Method and system for cuffless blood pressure estimation based on unsupervised pulse wave characterization learning

Abstract

The application provides a cuffless blood pressure estimation method and system based on unsupervised pulse wave feature learning. First, a photoplethysmography signal to be measured is acquired. The signal is input into a constructed blood pressure monitoring model to obtain a blood pressure waveform. The construction method of the model comprises: acquiring a sample data set, and performing data preprocessing and data enhancement to obtain a pre-training data set and a downstream fine-tuning data set; constructing an unsupervised pulse wave feature learning module and a downstream fine-tuning blood pressure estimation module; training the unsupervised pulse wave feature learning module by using a pre-training data set without labels, and migrating the parameters of a first encoder in the unsupervised pulse wave feature learning module to a second encoder in the downstream fine-tuning blood pressure estimation module; and training the downstream fine-tuning blood pressure estimation module by using a downstream fine-tuning data set with labels. The application can learn high-value physiological features related to blood pressure on a large amount of unlabeled PPG signal data set without relying on blood pressure labels.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a cuffless blood pressure estimation method and system based on unsupervised pulse wave representation learning. Background Technology

[0002] Continuous blood pressure monitoring technology has the potential to reduce the burden of care, improve outcomes, and thus lower mortality. However, the gold standard for continuous blood pressure measurement is an invasive method, namely catheter ablation, which carries potential risks of pain and infection. Among non-invasive blood pressure measurement methods, the most common oscillometric method requires an inflatable cuff to occlude blood vessels, making it unsuitable for long-term monitoring. Several non-invasive blood pressure measurement methods based on pulse transit time (PTT) theory have emerged. This new sensing technology requires placing a photodetector on the skin surface to acquire photoplethysmography (PPG) signals, while simultaneously connecting adhesive electrodes to acquire electrocardiogram (ECG) signals. These new methods have the potential for continuous measurement; however, ECG acquisition increases the power consumption of the device, and prolonged use with adhesive electrodes may cause skin irritation.

[0003] In recent decades, much research has focused on designing PPG morphological features to facilitate the use of machine learning models to capture their relationship with blood pressure. However, these methods require careful design of PPG morphological features and often encounter the NP-hard problem: too many features lead to redundancy and the potential curse of dimensionality, while too few features may overlook crucial details of the nonlinear coupling of vascular dynamics. Fortunately, recent advances in data-driven deep learning techniques offer a promising avenue for automatically identifying the optimal features that map PPG signals to blood pressure.

[0004] However, reported deep learning methods still face several challenges, including estimation performance, experimental paradigms, evaluation protocols, and application scenarios. First, a significant challenge lies in the inconsistency of test datasets. Many studies evaluate models on only a single dataset, failing to consider scenarios where models must infer from data with unknown distributions in real-world environments. Second, validation schemes based on random scrambling of samples need refinement. During data splitting, the model needs to differentiate between subjects to prevent encountering data from the same subjects used during training during testing. This avoids overly optimistic results. Therefore, a subject-based splitting paradigm should be adopted for validation to enhance model generalization and robustness. Third, existing studies either require additional ECG input or rely on users periodically providing cuff blood pressure calibration values. However, neither of these methods is conducive to continuous, long-term blood pressure measurements. Finally, all current studies employ fully supervised methods. The overall performance of these data-driven models largely depends on the size of the paired PPG-BP dataset. Unfortunately, due to the difficulty in obtaining blood pressure labels (requiring arterial cannulation or cuff recording), most of these studies have only been tested on small datasets, resulting in poor generalizability. Furthermore, few studies meet the minimum number of participants (at least 85 participants) required by the AAMI criteria, which also affects the reliability of the experimental results. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the purpose of this invention is to provide a cuffless blood pressure estimation method and system based on unsupervised pulse wave representation learning.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A cuffless blood pressure estimation method based on unsupervised pulse wave representation learning includes:

[0008] Acquire the photoplethysmography signal of the target under test;

[0009] The photoplethysmography signal to be measured is input into the constructed blood pressure monitoring model to obtain the blood pressure waveform of the target being measured.

[0010] The method for constructing the cuffless blood pressure waveform continuous estimation monitoring model includes:

[0011] Obtain a sample dataset; the sample dataset includes an unlabeled subset with photoplethysmography (PPG) signals and a paired subset with PPG signals and blood pressure waveform labels;

[0012] The sample dataset is preprocessed and augmented to obtain an unlabeled pre-training dataset and a labeled downstream fine-tuning dataset.

[0013] Construct an unsupervised pulse wave representation learning module and a downstream fine-tuning blood pressure estimation module;

[0014] The unsupervised pulse wave representation learning module is trained using the unlabeled pre-training dataset, and the parameters of the first encoder in the unsupervised pulse wave representation learning module are transferred to the second encoder of the downstream fine-tuning blood pressure estimation module.

[0015] The downstream fine-tuning blood pressure estimation module is trained using the labeled downstream fine-tuning dataset to obtain a trained blood pressure monitoring model.

[0016] Preferably, the sample dataset is preprocessed and augmented to obtain an unlabeled pre-training dataset and a labeled downstream fine-tuning dataset, including:

[0017] The sample dataset is sequentially resampled, outlier removal is performed, filtering and noise reduction is applied, phase alignment is performed, normalization is performed, and segmentation is performed to obtain the preprocessed dataset.

[0018] Gaussian noise, baseline drift, motion interference, power frequency interference, and respiratory sinus rhythm frequency modulation are sequentially added to the preprocessed dataset to obtain the unlabeled pre-training dataset and the labeled downstream fine-tuning dataset, respectively.

[0019] Preferably, the unsupervised pulse wave representation learning module includes: a temporal neighborhood coding submodule, a first encoder, a projector and a discriminator, and an unsupervised representation loss submodule;

[0020] The temporal neighborhood coding submodule is used to distinguish the neighborhood and non-neighborhood clipping windows of the photoplethysmography signal in the unlabeled pre-training dataset.

[0021] The first encoder is used to train based on the cropping window to obtain the parameters of the trained first encoder;

[0022] The projector and discriminator share weights and are used to converge sequences of photoplethysmography signals of similar states in the embedded space.

[0023] The unsupervised representation loss submodule is used to construct the objective function of the pre-training task and to identify the sequence of the photoplethysmography signal according to the objective function of the pre-training task.

[0024] Preferably, the downstream fine-tuning blood pressure estimation module includes: a second encoder and a downstream regressor;

[0025] The second encoder has the parameters of the first encoder; the second encoder is used to process the labeled downstream fine-tuning dataset to obtain representational features;

[0026] The downstream regressor is used to obtain blood pressure estimation data based on the characterization features.

[0027] A cuffless blood pressure estimation system based on unsupervised pulse wave representation learning includes:

[0028] The signal acquisition unit is used to acquire the photoplethysmography signal of the target under test;

[0029] The waveform monitoring unit is used to input the photoplethysmography signal to be measured into the constructed blood pressure monitoring model to obtain the blood pressure waveform of the target being measured.

[0030] The waveform monitoring unit includes:

[0031] The sample acquisition subunit is used to acquire a sample dataset; the sample dataset includes an unlabeled subset with photoplethysmography (PPG) signals and a paired subset with PPG signals and blood pressure waveform labels;

[0032] The sample processing subunit is used to perform data preprocessing and data augmentation on the sample dataset to obtain an unlabeled pre-training dataset and a labeled downstream fine-tuning dataset.

[0033] The model building subunit is used to build the unsupervised pulse wave representation learning module and the downstream fine-tuning blood pressure estimation module;

[0034] The first training subunit is used to train the unsupervised pulse wave representation learning module using the unlabeled pre-training dataset, and to transfer the parameters of the first encoder in the unsupervised pulse wave representation learning module to the second encoder of the downstream fine-tuning blood pressure estimation module.

[0035] The second training subunit is used to train the downstream fine-tuning blood pressure estimation module using the labeled downstream fine-tuning dataset to obtain a trained blood pressure monitoring model.

[0036] Preferably, the sample processing subunit includes:

[0037] The preprocessing module is used to sequentially resample, remove outliers, filter and denoise, phase align, normalize and segment the sample dataset to obtain the preprocessed dataset.

[0038] An enhancement module is used to sequentially add Gaussian noise, baseline drift, motion interference, power frequency interference, and respiratory sinus rhythm frequency modulation to the preprocessed dataset, thereby obtaining the unlabeled pre-training dataset and the labeled downstream fine-tuning dataset, respectively.

[0039] Preferably, the unsupervised pulse wave representation learning module includes: a temporal neighborhood coding submodule, a first encoder, a projector and a discriminator, and an unsupervised representation loss submodule;

[0040] The temporal neighborhood coding submodule is used to distinguish the neighborhood and non-neighborhood clipping windows of the photoplethysmography signal in the unlabeled pre-training dataset.

[0041] The first encoder is used to train based on the cropping window to obtain the parameters of the trained first encoder;

[0042] The projector and discriminator share weights and are used to converge sequences of photoplethysmography signals of similar states in the embedded space.

[0043] The unsupervised representation loss submodule is used to construct the objective function of the pre-training task and to identify the sequence of the photoplethysmography signal according to the objective function of the pre-training task.

[0044] Preferably, the downstream fine-tuning blood pressure estimation module includes: a second encoder and a downstream regressor;

[0045] The second encoder has the parameters of the first encoder; the second encoder is used to process the labeled downstream fine-tuning dataset to obtain representational features;

[0046] The downstream regressor is used to obtain blood pressure estimation data based on the characterization features.

[0047] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0048] This invention provides a cuffless blood pressure estimation method and system based on unsupervised pulse wave representation learning. The method includes: acquiring the photoplethysmography (PPG) signal of the target; inputting the PPG signal into a pre-constructed blood pressure monitoring model to obtain the monitored blood pressure waveform of the target; the method for constructing the cuffless blood pressure waveform continuous estimation monitoring model includes: acquiring a sample dataset; the sample dataset includes an unlabeled subset with PPG signals and a paired subset with PPG signals and blood pressure waveform labels; performing data preprocessing and data augmentation on the sample dataset to obtain an unlabeled pre-training dataset and a labeled downstream fine-tuning dataset; constructing an unsupervised pulse wave representation learning module and a downstream fine-tuning blood pressure estimation module; training the unsupervised pulse wave representation learning module using the unlabeled pre-training dataset, and transferring the parameters of the first encoder in the unsupervised pulse wave representation learning module to the second encoder of the downstream fine-tuning blood pressure estimation module; training the downstream fine-tuning blood pressure estimation module using the labeled downstream fine-tuning dataset to obtain a trained blood pressure monitoring model. This invention does not rely on blood pressure labels and can learn high-value blood pressure-related physiological characteristics on massive unlabeled PPG signal datasets. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A flowchart of the method provided in an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of the technical route provided in the embodiments of the present invention;

[0052] Figure 3 This is a schematic diagram of the model framework provided in an embodiment of the present invention;

[0053] Figure 4 A data processing flowchart provided for embodiments of the present invention;

[0054] Figure 5 This is a schematic diagram of model transfer provided in an embodiment of the present invention;

[0055] Figure 6 This is a schematic diagram of the PPG signal data enhancement strategy provided in an embodiment of the present invention;

[0056] Figure 7This is a blood pressure value estimation error diagram provided in an embodiment of the present invention; wherein... Figure 7 (a) is a graph showing the correlation of systolic blood pressure; Figure 7 (b) is the Bland-Altman plot of systolic blood pressure; Figure 7 (c) is a graph showing the correlation of diastolic blood pressure; Figure 7 (d) is the Bland-Altman plot of diastolic blood pressure. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] The purpose of this invention is to provide a cuffless blood pressure estimation method and system based on unsupervised pulse wave representation learning, which does not rely on blood pressure labels and can learn high-value blood pressure-related physiological representations on massive unlabeled PPG signal datasets.

[0059] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a cuffless blood pressure estimation method based on unsupervised pulse wave representation learning, comprising:

[0061] Step 100: Acquire the photoplethysmography signal of the target under test;

[0062] Step 200: Input the photoplethysmography signal to be measured into the constructed blood pressure monitoring model to obtain the blood pressure waveform of the target being measured;

[0063] Step 200 includes:

[0064] Step 201: Obtain the sample dataset; the sample dataset includes an unlabeled subset with photoplethysmography (PPG) signals and a paired subset with PPG signals and blood pressure waveform labels;

[0065] Step 202: Perform data preprocessing and data augmentation on the sample dataset to obtain an unlabeled pre-training dataset and a labeled downstream fine-tuning dataset;

[0066] Step 203: Construct an unsupervised pulse wave representation learning module and a downstream fine-tuning blood pressure estimation module;

[0067] Step 204: Train the unsupervised pulse wave representation learning module using the unlabeled pre-training dataset, and transfer the parameters of the first encoder in the unsupervised pulse wave representation learning module to the second encoder of the downstream fine-tuning blood pressure estimation module;

[0068] Step 205: Train the downstream fine-tuning blood pressure estimation module using the labeled downstream fine-tuning dataset to obtain a trained blood pressure monitoring model.

[0069] Specifically, due to the portability of photoplethysmography (PPG) sensors, PPG signals reflecting changes in vascular volume are readily available and have been widely used to measure heart rate, blood oxygen saturation, and other physiological parameters. However, no studies have yet discussed how to utilize these abundant unlabeled PPG signals to enhance blood pressure estimation results on limited PPG-BP paired datasets. Therefore, this invention proposes an unsupervised pulse wave representation learning framework for robust non-invasive blood pressure estimation, termed UPR-BP.

[0070] The proposed UPR-BP is a framework based on contrastive learning theory. It first pre-trains on a large dataset of unlabeled PPG signals to obtain valuable physiological representations, and then fine-tunes a Transformer-based unsupervised pre-trained encoder using paired PPG-BP data, achieving medical-grade blood pressure estimation accuracy using only 10% labeled data. Considering the challenges faced by previous research, this invention makes several important contributions that differ from existing methods:

[0071] First, UPR-BP is the first known unsupervised, cuffless blood pressure estimation framework. This approach overcomes the limitations of previous supervised methods, which often generalize poorly due to the limited size of labeled data. UPR-BP demonstrates robust performance even when evaluated on datasets from diverse sources, including bedside monitors and wearable watches, containing over 4,000 participants.

[0072] Secondly, UPR-BP employs Temporal Neighborhood Coding (TNC) to dynamically identify static properties in PPG signals. This allows the framework to capture subtle changes in physiological states, resulting in more accurate blood pressure estimation.

[0073] Finally, UPR-BP combines the invariance, variance, and covariance terms of the embedded encoding vector into a jointly optimized shared weight branch. This approach helps extract more meaningful physiological representations and prevents information collapse during the PPG signal representation learning process.

[0074] UPR-BP represents a significant advancement in the field of noninvasive blood pressure estimation. By leveraging unsupervised learning and advanced representation learning techniques, UPR-BP enables accurate and robust blood pressure estimation even with limited labeled data. This has the potential to improve the accessibility and affordability of continuous blood pressure monitoring, ultimately leading to better healthcare outcomes.

[0075] The technical approach in this embodiment includes the following specific steps, such as... Figure 2 As shown:

[0076] Step 1: In the unsupervised pre-training task, UPR-BP does not require synchronously acquired blood pressure labels; only PPG signals are needed as input. However, in the downstream blood pressure estimation task, a PPG-blood pressure paired dataset is required to acquire human PPG signals and simultaneously collect invasive blood pressure waveforms as reference signals for the model. This invention uses three paired PPG-blood pressure datasets to train and evaluate the performance of UPR-BP in blood pressure estimation.

[0077] To improve the quality of PPG representations learned by UPR-BP for downstream tasks, this invention employs cross-dataset and single-dataset evaluation protocols. The cross-dataset evaluation protocol involves unsupervised pre-training on the widely used, large-scale public MIMIC III dataset. Subsequently, a simple regressor is applied to the pre-trained encoder and fine-tuned on another dataset, thus evaluating the effectiveness of the PPG representation learning model for blood pressure estimation across different datasets. In contrast, in the single-dataset evaluation scheme, this invention randomly divides the MIMIC III dataset into two parts: an unlabeled PPG pre-training set and a PPG-blood pressure paired fine-tuning dataset based on subject records, in a 9:1 ratio. This invention uses the same evaluation procedure as the above scheme. This approach aims to evaluate the generalization of unsupervised pre-training to UPR-BP, ensuring its performance even on limited datasets. By employing these two evaluation procedures, this invention can effectively measure the performance of unsupervised methods using a small amount of paired PPG-blood pressure data and a large amount of unlabeled PPG data. In summary, this study utilizes three available blood pressure datasets, as described below.

[0078] 1) MIMIC (Multi-parameter Intelligent Monitoring in Intensive Care Unit) III Dataset. This dataset was jointly collected by Beth Israel Deaconess Medical Center and Philips Healthcare. The first version of the MIMIC III dataset contained 67,830 records from approximately 30,000 ICU patients, including respiratory, invasive blood pressure waveforms, electrocardiogram waveforms, and PPG waveforms. During unsupervised pre-training, this invention extracted data with durations exceeding 30 seconds from the PPG records in this dataset.

[0079] 2) Private Dataset. This dataset was collected from physiological parameter records of intensive care units in seven hospitals. The private dataset includes PPG and paired invasive blood pressure waveform data of 683 ICU patients.

[0080] 3) Aurora-BP Dataset. Unlike the previous two datasets collected from intensive care units, Microsoft collected data from 642 and 483 subjects using auscultation and oscillometric methods, respectively. PPG signals were obtained through an optical sensor embedded in a watch. However, according to the authors, the protocol, subject stratification, and analysis performed on this dataset were not designed or evaluated according to the benchmarks set by the U.S. Food and Drug Administration (FDA) for evaluating cuffless blood pressure devices, and the clinical applicability of these data cannot be determined. Therefore, this invention uses this dataset only as a reference for non-invasive blood pressure measurement to verify the general applicability of UPR-BP.

[0081] Although MIMIC III and private datasets provide paired PPG-blood pressure waveforms, accurately estimating noninvasive blood pressure from PPG signals remains a challenge due to the need to capture fine-grained discriminative features from limited paired signals. Supervised datasets can be paired samples. Including PPG signal and the corresponding target blood pressure value space here This represents a total of s PPG signal samples, including t time steps (equivalent to the sampling rate and d dimensions, which can be selected as PPG, PPG velocity sequence (VPG), and PPG acceleration sequence (APG)), and simultaneously measuring their corresponding systolic and diastolic blood pressure values. The UPR-BP proposed in this invention first learns signal representations on a massive amount of unlabeled PPG samples, and then performs a downstream fine-tuning process on a small supervised training set. The fine-tuned model can learn the mapping from PPG signal x to the corresponding blood pressure value y:

[0082]

[0083] The target loss can be defined by minimizing the mean squared error (MSE) between the reference blood pressure and the blood pressure estimated by the model.

[0084] like Figure 3 As shown, UPR-BP utilizes the local smoothness of sequences to define temporal neighborhoods with similar hemodynamic states. Although PPG, VGP, and APG reflect different physiological characteristics, they share the same physiological state, which this invention refers to as PPG invariance. To demonstrate this, this invention randomly selected two sequences and extracted the data from the cropping window W. u and W p The inner neighborhood represented, and the clipping window W u and Wd The non-neighborhood is represented. These clipped windows are then input into a transformer encoder with shared weights to obtain the corresponding sequence representation (Z). u Z p and Z d To facilitate the aggregation of sequence representations of similar states in the embedding space, UPR-BP employs VICReg and TNC. To achieve this, the present invention uses a projector and discriminator with shared weights. The Transformer encoder and projector architecture are as follows... Figure 3 As shown on the right.

[0085] Step 2: Constructing the Data Processing Flow. This invention utilizes the largest multicenter dataset to date to enhance the robustness and generalizability of the model. However, due to the different populations and equipment used in clinical settings, data distribution can vary significantly. Furthermore, the publicly available MIMIC III dataset presents significant challenges due to its varying record lengths and large amounts of missing data. To address these issues, this invention designs a standard data processing flow (such as...). Figure 4 The data is filtered to obtain reliable quality. The data processing flow includes resampling, outlier removal, filtering and noise reduction, normalization, phase alignment, enhancement, and segmentation. These steps are shown in the left figure. In the unsupervised pre-training phase, UPR-BP uses a Transformer encoder, projector, and discriminator to learn the physiological representation of the PPG signal. In the downstream supervised blood pressure estimation task, a regressor is built on top of the Transformer encoder to transfer the weights of the representation and achieve high-precision blood pressure estimation. The specific data processing flow is as follows:

[0086] 1) Resampling. Due to the different sampling frequencies of the datasets, this invention resamples the PPG signals and invasive blood pressure waveforms in the datasets to a uniform 125Hz.

[0087] 2) Outlier Removal. Physiological data collected in a clinical setting are inevitably subject to interference, which may lead to data anomalies such as missing signals and extreme blood pressure values. This invention removes missing signals using a first-order difference method and removes abnormal signals by creating an average template. Specifically, for signal segments belonging to the same signal segment (considered as the same patient), the average value of the signal segment is calculated to obtain the average template of the signal segment, and the average difference and correlation coefficient of all signal segments are calculated. Abnormal segments exceeding a threshold (3 times the average ± standard deviation) are removed.

[0088] 3) Filtering and noise reduction. Physiological signals have low frequency and amplitude, and their noise is mainly baseline drift and motion interference. Therefore, this invention uses a db8-based wavelet to decompose the PPG signal and invasive blood pressure signal into 9 levels through discrete wavelet transform. The approximation coefficients of level 1 and the detail coefficients of levels 7-9 are set to zero, and then the signal is reconstructed based on these decomposition coefficients to eliminate baseline noise in the PPG signal.

[0089] 4) Phase Alignment. Considering the potential delay of up to 500 milliseconds between the PPG signal and the blood pressure waveform in the dataset, the PPG signal is shifted using the maximum correlation coefficient between the PPG and blood pressure signals as a reference point to align the phases of the two signals.

[0090] 5) Normalization. The amplitude of the PPG signal is affected by ambient light intensity, sensor emitted light intensity, sensor photodiode sensitivity, blood oxygen concentration, venous volume, and subject skin color. To reduce individual differences, this invention uses a normalization method to scale the PPG signal within the [0,1] interval.

[0091] 6) Segmentation. This invention employs a fixed-length segmentation method, which divides the signal segment into fixed-length fragments using a sliding window (window length of 9 seconds, step size of 3 seconds).

[0092] Step 3: Design a PPG signal data augmentation strategy. Augmentation design for unsupervised pre-training data significantly impacts the accuracy of downstream fine-tuned blood pressure estimation. Current PPG-based blood pressure estimation research uses fully supervised models, and only a few effective augmentation designs have been attempted. Therefore, this invention specifically designs a data augmentation strategy for PPG signals to address the problem that PPG signals acquired by photoelectric sensors are susceptible to various interferences such as ambient light and motion. Specifically, as follows... Figure 6 The diagram illustrates the addition of Gaussian noise, power line noise, motion artifacts, baseline drift, and respiratory modulation. The data augmentation strategy uses five transformations to simulate noise that may be present in the pulse wave signal. The noise consists of five main components: 1) Gaussian random noise; 2) 50Hz power line interference; 3) respiratory artifacts and motion-induced baseline drift, manifesting as low-frequency interference; 4) high-frequency and non-smooth interference caused by muscle activity; and 5) respiratory modulation. The following data augmentation operations are dynamically executed independently with predetermined probabilities in the training loop:

[0093] 1) Add Gaussian noise. The enhancement operation is performed on the PPG signal. Gaussian noise with amplitude κ and frequency ω is added with probability ρ. This operation takes into account random noise common in PPG signals.

[0094] 2) Increase baseline drift. Add 0.05Hz baseline noise with probability ρ to the PPG signal. This operation simulates low-frequency interference caused by baseline drift due to breathing artifacts and motion.

[0095] 3) Increase motion interference. This affects the PPG signal. Add motion interference pulse trains with amplitude κ and frequency ω, and duration of 1-10%. This operation can capture noise caused by motion during PPG signal acquisition.

[0096] 4) Add power frequency interference. Add 50Hz power frequency interference with amplitude κ and frequency ω to the PPG signal with probability ρ. This operation takes into account the common power frequency noise in the real world.

[0097] 5) Increase the frequency modulation of the respiratory sinus rhythm (RSA). Add a 0.05Hz sinusoidal modulation signal with a mean of 1 and an amplitude of κ to the PPG signal. This operation simulates the effect of the respiratory sinus rhythm.

[0098] Step 4: Construct a model framework for cuffless blood pressure estimation based on an unsupervised pulse wave representation learning model (UPR-BP). This invention uses the term "PPG invariance" to refer to the PPG signal reflecting changes in vascular volume and the derived sequences (VPG and APG) representing the rate of change in blood flow. The static properties of these sequences are defined by the local smoothness of the sequences, and therefore possess invariance. The UPR-BP proposed in this invention aims to capture the characteristic representations of these PPG-invariant physiological states through a joint embedding architecture. These physiological representations can reflect changes in microcirculatory blood flow and can be used to assess changes in blood pressure and cardiac function.

[0099] Specifically, UPR-BP randomly selects two sequences from the PPG invariance and applies the enhanced Dickey-Fuller (ADF) statistical test to determine the neighborhood ranges with the same physiological state. These neighborhoods constitute a clipping window, with those of the same physiological state using W. u and W p This indicates that the non-neighborhood clipping window uses W. u and W d This is indicated. Then, UPR-BP uses a transformer encoder with shared weights to capture the spatiotemporal correlations within these clipping windows, thereby providing W... u W p and W d Generate deep feature representations (Z u Z p Z d ).

[0100] To learn valuable physiological representation information, UPR-BP introduces a joint optimization strategy of variance-covariance regularization (VICReg) and TNC. VICReg utilizes a twin-like projector to map neighboring representations in the PPG-invariant embedding space. This allows for the computation of loss functions related to physiological representation differences and embedding dependencies. Therefore, VICReg ensures that each dimension of the deep feature representation exhibits a certain variance (above a certain threshold to ensure diversity) while also ensuring that dimensions are uncorrelated (decoupled to ensure non-redundant information). TNC can distinguish the feature distribution within and outside the neighborhood clipping window in the high-dimensional embedding space. This helps UPR-BP learn feature representations of physiological state changes. The specific process will be described in detail in step 5.

[0101] Step 5: Construct the unsupervised pulse wave representation learning module in the cuffless blood pressure estimation model framework based on the unsupervised pulse wave representation learning model, and define the objective loss function for the pre-training task. This involves two sub-steps: first, obtaining neighborhood and non-neighborhood clipping windows for the multi-channel PPG sequence based on TNC; then, using VICReg to improve the information content of the learned deep representation vectors and reduce redundancy, while preventing model collapse.

[0102] PPG signals, acquired using photoelectric sensors, are a common and readily available method for describing hemodynamic information. PPGs contain rich hydrodynamic and vascular elasticity information, making them ideal signals for learning transferable physiological representations to improve blood pressure estimation. However, simultaneously recording invasive blood pressure waveforms or cuff blood pressure values ​​requires arterial cannulation or vascular occlusion, which is challenging. This results in limited sample sizes in clinical datasets for blood pressure estimation, making it difficult to meet the requirements of data-driven modeling. To extract valuable physiological representations, this invention proposes an unsupervised representation learning framework that maps multi-channel PPG-derived sequences into an embedding space. This enables deep models to capture dynamic temporal information while learning robust PPG representations from multiple perspectives.

[0103] Specifically, this invention randomly selects two different channels of PPG invariant sequences as inputs to UPR-BP (i.e., randomly selects PPG and VPG, PPG and APG, or VPG and APG), and uses UPR-BP to construct pruning windows for inner and non-neighborhoods. More specifically, this invention constructs a pruning window based on TNC to distinguish between neighborhood and non-neighborhood.

[0104] In time, the neighborhood N with center point u u The clipping window W is defined in the code. u The center point is derived from the Gaussian distribution. The sampling is performed in a medium-sized window of size σ. Here, γ is a parameter defining the neighborhood range, which depends on the static characteristics of the input sequence. Considering the dynamic changes in human physiological states, this invention uses the unit root test (i.e., the ADF statistical test) to derive the neighborhood range γ. For each clipping window W... u ∈X i We start with γ = 1 and gradually increase the value of γ until the p-value of the test exceeds a predetermined critical value. This means that the null hypothesis needs to be retained, indicating that the fragment is no longer statically invariant. Therefore, this invention defines fragments outside the neighborhood as... Crop window within a contrastive learning framework These can be considered negative samples. Similarly, as... Figure 3 As shown, the clipping window W p ~N u ∈X j For positive samples, This invention proposes that the potential states between non-adjacent clipping windows are different, which indirectly reflects the fluctuation of blood pressure over time.

[0105] Figure 3 The randomly selected W is shown u W p and W d Example. These three cropped PPG sequence windows maintain temporal dependencies on each other and are encoded as PPG representations. and Each branch contains r vectors of dimension q.

[0106] To maximize the information content (increase variance) and reduce redundancy (decrease covariance) of the learned deep representation vector Z, while preventing model collapse (decrease invariance), this invention proposes three regularization terms: an invariance term s(·), a variance term v(·), and a covariance term c(·), collectively referred to as VICReg. Specifically, to obtain the embedding vector... and The neighborhood representation of Z u and Z p It will be projected through a twin-like network. VICReg loss is defined as

[0107]

[0108] Where λ, μ and These are weighting coefficients or hyperparameters used to control the importance of each term in the loss function. In this invention, the experiment will use... Setting this parameter to the default value yields optimal experimental results. Furthermore, this invention performs a basic conditional grid search to determine the values ​​of λ and μ, with the constraint λ = μ > 1. This constraint ensures that invariant learning and regularization have equal weights during training.

[0109] The invariant term s(·) reduces the distance between embedding vectors obtained from different PPG sequences, similar to narrowing the distance between positive samples in contrastive learning. This process helps improve UPR-BP's ability to learn specific PPG invariant patterns under similar physiological conditions. Therefore, VICReg computes the embedding vector M... u and M p The Euclidean mean square distance between them is used to obtain s(·):

[0110]

[0111] To prevent the learned embedding vectors within a batch from collapsing into a single constant value, the variance term v(·) used in this invention is...

[0112]

[0113] Here, ε is a small scalar used to prevent numerical instability. Its purpose is to ensure that the variance remains above a given threshold ζ, thereby guaranteeing a certain degree of variability in the batch embedding results.

[0114] The most important covariance term c(·) in VICReg includes the square of the covariance matrix C(M):

[0115]

[0116] The calculation of C(M) is as follows:

[0117]

[0118] Where the unindexed letter M represents any adjacent pair of embedding vectors M u and M p Used to calculate their respective covariance matrices C(M) u ) and C(M p This regularization term encourages the off-diagonal elements of the covariance matrix to approach zero. In other words, it helps reduce redundant information between vectors (different samples) in a batch.

[0119] In contrastive learning, the discriminator needs to distinguish between representations within and outside the neighborhood. This distinction helps in identifying vascular dynamics under different physiological states. To this end, this invention employs a loss similar to TNC, calculating the loss by considering pairs of representations within and outside the neighborhood:

[0120]

[0121] In this study, a simple linear classifier was used as the discriminant. If Z and Z u If the pair is an adjacent pair in a PPG-derived sequence channel, the predicted value is 1; otherwise, it is 0. Similarly, an additional covariance term C(Z) is added. u Z) is defined as

[0122]

[0123] Where the letter Z without a subscript represents any PPG character Z. p and Z d These are used to calculate the PPG characterization of the center point Z, respectively. u The covariance matrix C(Z) u Z p ) and C(Z u Z d This provides the covariance term. It is achieved by associating Z... u and Z p and decorrelation Z u and Z d This improves characterization quality. Therefore, TNC can effectively capture the potential time-stationary states between PPG sequence channels, thus making it more sensitive to blood pressure fluctuations.

[0124] Therefore, the objective function of the joint optimization embedding structure proposed in this invention can be expressed as follows:

[0125]

[0126] Step 6: Construct a downstream fine-tuned blood pressure estimation model within the cuffless blood pressure estimation model framework based on an unsupervised pulse wave representation learning model, and define the objective loss function for the downstream blood pressure estimation fine-tuning task. After pre-training through a proxy task, as follows... Figure 5 As shown, the base model UPR-BP obtained from steps 5-6 typically retains only the encoder, discarding other parts of the model, such as the projector and discriminator. This invention performs a cross-dataset evaluation protocol to verify the performance of blood pressure estimation on the supervised PPG-blood pressure paired dataset. Specifically, this invention fine-tunes a regressor (·) based on pre-trained representations for estimating blood pressure values. Specifically, this invention trains a nonlinear regressor consisting of three fully connected (FC) layers. Two hidden layers each consist of 256 neurons with ReLU activation, while the output layer contains 2 neurons for estimating systolic and diastolic blood pressure values. The blood pressure estimation module can be viewed as a regression task and uses root mean square error loss:

[0127]

[0128] Step 7: Divide the data processed in Step 3 into training, validation, and test sets in a 7:1.5:1.5 ratio. The training and validation sets are used to train the system model established in Step 6 and select parameters. The test set is used to finally verify the generalization ability of the optimal system model saved in Step 6. Data splitting adopts an inter-individual splitting method to ensure that each data set does not contain data from the same patient, thereby avoiding information leakage. During network training, the Adam optimizer is used to update parameters with a learning rate of 1e-4, a weight decay of 0.999, a momentum of 0.8, and 1000 training iterations. The Xavier initializer is used to maintain the initial parameters as normally distributed.

[0129] Step 8: Acquire the user's PPG signal. Input the PPG signal waveform and necessary conditional information into the cuffless blood pressure waveform continuous estimation and monitoring system. The system can automatically output the user's systolic and diastolic blood pressure values. Alternatively, input the PPG signal data to be measured into the invasive blood pressure waveform estimation and monitoring system established in Step 6 based on photoplethysmography (PPG) for non-invasive blood pressure estimation. Figure 7 This is a graph showing the blood pressure estimation error of UPR-BP on the dataset established in this invention. The presented results demonstrate that the UPR-BP proposed in this invention has medical-grade blood pressure estimation accuracy. In the Bland-Altman plot, the horizontal axis shows the average blood pressure value between the estimated and reference invasive blood pressure measurements, and the vertical axis shows the difference in blood pressure values ​​between the two methods. The dashed line in the middle represents the average difference in blood pressure values ​​between the two methods. The dashed lines on the upper and lower sides represent the upper and lower limits of the 95% confidence interval for the average difference.

[0130] The beneficial effects of this invention are as follows:

[0131] 1. The input of this invention consists only of photoplethysmography (PPG) pulse waves. Therefore, the acquisition circuit only needs to acquire PPG signals. Compared with the traditional method based on pulse wave propagation speed, it eliminates the step of acquiring electrocardiogram (ECG) signals and does not require excessive derivative calculations and feature engineering. This makes it easy to integrate into devices such as wristbands, eliminating the need for blood pressure measurement equipment such as cuffs, freeing the device from the constraints of cuffs, and making it more portable.

[0132] 2. This invention proposes the first unsupervised pulse wave representation learning framework (UPR-BP) for robust cuffless blood pressure estimation. It does not rely on blood pressure labels and can learn high-value blood pressure-related physiological representations on massive unlabeled PPG signal datasets. UPR-BP utilizes contrastive learning theory to pre-train a Transformer encoder on large-scale unlabeled PPG data, and then fine-tunes the encoder using a small amount of paired PPG-blood pressure data to achieve medical-grade blood pressure estimation accuracy using only 10% of the labeled data.

[0133] 3. Dynamic identification of stationary characteristics: UPR-BP uses Temporal Neighborhood Coding (TNC) to dynamically identify stationary characteristics in PPG signals, thereby improving UPR-BP's ability to learn subtle changes in physiological state.

[0134] 4. Joint Optimization of Invariance, Variance, and Covariance: UPR-BP combines the invariance, variance, and covariance terms of the embedded encoding vector into a shared weight branch for joint optimization. This method produces more meaningful physiological representations and avoids the information collapse problem during PPG signal representation learning.

[0135] 5. Strong noise resistance. This invention designs a PPG signal data augmentation strategy that can generate PPG signals with different noise and interference, thereby helping the model learn more robust representations. Furthermore, this strategy can reduce model training time by generating new data, thus improving the efficiency of UPR-BP.

[0136] Robustness across datasets, meeting medical standards. This invention was validated using an individual-based data partitioning paradigm on a dataset of over 4,000 subjects from three different sources (including data collected by bedside monitors and wearable wristwatches). The experiments of this invention avoid the data leakage problems that may be caused by random shuffling paradigms. Furthermore, compared to existing small-sample studies, this invention strictly adheres to AAMI standards by setting up a test set of >85 subjects. Experimental results show that the estimation errors of UPR-BP for systolic and diastolic blood pressure meet clinical standards for medical devices.

[0137] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0138] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A cuffless blood pressure estimation method based on unsupervised pulse wave representation learning, characterized in that, include: Acquire the photoplethysmography signal of the target under test; The photoplethysmography signal to be measured is input into the constructed blood pressure monitoring model to obtain the blood pressure waveform of the target being measured. The method for constructing the blood pressure monitoring model includes: Obtain a sample dataset; the sample dataset includes an unlabeled subset with photoplethysmography (PPG) signals and a paired subset with PPG signals and blood pressure waveform labels; The sample dataset is preprocessed and augmented to obtain an unlabeled pre-training dataset and a labeled downstream fine-tuning dataset. Construct an unsupervised pulse wave representation learning module and a downstream fine-tuning blood pressure estimation module; The unsupervised pulse wave representation learning module includes: a temporal neighborhood coding submodule, a first encoder, a projector and a discriminator, and an unsupervised representation loss submodule; The temporal neighborhood coding submodule is used to distinguish between neighborhood and non-neighborhood clipping windows of the photoplethysmography signal in the unlabeled pre-training dataset; wherein, for any clipping window, the unit root test method is used to determine the neighborhood range with the same physiological state. The first encoder is used to train based on the cropping window to obtain the parameters of the trained first encoder; The projector and discriminator share weights and are used to converge sequences of photoplethysmography signals of similar states in the embedded space. The unsupervised representation loss submodule is used to construct the objective function of the pre-training task and to identify the sequence of the photoplethysmography signal according to the objective function of the pre-training task; wherein, the objective function is a weighted sum of the first loss and the second loss; the first loss includes three regularization terms: invariance term, variance term, and covariance term; the second loss is used to distinguish the representation of the clipping window within and outside the neighborhood to capture the vascular dynamic information under different physiological states; The downstream fine-tuning blood pressure estimation module includes: a second encoder and a downstream regressor; The second encoder has the parameters of the first encoder; the second encoder is used to process the labeled downstream fine-tuning dataset to obtain representational features; The downstream regressor is used to obtain blood pressure estimation data based on the characterization features; The unsupervised pulse wave representation learning module is trained using the unlabeled pre-training dataset, and the parameters of the first encoder in the unsupervised pulse wave representation learning module are transferred to the second encoder of the downstream fine-tuning blood pressure estimation module. The downstream fine-tuning blood pressure estimation module is trained using the labeled downstream fine-tuning dataset to obtain a trained blood pressure monitoring model.

2. The cuffless blood pressure estimation method based on unsupervised pulse wave representation learning according to claim 1, characterized in that, The sample dataset is preprocessed and augmented to obtain an unlabeled pre-training dataset and a labeled downstream fine-tuning dataset, including: The sample dataset is sequentially resampled, outlier removal is performed, filtering and noise reduction is applied, phase alignment is performed, normalization is performed, and segmentation is performed to obtain the preprocessed dataset. Gaussian noise, baseline drift, motion interference, power frequency interference, and respiratory sinus rhythm frequency modulation are sequentially added to the preprocessed dataset to obtain the unlabeled pre-training dataset and the labeled downstream fine-tuning dataset, respectively.

3. A cuffless blood pressure estimation system based on unsupervised pulse wave representation learning, characterized in that, include: The signal acquisition unit is used to acquire the photoplethysmography signal of the target under test; The waveform monitoring unit is used to input the photoplethysmography signal to be measured into the constructed blood pressure monitoring model to obtain the blood pressure waveform of the target being measured. The waveform monitoring unit includes: The sample acquisition subunit is used to acquire a sample dataset; the sample dataset includes an unlabeled subset with photoplethysmography (PPG) signals and a paired subset with PPG signals and blood pressure waveform labels; The sample processing subunit is used to perform data preprocessing and data augmentation on the sample dataset to obtain an unlabeled pre-training dataset and a labeled downstream fine-tuning dataset. The model building subunit is used to build the unsupervised pulse wave representation learning module and the downstream fine-tuning blood pressure estimation module; The first training subunit is used to train the unsupervised pulse wave representation learning module using the unlabeled pre-training dataset, and to transfer the parameters of the first encoder in the unsupervised pulse wave representation learning module to the second encoder of the downstream fine-tuning blood pressure estimation module. The second training subunit is used to train the downstream fine-tuning blood pressure estimation module using the labeled downstream fine-tuning dataset to obtain a trained blood pressure monitoring model. The unsupervised pulse wave representation learning module includes: a temporal neighborhood coding submodule, a first encoder, a projector and a discriminator, and an unsupervised representation loss submodule; The temporal neighborhood coding submodule is used to distinguish between neighborhood and non-neighborhood clipping windows of the photoplethysmography signal in the unlabeled pre-training dataset; wherein, for any clipping window, the unit root test method is used to determine the neighborhood range with the same physiological state. The first encoder is used to train based on the cropping window to obtain the parameters of the trained first encoder; The projector and discriminator share weights and are used to converge sequences of photoplethysmography signals of similar states in the embedded space. The unsupervised representation loss submodule is used to construct the objective function of the pre-training task and to identify the sequence of the photoplethysmography signal according to the objective function of the pre-training task; wherein, the objective function is a weighted sum of the first loss and the second loss; the first loss includes three regularization terms: invariance term, variance term, and covariance term; the second loss is used to distinguish the representation of the clipping window within and outside the neighborhood to capture the vascular dynamic information under different physiological states; The downstream fine-tuning blood pressure estimation module includes: a second encoder and a downstream regressor; The second encoder has the parameters of the first encoder; the second encoder is used to process the labeled downstream fine-tuning dataset to obtain representational features; The downstream regressor is used to obtain blood pressure estimation data based on the characterization features.

4. The cuffless blood pressure estimation system based on unsupervised pulse wave representation learning according to claim 3, characterized in that, The sample processing subunit includes: The preprocessing module is used to sequentially resample, remove outliers, filter and denoise, phase align, normalize and segment the sample dataset to obtain the preprocessed dataset. An enhancement module is used to sequentially add Gaussian noise, baseline drift, motion interference, power frequency interference, and respiratory sinus rhythm frequency modulation to the preprocessed dataset, thereby obtaining the unlabeled pre-training dataset and the labeled downstream fine-tuning dataset, respectively.

Images