Non-invasive blood pressure estimation method and system based on PPG and machine learning

Abstract

The invention belongs to the technical field of blood pressure monitoring, and discloses a non-invasive blood pressure estimation method and system based on PPG and machine learning, the method comprises the following steps: obtaining sample data of a plurality of objects, the sample data comprising blood pressure test data and multi-modal data corresponding to the blood pressure test data; constructing a training set based on the blood pressure test data and the multi-modal data corresponding to the blood pressure test data; a pre-constructed multi-modal space-time model is trained on the basis of the training set, a trained multi-modal space-time model is obtained, the trained multi-modal space-time model serves as a blood pressure estimation model, and the blood pressure estimation model is used for outputting a blood pressure estimation value of the measured object with actual multi-modal data of the measured object as input. The method has the advantages that the measurement experience is improved, the measurement performance is improved, the clinical application value is high, and popularization is easy.

Description

Technical Field

[0001] The present invention belongs to the technical field of blood pressure monitoring, and specifically relates to a non-invasive blood pressure estimation method and system based on PPG and machine learning. Background Art

[0002] Blood pressure monitoring is an important means of understanding and assessing an individual's cardiovascular health status, and is of great significance for maintaining people's health, preventing diseases, and improving the quality of life. In the field of blood pressure monitoring, the commonly used blood pressure measurement methods are as follows:

[0003] 1. Cuff-type blood pressure monitor: This monitor blocks blood flow through an inflated cuff and measures blood pressure as the air is gradually deflated. However, in actual applications, cuff-type blood pressure monitors still have at least the following problems: ① Continuous monitoring cannot be achieved, and each measurement must be separated by several minutes; ② The pressurization during the measurement process can cause physical discomfort; ③ A specific posture must be maintained, which restricts daily activities; ④ The measurement results are easily affected by the operation method.

[0004] 2. Ambulatory blood pressure monitor: This portable device provides 24-hour blood pressure monitoring. However, in actual use, this device still has at least the following problems: ① It still requires the use of a cuff, which affects sleep quality; ② The device is large, which affects daily activities; ③ Long-term wear causes significant discomfort; ④ The cost is high, making it difficult to widely adopt.

[0005] 4. Blood pressure estimation based on a single PPG feature: This method uses certain specific features of the PPG waveform (such as peak interval and propagation time) to estimate blood pressure. In practical applications, this method still has at least the following problems: ① The feature extraction is not comprehensive, ignoring other effective information in the signal; ② It does not consider individual differences and has limited generalization ability; ③ It is easily affected by environmental and movement interference and has poor stability; ④ The accuracy is not enough to meet medical-grade requirements;

[0006] 4. Pulse Transit Time (PTT): This method estimates blood pressure by measuring the time difference between the ECG R wave and the PPG peak. In practical applications, this method has at least the following problems: ① It requires simultaneous measurement of both ECG and PPG signals, which increases equipment complexity; ② It requires high individual calibration requirements; ③ It is difficult to ensure signal quality during exercise; and ④ The accuracy of long-term monitoring will gradually drift.

[0007] In summary, existing monitoring methods have the following common problems: lack of comprehensive analysis of multi-dimensional physiological characteristics, insufficient use of large-scale clinical data for model training, insufficient individual adaptation capabilities, limited anti-interference capabilities, and poor continuous monitoring experience. Summary of the Invention

[0008] The purpose of the present invention is to provide a non-invasive blood pressure estimation method and system based on PPG and machine learning, so as to solve the common problems of existing monitoring methods: lack of comprehensive analysis of multi-dimensional physiological characteristics, insufficient utilization of large-scale clinical data for model training, insufficient individual adaptation capability, limited anti-interference ability and poor continuous monitoring experience.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] In a first aspect, the present invention provides a non-invasive blood pressure estimation method based on PPG and machine learning, the method comprising:

[0011] Acquiring sample data of a plurality of subjects, the sample data comprising: blood pressure test data and multimodal data corresponding to the blood pressure test data, the multimodal data comprising at least: PPG raw signals, environmental data, motion data, and individual metadata;

[0012] Building a training set based on the blood pressure test data and multimodal data corresponding to the blood pressure test data;

[0013] The pre-constructed multimodal spatiotemporal model is trained based on the training set to obtain a trained multimodal spatiotemporal model, which is used as a blood pressure estimation model. The blood pressure estimation model is used to take the actual multimodal data of the subject as input and output the blood pressure estimation value of the subject.

[0014] Preferably, the method further comprises:

[0015] Acquire actual individual metadata, actual motion data, actual environmental data, and actual PPG signals of the measured subject;

[0016] Inputting the actual individual metadata, actual motion data, actual environmental data, and actual PPG signal of the subject into the blood pressure estimation model to estimate the blood pressure and obtain the estimated blood pressure value;

[0017] Determining whether the estimated blood pressure value reaches a preset blood pressure warning threshold, and if so, generating a blood pressure warning signal;

[0018] Visualize blood pressure warning signals.

[0019] Preferably, the multimodal spatiotemporal model includes:

[0020] A preprocessing module is used to preprocess the multimodal data to obtain processed multimodal data, where the processed multimodal data at least includes: processed PPG raw signals, processed environmental data, processed motion data, and processed individual metadata;

[0021] A feature extraction module is used to perform multi-scale feature extraction on the processed PPG raw signal based on a deep learning algorithm to obtain multi-scale features, and use the multi-scale features, the processed environmental data, the processed motion data, and the processed individual metadata as the original feature PPG raw signal PPG raw signal;

[0022] A graph model construction module, configured to construct a multimodal graph based on the original features, wherein the multimodal graph includes a plurality of nodes and edges between the nodes;

[0023] The weight distribution module is used to construct the node attention weight and the edge attention weight based on the attention mechanism, and perform weighted fusion based on the node attention weight and the edge attention weight to obtain the fusion feature;

[0024] The temporal evolution module is used to input the fused features into the coupled oscillator equation to capture the temporal phase synchronization between nodes in the multimodal graph to obtain the synchronization features, and then concatenate the synchronization features with the original features to obtain the concatenated features.

[0025] The prediction output module inputs the splicing features into the decision tree ensemble model, and the decision tree ensemble model outputs the prediction result.

[0026] Preferably, the multi-scale features include: time domain features, frequency domain features, high-order statistical features and cross-modal derivative features;

[0027] The time domain features include: peak interval, peak value, waveform area, first-order derivative, second-order derivative and local statistics;

[0028] The frequency domain features include: energy distribution of preset frequency bands, high-frequency transient noise and low-frequency physiological rhythm and motion interference dominant frequency bands;

[0029] The high-order statistical features include approximate entropy, sample entropy, waveform asymmetry, skewness and kurtosis;

[0030] The cross-modal derived features include: Granger causal strength.

[0031] Preferably, the types of the nodes include: physiological nodes, physical nodes and environmental nodes, and the edge relationships between the nodes include: time domain causal edges, frequency domain coupling edges and cross-modal influence edges.

[0032] Preferably, the preprocessing includes at least: filtering and denoising processing, segmentation processing and normalization processing.

[0033] Preferably, the method further comprises: enhancing the PPG original signal, comprising: enhancing the PPG original signal based on a pre-built physiological-motion coupling model and a vascular dynamics model, ambient light dynamic noise, and adversarial phase perturbation.

[0034] In a second aspect, the present invention provides a non-invasive blood pressure estimation system based on PPG and machine learning, for implementing the above-mentioned non-invasive blood pressure estimation method based on PPG and machine learning, the system comprising:

[0035] a data acquisition module, configured to acquire sample data of a plurality of subjects, the sample data including blood pressure test data and multimodal data corresponding to the blood pressure test data, the multimodal data including at least PPG raw signals, environmental data, motion data, and individual metadata;

[0036] A data construction module, configured to construct a training set based on the blood pressure test data and multimodal data corresponding to the blood pressure test data;

[0037] The model training module is used to train the pre-built multimodal spatiotemporal model based on the training set to obtain a trained multimodal spatiotemporal model, and use the trained multimodal spatiotemporal model as a blood pressure estimation model. The blood pressure estimation model is used to take the actual multimodal data of the subject as input and output the blood pressure estimation value of the subject.

[0038] In a third aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned non-invasive blood pressure estimation method based on PPG and machine learning when executing the computer program.

[0039] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned non-invasive blood pressure estimation method based on PPG and machine learning.

[0040] Beneficial effects:

[0041] 1. Improved measurement experience: Blood pressure estimation can be completed by obtaining the actual multimodal data of the measured subject. It is completely non-invasive, avoiding the discomfort caused by traditional cuff pressure, and can achieve 24-hour continuous monitoring without affecting daily activities and sleep. It is simple to operate and can complete measurements without professional guidance. It is applicable to a wide range of scenarios and can be monitored at any time in daily life.

[0042] 2. Improved measurement performance: Using multimodal data to train a multimodal spatiotemporal model can establish nonlinear relationships between multimodal data and blood pressure, such as temporal and spatial dependencies. This significantly improves accuracy and enhances measurement stability. Furthermore, by taking individual metadata into account, the system offers strong individualized adaptability. The measurement results are highly repeatable and unaffected by the operating method.

[0043] 3. High clinical application value: Supports long-term continuous monitoring, helps detect abnormal blood pressure fluctuations; can provide early warning of potential risks, facilitating timely intervention; provides blood pressure trend analysis to assist clinical decision-making; strong data traceability facilitates doctors to evaluate treatment effects;

[0044] 4. Easy to promote: low hardware cost, easy to mass-produce; high integration, can be embedded in existing wearable devices; low power consumption, suitable for long-term use. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following detailed description, they are used to explain the embodiments of the present invention, but do not constitute a limitation of the embodiments of the present invention. In the accompanying drawings:

[0046] Figure 1 is a flow chart of a non-invasive blood pressure estimation method based on PPG and machine learning provided by one embodiment of the present invention;

[0047] Figure 2 4 is a block diagram of a non-invasive blood pressure estimation system based on PPG and machine learning provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the present invention will be briefly introduced below in conjunction with the drawings and the description of the embodiments or the prior art. Obviously, the following description of the structure of the drawings is only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0049] Example 1

[0050] Figure 1 FIG. 1 is a flow chart of a non-invasive blood pressure estimation method based on PPG and machine learning provided by an embodiment of the present invention. Figure 1 As shown, this embodiment provides a non-invasive blood pressure estimation method based on PPG and machine learning, the method comprising:

[0051] Step S10: Acquire sample data of several subjects, the sample data including: blood pressure test data and multimodal data corresponding to the blood pressure test data, the multimodal data including but not limited to: PPG raw signals, environmental data, motion data, and individual metadata;

[0052] In this embodiment, the PPG raw signal can be collected using a high-precision photoelectric sensor with a sampling frequency of 50 Hz; the blood pressure test data can be standardized using a traditional cuff-type blood pressure monitor and an ambulatory blood pressure monitor.

[0053] In this embodiment, the environmental data includes but is not limited to: light intensity and temperature. A light intensity sensor may be used to collect light intensity, and a temperature sensor may be used to collect temperature.

[0054] In this embodiment, the motion data includes but is not limited to acceleration, cadence, etc., and a motion sensor (such as a three-axis accelerometer) can be used to collect the motion data.

[0055] In this embodiment, individual metadata includes but is not limited to basic data such as age, gender, height, and weight. Individual metadata can be obtained through recording.

[0056] Step S30: constructing a training set based on the blood pressure test data and the multimodal data corresponding to the blood pressure test data; in this embodiment, the multimodal data is used as the training data of the training set, and the blood pressure test data is used as the label of the training set.

[0057] Step S40: Train the pre-constructed multimodal spatiotemporal model based on the training set to obtain a trained multimodal spatiotemporal model, and use the trained multimodal spatiotemporal model as a blood pressure estimation model. The blood pressure estimation model is used to take the actual multimodal data of the subject as input and output the blood pressure estimation value of the subject.

[0058] As a further optimization of this embodiment, the multimodal spatiotemporal model includes: a preprocessing module, a feature extraction module, a graph model construction module, a weight distribution module, a time series evolution module and a prediction output module.

[0059] The preprocessing module is used to preprocess the multimodal data to obtain processed multimodal data, where the processed multimodal data at least includes: processed PPG raw signals, processed environmental data, processed motion data, and processed individual metadata;

[0060] Preprocessing of the raw PPG signal includes filtering and denoising, segmentation (usually 8-10 seconds per segment), and normalization. Preprocessing of environmental data, motion data, and individual metadata includes, but is not limited to, data cleaning and normalization.

[0061] Among them, filtering and denoising processing: using a bandpass filter to remove environmental noise, baseline drift correction and motion artifact removal from the PPG original signal.

[0062] The function expression of the bandpass filter is:

[0063]

[0064] Where H(z) is the response function of the bandpass filter, a 1k is the first denominator coefficient, a 2k is the second denominator coefficient, b 0k is the gain term of the current second-order section, b 1k is the frequency of the control zero point, b 2k is the bandwidth of the control zero, K is the number of second-order sections, k is the kth second-order section, and z is a complex variable representing the unit delay (z -1 =e -jw ), corresponding to the PPG original signal of the “last sampling point” in the discrete-time system.

[0065] The bandpass range of the bandpass filter is: Low cutoff frequency: f l =0.5Hz (removes DC offset and breathing noise); high cutoff frequency: f h =5Hz (remove myoelectric noise).

[0066] In this embodiment, a 1k and a 2k Used to determine the pole of the filter (the denominator of the response function), that is, to control the frequency and damping characteristics of the pole, affecting the "peak" position of the frequency response; 1+a 1k ·z -1 +a 2k ·z -2 Used to ensure that the filter is stable (all poles must be inside the unit circle).

[0067] Baseline drift correction: eliminate the long-term drift trend of the signal;

[0068] y(n)=x(n)-median(x(n));

[0069] Where median(x(n)) represents the median of all sample points of the PPG original signal x(n). The median is subtracted from the PPG original signal x(n) so that the corrected signal is returned to zero based on the median.

[0070] y_norm(n)=[y(n) / max(|y(n)|)]×2000;

[0071] Where y_norm(n) is the original PPG signal after baseline drift correction.

[0072] Motion artifact removal: Identify and filter out interference caused by movement;

[0073] MA_score=∑|d 2 x(n) / dt 2 |>threshold;

[0074] Wherein, MA_score is the sum of the absolute values of the second-order derivatives of the PPG original signal, threshold is the adaptive threshold, and x(n) is the signal value of the PPG original signal.

[0075] 1. Acceleration characteristics: Motion artifacts usually appear as sudden changes in the signal, which are more obvious in the second-order derivative. The second-order derivative can capture changes in signal curvature and is particularly sensitive to sudden changes.

[0076] 2. Cumulative effect: By summing the absolute values of the second-order derivatives within a time window, the instability of the signal can be cumulatively evaluated. Normal PPG raw signals have a relatively smooth change pattern, and the absolute value of the second-order derivatives is small. However, for signals containing motion artifacts, the absolute value of the second-order derivatives will increase significantly.

[0077] 3. Threshold judgment: Setting an appropriate threshold can distinguish normal physiological fluctuations from abnormal motion artifacts. When the MA_score exceeds the threshold, it indicates that the current signal segment may be contaminated by motion artifacts and needs to be marked or filtered.

[0078] In blood pressure measurement based on PPG raw signals, motion artifacts are one of the main factors affecting measurement accuracy. Using this second-order derivative-based method has the following advantages:

[0079] 1. Real-time detection of signal quality and elimination of low-quality signal segments;

[0080] 2. Improve the reliability and accuracy of blood pressure estimation;

[0081] 3. Reduce errors caused by body movement.

[0082] The advantage of this method is that it is computationally simple and can be implemented on devices with limited resources. It also has good recognition capabilities for short-term motion interference.

[0083] The feature extraction module is used to perform multi-scale feature extraction on the processed PPG raw signal based on the deep learning algorithm to obtain multi-scale features, and use the multi-scale features, processed environmental data, processed motion data and processed individual metadata as the original feature PPG raw signal.

[0084] In this embodiment, the deep learning algorithm adopts 1D-CNN (one-dimensional convolutional neural network) with different convolution kernel sizes; the multi-scale features of this embodiment include: time domain features, frequency domain features, high-order statistical features and cross-modal derivative features.

[0085] In this embodiment, the time domain features include but are not limited to: peak interval, peak value, waveform area, first-order derivative, second-order derivative, and local statistics;

[0086] Among them, local statistics include but are not limited to: mean, standard deviation;

[0087] Mean: μ = (1 / N)∑x(n);

[0088] Where n is the discrete time index (n = 0, 1, 2, …, N-1) corresponding to the nth sampling point, μ is the average amplitude of all sampling points of x(n), and x(n) is the value of the PPG raw signal at the nth sampling point (which may be in volts, ADC value, or normalized dimensionless value).

[0089] Standard deviation: σ = √[(1 / N)∑(x(n)-μ) 2 ];

[0090] Where μ is the average amplitude of all sampling points of x(n), reflecting the DC component (baseline level) of the signal.

[0091] For the PPG raw signal, μ may correspond to the static background value of the blood volume. If the DC offset of the PPG raw signal has been removed (such as after high-pass filtering), μ may be close to zero.

[0092] Peak detection: P(n) = {x(n)|x(n)>x(n±q), k∈[1,Q]};

[0093] q is a variable integer, indicating that the current sampling point x(n) needs to be compared with the qth point on its left and right sides (i.e., x(nq) and x(n+q)). This ensures that x(n) is the maximum value in the local range and prevents false detection due to noise or small fluctuations (for example, avoiding misidentifying small jitter as a peak).

[0094] q∈[1,Q], that is, the minimum value of q is 1 (adjacent points) and the maximum value is the window radius Q.

[0095] Pulse interval: IPI(n)=t(P n+1)-t(P n );

[0096] Where, P n Indicates the position (time point or sampling point index) of the nth detected pulse wave peak, P n+1 Indicates the position of the n+1th (i.e., the next adjacent) pulse wave peak, t(P n ) and t(P n+1 ) are the timestamps of the corresponding peaks (the unit may be seconds or sampling point number).

[0097] First-order derivative d1(n) (reflects the rate of change): d1(n) = [x(n+1)-x(n)] / Δt;

[0098] Second-order derivative d 2 (n) (reflects acceleration): d 2 (n) = [d1(n+1) - d1(n)] / Δt;

[0099] Where Δt is the time variation, x(n) is the signal value of the PPG original signal at the nth sampling point, x(n+1) is the signal value of the PPG original signal at the n+1th sampling point, d1(n) is the first-order derivative of the PPG original signal at the nth sampling point, d1(n+1) is the first-order derivative of the PPG original signal at the n+1th sampling point, and d 2 (n) is the second-order derivative of the PPG original signal at the n-th sampling point.

[0100] Among them, high-order statistical features include but are not limited to: approximate entropy, sample entropy, waveform asymmetry, skewness and kurtosis;

[0101] Skewness:

[0102] Where X is a random variable (signal sampling value), μ1 is the mean, μ1=E[X], σ1 is the standard deviation, σ1=E[(X-μ1) 2 ], E[·] is the expected operation (i.e., the mean for discrete signals).

[0103] Kurtosis:

[0104] Note: Excess Kurtosis is usually calculated as: K excess =A-3 (make the kurtosis of normal distribution 0)

[0105] Waveform asymmetry:

[0106] Where, t sys The systolic time is the time interval from the rise of the PPG waveform to the peak value, tdia Diastolic time refers to the time interval from the beginning of the decline of the PPG waveform peak to the beginning of the next waveform.

[0107] In this embodiment, the frequency domain features include: energy distribution of a preset frequency band, high-frequency transient noise and low-frequency physiological rhythm and motion interference dominant frequency band.

[0108] In this embodiment, the cross-modal derived features include Granger causal strength. Granger causality is a statistical causal relationship concept proposed by economist Clive Granger in 1969 and is used for time series data analysis.

[0109] The graph model construction module is used to construct a multimodal graph based on the original features, where the multimodal graph includes a number of nodes and edges between the nodes.

[0110] In this embodiment, the types of nodes include: physiological nodes, physical nodes and environmental nodes; wherein each physiological node represents a multi-scale feature, each physical node represents a set of motion data, and each environmental node represents a set of environmental data.

[0111] The edge relationships between the nodes include: time-domain causal edges, frequency-domain coupling edges, and cross-modal influence edges.

[0112] The weight distribution module is used to construct the attention weight of the node and the attention weight of the edge based on the attention mechanism, and perform weighted fusion based on the attention weight of the node and the attention weight of the edge to obtain the fusion feature.

[0113] In this embodiment, the expression of the attention weight of the node is:

[0114]

[0115] Where, α ij is the attention weight of node j to node i, that is, the importance of node j to node i, W is a trainable linear transformation matrix, h i is the original feature vector of node i, h j is the original feature vector of node j, the original feature vector is the feature corresponding to the multimodal data, for example: the frequency domain feature of the original PPG signal, σ() is the activation function, a is the trainable attention mechanism parameter vector, T is the transpose symbol, N i is the set of neighbor nodes of node i, b is the bth node in the set of neighbor nodes of node i, exp() is the natural exponential function, [Wh i :Wh j ] represents the vector Wh i With Wh j Splicing, [Whi :Wh k ] represents the vector Wh i With vector Wh k Perform splicing.

[0116] In this embodiment, the expression of the edge attention weight is:

[0117]

[0118] f r (t): A time-dependent function that captures the a priori regularity of edge weights over time. For example, in a sports scene, the weight of a Granger causal edge (Sports → PPG) may increase significantly during periods of intense exercise (t).

[0119] Loss function L for edge weight w r The gradient of , which reflects the sensitivity of the current weight to the model prediction error; Function: The larger the gradient, the more obvious the effect of adjusting the weight on reducing the loss.

[0120] Then, all the original features are weighted and summed using the node attention weight and the edge attention weight to obtain the fused feature.

[0121] The temporal evolution module is used to input the fused features into the coupled oscillator equation to capture the temporal phase synchronization between nodes in the multimodal graph, obtain the synchronization features, and then splice the synchronization features with the original features to obtain the spliced features.

[0122] In this embodiment, the timing evolution module adopts a bidirectional LSTM structure, which can simultaneously consider the forward and backward time dependencies of the PPG original signal and capture more comprehensive timing features. The timing evolution module uses bidirectional LSTM to perform timing modeling on the fusion features and extract the forward and backward time dependency features (such as signal trends, periodic fluctuations, etc.). The output features of LSTM will be used as the input of the coupled oscillator equation to dynamically adjust the phase synchronization parameters between nodes. Specifically, the hidden state h generated by LSTM is t With cell state c t It is mapped into the coupling strength coefficient and phase offset in the oscillator equation, thereby achieving adaptive control of the synchronization behavior between nodes.

[0123] The functional expression of the coupled oscillator equation is:

[0124]

[0125] Where K ij (h t ) is the LSTM hidden state h tThe generated coupling strength between node i and node j, θ j is the phase angle of node j, θ i is the phase angle of node i, w i is the natural frequency of node i, φ ij (c t ) is the LSTM cell state c t The generated phase correction term directly reflects the mathematical connection between LSTM and the coupled oscillator equation. In this embodiment, the coupled oscillator equation is responsible for modeling the dynamic evolution of physiological signals in the time dimension, capturing the phase synchronization between multiple nodes (such as PPG, movement, and environment). For example, the synchronization of the periodic fluctuations of the raw PPG signal with the movement rhythm (such as cadence) may reflect the real-time impact of exercise on the cardiovascular system.

[0126] In this embodiment, the attention mechanism is responsible for dynamically allocating weights in spatial dimensions (nodes and edges), highlighting key nodes (such as motion nodes during intense exercise) and key relationships (such as Granger causal edges).

[0127] The timing module solves the problem of “when is it relevant” (temporal dynamics), and the attention mechanism solves the problem of “where is important” (spatial weight). The combination of the two realizes “joint spatiotemporal modeling”.

[0128] In this embodiment, the temporal evolution of multimodal nodes is modeled as a coupled oscillation system through a phase synchronization mechanism to capture the periodicity of physiological signals (such as heart rate variability) and external interference (such as motion artifacts).

[0129] Example: When the user runs, the phase of the motion node changes rapidly, through the coupling term sin(θ j -θ i ) drives node phase adjustment to simulate the real-time impact of exercise on the cardiovascular system.

[0130] In this embodiment, node-level attention: according to the phase synchronization state (such as θ j and θ i ), dynamically assign node weights.

[0131] If the PPG node is highly synchronized with the motion node (θ j -θ i ≈0), which may indicate significant motion interference, and the attention weight α ij Increase.

[0132] Relation-level attention: by edge weight w r Adjust the coupling strength K to strengthen key interactions.

[0133] For example: If the weight w of the Granger causal edge (motion→PPG) rIf it is higher, the corresponding coupling term K increases and the synchronization process becomes more sensitive.

[0134] The prediction output module inputs the splicing features into a decision tree ensemble model, and the decision tree ensemble model outputs a prediction result.

[0135] In this embodiment, the method further includes: performing a credibility evaluation on the prediction result, and the credibility evaluation function is:

[0136] C=w1Q1+w2Q2+w3Q3+w4Q4+∈.

[0137] Where C is the credibility assessment score, w1, w2, w3, w4 are weights, Q1, Q2, Q3, Q4 are quality indicators, and ∈ is the noise term.

[0138] This research analyzes the nonlinear interactions between different modal features in multimodal data (such as physiological signals, motion sensor data, and environmental data) to identify the synergistic enhancement or inhibition effects (i.e., "resonance") that form under certain conditions. This effect reflects the coupling mechanism of different modal data in a dynamic system, rather than a simple linear superposition.

[0139] As a further optimization of this embodiment, the method further includes:

[0140] Step a10: Acquire actual individual metadata, actual motion data, actual environment data, and actual PPG signal of the measured object;

[0141] Step a20: Inputting the subject's actual individual metadata, actual motion data, actual environmental data, and actual PPG signal into a blood pressure estimation model to estimate blood pressure and obtain estimated blood pressure values; the estimated blood pressure values include systolic pressure and diastolic pressure;

[0142] Step a30: determining whether the estimated blood pressure value reaches a preset blood pressure warning threshold; if so, generating a blood pressure warning signal;

[0143] Step a40: Visually display the blood pressure warning signal.

[0144] As a further optimization of this embodiment, the method further includes: generating a blood pressure change trend graph based on the blood pressure estimation value, and analyzing the short-term fluctuation pattern, wherein the blood pressure change trend graph includes indicators such as a fluctuation index and a trend slope.

[0145] The expression of volatility index is:

[0146] In the formula, VI (Variability Index) reflects the relative fluctuation of blood pressure values within a specific time period. The larger the value, the more severe the blood pressure fluctuation. VI is used to quantify the short-term instability of blood pressure and assist in judging abnormal physiological states (such as hypertensive crisis, orthostatic hypotension, etc.). σ(BP t ) is the blood pressure value BP t The standard deviation in the time window t; σ(BP t ) is calculated by calculating the standard deviation of the blood pressure sampling values in a selected time period (such as the past 5 minutes) to represent the absolute fluctuation amplitude of blood pressure; μ(BP t ) is the blood pressure value BP t The mean value in the time window t, μ(BP t ) is calculated by averaging the blood pressure sampling values over the same time period as the benchmark reference value for the fluctuation amplitude.

[0147] The expression for the trend slope is:

[0148] Where k0 is the slope of the blood pressure change trend, in mmHg / s.

[0149] As a further optimization of this embodiment, the method also includes: enhancing the PPG original signal, including: enhancing the PPG original signal based on a pre-built physiological-motion coupling model and vascular dynamics model, ambient light dynamic noise, and adversarial phase perturbation.

[0150] Among them, the enhancement steps of the physiological-motor coupling model on the PPG original signal are:

[0151] 1. Extract typical motion patterns (such as walking, running, and arm swinging) from real accelerometer data;

[0152] 2. Model the interference of motion on the PPG raw signal through a linear time-varying system.

[0153] PPG noisy =PPG clean ·(1+δ·Acc norm );

[0154] Where δ is the exercise intensity coefficient (positively correlated with the accelerometer amplitude), PPG clean It is a clean PPG original signal, that is, an ideal signal that is not disturbed by movement and only reflects physiological information (such as heart rate and blood volume changes). It is usually generated by the photoelectric volume change caused by heart beat. noisy The PPG raw signal contaminated by noise is the actual collected signal, which contains motion artifacts (such as limb movement, sensor sliding, etc.) and other environmental noise. Acc normis the normalized accelerometer signal.

[0155] By modeling motion through a linear time-varying system, realistic motion interference data can be generated, avoiding the limitations of simple noise addition.

[0156] Among them, the steps of enhancing the original PPG signal by the vascular dynamics model are:

[0157] 1. Generate PPG waveforms under different blood pressures based on the Windkessel model:

[0158]

[0159] Where P(t) is the intravascular pressure, P0 is the initial pressure, Q(t) is the blood flow rate, R is the peripheral resistance, and τ is the vascular compliance parameter.

[0160] 2. Generate PPG raw signals in vasoconstriction (low temperature) or vasodilation (high temperature) states by adjusting R and τ.

[0161] Value: Enhances the model's ability to learn blood pressure-waveform correlations and reduces reliance on clinically labeled data.

[0162] Among them, the steps of ambient light dynamic noise on PPG original signal are:

[0163] 1. Collect light sensor data under different lighting conditions (such as light intensity and strobe frequency);

[0164] 2. Injecting optical noise into the PPG raw signal through pulse modulation:

[0165]

[0166] Where m is the mth frequency component, A m is the amplitude of the mth frequency component (associated with the light intensity sensor reading), M is the total number of frequency components, f m is the value of the mth frequency component (such as 50Hz LED flicker, 0.1-5Hz sunlight fluctuation), and β is the coefficient.

[0167] Value: Improves the model's anti-interference ability in complex lighting environments.

[0168] Among them, the steps of enhancing the original PPG signal by adversarial phase perturbation are:

[0169] 1. Perform Hilbert transform on the PPG raw signal to extract the instantaneous phase φ(t).

[0170] 2. Inject random phase offset to get the processed phase:

[0171]

[0172] Where ∈ is the perturbation intensity (usually 0.1-0.3 radians), is a probability distribution, indicating that the random variable obeys a normal distribution with a mean of 0 and a variance of 1. Its probability density function is:

[0173]

[0174] 3. Reconstruct the time domain signal through inverse transform.

[0175] Value: Increases the model’s robustness to poor sensor contact or signal asynchrony.

[0176] As a further optimization of this embodiment, lightweight adversarial training is adopted during the training process. The steps of lightweight adversarial training are as follows:

[0177] 1. Calculate the gradient of the input PPG original signal to the loss function

[0178] 2. Generate adversarial examples x adv :

[0179]

[0180] γ: Perturbation step size (must be limited to within 5% of the signal amplitude).

[0181] Optimization constraints: Projected gradient descent (PGD) is used to ensure that the perturbation conforms to the physical rationality of the original PPG signal (such as not destroying the periodicity of the waveform).

[0182] Value: Defends against unknown noise patterns in real scenes and avoids the computational overhead of complex GAN training.

[0183] The present invention has the following advantages:

[0184] 1. Improved measurement experience: Blood pressure estimation can be completed by obtaining the actual multimodal data of the measured subject. It is completely non-invasive, avoiding the discomfort caused by traditional cuff pressure, and can achieve 24-hour continuous monitoring without affecting daily activities and sleep. It is simple to operate and can complete measurements without professional guidance. It is applicable to a wide range of scenarios and can be monitored at any time in daily life.

[0185] 2. Improved measurement performance: Using multimodal data to train a multimodal spatiotemporal model can establish nonlinear relationships between multimodal data and blood pressure, such as temporal and spatial dependencies. This significantly improves accuracy and enhances measurement stability. Furthermore, by taking individual metadata into account, the system offers strong individualized adaptability. The measurement results are highly repeatable and unaffected by the operating method.

[0186] 3. High clinical application value: Supports long-term continuous monitoring, helps detect abnormal blood pressure fluctuations; can provide early warning of potential risks, facilitating timely intervention; provides blood pressure trend analysis to assist clinical decision-making; strong data traceability facilitates doctors to evaluate treatment effects;

[0187] 4. Easy to promote: low hardware cost, easy to mass-produce; high integration, can be embedded in existing wearable devices; low power consumption, suitable for long-term use.

[0188] Example 2

[0189] Figure 2 FIG is a block diagram of a non-invasive blood pressure estimation system based on PPG and machine learning provided by an embodiment of the present invention. Figure 2 As shown, this embodiment provides a non-invasive blood pressure estimation system based on PPG and machine learning, which is used to implement the non-invasive blood pressure estimation method based on PPG and machine learning in Example 1. The system includes:

[0190] a data acquisition module, configured to acquire sample data of a plurality of subjects, the sample data including blood pressure test data and multimodal data corresponding to the blood pressure test data, the multimodal data including at least PPG raw signals, environmental data, motion data, and individual metadata;

[0191] A data construction module, configured to construct a training set based on the blood pressure test data and multimodal data corresponding to the blood pressure test data;

[0192] The model training module is used to train the pre-built multimodal spatiotemporal model based on the training set to obtain a trained multimodal spatiotemporal model, and use the trained multimodal spatiotemporal model as a blood pressure estimation model. The blood pressure estimation model is used to take the actual multimodal data of the subject as input and output the blood pressure estimation value of the subject.

[0193] This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the non-invasive blood pressure estimation method based on PPG and machine learning in the first embodiment is implemented.

[0194] This embodiment further provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the non-invasive blood pressure estimation method based on PPG and machine learning in the first embodiment is implemented.

[0195] The present invention has the following advantages:

[0196] 1. Improved measurement experience: Blood pressure estimation can be completed by obtaining the actual multimodal data of the measured subject. It is completely non-invasive, avoiding the discomfort caused by traditional cuff pressure, and can achieve 24-hour continuous monitoring without affecting daily activities and sleep. It is simple to operate and can complete measurements without professional guidance. It is applicable to a wide range of scenarios and can be monitored at any time in daily life.

[0197] 2. Improved measurement performance: Using multimodal data to train a multimodal spatiotemporal model can establish nonlinear relationships between multimodal data and blood pressure, such as temporal and spatial dependencies. This significantly improves accuracy and enhances measurement stability. Furthermore, by taking individual metadata into account, the system offers strong individualized adaptability. The measurement results are highly repeatable and unaffected by the operating method.

[0198] 3. High clinical application value: Supports long-term continuous monitoring, helps detect abnormal blood pressure fluctuations; can provide early warning of potential risks, facilitating timely intervention; provides blood pressure trend analysis to assist clinical decision-making; strong data traceability facilitates doctors to evaluate treatment effects;

[0199] 4. Easy to promote: low hardware cost, easy to mass-produce; high integration, can be embedded in existing wearable devices; low power consumption, suitable for long-term use.

[0200] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0201] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A system that specifies the functions of a box or boxes.

[0202] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A non-invasive blood pressure estimation method based on PPG and machine learning, characterized in that: The method comprises: Acquiring sample data of a plurality of subjects, the sample data comprising: blood pressure test data and multimodal data corresponding to the blood pressure test data, the multimodal data comprising at least: PPG raw signals, environmental data, motion data, and individual metadata; Building a training set based on the blood pressure test data and multimodal data corresponding to the blood pressure test data; The pre-constructed multimodal spatiotemporal model is trained based on the training set to obtain a trained multimodal spatiotemporal model, which is used as a blood pressure estimation model. The blood pressure estimation model is used to take the actual multimodal data of the subject as input and output the blood pressure estimation value of the subject.

2. The non-invasive blood pressure estimation method based on PPG and machine learning according to claim 1, characterized in that: The method further comprises: Acquire actual individual metadata, actual motion data, actual environmental data, and actual PPG signals of the measured subject; Inputting the actual individual metadata, actual motion data, actual environmental data, and actual PPG signal of the subject into the blood pressure estimation model to estimate the blood pressure and obtain the estimated blood pressure value; Determining whether the estimated blood pressure value reaches a preset blood pressure warning threshold, and if so, generating a blood pressure warning signal; Visualize blood pressure warning signals.

3. The non-invasive blood pressure estimation method based on PPG and machine learning according to claim 1 or 2, characterized in that: The multimodal spatiotemporal model includes: A preprocessing module is used to preprocess the multimodal data to obtain processed multimodal data, where the processed multimodal data at least includes: processed PPG raw signals, processed environmental data, processed motion data, and processed individual metadata; A feature extraction module is used to extract multi-scale features from the processed PPG raw signal based on a deep learning algorithm to obtain multi-scale features, and use the multi-scale features, processed environmental data, processed motion data, and processed individual metadata as raw features; A graph model construction module, configured to construct a multimodal graph based on the original features, wherein the multimodal graph includes a plurality of nodes and edges between the nodes; The weight distribution module is used to construct the node attention weight and the edge attention weight based on the attention mechanism, and perform weighted fusion based on the node attention weight and the edge attention weight to obtain the fusion feature; The temporal evolution module is used to input the fused features into the coupled oscillator equation to capture the temporal phase synchronization between nodes in the multimodal graph to obtain the synchronization features, and then concatenate the synchronization features with the original features to obtain the concatenated features. The prediction output module inputs the splicing features into the decision tree ensemble model, and the decision tree ensemble model outputs the prediction result.

4. The non-invasive blood pressure estimation method based on PPG and machine learning according to claim 3, characterized in that: The multi-scale features include: time domain features, frequency domain features, high-order statistical features and cross-modal derivative features; The time domain features include: peak interval, peak value, waveform area, first-order derivative, second-order derivative and local statistics; The frequency domain features include: energy distribution of preset frequency bands, high-frequency transient noise and low-frequency physiological rhythm and motion interference dominant frequency bands; The high-order statistical features include approximate entropy, sample entropy, waveform asymmetry, skewness and kurtosis; The cross-modal derived features include: Granger causal strength.

5. The non-invasive blood pressure estimation method based on PPG and machine learning according to claim 4, characterized in that: The types of the nodes include: physiological nodes, physical nodes and environmental nodes, and the edge relationships between the nodes include: time domain causal edges, frequency domain coupling edges and cross-modal influence edges.

6. The non-invasive blood pressure estimation method based on PPG and machine learning according to claim 3, characterized in that: The preprocessing includes at least: filtering and denoising processing, segmentation processing and normalization processing.

7. The non-invasive blood pressure estimation method based on PPG and machine learning according to claim 1, characterized in that: The method further includes: enhancing the PPG original signal, including: enhancing the PPG original signal based on a pre-built physiological-motion coupling model, a vascular dynamics model, ambient light dynamic noise, and adversarial phase perturbation.

8. A non-invasive blood pressure estimation system based on PPG and machine learning, for implementing the non-invasive blood pressure estimation method based on PPG and machine learning according to any one of claims 1 to 7, characterized in that: The system comprises: a data acquisition module, configured to acquire sample data of a plurality of subjects, the sample data including blood pressure test data and multimodal data corresponding to the blood pressure test data, the multimodal data including at least PPG raw signals, environmental data, motion data, and individual metadata; A data construction module, configured to construct a training set based on the blood pressure test data and multimodal data corresponding to the blood pressure test data; The model training module is used to train the pre-built multimodal spatiotemporal model based on the training set to obtain a trained multimodal spatiotemporal model, and use the trained multimodal spatiotemporal model as a blood pressure estimation model. The blood pressure estimation model is used to take the actual multimodal data of the subject as input and output the blood pressure estimation value of the subject.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the non-invasive blood pressure estimation method based on PPG and machine learning according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the non-invasive blood pressure estimation method based on PPG and machine learning according to any one of claims 1 to 7 is implemented.

Images