PPG blood pressure estimation method and system combining deep learning and clinical prior knowledge

Abstract

The application discloses a PPG blood pressure estimation method and system combining deep learning and clinical prior knowledge, and the PPG blood pressure estimation method combining deep learning and clinical prior knowledge first acquires a pulse wave signal through a signal acquisition module; then, the pulse wave signal is preprocessed, and a first branch pulse wave signal and a second branch pulse wave signal are respectively output; then, the first branch pulse wave signal is subjected to dimension conversion and multi-dimensional features are extracted through a deep learning network, and clinical prior knowledge features in the second branch pulse wave signal are extracted; finally, a blood pressure estimation model of a corresponding time period is constructed according to the multi-dimensional features and the clinical prior knowledge features, and the blood pressure value of the corresponding time period is predicted and output through the blood pressure estimation model, so that the correlation between the extracted PPG features and BP and the blood pressure measurement accuracy are improved through the double-branch feature extraction combining deep learning and clinical prior knowledge, the effectiveness and interpretability of the blood pressure estimation method are ensured, and the reliability of the estimated blood pressure value is improved.

Description

Technical Field

[0001] This invention relates to the field of blood pressure monitoring technology, and in particular to a PPG blood pressure estimation method and system that combines deep learning and clinical prior knowledge. Background Technology

[0002] Traditional blood pressure monitoring methods include non-invasive and invasive methods. Non-invasive blood pressure measurement requires a cuff and inflation, which results in a long measurement time and makes it difficult to achieve long-term continuous blood pressure monitoring. Although invasive blood pressure measurement can monitor blood pressure continuously for a long time, it is invasive and increases the risk of infection or even more serious complications.

[0003] Existing blood pressure monitoring methods based on photoplethysmograms (PPGs) can avoid the inflation and deflation of cuffs, achieving cuffless blood pressure measurement. These methods use PPGs for blood pressure measurement, first extracting various features from the PPG signal, then fitting a regression line using methods such as linear regression, and using the extracted features for blood pressure estimation. Alternatively, methods based on neural networks and deep learning can be used, with PPG as network input and continuous arterial blood pressure as labels for supervised learning, outputting continuous blood pressure (BP) values ​​to construct an end-to-end method. However, the accuracy of blood pressure measurement methods based solely on PPG features depends on the effectiveness of the extracted features, resulting in low accuracy and poor stability. Furthermore, using deep learning to construct an end-to-end measurement model relies excessively on the amount of data. Additionally, the standard blood pressure values ​​used in these methods are continuous arterial blood pressure waveforms, which require invasive measurement and pose a risk of infection. Summary of the Invention

[0004] The purpose of this invention is to design a PPG blood pressure estimation method that combines deep learning and clinical prior knowledge in order to solve the above problems.

[0005] The present invention achieves the above objectives through the following technical solutions:

[0006] A PPG blood pressure estimation method combining deep learning and prior clinical knowledge includes the following steps:

[0007] Acquire pulse wave signals;

[0008] The pulse wave signal is preprocessed to output the first branch pulse wave signal and the second branch pulse wave signal respectively.

[0009] The pulse wave signal of the first branch is dimensionally transformed, and multidimensional features are extracted through a deep learning network.

[0010] Extract clinical prior knowledge features from the pulse wave signal of the second branch;

[0011] Based on the multidimensional features and the clinical prior knowledge features, a blood pressure estimation model for the corresponding time period is constructed, and the blood pressure value for the corresponding time period is predicted and output through the blood pressure estimation model.

[0012] This invention also proposes a PPG blood pressure estimation system combining deep learning and clinical prior knowledge. This system is used to implement the aforementioned PPG blood pressure estimation method combining deep learning and clinical prior knowledge. The PPG blood pressure estimation system includes:

[0013] The signal acquisition module is used to acquire pulse wave signals;

[0014] The preprocessing module is used to preprocess the pulse wave signal;

[0015] The dual-branch feature extraction module is used to extract multidimensional features and clinical prior knowledge features through branch extraction.

[0016] The automatic blood pressure estimation model construction module is used to automatically construct a blood pressure estimation model for the corresponding time period based on the multidimensional features and clinical prior knowledge features, and output the blood pressure value for the corresponding time period.

[0017] The beneficial effects of this invention are as follows:

[0018] The PPG blood pressure estimation method combining deep learning and clinical prior knowledge first acquires pulse wave signals through a signal acquisition module. Then, the pulse wave signals are preprocessed to output first-branch and second-branch pulse wave signals respectively. Next, the first-branch pulse wave signal undergoes dimensionality transformation, and multi-dimensional features are extracted using a deep learning network. Clinical prior knowledge features are also extracted from the second-branch pulse wave signal. Finally, a blood pressure estimation model for the corresponding time period is constructed based on the multi-dimensional features and clinical prior knowledge features. The blood pressure value for the corresponding time period is predicted and output using this model. This dual-branch feature extraction method, combining deep learning and clinical prior knowledge, improves the correlation between the extracted PPG features and blood pressure (BP) and the accuracy of blood pressure measurement, ensuring the effectiveness and interpretability of the blood pressure estimation method and enhancing the reliability of the estimated blood pressure values. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the PPG blood pressure estimation method of this invention that combines deep learning and clinical prior knowledge.

[0020] Figure 2 This is an overall block diagram of the PPG blood pressure estimation method of the present invention, which combines deep learning and clinical prior knowledge.

[0021] Figure 3The image shows the convolutional signal waveform of the PPG blood pressure estimation method of this invention, which combines deep learning and clinical prior knowledge.

[0022] Figure 4 This is a peak systolic interval plot of the PPG blood pressure estimation method of this invention, which combines deep learning and clinical prior knowledge.

[0023] Figure 5 This is a schematic diagram of PPG segmentation in the PPG blood pressure estimation method of the present invention, which combines deep learning and clinical prior knowledge.

[0024] Figure 6 is a schematic diagram of the dimensional transformation of the PPG blood pressure estimation method of the present invention that combines deep learning and clinical prior knowledge;

[0025] Figure 7 is a schematic diagram of waveform feature extraction of clinical prior knowledge in the PPG blood pressure estimation method of this invention, which combines deep learning and clinical prior knowledge. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0027] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0028] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0029] In the description of this invention, it should be understood that the terms "upper," "lower," "inner," "outer," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0030] Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0031] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, terms such as "set" and "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0032] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0033] like Figure 1 As shown in Figure 7, the PPG blood pressure estimation method combining deep learning and prior clinical knowledge includes the following steps:

[0034] Acquire pulse wave signals;

[0035] The pulse wave signal is preprocessed to output the first branch pulse wave signal and the second branch pulse wave signal respectively.

[0036] The pulse wave signal of the first branch is dimensionally transformed, and multidimensional features are extracted through a deep learning network.

[0037] Extract clinical prior knowledge features from the pulse wave signal of the second branch;

[0038] Based on the multidimensional features and the clinical prior knowledge features, a blood pressure estimation model for the corresponding time period is constructed, and the blood pressure value for the corresponding time period is predicted and output through the blood pressure estimation model.

[0039] In one embodiment, the step of performing dimensionality transformation on the first branch pulse wave signal and extracting multidimensional features through a deep learning network specifically includes:

[0040] The pulse wave signal of the first branch is converted into a two-dimensional space to form a visibility map;

[0041] Based on the VGG19 model, local features and texture information are extracted from the visibility map;

[0042] By combining the convolution module and the MobileViT v2 module, the global topology in the visibility graph is extracted.

[0043] In one embodiment, the clinical prior knowledge features include temporal features, morphological features, and statistical features.

[0044] In one embodiment, the step of constructing a blood pressure estimation model for the corresponding time period based on the multidimensional features and the clinical prior knowledge features specifically includes:

[0045] The multidimensional features are fused with the clinical prior knowledge features to output the feature vector set for each time period.

[0046] Based on the feature vector set of each time period, a blood pressure estimation model for the corresponding time period is automatically constructed.

[0047] In one embodiment, the step of preprocessing the pulse wave signal specifically includes:

[0048] First, eliminate noise and discontinuities / saturation signals in the pulse wave signal, and then filter out abnormal segments in the pulse wave signal;

[0049] The pulse wave signal is then subjected to fine segmentation processing.

[0050] In this embodiment, a single-channel pulse wave sensor is first used to collect pulse wave signals at the fingertip, which results in a short measurement time and simpler data acquisition. Then, a preprocessing module is used to eliminate noise and discontinuous / saturated signals in the pulse wave signal and to filter out abnormal segments in the pulse wave signal to improve the quality of the PPG signal. Finally, the pulse wave signal is segmented.

[0051] Specifically, the elimination of noise and discontinuities / saturation signals in the pulse wave signal is as follows: For PPG data in the morning, afternoon, and evening time periods, a 5th-order Butterworth bandpass filter with a passband of 0.5 Hz-5 Hz is first used to eliminate baseline drift and high-frequency noise; then, discontinuous signals and signals with saturation amplitude are removed: the PPG signal is divided into 10-second segments. In each 10-second PPG segment, if any three adjacent sampling points have equal values, then the segment contains discontinuous or saturated PPG data, and this 10-second PPG segment is deleted.

[0052] In this embodiment, the process of filtering out abnormal segments in the pulse wave signal is as follows: First, the amplitude of the PPG signal in all segments is truncated to remove amplitude values ​​less than zero, leaving PPG signals with positive amplitudes for processing; then, two different convolution kernels are defined:

[0053]

[0054]

[0055] Among them, f sThe sampling frequency of the PPG is w1, and w2 are two manually defined time variables used to highlight each cardiac cycle (beat) and the PPG systolic peak, respectively. The values ​​of w1 and w2 are approximately equal to the duration of each cardiac cycle and each systolic peak, respectively, and are set to 0.667s and 0.111s. Then, these two convolution kernels are applied to the filtered PPG segment to generate two convolutional signals, namely the convolutional signals PPG. beat and PPG peak :

[0056]

[0057]

[0058] Among them, PPG seg This represents the filtered PPG fragment. Figure 3 In the convolution signal PPG beat and PPG peak Using different colors to show that convolution produces higher amplitudes within each cardiac cycle and systolic peak, respectively.

[0059] To further pinpoint each cardiac contraction peak, a threshold curve, denoted as TH, is calculated using the following formula. peak :

[0060]

[0061] Where α is related to PPG beat The relevant offset is defined as:

[0062]

[0063] Where z is PPG seg The arithmetic mean curve. In this embodiment, the coefficient β is manually set to 0.03.

[0064] Figure 4 PPG was also displayed. seg PPG peak and TH peak TH peak The sawtooth waveform indicates the position of each cardiac cycle in the PPG, while the PPG peak Each peak corresponds to a specific location of the contraction peak; therefore, we search for PPG. peak Greater than TH peak The interval is considered as the interval containing the PPG contraction peak. By counting the intervals within each segment, the number of contraction peaks in each PPG segment is obtained.

[0065] Normally, the normal heart rate range for adults is 60-100 beats per minute. Under normal circumstances, a 10-second PPG segment should have at least 10 cardiac cycles, resulting in 10 systolic peaks. Since the systolic peaks at the beginning and end of each segment may not be fully included in the segment, we set a threshold of 8 for the number of systolic peaks, retaining only PPG segments with more than 8 systolic peaks to avoid including poor-quality segments in the PPG signal.

[0066] Through the above operations, 3112 10-second PPG segments were finally obtained for the morning, afternoon and evening time periods. The mean and standard deviation of SBP (systolic pressure) and DBP (diastolic pressure) for the three time periods were 108.05±14.83 mmHg (SBP) and 70.02±10.36 mmHg (DBP), 108.01±14.25 mmHg (SBP) and 71±9.70 mmHg (DBP), and 112.17±9.71 mmHg (SBP) and 73.90±12.57 mmHg (DBP).

[0067] After the above operations, the PPG signal has been segmented into 10-second segments, ensuring the existence of a normal waveform. To obtain more detailed information within smaller time intervals, the PPG signal is further segmented in this embodiment; specifically, a systolic peak is added before and after each cardiac cycle to capture the PPG signal and its changing trend for each cycle. Figure 5 To further segment the 10-second PPG fragment, a PPG segmentation strategy defines a Region of Interest (ROI) for every three PPG contraction peaks. Each PPG ROI includes a complete cardiac cycle, as well as the preceding descending limb and the following ascending limb of the previous and next cycles, forming a PPG window.

[0068] Since one-dimensional analysis alone is insufficient to reveal the hidden patterns and dynamic characteristics in PPG signals, it is necessary to convert PPG signals to two-dimensional space for analysis. The PPG window is converted into a corresponding two-dimensional image based on the VG (Visibility Graph). Here, VG is an undirected graph generated based on the natural visibility between sample amplitudes in the time series. Natural visibility refers to the connectivity between nodes in the one-dimensional sequence.

[0069] Specifically, the amplitudes of all extracted PPG windows are first remapped to the range of 0 to 1. To determine the natural visibility between points in the sequence, the amplitude peaks of the PPG are considered as nodes with different heights, as shown in Figure 6(a). The visibility between two nodes is determined by the line segment connecting them. If the line segment connecting two nodes intersects the amplitudes of other nodes, then these two nodes do not have natural visibility; conversely, if the line segment connecting two nodes does not intersect the amplitudes of any other sample, then these two points have natural visibility. In Figure 6(a), the two nodes at the ends of the red line segment have natural visibility because the connection does not cross other amplitudes; while the two nodes at the ends of the blue line segment do not have natural visibility because their connection crosses other amplitudes. The line segment connecting any node to the other nodes is considered as an undirected edge, and each node and all undirected edges form an undirected graph.

[0070] Then, a binary image is generated based on the adjacency matrix of this undirected graph, where the x and y coordinates represent the indices of the nodes in the undirected graph. The value of each pixel in the binary image depends on the natural visibility between two nodes whose indices are represented by their x and y coordinates. If there is natural visibility between the two nodes, the pixel value is set to 1; otherwise, it is set to 0. Therefore, all PPG sequences are transformed into a symmetric image, VG, through the adjacency matrix of their undirected graphs, as shown in Figure 6, where the red and blue edges in (a) correspond to the white and black pixels in (b), respectively.

[0071] To improve the correlation between the extracted PPG features and BP, as well as the accuracy of blood pressure measurement, this embodiment employs a dual-branch feature extraction module that combines deep learning and clinical prior knowledge. Specifically, it extracts multidimensional features from one branch of the PPG signal while simultaneously extracting clinical prior knowledge features from the other branch of the PPG signal. This ensures the effectiveness and interpretability of the blood pressure estimation method and improves the reliability of the estimated blood pressure values.

[0072] This embodiment uses a transfer learning model in deep learning for feature extraction, namely, it uses two complementary network architectures, the VGG19 model and the MobileViT v2 module, to extract local connectivity patterns, global topology and multi-scale information from VG.

[0073] Understandably, to meet the model input requirements for three-channel RGB images, each VG needs to be copied three times to create a three-channel image, and then the image size is adjusted to 224×224 using bilinear interpolation. Specifically, the first network is based on the modified VGG19 model, which extracts rich local features and texture information from the VG image to obtain a 4096-dimensional feature vector, in order to understand the local changes of PPG signals related to DBP. The second network combines the convolutional module with the attention mechanism of MobileViTv2 to capture the global topological structure in the VG image, obtaining an 896-dimensional feature vector, in order to understand the overall dynamic features of PPG signals related to SBP. The two complementary feature vectors extracted by the two networks will be used as regression inputs for the DBP and SBP estimation models, respectively.

[0074] To address the limitations of deep learning networks in terms of poor feature interpretation and lack of clinical prior knowledge application, this embodiment extracts a series of physiologically significant and clinically interpretable features from PPG waveforms. Specifically, 72 feature parameters are extracted from each PPG window, including temporal features, morphological features, and statistical features.

[0075] Specifically, time-domain features can directly reflect the basic characteristics of PPG waveforms, such as amplitude, duration, and slope, all of which are related to the physiological mechanisms of blood pressure. As shown in Figure 7(a), 12 time-domain features were extracted from the PPG waveform based on amplitude, duration, and specific slope position. These features include PE (peak height), VD (valley depth), PD (period duration), AS (ascending slope), DS (descending slope), WH (waveform height), WD (waveform depth), AD (ascending duration), DD (descending duration), AA (ascending area), DA (descending area), and the K value. The K value is calculated from the maximum, average, and minimum values ​​of the PPG signal using a formula.

[0076]

[0077] Among them, PPG period This represents a complete PPG waveform cycle.

[0078] Morphological features can describe the geometry of the waveform in more detail and capture subtle changes that may reflect the dynamic characteristics of the vascular system. Morphological features include 8 feature parameters in 6 groups, that is, the PPG waveform is described by 48 feature parameters. The PPG waveform is vertically divided into two parts from the PEAK point. For the left part, WH is divided into 10 parts on the decomposition line. The horizontal distance from each WH segment point to the left waveform is calculated to obtain SW[x]. According to the distance from the WD segment point to the right waveform, DW[x] of the right part is calculated. The sum and ratio of SW[x] and DW[x] can be used to obtain the sum SSD of SW[x] and DW[x] and the ratio RSD of SW[x] and DW[x].

[0079] Statistical features can reflect the overall trend and variability of PPG signals, which are related to long-term blood pressure regulation mechanisms. This embodiment extracts six statistical features using the principles of central tendency, dispersion, and shape, describing the PPG signal as a probability distribution. The statistical features include mean (ME), variance (VAR), margin (MF), form factor (FF), skewness (SK), and kurtosis (KU). The calculation formulas for the six statistical features are as follows:

[0080]

[0081]

[0082]

[0083]

[0084]

[0085]

[0086] in, These are sampling points in the PPG cycle, where n represents the length of the signal cycle; statistical features not only supplement local information but also capture long-term changes related to autonomic nervous system regulation and vascular compliance.

[0087] To address the risks of model overfitting, the complexity of using traditional linear regression models due to nonlinear interactions between features, and the added complexity of dedicated models for SBP and DBP estimation at different time periods, this embodiment employs AutoML (Automated Machine Learning) as the modeling strategy. AutoML handles complex datasets and regression tasks with minimal human intervention and is easy to deploy. It optimizes the machine learning pipeline by exploring combinations of preprocessing steps, models, and hyperparameters. In this embodiment, TPOT (Tree-based Pipeline Optimization Tool) is preferred as the AutoML tool to simplify the model selection and parameter tuning process.

[0088] Understandably, the first step is to concatenate multidimensional features and clinical prior knowledge features to obtain a feature vector set for blood pressure estimation in each time period. Then, TPOT generates a random population based on the feature vector set for each time period, with each individual representing a random machine learning scheme. Specifically, this embodiment uses five-fold cross-validation to evaluate the fitness of each scheme, with MSE (Mean-squared error) as the evaluation metric. Fitness is achieved by TPOT simulating natural selection to select schemes with better performance for propagation, generating new schemes through crossover and mutation. Crossover combines parts of two schemes into a new scheme, while mutation randomly changes a part of the scheme. TPOT iteratively optimizes the pipeline through selection, crossover, and mutation to output the best-performing pipeline. Finally, the optimized scheme will be used to train dedicated models for SBP and DBP estimation in each time period. Each dedicated model can then output the predicted values ​​of SBP and DBP for the corresponding time period, thus addressing the problem of diurnal variation in BP by modeling different time periods separately, thereby improving the stability and practicality of BP measurement.

[0089] This invention presents a PPG blood pressure estimation method that combines deep learning and clinical prior knowledge. First, it acquires pulse wave signals through a signal acquisition module. Then, it preprocesses the pulse wave signals, outputting first-branch and second-branch pulse wave signals respectively. Next, it transforms the dimensions of the first-branch pulse wave signal and extracts multi-dimensional features using a deep learning network, while extracting clinical prior knowledge features from the second-branch pulse wave signal. Finally, it constructs a blood pressure estimation model for the corresponding time period based on the multi-dimensional features and clinical prior knowledge features, and predicts and outputs the blood pressure value for the corresponding time period using this model. This dual-branch feature extraction, combining deep learning and clinical prior knowledge, improves the correlation between the extracted PPG features and blood pressure (BP) and the accuracy of blood pressure measurement, ensuring the effectiveness and interpretability of the blood pressure estimation method and enhancing the reliability of the estimated blood pressure values.

[0090] This invention also proposes a PPG blood pressure estimation system that combines deep learning and clinical prior knowledge. This PPG blood pressure estimation system is used to implement the above-mentioned PPG blood pressure estimation method that combines deep learning and clinical prior knowledge. The specific method of this PPG blood pressure estimation method that combines deep learning and clinical prior knowledge is as described in the above embodiments. Since this PPG blood pressure estimation system that combines deep learning and clinical prior knowledge adopts all the technical solutions of all the above embodiments, it has at least all the beneficial effects brought about by the technical solutions of the above embodiments, which will not be elaborated here.

[0091] In one embodiment, the PPG blood pressure estimation system combining deep learning and clinical prior knowledge includes:

[0092] The signal acquisition module is used to acquire pulse wave signals;

[0093] The preprocessing module is used to preprocess the pulse wave signal;

[0094] The dual-branch feature extraction module is used to extract multidimensional features and clinical prior knowledge features through branch extraction.

[0095] The automatic blood pressure estimation model construction module is used to automatically construct a blood pressure estimation model for the corresponding time period based on the multidimensional features and clinical prior knowledge features, and output the blood pressure value for the corresponding time period.

[0096] In one embodiment, the signal acquisition module is a single-channel pulse wave sensor.

[0097] In this embodiment, the PPG blood pressure estimation system, which combines deep learning and clinical prior knowledge, acquires pulse signals through a signal acquisition module and realizes blood pressure estimation based on pulse waves through a preprocessing module, a bi-branch feature extraction module, and an automatic blood pressure estimation model construction module.

[0098] Blood pressure measurement methods based on PTT (Pulse Transmit Time) have explored the relationship between PWV (Pulse Wave Velocity) and blood pressure, constructing various physical models to represent blood pressure using PWV, thereby achieving cuffless blood pressure measurement. PWV is generally quantified using PTT, and PTT calculation requires acquiring two or more physiological signals, typically PPG and another physiological signal, such as ECG (Electrocardiogram). PTT is calculated using the time difference between the peak values ​​of the two synchronous signals. Multi-wavelength PPG can also be used to replace multiple physiological signals, but this requires complex deconstruction and reconstruction of the multi-wavelength PPG. These methods require simultaneous acquisition of multiple physiological signals from different or different sites or complex signal processing, resulting in problems such as complex signal acquisition, low signal quality, and complex signal processing. Therefore, the signal acquisition module in this embodiment is preferably a single-channel pulse wave sensor, that is, a single pulse heart rate sensor is used to acquire a single-channel PPG signal at the fingertip, and a single measurement does not exceed 5 minutes to ensure the simplicity of data acquisition. Moreover, the acquired single-channel PPG does not require complex processing. It only needs to go through conventional noise reduction and filtering methods to ensure the quality of the data used, thereby ensuring the speed of the signal preprocessing process, shortening the data processing time, and avoiding the simultaneous acquisition of multiple physiological signals to reduce the number of sensors in the signal acquisition module.

[0099] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A PPG blood pressure estimation method combining deep learning and clinical prior knowledge, characterized in that, The PPG blood pressure estimation method that combines deep learning and prior clinical knowledge includes the following steps: Acquire pulse wave signals; The pulse wave signal is preprocessed to output the first branch pulse wave signal and the second branch pulse wave signal respectively. The pulse wave signal of the first branch is dimensionally transformed, and multidimensional features are extracted through a deep learning network. Extract clinical prior knowledge features from the pulse wave signal of the second branch; Based on the multidimensional features and the clinical prior knowledge features, AutoML (Automatic Machine Learning) is used as the modeling strategy to construct a blood pressure estimation model for the corresponding time period, and the blood pressure value for the corresponding time period is predicted and output through the blood pressure estimation model. The specific steps for preprocessing the pulse wave signal are as follows: First, noise and discontinuities / saturation signals in the pulse wave signal are eliminated, and abnormal segments in the pulse wave signal are screened out. Specifically, eliminating noise and discontinuities / saturation signals in the pulse wave signal involves: for PPG data in the morning, afternoon, and evening time periods, a 5th-order Butterworth bandpass filter with a passband of 0.5 Hz-5 Hz is used to eliminate baseline drift and high-frequency noise; then, discontinuities and signals with saturated amplitudes are removed: the PPG signal is divided into 10-second segments. In each 10-second PPG segment, if any three adjacent sampling points have equal values, the segment contains discontinuous or saturated PPG data, and this 10-second PPG segment is deleted. Screening out abnormal segments in the pulse wave signal involves: truncating the amplitude of the PPG signal in all segments to remove amplitude values ​​less than zero, leaving PPG signals with positive amplitudes; two different convolution kernels are defined. , , where f s The sampling frequency of the PPG is w1, and w2 are two manually defined time variables used to highlight each cardiac cycle and the PPG contraction peak, respectively. The values ​​of w1 and w2 are equal to the duration of each cardiac cycle and each contraction peak, respectively. Then, these two convolution kernels are applied to the filtered PPG segment to generate two convolutional signals, namely the convolutional signal PPG. beat and PPG peak : , Among them, PPG seg The PPG segments represent filtered data, with convolution producing higher amplitudes within each cardiac cycle and systolic peak range; to further pinpoint each cardiac systolic peak, [the following is used]... Calculate a threshold curve, where α is the relationship between PPG and PPG. beat The relevant offset is represented as: z is PPG seg Find the arithmetic mean curve of PPG. peak Greater than TH peak The interval is considered as the interval containing PPG contraction peaks. By counting the intervals within each segment, the number of contraction peaks in each PPG segment is obtained. The threshold for the number of contraction peaks is set to 8, and only PPG segments with more than 8 contraction peaks are retained. β is a coefficient. The pulse wave signal is then finely segmented; specifically, a systolic peak is added before and after each cardiac cycle, and an ROI is defined for every three PPG systolic peaks. Each PPG ROI includes a complete cardiac cycle, as well as the previous and next descending branches of the previous and next cycles, forming a PPG window. The specific steps of performing dimensionality transformation on the first branch pulse wave signal and extracting multidimensional features through a deep learning network are as follows: The pulse wave signal of the first branch is converted into a two-dimensional space to form a visibility map. Specifically, the amplitudes of all extracted PPG windows are remapped to the range of 0 to 1. The amplitude peaks of the PPG are regarded as nodes with different heights. If the line segment connecting two nodes intersects the amplitudes of other nodes, then these two nodes do not have natural visibility. Conversely, if the line segment connecting two nodes does not intersect the amplitudes of any other sample, then these two points have natural visibility. The line segment connecting any node to the other nodes is regarded as an undirected edge, and each node and all undirected edges form an undirected graph. A binary image is generated based on the adjacency matrix of the undirected graph, where the horizontal and vertical coordinates represent the index of the node in the undirected graph. The value of each pixel in the binary image depends on the natural visibility between two nodes whose horizontal and vertical coordinates are used as indices. If there is natural visibility between the two nodes, the pixel value is set to 1; otherwise, it is set to 0. Based on the VGG19 model, local features and texture information are extracted from the visibility map; By combining the convolution module and the MobileViT v2 module, the global topology in the visibility graph is extracted; The clinical prior knowledge features include temporal features, morphological features, and statistical features; Time-domain features: Twelve time-domain features were extracted from the PPG waveform based on amplitude, duration, and specific slope position. These features include PE, VD, PD, AS, DS, WH, WD, AD, DD, AA, DA, and the K value. The K value is calculated from the maximum, average, and minimum values ​​of the PPG signal using a formula. Among them, PPG period This represents a complete PPG waveform cycle; Morphological characteristics: The PPG waveform is described by 48 characteristic parameters; the PPG waveform is vertically divided into two parts from the PEAK point. For the left part, WH is divided into 10 parts on the decomposition line. The horizontal distance from each WH segment point to the left waveform is calculated to obtain SW[x]; based on the distance from the WD segment point to the right waveform, DW[x] of the right part is calculated. The sum and ratio of SW[x] and DW[x] can be used to obtain the sum SSD of SW[x] and DW[x] and the ratio RSD of SW[x] and DW[x]. Statistical characteristics: Six statistical characteristics were extracted using the principles of central tendency, discreteness, and shape, describing the PPG signal as a probability distribution. These characteristics include mean (ME), variance (VAR), marginal factor (MF), form factor (FF), skewness (SK), and kurtosis (KU). , , , , , ;in, These are sampling points within the PPG cycle. This indicates the length of the signal period.

2. The PPG blood pressure estimation method combining deep learning and clinical prior knowledge according to claim 1, characterized in that, The specific steps for constructing a blood pressure estimation model for the corresponding time period based on the multidimensional features and the clinical prior knowledge features are as follows: The multidimensional features are fused with the clinical prior knowledge features to output the feature vector set for each time period. Based on the feature vector set of each time period, a blood pressure estimation model for the corresponding time period is automatically constructed.

3. A PPG blood pressure estimation system combining deep learning and prior clinical knowledge, characterized in that, The PPG blood pressure estimation system combining deep learning and clinical prior knowledge is used to implement the PPG blood pressure estimation method combining deep learning and clinical prior knowledge as described in any one of claims 1-2, wherein the PPG blood pressure estimation system combining deep learning and clinical prior knowledge includes: The signal acquisition module is used to acquire pulse wave signals; The preprocessing module is used to preprocess the pulse wave signal; The dual-branch feature extraction module is used to extract multidimensional features and clinical prior knowledge features through branch extraction. The automatic blood pressure estimation model construction module is used to automatically construct a blood pressure estimation model for the corresponding time period based on the multidimensional features and clinical prior knowledge features, and output the blood pressure value for the corresponding time period.

4. The PPG blood pressure estimation system combining deep learning and clinical prior knowledge according to claim 3, characterized in that, The signal acquisition module is a single-channel pulse wave sensor.

Images