protocol.

[0136] The host computer, after obtaining high-quality PPG signal data from the user via the UART communication protocol, parses the data according to the protocol and inputs it into an intelligent waveform segmentation and feature extraction algorithm to obtain 20-dimensional high-quality waveform features. Then, feature combinations are selected according to the requirements of the personalized blood pressure prediction model. Finally, the selected feature combinations are input into the model to obtain a high-precision blood pressure value. The real-time PPG waveform and predicted blood pressure value collected by the user are displayed on the front-end user interface.

[0137] The signal acquisition module uses TI's AFE4490 fully integrated analog front-end chip, mainly including a programmable integrated LED driver, a 22-bit integrated analog-to-digital converter (ADC), a low-noise detector receiving channel, and an LED fault diagnosis circuit. It can perfectly achieve programmable drive current intensity control for narrow-spectrum light sources, high-precision analog-to-digital conversion of narrow-band response organic photodetector output signals, and TIA transimpedance amplification and hardware low-pass filtering of PPG signals. Through flexible pulse sequencing and timing control, it can acquire high-quality PPG signals from users.

[0138] Specifically, the narrow-spectrum light source in the signal acquisition module is made of Mn4+-doped fluorinated gallium salt single-crystal red light-emitting material, with an emission peak of 627nm and a full width at half maximum (FWHM) of 3.2nm. The power is controlled within a safe range for human use. This light source is connected to the programmable integrated LED driver of the AFE4490 chip, which provides the driving current through a programmable current source within the chip. The AFE4490 chip supports high-precision LED driving current adjustment within the 0-200mA range, while also featuring a 110dB dynamic range (achieving low noise at low LED currents). It can flexibly adjust the light source intensity according to different users' skin characteristics and environmental conditions to ensure optimal signal quality.

[0139] The narrowband response organic photodetector has a peak response of 655nm and a full width at half maximum (FWHM) of 85nm, and is connected to the low-noise detector receiver channel of the AFE4490 chip. This channel includes a transimpedance amplifier (TIA) that converts the weak current signal generated by the photodetector into a voltage signal, and provides seven programmable gain settings (feedback resistor RF range: 10kΩ, 25kΩ, 50kΩ, 100kΩ, 250kΩ, 500kΩ, 1MΩ). After amplification by the TIA, the signal is further amplified by a programmable gain amplifier (PGA), with a gain range of 0dB / 3.5dB / 6dB / 9.5dB / 12dB (corresponding to linear gains of 1× / 1.5× / 2× / 3× / 4×). Then, the signal undergoes preliminary filtering by a low-pass filter with a fixed 500Hz cutoff frequency to remove high-frequency noise. Finally, the processed signal is fed into a 22-bit high-precision ADC for digital conversion. The sampling rate is determined by the pulse repetition frequency (PRF), up to 4×PRF (PRF range: 62.5SPS to 5k SPS, corresponding to a maximum sampling rate of 20kSPS), ensuring that subtle changes in the PPG signal are captured.

[0140] The AFE4490 chip also offers flexible timing control functions, which can precisely control the LED lighting time and sampling time, enabling time-division multiplexing operation, reducing power consumption and improving signal quality. The chip's built-in LED fault diagnosis circuit can monitor the LED's operating status in real time, ensuring reliable system operation.

[0141] The PPG signal processing module employs a high-performance microcontroller with a clock frequency of 20MHz, featuring 2KB of RAM and 32KB of Flash memory, sufficient to handle complex signal processing tasks. The microcontroller communicates with the AFE4490 chip via the SPI protocol, writing relevant register configuration instructions during system initialization to control the driving of the narrowband light source and narrowband organic photodetector to read signals sequentially. SPI communication uses full-duplex mode with a clock frequency of 2MHz to ensure efficient data transmission.

[0142] After receiving the raw PPG data from the AFE4490 chip, the microcontroller performs preliminary data processing, including data format conversion and simple filtering. The processed PPG signal data is then transmitted to the host computer via the UART communication protocol. The UART communication uses a baud rate of 115200 to ensure stable and reliable data transmission.

[0143] The host computer consists of two parts: a front-end and a back-end. The front-end adopts a modern UI design with an intuitive user interface, capable of displaying the acquired PPG waveform and real-time updated blood pressure values. The interface also provides various control buttons, allowing users to start / stop data acquisition, freeze the waveform window, save historical data, and perform other operations.

[0144] The backend reads and parses the PPG signal data transmitted from the microcontroller via the UART protocol, and performs data preprocessing on the PPG data, including software bandpass filtering and baseline correction. The software bandpass filtering uses a Butterworth filter with a cutoff frequency set to 0.8-10Hz to effectively remove high-frequency noise; baseline correction uses a moving average method with a 5-second window size to effectively eliminate baseline drift. The preprocessed PPG signal data is input into an intelligent waveform segmentation and feature extraction algorithm to obtain 20-dimensional high-quality waveform features. Feature combinations are then selected according to the requirements of the personalized blood pressure prediction model, and finally, the selected feature combinations are input into the model to obtain a high-precision blood pressure prediction value.

[0145] The system's core advantage lies in the collaborative operation of a narrow-spectrum light source and a narrow-band response organic photodetector, effectively reducing ambient light interference and motion artifacts. This results in high-quality PPG signals that are unaffected by ambient light interference. Simultaneously, by fully utilizing the high-quality PPG signals acquired from users, the system employs techniques such as small-sample dataset expansion, intelligent waveform feature extraction, and recursive feature elimination cross-validation algorithms to optimize feature subsets, constructing a high-precision, personalized blood pressure prediction model for each user. This leads to more accurate blood pressure predictions and ultimately achieves high-precision, all-weather PPG blood pressure health monitoring. Compared to traditional blood pressure monitoring methods, this system eliminates the need for a cuff, simplifying operation, providing a better user experience, and enabling continuous monitoring for more comprehensive blood pressure health management.

[0146] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the embodiments described above. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A PPG blood pressure monitoring method, characterized in that, include: User PPG signals were collected using a narrow-spectrum light source and a narrow-band response organic photodetector to form a dataset. The user's PPG signal is divided into sliding window segments to expand the dataset; The expanded dataset is preprocessed by using an intelligent waveform segmentation and feature extraction algorithm to extract 20-dimensional waveform features for each window, discarding incomplete waveforms, and then extracting the mean of the complete waveform feature data within each window to obtain 20-dimensional features for each window. The initial blood pressure prediction model was trained using a variety of machine learning algorithms and 20-dimensional features. Based on the recursive feature elimination cross-validation algorithm, the 20-dimensional features are optimized to obtain the optimal feature subset combination for each model; The initial blood pressure prediction model is retrained using the optimal feature subset combination, and its performance is compared with the training results obtained in the previous step to finally obtain a personalized blood pressure prediction model. The real-time acquired PPG signal is processed by intelligent waveform segmentation and feature extraction algorithms to extract features. After constructing a feature subset combination according to the requirements of the personalized blood pressure prediction model, the feature is finally input into the personalized blood pressure prediction model to obtain the blood pressure prediction value. The narrow-spectrum light source is specifically a human-safe power narrow-spectrum red light source with an emission peak of 627 nm and a half-width of 3.2 nm, prepared from Mn4+-doped fluorinated gallate single-crystal red light material. Narrowband-response photodetectors are fabricated by combining an optical depletion layer with a specific absorption spectrum and an organic photoactive layer, resulting in a narrowband-response organic photodetector with a response peak of 655 nm. The fabrication conditions for the optical depletion layer and the organic photoactive layer are as follows: Optical depletion layer: P3HT:J71:IEICO-4F, soluble in chloroform, spin-coated to form a film of about 400~500 nm, and heat-annealed at 100℃ / 10 min; Organic photoactive layer: PM6:L8-BO, soluble in chloroform, spin-coated to form a film of about 150~200 nm, heat-annealed at 100℃ / 10min.

2. The PPG blood pressure monitoring method according to claim 1, characterized in that, The process of performing sliding window segmentation on the user's PPG signal to expand the dataset involves the following steps: An 875-point sliding window was used to segment the 5000-point PPG signal acquired each time; Each window inherits the SBP / DBP tags from the original file; Valid windows are automatically filtered through waveform quality assessment; Achieve a 5-10x scaling of the dataset from file level to window level.

3. The PPG blood pressure monitoring method according to claim 1, characterized in that, The intelligent waveform segmentation and feature extraction algorithm includes: Intelligent segmentation of PPG waveforms based on valley detection; Tidal peak value is identified by first derivative analysis; Automatically detect and discard incomplete waveforms; Extract 20-dimensional waveform features; The mean value of the complete waveform features is calculated to generate a 20-dimensional feature window.

4. The PPG blood pressure monitoring method according to claim 1, characterized in that, The PPG signal includes the main wave, tidal wave, descending mid-slope wave, and diabetic wave.

5. The PPG blood pressure monitoring method according to claim 4, characterized in that, The 20-dimensional features for each window include peak features, slope features, ratio features, time features, and time ratio features.

6. The PPG blood pressure monitoring method according to claim 1, characterized in that, The various machine learning algorithms include SVR, GBR, RF, LR, and XGB.

7. The PPG blood pressure monitoring method according to claim 5, characterized in that, The 20-dimensional waveform features include: Peak characteristics: systolic blood pressure peak, tidal wave peak, diastolic blood pressure peak, trough value; Slope characteristics: the rising / falling slope between each peak; Ratio characteristics: the relative relationship between peaks and troughs; Temporal characteristics: pulse width, diastolic time interval; Time ratio characteristic: the relative relationship of time intervals.

8. A system for implementing the PPG blood pressure monitoring method according to any one of claims 1-7, characterized in that, include: The signal acquisition module uses a fully integrated analog front-end module, which connects a narrow-spectrum light source and a narrow-band response organic photodetector to acquire user PPG signals. The PPG signal processing module is used to control and drive the narrow-spectrum light source and narrow-band response organic photodetector to read signals and features in a time sequence, and to transmit the collected PPG signals and blood oxygen data to the host computer via the UART protocol. The host computer preprocesses and extracts features from the PPG signal, then inputs it into the personalized blood pressure prediction model to obtain the blood pressure value.

Images