A personalized blood pressure monitoring method based on a population prototype library

Abstract

The application discloses a kind of individualized blood pressure monitoring methods based on group prototype library, belong to biomedical signal processing and artificial intelligence technical field.Aiming at the problem that existing model is poor in generalization ability on new user, and needs a lot of calibration data, the application first constructs group prototype library: collect the pressure pulse wave signals and reference blood pressure of multiple users, extract waveform feature prototype through pre-training encoder, and store the individualized prediction model of each user in association.For new user, based on the theory of few-sample learning, collect 3-5 calibration data to extract target features, calculate the similarity with the prototype in the library to select K neighbor models;Further, evaluate the prediction uncertainty of each neighbor model on the calibration data, calculate the weighted fusion weight based on the dual mechanism of similarity and uncertainty, and dynamically adjust the weight parameter according to the data volume to construct the individualized model of target user.

Description

Technical Field

[0001] This invention relates to the fields of biomedical signal processing and artificial intelligence, specifically to a non-invasive blood pressure monitoring method based on pressure pulse wave signals, and more particularly to a small-sample personalized calibration method based on population prototype matching. Background Technology

[0002] With the deep integration of the Internet of Things, artificial intelligence, and healthcare, non-invasive continuous blood pressure monitoring technology based on biomedical signals has become a research hotspot in the field of intelligent health monitoring. Among them, pressure pulse wave technology is of great value in the daily management of hypertension because it can capture the dynamic trend of blood pressure changes. Pressure pulse wave signals have a high signal-to-noise ratio and rich physiological characteristics, and their waveform characteristics are clearly correlated with blood pressure, providing a good foundation for accurate modeling.

[0003] However, this technology still faces a core bottleneck in its transition from laboratory to practical application: insufficient inter-individual generalization ability of the model. Due to significant differences in physiological parameters such as age, vascular elasticity, and underlying diseases among different users, the mapping relationship between their pressure pulse wave morphology and blood pressure values ​​exhibits highly personalized characteristics. General blood pressure prediction models trained on large-scale population data can achieve clinically acceptable accuracy in known user groups through empirical risk minimization. However, when directly applied to entirely new users, the prediction error increases dramatically due to data distribution bias—a phenomenon known as the "inter-user generalization gap."

[0004] To bridge this gap, existing technological approaches mainly fall into two categories: First, enhancing the generalization ability of a general model by collecting massive amounts of diverse training data, attempting to cover as many physiological variations as possible. However, collecting pressure pulse wave data is costly and complex, especially requiring the simultaneous acquisition of high-precision reference blood pressure values. Building a massive database covering the entire population presents significant economic and operational challenges. Second, building a personalized model for each user requires collecting dozens or even hundreds of sets of synchronized calibration data for new users, followed by independent model training or fine-tuning. While this method achieves high personalization, the lengthy data collection cycle and cumbersome calibration process severely damage the user experience, contradicting the "wear and use" philosophy of wearable devices.

[0005] In recent years, the rise of transfer learning and few-shot learning methods has provided new ideas for solving the above problems. These methods aim to transfer knowledge from general models to new users, thereby reducing dependence on the amount of data for the target user. However, when faced with blood pressure monitoring tasks with significant individual differences, existing transfer methods still require a certain amount of calibration data to ensure prediction stability, and relying solely on feature similarity matching is prone to mismatches due to data sparsity.

[0006] Therefore, there is an urgent need for a technical solution that can quickly generate high-precision personalized models with only a very small amount of calibration data. Through efficient transfer learning or few-shot learning mechanisms, the generalization gap between users can be bridged with extremely low calibration costs, enabling rapid personalized adaptation of blood pressure monitoring models. Summary of the Invention

[0007] (a) Purpose of the invention This invention aims to address the technical bottlenecks of existing blood pressure monitoring technologies based on pressure pulse waves, such as poor generalization ability among new users and the need for large amounts of data for personalized calibration. It provides a personalized blood pressure monitoring method based on a population prototype library. By constructing a "knowledge base" containing multi-user feature prototypes and personalized models, it enables the rapid generation of high-precision personalized models with only a small amount of calibration data, significantly improving the user experience and measurement accuracy of blood pressure monitoring devices.

[0008] (II) Technical Solution To achieve the above objectives, this invention proposes a personalized blood pressure monitoring method based on a group prototype library. The core idea of ​​this invention is to construct a "knowledge base" containing multi-user feature prototypes and personalized models. By quickly matching the similarity between new users and prototypes in the library, and intelligently fusing models from multiple similar prototypes, a personalized prediction model for new users can be generated using only a small amount of data.

[0009] like Figure 1 As shown, the method includes the following steps performed sequentially: Step 1: Construct a group prototype library by acquiring pressure pulse wave signal data and reference blood pressure values ​​from M historical users. Each historical user contains L sets of synchronously collected data. A personalized blood pressure prediction model is trained using all the data from each user. Simultaneously, a pre-trained feature encoder is utilized. (like Figure 3 (As shown) Extract features from all pulse wave signals of the user and calculate the mean value to obtain a feature prototype vector representing its waveform style. (like Figure 4 (As shown). Finally, a group prototype library is built. .

[0010] Step 2, New User Feature Extraction: Collect N sets of calibration measurement data (N is typically 3-5) from the target user. Each set includes a synchronously acquired pressure pulse wave signal and its reference blood pressure value. Use the feature encoder described in Step 1. Extract the pulse wave features from the calibration data of the target user, and calculate the mean of all feature vectors to obtain the target feature vector t̄. This vector serves as the unique identifier of the user in the feature space.

[0011] Step 3, Similarity matching and neighbor selection: Calculate the target feature vector t̄ and the prototype vectors in the prototype library. similarity .

[0012] Sort the prototypes by similarity from highest to lowest, and select the K prototypes with the highest similarity to form the neighbor prototype set. , where K is a preset integer, and 1 ≤ K ≤ M. This set represents the group of historical users who are most similar to the current new user in terms of waveform characteristics.

[0013] Step 4, Personalized model integration and construction, for the neighbor prototype set N K For each prototype-model pair, the system performs a dual evaluation to determine its contribution to personalized modeling: Forecast uncertainty assessment: computational model Predicted values ​​on target user calibration data Compared with reference value The absolute error sequence between And calculate the sample standard deviation of the sequence as the uncertainty. : This metric reflects the model's reliability in adapting to the current individual. It inherits the feature similarity calculated in step 3. This serves as a measure of physiological prior matching. By fusing these two pieces of information, ensemble weights are assigned to each neighbor model using a predefined weighting function: Where α and β are adjustment parameters greater than 0.

[0014] Finally, by performing a weighted linear combination of the prediction outputs of all neighboring models, a personalized blood pressure prediction model for the target user is constructed in real time. Step 5, Personalized Blood Pressure Monitoring and Iterative Optimization: Utilizing the constructed personalized blood pressure prediction model, real-time blood pressure values ​​are predicted based on continuously collected pressure pulse wave signals from the target user. To ensure continuous model accuracy, the system establishes an adaptive update mechanism: as monitoring data accumulates, the system periodically or based on performance evaluation triggers an optimization process, repeatedly executing steps 2 to 4. This involves iteratively updating and adjusting the parameters of the personalized prediction model using newly added data, allowing it to dynamically adapt to long-term changes in the user's physiological state. Simultaneously, when the amount of data accumulated by a target user exceeds a preset threshold, they can be converted into historical users, and the updated prototype-model pair can be added to the group prototype library, achieving dynamic expansion of the knowledge base.

[0015] (III) Beneficial Effects Compared with existing technologies, the personalized calibration method for blood pressure monitoring based on population prototype matching provided by this invention has the following significant advantages: First, high-precision personalization can be achieved with minimal calibration. This invention sets the calibration data volume to 3-5 times, a statistically sufficient value for estimating the mean and variance, and its effectiveness has been widely verified in engineering fields where individual differences are significant. Through a group prototype library and a similarity matching mechanism, this invention can quickly generate a high-precision personalized model upon first use by the user, with a prediction error significantly lower than the control group even with only a minimal calibration.

[0016] Second, a dual-weighting mechanism enhances model robustness and adaptability. This invention introduces prediction uncertainty as a second evaluation dimension, which, together with feature similarity, guides model integration, avoiding the risk of mismatches that may arise from relying solely on waveform features. By penalizing unstable models and rewarding reliable models, the final model integrates the depth of collective knowledge and individual validation.

[0017] Third, closed-loop iteration and dynamic parameter tuning endow the system with continuous evolution capabilities. This invention dynamically adjusts weight parameters based on the amount of data, achieving a smooth transition from "stability first" to "precise matching first"; simultaneously, personalized models can be iteratively updated as data accumulates, and excellent models can be fed back into the prototype library. This enables the system to have the ability to continuously learn and self-evolve, and its performance can continuously improve over time. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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 the structures shown in these drawings without creative effort.

[0019] Figure 1 This is an overall flowchart of the personalized blood pressure monitoring method based on a population prototype library according to the present invention; Figure 2 This is a flowchart illustrating the process of constructing a group prototype library and personalizing calibration for new users as described in this invention. Figure 3 This is a schematic diagram of the feature encoder pre-training process described in this invention; Figure 4 This is a schematic diagram illustrating the calculation of the user feature prototype vector according to the present invention; Figure 5 This is a schematic diagram of the similarity matching and neighbor selection process described in this invention; Figure 6This is a schematic diagram of the dual-weighted fusion process for personalized model integration as described in this invention; Figure 7 This is a comparison of the blood pressure prediction accuracy of the method described in this invention with that of a single similarity matching method. Detailed Implementation

[0020] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Figures 1-7 The diagrams shown are flowcharts and experimental results of a preferred embodiment of the present invention, intended to help understand the technical concept of the present invention, and not to limit the scope of protection of the present invention in any way.

[0021] Example 1: Rapid Personalized Blood Pressure Monitoring Based on a Population Prototype Library This embodiment fully demonstrates the entire process of the method of the present invention, such as... Figure 1 As shown, it mainly includes five core steps: building a group prototype library, new user calibration and feature extraction, similarity matching and neighbor selection, personalized model integration and construction, and personalized blood pressure monitoring and iterative optimization.

[0022] Step 1: Building the group prototype library: like Figure 2 As shown on the left, the construction of the group prototype library includes two sub-steps: personalized model training and feature prototype vector extraction.

[0023] (1) Personalized model training This embodiment collected pressure pulse wave data from 500 historical users (M=500). Each user provided approximately 200 sets of synchronously acquired pressure pulse wave signals and reference blood pressure values. The reference blood pressure values ​​were obtained using a medical-grade upper arm electronic blood pressure monitor, with systolic blood pressure ranging from 90-180 mmHg and diastolic blood pressure ranging from 50-110 mmHg.

[0024] For the m-th user, a lightweight neural network is trained using all 200 sets of data as a personalized blood pressure prediction model. The network structure is as follows: Input layer: Receives a 3-second pressure pulse wave signal (sampling rate 200Hz, i.e., 600 sampling points). Hidden Layer 1: A one-dimensional convolutional layer with 32 kernels, a kernel size of 5, a stride of 1, and the activation function ReLU. Hidden layer 2: One-dimensional convolutional layer with 64 kernels, kernel size 5, stride 1, and ReLU activation function. Hidden layer 3: Global average pooling layer Hidden layer 4: Fully connected layer, 64 neurons, activation function ReLU Output layer: Fully connected layer, 2 neurons, outputting systolic and diastolic blood pressure respectively. Training parameters: batch size 32, learning rate 0.001, optimizer Adam, loss function mean squared error, training epochs 100, early stopping strategy to prevent overfitting.

[0025] (2) Feature encoder pre-training like Figure 3 As shown, a system is constructed using an encoder. and decoder The network consists of an autoencoder. The encoder structure is the same as the convolutional part of the personalized model described above (hidden layers 1-3), outputting a 64-dimensional feature vector. The decoder is a symmetric structure of the encoder, upsampling the 64-dimensional feature vector to reconstruct it back to the original signal dimension.

[0026] The autoencoder was trained using 100,000 unlabeled pressure pulse wave data points from all 500 users. The training objective was to minimize the reconstruction error. After training, the decoder was discarded, and the encoder was retained. This encoder can map any segment of pressure pulse wave signal into a 64-dimensional feature vector, which encodes the morphological features of the signal.

[0027] (3) Calculation of feature prototype vector For all 200 pulse wave signals of the m-th user, the encoder E(⋅) is used to extract features one by one, resulting in 200 64-dimensional feature vectors. Then, the mean of these vectors is calculated as the feature prototype vector for that user. : Finally, a group prototype library P is constructed, containing 500 records.

[0028] Step 2: New User Calibration and Feature Extraction like Figure 2 As shown on the right, new users are guided to perform three calibration measurements upon first use. During each measurement, the user wears both the pressure pulse wave acquisition device and a standard cuff blood pressure monitor simultaneously. Three sets of data were collected: .in This is a pressure pulse wave signal. These are the corresponding reference values ​​for systolic and diastolic blood pressure.

[0029] Using a pre-trained encoder Extract the feature vectors of these three pulse wave signals to obtain The target feature vector is obtained by calculating the mean. . Step 3: Similarity Matching and Neighbor Selection like Figure 5 As shown, the target feature vector is calculated. With 500 prototype vectors in the group prototype library Euclidean distance Convert Euclidean distance to similarity: scale parameter Set to the average of the pairwise Euclidean distances between all prototype vectors in the prototype library.

[0030] Sort the prototypes by similarity from highest to lowest, and select the K=5 prototypes with the highest similarity to obtain the neighbor prototype set N. K . Step 4: Personalized Model Integration like Figure 6 As shown, the present invention employs a dual-weighted fusion mechanism, which comprehensively considers feature similarity and prediction uncertainty.

[0031] (1) Calculate the uncertainty of prediction For each model in the neighbor set It was applied to the three calibrated pulse wave signals of the new user. , to obtain the predicted value The root mean square error is calculated as an indicator of uncertainty. in , This is the mean of the absolute error sequence.

[0032] (2) Calculate the integration weights The weighting formula is used (in this embodiment, α=1, β=1): (3) Constructing personalized models The personalized blood pressure prediction model for new users is defined as a weighted ensemble of five neighboring models. For any new pulse wave signal x, the predicted blood pressure value is calculated as follows: Step 5: Personalized monitoring and iterative optimization After new users complete calibration, they enter the daily monitoring phase. The system collects pressure pulse wave signals in real time, inputs them into the integrated model mentioned above, and continuously outputs blood pressure values. After one week of continuous use, the system has collected a total of 50 sets of valid monitoring data. The system automatically triggers an iterative optimization process: Use these 50 sets of data as the new calibration dataset and repeat steps 2 to 4. At this point, the amount of calibration data N=50, and according to the dynamic parameter tuning rules, α=2.0 and β=0.5 are set to make the weight allocation more inclined towards the model with high feature similarity; Generate an optimized personalized model and recalculate the user's feature prototype vector. The updated prototype-model pairs are added as new entries to the group prototype library, expanding the prototype library to 501 entries and enabling dynamic updates of the knowledge base.

[0033] Example 2: Detailed Explanation of Dynamic Parameter Tuning Strategy This embodiment details the dynamic adjustment strategy of parameters α and β in the method of the present invention, and combines it with... Figure 7 The experimental results shown verify its effectiveness.

[0034] Scenario 1: Very small amount of calibration data (N=3) We set the adjustment parameters α=0.3 and β=2.0 to ensure that the weights are primarily driven by uncertainty. Consider two candidate neighbor models: Model A: While exhibiting high similarity to the waveform features of the target user, its prediction errors differ significantly across the three calibrations, indicating poor prediction consistency. The calculated uncertainty index... High; Model B: While its waveform characteristics are slightly less similar to those of the target user, its prediction errors across three calibrations are quite close, indicating good prediction stability. The calculated uncertainty index... Lower.

[0035] Using weight formula Calculation, due to and Model B received significantly higher ensemble weights than Model A. This result indicates that when calibration data is extremely limited, the system prioritizes trusting models that perform more stably on a small amount of data, even if their waveform similarity is slightly lower, to avoid mismatches caused by isolated outliers.

[0036] Scenario 2: Medium data volume (N=8) Setting α=1.0 and β=1.0 ensures that similarity and uncertainty have equal importance in the weighting decision. In this case, the weights are dominated by ratios, balancing the matching degree of waveform features and the stability of model predictions. This approach is suitable for scenarios where users already have a certain amount of calibration data but still need to leverage collective knowledge.

[0037] Scenario 3: Sufficient data volume (N=15) Setting α=2.0 and β=0.5 allows the weights to be primarily driven by feature similarity. When users accumulate sufficient calibration data, the model can accurately characterize their physiological features. At this point, historical user models that are highly similar to the user's waveform should be trusted, as these models represent similar vascular physiological structures and blood pressure regulation mechanisms.

[0038] To verify the technical effectiveness of the method of this invention, a comparative test was conducted on 50 new users. Each user provided multiple sets of synchronously acquired pressure pulse wave signals and reference blood pressure values, with the first three sets serving as initial calibration data and the remainder as a validation set. The mean absolute error (MAE) was used as the evaluation index for prediction accuracy.

[0039] Control group: Only K=5 most similar historical user models were selected based on feature similarity, and ensembled with equal weights. The prediction result was the average of the outputs of the 5 models.

[0040] Experimental group (the method described in this invention): The dual-weighted integration method based on similarity and uncertainty as described in claim 1 is used to dynamically adjust the α and β parameters according to the amount of calibration data.

[0041] Experimental results are as follows Figure 7 As shown: In regions with limited data (N≤5): the method of this invention increases the β value to make uncertainty dominate the weight. In this region, the prediction error is significantly lower than that of the control group, and the error fluctuation range is smaller, indicating its superiority and stability under extremely limited calibration data. This result is highly consistent with the description of Scenario 1 in Example 2.

[0042] In regions with moderate amounts of data (N=5-10): as the amount of data increases, the system gradually balances the weights of similarity and uncertainty, the error curve continues to decrease, the gap between the two methods remains stable, and the method of this invention is always superior to the control group.

[0043] In regions with sufficient data (N≥10): By increasing the α value, the similarity weight of the method increases, further reducing the error of the method and continuously expanding its advantage over the control group, thus verifying its ability to fully utilize group knowledge when data is sufficient.

[0044] Figure 7 Overall, the effectiveness and superiority of the dynamic parameter tuning strategy of this invention are verified: with a very small amount of calibration data, the method of this invention avoids the risk of overfitting through the uncertainty-dominated mechanism; as the amount of data accumulates, it gradually transitions to the similarity-dominated mechanism, making full use of group knowledge to achieve accurate adaptation; the error curve decreases smoothly throughout the process, and the stability is better than the control group.

[0045] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A personalized blood pressure monitoring method based on a population prototype library, characterized in that, Includes the following steps: Step 1: Construct a group prototype library by acquiring pressure pulse wave signal data from M historical users. Each historical user's data contains L sets of synchronously acquired pressure pulse wave signals. and corresponding reference blood pressure values Use all the data A personalized blood pressure prediction model was obtained through training. The model takes a single-segment pressure pulse wave signal as input, and then processes all pressure pulse wave signals... Input a pre-trained feature encoder A set of high-dimensional feature vectors is obtained. The mean value is then used as the waveform feature prototype vector for that user. The calculation method is as follows: waveform feature prototype vectors of each historical user Corresponding personalized blood pressure prediction model Corresponding storage, building a group prototype library ; Step 2, New User Calibration and Feature Extraction: Collect N sets of calibration measurement data from the target user, each set including pressure pulse wave signal. and its reference blood pressure value ; Using the feature encoder described in step 1 Extract the pulse wave features from the target user's calibration data and calculate the mean of all feature vectors to obtain the target feature vector t̄; Step 3, Similarity matching and neighbor selection: Calculate the target feature vector t̄ and each prototype vector in the group prototype library P. similarity between Sort the prototypes by similarity from highest to lowest, and select the K prototypes with the highest similarity to form the neighbor prototype set. , where K is a preset integer, and 1 ≤ K ≤ M; Step 4, Personalized model integration and construction, for the neighbor prototype set N K Each matching prototype in Calculate its corresponding model Predictive uncertainty on target user calibration data Based on similarity With uncertainty Calculate the integration weights of each prototype A personalized blood pressure prediction model for the target user is constructed through weighted fusion, where the weight calculation formula is as follows: Where α and β are adjustment parameters greater than 0; Step 5, Personalized blood pressure monitoring and iterative optimization: Using the constructed personalized blood pressure prediction model, predict the blood pressure value of the pressure pulse wave signal subsequently collected from the target user; As the amount of data from the target user accumulates, repeat steps 2-4 periodically, and use the accumulated new data to iteratively optimize the personalized blood pressure prediction model.

2. The method according to claim 1, characterized in that, The feature encoder It was obtained through training in the following manner: Build by encoder and decoder The self-encoder network architecture is composed of; A large amount of unlabeled pressure pulse wave signal data was used to minimize the reconstruction loss between the input signal and the reconstructed signal. The primary training objective is to achieve this; a temporal consistency constraint loss is also introduced. The feature encoder is required to maintain continuity in the feature space for features extracted from adjacent pulse wave cycles; the feature encoder The total loss function is: , where λ is a hyperparameter that adjusts the weights of the two loss terms.

3. The method according to claim 1, characterized in that, The similarity described in step 3 The calculation formula is: in Represents the target feature vector With prototype vector The Euclidean distance between them, where σ is the scale parameter, is set by the average Euclidean distance between all pairwise prototype vectors in the group prototype library P.

4. The method according to claim 1, characterized in that, The uncertainty described in step 4 Calculated using statistical methods: Computational model Predicted value sequence on N sets of calibration data Compared with reference blood pressure value sequence absolute error between Calculate the sample standard deviation of the absolute error sequence as the uncertainty. in This is the mean of the absolute error sequence.

5. The method according to claim 1, characterized in that, The weighted fusion described in step 4 is implemented using an output layer fusion method: For an input pulse wave signal x, the personalized blood pressure prediction model for the target user outputs a weighted sum of the predictions from each neighboring model, i.e.: 。 6. The method according to claim 1, characterized in that, This includes the dynamic update steps for the group prototype library: When the amount of monitoring data accumulated by a target user exceeds a preset threshold Q, the user is converted into a historical user, and its updated feature prototype vector and corresponding personalized model are added as new prototypes to the group prototype library P. When the number of prototypes in the group prototype library exceeds a preset limit, some prototypes are deleted according to the elimination criteria. The elimination criteria include at least one of the following: timeliness criteria based on usage frequency, redundancy criteria based on feature similarity, or accuracy criteria based on model generalization ability.

7. The method according to claim 1, characterized in that, The adjustment parameters α and β mentioned in step four are dynamically adjusted based on the amount of calibration data N of the target user: When N ≤ 3, set α = 0.5, β = 2.0; when 3 < N ≤ 10, set α = 1.0, β = 1.0; when N > 10, set α = 2.0, β = 0.5.

Images