Blood pressure signal measuring method based on medical big data and multiple training strategies
By combining unsupervised time-frequency feature learning and transfer learning, the invasiveness and insufficient accuracy of non-invasive blood pressure measurement is solved, and the accurate measurement of personalized non-invasive blood pressure signals is achieved, which is suitable for scenarios with limited individual data.
Patent Information
- Application Number
- CN202510379994.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-28
- Publication Date
- 2025-07-04
AI Technical Summary
The non-invasive blood pressure measurement methods in the prior art have problems such as strong invasiveness, inconvenient operation, insufficient accuracy and difficulty in data labeling. Especially when the amount of personalized data is limited, it is difficult to achieve accurate blood pressure estimation.
Combined with unsupervised time-frequency feature learning and transfer learning, by constructing a blood pressure signal measurement model based on medical big data, unsupervised pre-training is performed using labeled photoelectric volume pulse wave signals, supervised pre-training is performed in combination with labeled signals, and a small amount of personalized data is used for model fine-tuning in the fine-tuning stage to achieve personalized non-invasive blood pressure signal measurement.
Effectively extract the shallow and deep features of the photoelectric volume pulse wave signal, improve the accuracy and performance of blood pressure signal measurement, and realize the accurate measurement of personalized non-invasive blood pressure signals, which is suitable for scenarios with limited individual data.
Smart Images
Figure CN120256915A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of biomedical signal processing, and particularly relates to a blood pressure signal measurement method based on medical big data and various training strategies. Background Art
[0002] Blood pressure (BP) is one of the most important physiological parameters for measuring human cardiovascular health, and plays an important role in preventing cardiovascular diseases, regulating the balance of the internal environment, and maintaining life and health. Hypertension affects 30% of American adults and causes more than 410,000 deaths each year. Currently, hypertension has become a common chronic disease, with the prevalence increasing year by year and becoming a major cause of premature death in humans. To accurately diagnose and treat hypertension, it is necessary to measure blood pressure regularly.
[0003] Traditional cuff blood pressure monitors have certain limitations to some extent. It requires attaching a cuff around the patient's upper arm. The measurement process of the cuff blood pressure monitor requires manual operation. Improper operation will affect the measurement results and continuous monitoring cannot be achieved. An arterial catheter can continuously measure blood pressure, but it has strong invasiveness and is not practical in daily life. Therefore, it is particularly important to incorporate non-invasive and continuous blood pressure measurement into daily life in order to better understand and detect abnormal blood pressure fluctuations. Research on non-invasive continuous blood pressure monitoring technology has now become a hot issue in the field of medical devices.
[0004] In the existing technology, many methods extract the pulse transit time (PTT) by combining electrocardiogram (ECG) or photoplethysmogram signal (PPG), and then estimate the blood pressure value through physical modeling. The estimated blood pressure based on this method requires multiple sensors and frequent calibration. At the same time, most studies using this method use relatively simple linear models for blood pressure prediction, and the prediction effects on different data sets are not ideal.
[0005] In addition, many methods estimate blood pressure by analyzing the PPG waveform, select features with specific physical meanings, and fit the extracted features with blood pressure values based on traditional machine learning algorithms or shallow artificial neural networks. Manually designing and extracting features for model training often requires a lot of relevant knowledge and is highly subjective, which limits the feature extraction and non-linear processing capabilities of the neural network. Moreover, feature extraction is relatively cumbersome and is easily affected by various factors.
[0006] The deep learning method can automatically extract features and identify patterns from high-dimensional non-linear raw data by adopting a multi-layer neural network architecture, which can avoid the drawbacks of manual feature selection to a certain extent, and can learn a deep signal feature extraction method suitable for the specific task of blood pressure estimation.
[0007] Some researchers have proposed applying transfer learning to the training process of blood pressure signal measurement models. Transfer learning obtains knowledge from solving one problem (i.e., the source domain) and applies it to a different but related target problem (i.e., the target domain), where the target problem usually contains a small number of data samples to train the model. For example, the solution represented by the reference "Leitner J, Chiang P H, Dey S. Personalized Blood Pressure Estimation Using Photoplethysmography: A Transfer Learning Approach[J]. IEEE Journal of Biomedical and Health Informatics, 26(1): 218-228." proposes a deep learning method for personalized blood pressure (BP) estimation based on PPG signals, and solves the problem of limited PPG and BP data of the target individual to be measured by specific layers of a personalized pre-trained model. However, this method has the following problems: (1) The amount of target individual data used for personalized fine-tuning is large, and it is often difficult to generate enough personalized data in clinical settings. (2) The blood pressure estimation accuracy is poor.
[0008] With the development of intelligent wearable devices, PPG signals have become one of the most easily obtained physiological data. However, it is very difficult, expensive, and time-consuming to collect large-scale annotated labels of individual datasets by experts. To solve the above problems, for example, the solution represented by the reference "Zhang X, Zhao Z, Tsiligkaridis T, et al. Self-Supervised Contrastive Pre-Training for Time Series via Time-Frequency Consistency[C] / / 36th Conference on Neural Information Processing Systems (NeurIPS), 2022." proposes a self-supervised pre-training strategy for time series based on a time-frequency consistency (TF-C) model, introducing TF-C as a mechanism to support knowledge transfer between time series datasets, making the time-based and frequency-based representations and their local neighborhoods close in the latent space, and not requiring any labels during pre-training. The TF-C attribute can be used as a general attribute for pre-training and is applicable to various time series datasets. However, this method has the following problem: Only time-frequency features are extracted and the blood pressure signal cannot be directly output. Therefore, it is very necessary to combine transfer learning and TF-C to design a blood pressure signal measurement method based on medical big data and various training strategies. Summary of the Invention
[0009] To overcome the deficiencies in the above-mentioned existing technologies, the purpose of the present invention is to provide a blood pressure signal measurement method based on medical big data and multiple training strategies. By combining multiple training strategies based on a blood pressure signal measurement model, information integration at different levels and angles can be achieved, and accurate measurement of personalized non-invasive blood pressure signals can be realized through a large-sample general dataset and a small amount of labeled data of the individual to be measured.
[0010] To achieve the above objective, the technical solution adopted by the present invention is as follows:
[0011] A blood pressure signal measurement method based on medical big data and multiple training strategies, comprising the following steps:
[0012] Step 1: Collect the photoplethysmogram signal and the invasive continuous blood pressure waveform signal of the human body;
[0013] Step 2: Construct a blood pressure signal measurement model based on medical big data and multiple training strategies; the blood pressure signal measurement model includes a signal preprocessing module, an unsupervised time-frequency feature learning module, and a supervised neural network module;
[0014] Step 3: Use the unlabeled photoplethysmogram signal to perform unsupervised pre-training on the unsupervised time-frequency feature learning module; the unlabeled photoplethysmogram signal is a photoplethysmogram signal without a blood pressure signal label;
[0015] Step 4: Use the labeled photoplethysmogram signal to perform supervised pre-training on the supervised neural network module; the labeled photoplethysmogram signal is a photoplethysmogram signal with a blood pressure signal label;
[0016] Step 5: Collect the photoplethysmogram signal and the invasive continuous blood pressure waveform signal of the individual to be measured, and use the labeled photoplethysmogram signal of the individual to be measured to fine-tune the supervised neural network module obtained by supervised pre-training to obtain a trained personalized blood pressure signal model;
[0017] Step 6: Input the photoplethysmogram signal of the individual to be measured into the personalized blood pressure signal measurement model to calculate the blood pressure signal of the individual to be measured.
[0018] The photoplethysmogram signal and the invasive continuous blood pressure waveform signal in Step 1 come from multiple individuals.
[0019] The signal preprocessing module in Step 2 is specifically:
[0020] Resample the collected photoplethysmogram (PPG) signals and invasive continuous blood pressure waveforms of the human body; after resampling, denoise them; after denoising, segment them; after segmentation, normalize them. The unlabeled PPG signals after normalization are used as the unsupervised pre-training dataset, and the labeled PPG signals are used as the supervised pre-training dataset.
[0021] The unsupervised time-frequency feature learning module in step 2 is constructed based on the TF-C model, and simultaneously learns the representations containing time-domain and frequency-domain information, so as to effectively extract time-frequency features. Its input is the unsupervised pre-training dataset, and the output is the shallow features extracted from the unlabeled PPG signals, specifically including:
[0022] The time encoder module. Each segment of the unlabeled PPG signal generates a representation vector through time-domain encoding, and uses contrastive learning for unsupervised representation learning in the time domain, making the representation of the unlabeled PPG signal segment and the result of its data augmentation similar, and far from the representations of other unlabeled PPG signal segments.
[0023] The frequency encoder module. The unlabeled PPG signal segment is transformed into a frequency spectrum through Fourier transform, and each frequency spectrum of the unlabeled PPG signal segment generates a representation vector through frequency-domain encoding, and uses contrastive learning for unsupervised representation learning in the frequency domain; making the frequency spectrum obtained from the unlabeled PPG signal segment and the result of its data augmentation in the frequency domain similar, and far from the frequency spectra of other unlabeled PPG signal segments.
[0024] The time-frequency consistency module maps both the time-domain and frequency-domain representations to the joint time-frequency space, and uses contrastive learning for unsupervised representation learning in the joint time-frequency space to achieve the consistency of the time domain and the frequency domain; making the time-domain representation of the unlabeled PPG signal segment and the frequency-domain representation of the frequency spectrum it obtains similar, the time-domain representation of the unlabeled PPG signal segment far from the time-domain representations of other unlabeled PPG signal segments, and the frequency-domain representation of the frequency spectrum obtained from the unlabeled PPG signal segment far from the frequency-domain representations of other unlabeled PPG signal segments.
[0025] The supervised neural network module in step 2 is composed of a neural network, which learns the internal representation of the features extracted by the unsupervised time-frequency feature learning module and captures local features or sequence dependencies. Its input is the blood pressure label and the shallow features extracted by the unsupervised time-frequency feature learning module from the supervised pre-training dataset, and the output is the predicted blood pressure signal, including:
[0026] The convolution module, i.e., Conv, extracts localized features from the output of the unsupervised time-frequency feature learning module.
[0027] The gated recurrent unit module, i.e., GRU, captures and transmits the long-term dependencies in the time series of photoplethysmogram signals through the update gate and reset gate mechanisms;
[0028] The blood pressure value estimation module, i.e., FC, performs the final estimation of blood pressure based on the extracted features, that is, converts the extracted features into blood pressure through a fully connected layer.
[0029] The specific content of step three is as follows:
[0030] Initialize the parameters of the unsupervised time-frequency feature learning module, and pre-train the unsupervised time-frequency feature learning module using photoplethysmogram signals from multiple individuals.
[0031] The specific content of step four is as follows:
[0032] Initialize the parameters of the supervised neural network module, and pre-train the supervised neural network module using photoplethysmogram signals and invasive continuous blood pressure waveform signals from multiple individuals.
[0033] The specific content of step five is as follows:
[0034] Obtain the photoplethysmogram signal and invasive continuous blood pressure waveform signal of the individual to be measured and perform signal preprocessing to obtain the labeled photoplethysmogram signal as the supervised fine-tuning data set. Load the pre-trained supervised neural network module, where the parameters of the module are used as the initial parameters in the fine-tuning step. Use the photoplethysmogram signal and invasive continuous blood pressure waveform signal of a small number of individuals to be measured to fine-tune the supervised neural network module to obtain a trained personalized blood pressure signal measurement model.
[0035] The present invention also includes a system that can run the above-mentioned blood pressure signal measurement method based on medical big data and multiple training strategies.
[0036] The present invention also includes a device, including:
[0037] A memory: used to store a computer program for implementing the above-mentioned blood pressure signal measurement method based on medical big data and multiple training strategies;
[0038] A processor: used to implement the above-mentioned blood pressure signal measurement method when executing the computer program.
[0039] The present invention also includes a computer-readable storage medium, and the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned blood pressure signal measurement method based on medical big data and multiple training strategies is implemented.
[0040] Compared with the prior art, the present invention has the following advantages:
[0041] (1) The blood pressure signal measurement model designed by the present invention combines multiple training strategies, which can achieve information integration at different levels and angles, improve the model's understanding ability of physiological signals, effectively extract the shallow and deep features of the photoplethysmogram signal time series, learn the deep signal feature extraction method suitable for the specific task of blood pressure signal measurement, and improve the performance of blood pressure signal measurement.
[0042] (2) The unsupervised time-frequency feature learning module of the present invention is constructed based on the TF-C model, and simultaneously learns the representations containing time-domain and frequency-domain information to effectively extract time-frequency features; at the same time, the supervised neural network module is based on transfer learning and automatically learns the internal representations of the time-frequency features extracted by the unsupervised time-frequency feature learning module to capture local features or sequence dependencies; the two achieve the personalization of the model based on TF-C and transfer learning, which can effectively solve the problem of limited data of a single individual.
[0043] (3) In the fine-tuning stage of the present invention, only a small amount of photoplethysmogram signals and blood pressure signals of the individual to be measured are required as training data to complete the personalization of the model, which has a relatively broad application scenario.
[0044] In summary, the present invention is based on a blood pressure signal measurement model, combines multiple training strategies to achieve information integration at different levels and angles, and constructs a personalized blood pressure signal measurement model by combining pre-training on a large-sample general dataset and fine-tuning with a small amount of labeled data of the individual to be measured, effectively solving the problem of limited data of a single individual to be measured, so as to achieve accurate measurement of personalized non-invasive blood pressure signals. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 It is a schematic flow chart of the present invention.
[0046] Figure 2 It is a schematic structural diagram of the blood pressure signal measurement model in step two of the present invention.
[0047] Figure 3 It is a schematic structural diagram of the supervised neural network module of the present invention.
[0048] Figure 4 It is a schematic flow chart of the blood pressure signal measurement of the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0049] The present invention will be described in detail below with reference to the accompanying drawings.
[0050] Refer to Figure 1 , a blood pressure signal measurement method based on medical big data and multiple training strategies, refer to Figure 1 , including the following steps:
[0051] Step 1: Collect the photoplethysmogram (PPG) signal and the invasive continuous blood pressure waveform signal of the human body;
[0052] The PPG signal and the invasive continuous blood pressure waveform signal come from multiple individuals. The data can be downloaded from public datasets or collected independently.
[0053] In this embodiment, the publicly available MIMIC-Ⅲ (Medical Information Mart for Intensive Care) dataset is adopted. Since the signal records in this dataset are not complete, the present invention screened several pairs of PPG-BP data of individuals for the blood pressure signal measurement task.
[0054] Step 2: Construct a blood pressure signal measurement model based on medical big data and various training strategies;
[0055] In this embodiment, referring to Figure 2 , the blood pressure signal measurement model includes a signal preprocessing module, an unsupervised time-frequency feature learning module, and a supervised neural network module.
[0056] Construct the signal preprocessing module, specifically:
[0057] Resample the collected PPG signal and the invasive continuous blood pressure waveform of the human body; after resampling, denoise them; after denoising, segment them; after segmentation, normalize them. The unlabeled PPG signal after normalization is used as the unsupervised pre-training dataset, and the labeled PPG signal is used as the supervised pre-training dataset.
[0058] Construct the unsupervised time-frequency feature learning module. Based on the TF-C model, construct the unsupervised time-frequency feature learning module to simultaneously learn the representations containing time-domain and frequency-domain information, so as to effectively extract time-frequency features. The input of the unsupervised time-frequency feature learning module is the unsupervised pre-training dataset, and its output is the shallow features extracted from the unlabeled PPG signal, specifically including:
[0059] The time encoder module. Each unlabeled PPG signal segment generates a representation vector through time-domain encoding, and unsupervised representation learning is performed using contrastive learning in the time domain. Make the representation of the unlabeled PPG signal segment and the result of its data augmentation similar, and far from other unlabeled PPG signal segments.
[0060] The frequency encoder module generates a spectrum from the unlabeled photoplethysmography signal fragments through Fourier transform. Each unlabeled photoplethysmography signal fragment spectrum is encoded in the frequency domain to generate a representation vector, and unsupervised representation learning is performed in the frequency domain using contrastive learning. The spectrum obtained from the unlabeled photoplethysmography signal fragment is similar to the representation of the data enhancement result in the frequency domain, and is far away from the spectrum of other unlabeled photoplethysmography signal fragments;
[0061] The time-frequency consistency module maps both time domain and frequency domain representations to the joint time-frequency space, and uses contrastive learning to perform unsupervised representation learning in the joint time-frequency space to achieve consistency between the time domain and the frequency domain; the time domain representation of the unlabeled photoplethysmography signal fragment is similar to the frequency domain representation of the spectrum obtained, the time domain representation of the unlabeled photoplethysmography signal fragment is far away from the time domain representation of other unlabeled photoplethysmography signal fragments, and the frequency domain representation of the spectrum obtained from the unlabeled photoplethysmography signal fragment is far away from the frequency domain representation of other unlabeled photoplethysmography signal fragments.
[0062] Construct a supervised neural network module, specifically:
[0063] The supervised neural network module is composed of a neural network, automatically learning the intrinsic representation of the features extracted by the unsupervised time-frequency feature learning module, and capturing local features or sequence dependencies. The supervised neural network module inputs the blood pressure label and the shallow features extracted by the unsupervised time-frequency feature learning module from the supervised pre-training data set, and its output is the predicted blood pressure.
[0064] In this embodiment, refer to Figure 3 , construct a supervised neural network module, specifically:
[0065] The convolution module, i.e., Conv, is configured with multiple convolutional layers to extract localized features from the output of the unsupervised time-frequency feature learning module, where the initial layers extract lower-level features and the deeper layers extract higher-level features;
[0066] The gated recurrent unit module, or GRU, captures and propagates long-term dependencies in the time series of photoplethysmography signals through update gate and reset gate mechanisms;
[0067] The blood pressure value estimation module, namely FC, performs the final estimation of blood pressure based on the features extracted by the previous modules, that is, converts the extracted features into blood pressure through the fully connected layer.
[0068] Step 3: using an unlabeled photoplethysmogram signal to perform unsupervised pre-training on an unsupervised time-frequency feature learning module; the unlabeled photoplethysmogram signal is a photoplethysmogram signal without a blood pressure signal label;
[0069] In this embodiment, refer toFigure 2 When there is no supervision, the pre-training of the time-frequency feature learning module includes the following steps: initializing the parameters of the unsupervised time-frequency feature learning module, and using the photoplethysmogram signals from multiple individuals to pre-train the unsupervised time-frequency feature learning module.
[0070] Step Four: Use the labeled photoplethysmogram signals to perform supervised pre-training on the supervised neural network module; the labeled photoplethysmogram signals are the photoplethysmogram signals with blood pressure signal labels.
[0071] In this embodiment, referring to Figure 2 When there is supervision, the pre-training of the supervised neural network module includes the following steps: initializing the parameters of the supervised neural network module, and using the photoplethysmogram signals and invasive continuous blood pressure waveform signals from multiple individuals to pre-train the supervised neural network module.
[0072] Step Five: Collect the photoplethysmogram signals and invasive continuous blood pressure waveform signals of the target individual, and use the labeled photoplethysmogram signals of the target individual to fine-tune the supervised neural network module obtained by supervised pre-training to obtain a trained personalized blood pressure signal measurement model.
[0073] In this embodiment, referring to Figure 2 Collect the photoplethysmogram signals and invasive continuous blood pressure waveform signals of the target individual and perform signal preprocessing to obtain labeled photoplethysmogram signals as the supervised fine-tuning data set. Load the supervised neural network module obtained by pre-training, where the parameters of the module are used as the initial parameters in the fine-tuning step, and use the photoplethysmogram signals and invasive continuous blood pressure waveform signals of the target individual to fine-tune the supervised neural network module to obtain a trained personalized blood pressure signal measurement model.
[0074] Step Six: Input the photoplethysmogram signals of the target individual into the personalized blood pressure signal measurement model to estimate the blood pressure signals of the target individual.
[0075] In this embodiment, referring to Figure 4 Use the trained personalized blood pressure signal measurement model to process the photoplethysmogram signals of the target individual in real time and calculate the blood pressure signals of the target individual.
[0076] The present invention also includes a system capable of running the above-mentioned blood pressure signal measurement method based on medical big data and multiple training strategies.
[0077] The present invention also includes a device, including:
[0078] A memory: used to store a computer program for implementing the above-mentioned blood pressure signal measurement method based on medical big data and multiple training strategies.
[0079] Processor: When used to execute the computer program, it implements the above-mentioned method for measuring blood pressure signals based on medical big data and multiple training strategies.
[0080] The present invention further includes a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, it implements the above-mentioned method for measuring blood pressure signals based on medical big data and multiple training strategies.
[0081] In summary, in the present invention, first, the photoplethysmogram signal and invasive blood pressure waveform signal of the human body are collected to construct a measurement model; then, unsupervised pre-training is performed using unlabeled signals, and supervised pre-training is performed using labeled signals; then, the supervised neural network module is fine-tuned to obtain a personalized blood pressure measurement model; finally, the signal of the target individual is input into the model to estimate its blood pressure. This method can achieve accurate measurement of personalized non-invasive blood pressure signals in the case of a large-sample general dataset and a small amount of labeled datasets.
Claims
1. A blood pressure signal measurement method based on medical big data and multiple training strategies, characterized in that It includes the following steps: Step 1: Collect the photoplethysmogram (PPG) signal and the invasive continuous blood pressure waveform signal of the human body; Step 2: Construct a blood pressure signal measurement model based on medical big data and various training strategies; the blood pressure signal measurement model includes a signal preprocessing module, an unsupervised time-frequency feature learning module, and a supervised neural network module; Step 3: Use the unlabeled PPG signal to perform unsupervised pre-training on the unsupervised time-frequency feature learning module; the unlabeled PPG signal is the PPG signal without a blood pressure signal label; Step 4: Use the labeled PPG signal to perform supervised pre-training on the supervised neural network module; the labeled PPG signal is the PPG signal with a blood pressure signal label; Step 5: Collect the PPG signal and the invasive continuous blood pressure waveform signal of the target individual to be measured, and use the labeled PPG signal of the target individual to be measured to fine-tune the supervised neural network module obtained by supervised pre-training to obtain a trained personalized blood pressure signal model; Step 6: Input the PPG signal of the target individual to be measured into the personalized blood pressure signal measurement model to calculate the blood pressure signal of the target individual to be measured; The PPG signal and the invasive continuous blood pressure waveform signal in Step 1 come from multiple individuals.
2. The blood pressure signal measurement method based on medical big data and various training strategies according to claim 1, wherein, The signal preprocessing module in Step 2 is specifically: Resample the collected PPG signal and invasive continuous blood pressure waveform of the human body; after resampling, denoise it; After denoising, segment it, and after segmentation, normalize it. The unlabeled PPG signal after normalization is used as the unsupervised pre-training dataset, and the labeled PPG signal is used as the supervised pre-training dataset.
3. A blood pressure signal measurement method based on medical big data and multiple training strategies according to claim 1, characterized in that The unsupervised time-frequency feature learning module in Step 2 is constructed based on the TF-C model, and simultaneously learns the representations containing time-domain and frequency-domain information to effectively extract time-frequency features. Its input is the unsupervised pre-training dataset, and the output is the shallow features extracted from the unlabeled PPG signal, specifically including: A time encoder module. Each unlabeled PPG signal segment generates a representation vector through time-domain encoding, and performs unsupervised representation learning using contrastive learning in the time domain, making the representation of the unlabeled PPG signal segment and the result of its data augmentation similar, and far from other unlabeled PPG signal segments; A frequency encoder module. Generate the frequency spectrum of the unlabeled PPG signal segment through Fourier transform. Each frequency spectrum of the unlabeled PPG signal segment generates a representation vector through frequency-domain encoding, and performs unsupervised representation learning using contrastive learning in the frequency domain; making the frequency spectrum obtained from the unlabeled PPG signal segment and the result of its data augmentation in the frequency domain similar, and far from the frequency spectra of other unlabeled PPG signal segments; The time-frequency consistency module maps both time domain and frequency domain representations to the joint time-frequency space, and uses contrastive learning to perform unsupervised representation learning in the joint time-frequency space to achieve consistency between the time domain and the frequency domain; the time domain representation of the unlabeled photoplethysmography signal fragment is similar to the frequency domain representation of the spectrum obtained, the time domain representation of the unlabeled photoplethysmography signal fragment is far away from the time domain representation of other unlabeled photoplethysmography signal fragments, and the frequency domain representation of the spectrum obtained from the unlabeled photoplethysmography signal fragment is far away from the frequency domain representation of other unlabeled photoplethysmography signal fragments.
4. A blood pressure signal measurement method based on medical big data and multiple training strategies according to claim 1, characterized in that, The supervised neural network module in step 2 is composed of a neural network, which learns the intrinsic representation of the features extracted by the unsupervised time-frequency feature learning module, captures local features or sequence dependencies, and its input is the blood pressure label and the shallow features extracted by the unsupervised time-frequency feature learning module from the supervised pre-training data set, and the output is the predicted blood pressure signal, including: The convolution module, i.e. Conv, extracts localized features from the output of the unsupervised time-frequency feature learning module. The gated recurrent unit module, or GRU, captures and propagates long-term dependencies in the time series of photoplethysmography signals through update gate and reset gate mechanisms; The blood pressure value estimation module, namely FC, makes the final estimation of blood pressure based on the extracted features, that is, converts the extracted features into blood pressure through the fully connected layer.
5. A blood pressure signal measurement method based on medical big data and multiple training strategies according to claim 1, characterized in that The step three is specifically as follows: Initialize the parameters of the unsupervised time-frequency feature learning module, and use the photoplethysmography signals from multiple individuals to pre-train the unsupervised time-frequency feature learning module.
6. A blood pressure signal measurement method based on medical big data and multiple training strategies according to claim 1, characterized in that, The step 4 is specifically as follows: Initialize the parameters of the supervised neural network module, and pre-train the supervised neural network module using photoplethysmographic signals and invasive continuous blood pressure waveform signals from multiple individuals.
7. A blood pressure signal measurement method based on medical big data and multiple training strategies according to claim 1, characterized in that, The step five is specifically as follows: The photoplethysmogram signal and invasive continuous blood pressure waveform signal of the target individual to be measured are obtained and preprocessed to obtain labeled photoplethysmogram signals as a supervised fine-tuning data set, and the pre-trained supervised neural network module is loaded, wherein the parameters of the module are used as initial parameters in the fine-tuning step, and the supervised neural network module is fine-tuned using a small amount of photoplethysmogram signals and invasive continuous blood pressure waveform signals of the target individual to be measured to obtain a trained personalized blood pressure signal measurement model.
8. A system, characterized in that, A blood pressure signal measurement method based on medical big data and multiple training strategies can be run as described in any one of claims 1-7.
9. A device, characterized in that, include: Memory: used to store a computer program for implementing a blood pressure signal measurement method based on medical big data and multiple training strategies as described in any one of claims 1 to 7; Processor: used to implement a blood pressure signal measurement method based on medical big data and multiple training strategies as described in any one of claims 1-7 when executing the computer program.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the blood pressure signal measurement method based on medical big data and multiple training strategies as described in any one of claims 1-7.