Cuff-free blood pressure continuous monitoring method based on ICG, PPG and ECG multi-mode physiological signals

Abstract

The invention relates to the technical field of blood pressure measurement, in particular to a cuff-free blood pressure continuous monitoring method based on ICG, PPG and ECG multi-mode physiological signals. According to the method, deep mining and fusion are carried out based on multi-modal physiological signals such as a chest impedance signal ICG, a photoelectric volume pulse wave signal PPG and an electrocardiosignal ECG of a tested object synchronously collected by multi-modal hemodynamics monitoring equipment, and a richer multi-modal feature matrix related to blood pressure changes is constructed; after deep learning fusion training is carried out on the constructed cuff-free blood pressure prediction model, prediction output of a cuff-free blood pressure prediction result can be directly carried out according to ICG, PPG and ECG multi-mode physiological signals of a to-be-detected tested object by means of the cuff-free blood pressure prediction model, the generalization ability and monitoring precision of continuous monitoring of the cuff-free blood pressure can be effectively improved, and the accuracy of continuous monitoring of the cuff-free blood pressure can be improved. The continuous, non-invasive and high-precision blood pressure monitoring requirements are met, and the clinical application value is more remarkable.

Description

Technical Field

[0001] This invention relates to the field of blood pressure measurement technology, and in particular to a cuffless continuous blood pressure monitoring method based on multimodal physiological signals of ICG, PPG and ECG. Background Technology

[0002] Cardiovascular disease has become the leading cause of death worldwide, making continuous blood pressure monitoring crucial for disease early warning and treatment. Arterial blood pressure (BP) is the lateral pressure exerted by blood against the arterial wall per unit area, and its formation depends on the dynamic balance between cardiac pumping and peripheral resistance. The specific mechanisms influencing BP can be categorized into the following four aspects:

[0003] (a) Cardiac output

[0004] Cardiac output (CO) is determined by both heart rate (HR) and stroke volume (SV). The rapid ejection phase during systole (accounting for 60%–70% of SV) causes a sudden increase in aortic pressure, directly determining systolic blood pressure (SBP); the slow perfusion of the remaining diastolic blood flow to peripheral tissues maintains diastolic blood pressure (DBP). Prolonged elevated CO (such as in hyperthyroidism) can induce persistent hypertension, while insufficient CO (such as in acute myocardial infarction) can lead to hypotension or even shock.

[0005] (II) Peripheral Resistance

[0006] Peripheral resistance (PR) is primarily determined by the degree of arteriolar / microarterial constriction and blood viscosity. When PR is elevated (e.g., due to vasospasm or atherosclerosis), diastolic blood flow obstruction significantly increases diastolic blood flow (DBP), while increased systolic ejection resistance also slightly increases systolic blood flow (SBP). The core pathophysiological mechanism of essential hypertension is related to persistently elevated PR, and clinical antihypertensive strategies often target and reduce PR using drugs such as calcium channel blockers.

[0007] (III) Arterial compliance

[0008] Elastic arteries (such as the aorta) store blood potential energy through systolic expansion and maintain blood flow continuity through diastolic elastic retraction. When the arterial compliance (AC) of young individuals is good, SBP can be effectively buffered (<120mmHg), and DBP is stable at about 80mmHg. When arterial stiffness (such as aging or diabetes) causes compliance to decrease, systolic pressure buffering is lost (SBP>160mmHg), and diastolic elastic retraction is insufficient (DBP<70mmHg), resulting in simple systolic hypertension with increased pulse pressure, which is common in the elderly population.

[0009] (iv) Circulating blood volume-vessel volume matching

[0010] Effective circulating blood volume needs to be dynamically balanced with total vessel volume. Acute blood loss (such as trauma) or abnormal dilation of blood vessels (such as septic shock) can cause relative lack of circulating blood volume, leading to hypotension; while volume overload (such as salt-sensitive hypertension) or pathological contraction of blood vessels (such as activation of the renin-angiotensin system) can cause blood pressure to rise.

[0011] Traditional cuff blood pressure monitors rely on intermittent inflation and pressure, which has the defects of poor comfort, inability to capture dynamic blood pressure fluctuations, etc. Although there are continuous monitoring techniques based on photoplethysmography (PPG), single-mode signals are easily affected by individual vascular characteristics and motion artifacts, resulting in insufficient cross-user generalization ability, challenges in calibration of ECG-PPG dual-mode, and the need for frequent calibration of models based on pulse wave transit time (PTT). Due to the limitations of the application scenarios or acquisition devices used by the methods, the changes in thoracic impedance (ICG) are not combined, and the effects of dynamic changes in vascular tension are ignored; or although multiple physiological signals (ICG, ECG, PPG) are obtained, they are not fully coupled and analyzed, resulting in redundant feature dimensions and weak physiological relevance, poor individual applicability, and no help for improving the accuracy of continuous blood pressure monitoring, making it difficult to meet the clinical application requirements of continuous, non-invasive, and high-precision blood pressure monitoring. SUMMARY

[0012] To overcome the shortcomings of the prior art, the present application provides a sleeveless blood pressure continuous monitoring method based on ICG, PPG and ECG multi-modal physiological signals, which trains the established sleeveless blood pressure prediction model by respectively constructing a multi-modal feature matrix pair corresponding to each group of multi-modal physiological signals, and then uses it to implement sleeveless continuous blood pressure monitoring, thereby improving the accuracy of sleeveless continuous blood pressure monitoring, and solving the technical problems that existing methods cannot truly realize the conversion application of sleeveless continuous blood pressure monitoring, and the blood pressure monitoring accuracy of existing methods is unstable, the individual applicability is poor, and it is difficult to meet the clinical application requirements.

[0013] To solve the above technical problems, the technical scheme adopted by the present application is as follows:

[0014] A cuffless blood pressure continuous monitoring method based on ICG, PPG and ECG multi-modal physiological signals, comprising the following steps:

[0015] S1: acquiring multi-modal physiological signals of a test subject as a sample, and reference blood pressure values corresponding to each group of multi-modal physiological signals; the multi-modal physiological signals include chest impedance signals ICG, photoplethysmography signals PPG and electrocardiogram signals ECG synchronously collected for the test subject;

[0016] S2: after time sequence cycle alignment processing according to the synchronous collection time sequence information of the multi-modal physiological signals, pre-processing and signal quality evaluation and screening are performed on the multi-modal physiological signals, and then feature point positioning processing of the corresponding time sequence cycle is performed on the screened multi-modal physiological signals;

[0017] S3: according to the related change relationship between each group of multi-modal physiological signals and their respective corresponding reference blood pressure values, related features reflecting pulse conduction and related features reflecting hemodynamic state in the multi-modal physiological signals of the corresponding time sequence cycle are mined, and a multi-modal feature matrix corresponding to each group of multi-modal physiological signals is constructed;

[0018] S4: a cuffless blood pressure prediction model adopting a multi-head attention mechanism is constructed, the multi-modal feature matrix corresponding to the multi-modal physiological signals is used as the training input data of the cuffless blood pressure prediction model, the reference blood pressure values corresponding to the multi-modal physiological signals are used as labels, and the cuffless blood pressure prediction model is subjected to deep learning training, so as to obtain a trained cuffless blood pressure prediction model;

[0019] S5: acquiring multi-modal physiological signals of a test subject to be measured, processing the multi-modal physiological signals of the test subject to be measured according to steps S2 and S3 to obtain a corresponding multi-modal feature matrix, and then inputting the multi-modal feature matrix into the trained cuffless blood pressure prediction model to obtain a cuffless blood pressure prediction result predicted by the cuffless blood pressure prediction model.

[0020] As a preferred solution, in step S2, the pre-processing of the multi-modal physiological signals specifically includes:

[0021] S2101: the multi-modal physiological signals are subjected to denoising processing respectively by using a wavelet decomposition method to improve the signal-to-noise ratio; after the denoising processing, the electrocardiogram signal ECG retains a frequency band of 0.114-12.9 Hz, the chest impedance signal ICG retains a frequency band of 0.381-12.9 Hz, and the photoplethysmography signal PPG retains a frequency band of 0.381-25.8 Hz;

[0022] S2102: the multi-modal physiological signals after the denoising processing are subjected to standardization processing to unify the dimension.

[0023] As a preferred solution, the signal quality evaluation and screening of the multi-modal physiological signals in step S2 specifically includes:

[0024] S2201: data cleaning is performed on the pre-processed multi-modal physiological signals, and the following are screened out: flawed signals with signal peak values or valley values exceeding a preset threshold range, invalid signals with mutual conflicts or mismatched time sequence periods between the multi-modal physiological signals of the same subject as a sample, and different repeated signals with completely identical data, and the remaining multi-modal physiological signals are determined as usable signals;

[0025] S2202: the multi-modal physiological signals determined as usable are further screened, and the multi-modal physiological signals with the thoracic impedance signal ICG, the photoplethysmogram signal PPG and the electrocardiogram signal ECG of the same subject as a sample all determined as usable and time sequence periods aligned are reserved, otherwise, they are screened out; thus, the multi-modal physiological signals reserved after the final screening are obtained.

[0026] As a preferred solution, the feature point positioning processing of the corresponding time sequence periods on the screened multi-modal physiological signals in step S2 specifically includes:

[0027] S2301: for the thoracic impedance signal ICG, the photoplethysmogram signal PPG and the electrocardiogram signal ECG of the same subject as a sample, the feature points of each cardiac cycle in the electrocardiogram signal ECG are positioned first, including the P-wave peak point, the Q point, the R-wave peak point, the S point, the T-wave peak point and the T-wave end point in each cardiac cycle of the electrocardiogram signal ECG;

[0028] S2302: the feature points of the thoracic impedance signal ICG are positioned in the signal period waveform of the thoracic impedance signal ICG corresponding to each cardiac cycle of the electrocardiogram signal ECG, including the aortic valve opening point B, the wave peak point C and the aortic valve closing point X in each signal period waveform of the thoracic impedance signal ICG;

[0029] S2303: the feature points of the photoplethysmogram signal PPG are positioned in the signal period waveform of the photoplethysmogram signal PPG corresponding to each cardiac cycle of the electrocardiogram signal ECG, including the pulse wave starting point b, the systolic peak point c, the re-pulse wave starting point f, the re-pulse wave peak point g, the maximum point a of the rising branch slope and the minimum point e of the falling branch slope in each signal period waveform of the photoplethysmogram signal PPG;

[0030] S2304: the feature point positioning processing of the corresponding time sequence periods is performed according to steps S2301-S2303 on the multi-modal physiological signals of each different subject as a sample after the screening.

[0031] As a preferred solution, in step S2301, the specific way of locating the feature points of the electrocardiosignal ECG in each cardiac cycle is as follows:

[0032] In a single cardiac cycle of the electrocardiosignal ECG, the R-wave peak point is located by the sliding window peak detection from the starting position of the cardiac cycle, and the Q point is located by searching back from the R-wave peak point to the first local minimum point, and the S point is located by searching forward from the R-wave peak point to the first local maximum point; then, the P-wave peak point and the T-wave peak point are located by the first derivative zero-crossing detection of the forward waveform of the S point, and finally the T-wave end point is located by identifying the inflection point at the end of the T wave.

[0033] In this way, the feature points in each cardiac cycle of the electrocardiosignal ECG are located respectively.

[0034] As a preferred solution, in step S2302, the specific way of locating the feature points of the thoracic impedance signal ICG in the signal cycle waveform of the thoracic impedance signal ICG corresponding to each cardiac cycle of the electrocardiosignal ECG is as follows:

[0035] According to the aligned time sequence cycle, the signal cycle of the thoracic impedance signal ICG corresponding to each cardiac cycle of the electrocardiosignal ECG is determined.

[0036] Then, in a single signal cycle of the thoracic impedance signal ICG, the main wave peak point C is located by the sliding window peak detection from the starting position of the signal cycle; then, the aortic valve opening point B is located by searching back from the main wave peak point C to the first local maximum point, and the aortic valve closing point X is located by searching forward from the main wave peak point C to the lowest valley point in the signal cycle.

[0037] In this way, the feature points of each signal cycle of the thoracic impedance signal ICG are located respectively.

[0038] As a preferred solution, in step S2303, the specific way of locating the feature points of the photoplethysmogram signal PPG in the signal cycle waveform of the photoplethysmogram signal PPG corresponding to each cardiac cycle of the electrocardiosignal ECG is as follows:

[0039] According to the aligned time sequence cycle, the signal cycle of the photoplethysmogram signal PPG corresponding to each cardiac cycle of the electrocardiosignal ECG is determined.

[0040] Then, in a single signal cycle of the photoelectric plethysmogram signal PPG, the signal cycle starting position is taken as the starting point, the minimum point in the signal cycle is located by sliding window valley detection to locate the pulse wave starting point b, and the maximum point in the signal cycle is located by sliding window peak detection to locate the systolic peak point c; then, by calculating the first derivative extreme value, the maximum slope corresponding position is located to the rising branch slope maximum point a between the pulse wave starting point b and the systolic peak point c, and the minimum slope corresponding position is located to the descending branch slope minimum point e after the systolic peak point c; then, the first derivative zero-crossing point calculation detection is performed again in the waveform region after the descending branch slope minimum point e, it is judged whether there is a local area maximum point of the first derivative zero-crossing, if there is, the local area maximum point is determined as the located double pulse wave peak point g, if there is no local area maximum point of the first derivative zero-crossing, the second derivative zero-crossing local maximum point is further calculated and found, and is located as the double pulse wave peak point g; finally, between the descending branch slope minimum point e and the double pulse wave peak point g, the local area minimum point is located as the double pulse wave starting point f.

[0041] In this way, the feature points of each signal cycle of the photoelectric plethysmogram signal PPG are located respectively.

[0042] As a preferred scheme, in step S3, the related features reflecting pulse conduction in the multi-modal physiological signal corresponding to the time sequence cycle include one or more of the following:

[0043] a pulse wave propagation time PTTa from the R-wave peak to a maximum point a of a rising slope of the photoplethysmographic signal PPG, a pulse wave propagation time PTTb from the R-wave peak to a starting point b of a pulse wave of the photoplethysmographic signal PPG, a pulse wave propagation time PTTc from the R-wave peak to a systolic peak point c of the photoplethysmographic signal PPG, a time span from the starting point b of the pulse wave to the systolic peak point c, a time span from the starting point b of the pulse wave to a starting point f of a dicrotic wave, a time span from the maximum point a of the rising slope to a minimum point e of a falling slope of the photoplethysmographic signal PPG, a time span from the systolic peak point c of the photoplethysmographic signal PPG to a starting point b of a pulse wave of a next signal cycle, a time span from the starting point f of the dicrotic wave of the photoplethysmographic signal PPG to an end point of a current signal cycle, a time span from a start point of the photoplethysmographic signal PPG to the end point of the current signal cycle, a time span of 1 / 6 of a maximum amplitude in a waveform of the photoplethysmographic signal PPG, a time span of 2 / 6 of the maximum amplitude in the waveform of the photoplethysmographic signal PPG, a time span of 3 / 6 of the maximum amplitude in the waveform of the photoplethysmographic signal PPG, a time span of 4 / 6 of the maximum amplitude in the waveform of the photoplethysmographic signal PPG, a time span of 5 / 6 of the maximum amplitude in the waveform of the photoplethysmographic signal PPG, a rising slope of the photoplethysmographic signal PPG, a falling slope of the photoplethysmographic signal PPG, an area under a curve of the photoplethysmographic signal PPG between the starting point b of the pulse wave and the end point of the current signal cycle in a same cardiac cycle, an area under the curve of the photoplethysmographic signal PPG between the maximum point a of the rising slope and the end point of the current signal cycle in the same cardiac cycle, an area under the curve of the photoplethysmographic signal PPG between the systolic peak point c and the end point of the current signal cycle in the same cardiac cycle, a ratio of an amplitude of the maximum point a of the rising slope to an amplitude of the systolic peak point c in the same cardiac cycle, a ratio of an amplitude of the minimum point e of the falling slope to the amplitude of the systolic peak point c in the same cardiac cycle, a ratio of an amplitude of the starting point f of the dicrotic wave to the amplitude of the systolic peak point c in the same cardiac cycle;

[0044] The relevant features in the multi-modal physiological signals corresponding to the timing cycle that reflect the hemodynamic state include one or more of the following:

[0045] a heart rate HR, a pre-ejection period PET, a left ventricular ejection time LVET, a thoracic fluid content TFC, a systolic time ratio STR, a stroke volume SV, a cardiac output CO, a body surface area BSA, a stroke volume index SVI, a cardiac index CI, an arterial compliance AC, an acceleration index of myocardial contraction ACI, a velocity index VI, a time span Qdz of a Q point of an electrocardiogram ECG to a C point of a peak of an impedance cardiogram ICG;

[0046] The set of the related features reflecting the pulse conduction and the related features reflecting the hemodynamic state corresponding to each group of multi-modal physiological signals constitutes a multi-modal feature matrix corresponding to the group of multi-modal physiological signals.

[0047] As a preferred solution, when acquiring the multi-modal physiological signals of the subject, the basic vital sign features of the subject are also acquired; the basic vital sign features include the height, weight, age and gender features of the subject; when constructing the multi-modal feature matrix corresponding to the multi-modal physiological signals, the basic vital sign features of the same subject are also added to the multi-modal feature matrix as set elements.

[0048] As a preferred solution, in step S4, the cuffless blood pressure prediction model comprises a first normalization layer, a multi-head self-attention network layer, a second normalization layer and a multi-layer perception network layer which are sequentially cascaded.

[0049] The first normalization layer is used to linearly transform the input multi-modal feature matrix into corresponding query vectors, key vectors and value vectors:

[0050] Q i =x i W Q , K i =x i W K , V i =x i W V ;

[0051] wherein x i ∈R M×N represents any i-th multi-modal feature matrix, M represents the data dimension of the single related feature contained in the multi-modal feature matrix, and N represents the total number of related features contained in the multi-modal feature matrix; Q i , K i , V i are respectively the query vector, the key vector and the value vector corresponding to the i-th multi-modal feature matrix x i ; are respectively the linear conversion matrices of the query, the key and the value in the multi-head self-attention mechanism; d k is the dimension of the attention head in the multi-head self-attention network layer;

[0052] The multi-head self-attention network layer is used to extract the self-attention feature vector of the multi-modal feature matrix:

[0053]

[0054] wherein H i represents the i-th multi-modal feature matrix x ia corresponding self-attention feature vector; denotes the output of the lth attention head in the multi-head self-attention network layer, l = 1, 2, …, L, L represents the number of attention heads in the multi-head self-attention network layer; Concat(·) is a concatenation function; W O is an output weight matrix of the multi-head self-attention network layer; Attention(·) represents an attention mechanism operation; Softmax(·) represents a normalized exponential function operation;

[0055] The second layer normalization layer takes the addition of the self-attention feature vector and the multi-modal feature matrix as input, and performs linear transformation processing on the feature data dimension to obtain an attention superposition feature vector corresponding to the multi-modal feature matrix:

[0056] y i = LN(H i +x i );

[0057] wherein y i denotes the attention superposition feature vector corresponding to the i th multi-modal feature matrix x i ; LN(·) represents linear transformation of the second layer normalization layer;

[0058] The multi-layer perceptron network layer is used to fuse the attention superposition feature vectors of the multi-modal feature matrices to obtain a prediction output of a cuffless blood pressure prediction result, and the prediction result includes predicted values of systolic blood pressure SBP, diastolic blood pressure DBP and mean arterial pressure MAP; the multi-layer perceptron network layer includes a plurality of hidden layers and an output layer, and the output layer includes 3 output neurons corresponding to the predicted values of the systolic blood pressure SBP, the diastolic blood pressure DBP and the mean arterial pressure MAP, and the processing process is represented as:

[0059]

[0060] wherein, respectively represent the outputs of the 1st hidden layer, the (j-1)th hidden layer, the jth hidden layer and the Jth hidden layer in the multi-layer perceptron network layer when the attention superposition feature vectors y i are input, j = 2, 3, …, J, J represents the number of hidden layers in the multi-layer perceptron network layer; W (1) , W (j) respectively represent the hidden layer weights of the 1st hidden layer and the jth hidden layer in the multi-layer perceptron network layer; b (1) , b (j) respectively represent the hidden layer biases of the 1st hidden layer and the jth hidden layer in the multi-layer perceptron network layer; σ(·) represents a ReLU activation function operation; O i,m denotes the output of the output layer for the attention superposition feature vector yi Output of the mth output neuron of the output layer, mx1, 2, 3, O i,1 , O i,2 , O i,3 respectively correspond to the predicted values of systolic blood pressure SBP, diastolic blood pressure DBP and mean arterial pressure MAP; respectively represent the output layer weight and the output layer bias of the mth output neuron in the output layer.

[0061] Compared with the prior art, the present application has the following beneficial effects:

[0062] 1. In the present application, the multi-modal physiological signals such as thoracic impedance signal ICG, photoplethysmogram PPG, electrocardiogram ECG, etc. synchronously collected by the multi-modal hemodynamic monitoring device are used to participate in the continuous monitoring of the cuffless blood pressure. The synchronous collection mechanism ensures the strict alignment of multi-dimensional physiological data in the time sequence cycle dimension, providing a basis for the time sequence cycle consistency of subsequent feature fusion. In the preprocessing and signal quality assessment and screening stage, targeted noise reduction algorithms and screening mechanisms are used to significantly improve the signal-to-noise ratio of the original signal and effectively eliminate baseline drift, signal error conflicts, repeated signals and other disturbance factors that may exist in the multi-modal physiological signal, providing a basis for constructing high-quality data for feature extraction.

[0063] 2. In the present application, further multi-dimensional feature extraction is performed on the multi-modal signals after preprocessing and signal quality assessment and screening, covering hemodynamic parameters such as pre-ejection time and left ventricular ejection time of ICG, vascular elasticity indicators such as pulse wave transit time and amplitude variation coefficient of PPG, and electrophysiological characteristics such as R-wave interval and QRS wave group morphology of ECG. A multi-modal feature matrix is constructed to structureally fuse physiological features of different modalities and different dimensions, forming a composite feature space that can fully reflect the state of the cardiovascular system, effectively making up for the limitations of single features in representing blood pressure state.

[0064] 3. In the present application, a cuffless blood pressure prediction model based on multi-modal feature matrix construction is adopted, with reference blood pressure value as the label, to perform deep learning training on the cuffless blood pressure prediction model, realizing accurate mapping from multi-modal physiological features to blood pressure values. In the prediction stage, the ICG, PPG and ECG multi-modal physiological signals of the subject to be tested can be directly used to predict the cuffless blood pressure value in real time and continuously. This end-to-end prediction mechanism can effectively improve the generalization ability and monitoring accuracy of cuffless blood pressure continuous monitoring, meet the needs of continuous, non-invasive and high-precision blood pressure monitoring, break through the bottleneck of insufficient accuracy and poor individual applicability of traditional cuffless blood pressure monitoring devices, and has more significant clinical application value. BRIEF DESCRIPTION OF DRAWINGS

[0065] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below with reference to the drawings, in which:

[0066] Figure 1 The flow chart of the method for continuous monitoring of blood pressure without a cuff based on ICG, PPG and ECG multi-modal physiological signals of the application.

[0067] Figure 2 The work flow diagram of multi-modal physiological signal acquisition, multi-modal feature extraction and training of the blood pressure prediction model without a cuff in the method of the application.

[0068] Figure 3 The flow chart of the specific implementation steps of multi-modal physiological signal acquisition, multi-modal feature extraction and training of the blood pressure prediction model without a cuff in the method of the application.

[0069] Figure 4 The display diagram of ECG signals, ICG signals, PPG signals and their feature points. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below with reference to the drawings, in which:

[0071] In view of the problems of poor individual difference adaptability and dependence on frequent calibration existing in the prior single-mode or dual-mode methods, the application provides a method for continuous monitoring of blood pressure without a cuff based on ICG, PPG and ECG multi-modal physiological signals. The method deeply mines and fuses multi-modal physiological signals such as chest impedance signals ICG, photoplethysmogram signals PPG and electrocardiogram signals ECG of a subject synchronously acquired by a multi-modal hemodynamic monitoring device, constructs a multi-modal feature matrix containing more abundant multi-modal features related to blood pressure changes, and after deep learning fusion training of a constructed blood pressure prediction model without a cuff, the blood pressure prediction model without a cuff can be used to directly predict the output of the blood pressure prediction result according to the ICG, PPG and ECG multi-modal physiological signals of a subject to be tested, thereby effectively improving the precision and applicability of the blood pressure prediction model without a cuff.

[0072] Based on the above technical idea, the present application provides a non-tubing blood pressure continuous monitoring method based on ICG, PPG and ECG multi-modal physiological signals, the flowchart of which is shown as follows Figure 1 The method comprises the following steps:

[0073] S1: Obtain multi-modal physiological signals of a test subject as a sample, and reference blood pressure values corresponding to each group of multi-modal physiological signals; the multi-modal physiological signals include chest impedance signals ICG, photoplethysmography signals PPG and electrocardiogram signals ECG synchronously collected for the test subject;

[0074] S2: After time sequence cycle alignment processing according to the synchronous collection time sequence information of the multi-modal physiological signals, pre-process and signal quality evaluation and screening are performed on the multi-modal physiological signals, and then feature point positioning processing of the corresponding time sequence cycle is performed on the screened multi-modal physiological signals;

[0075] S3: According to the correlation change relationship between each group of multi-modal physiological signals and their respective corresponding reference blood pressure values, relevant features reflecting pulse conduction and relevant features reflecting hemodynamic state in the multi-modal physiological signals of the corresponding time sequence cycle are mined, and a multi-modal feature matrix corresponding to each group of multi-modal physiological signals is constructed;

[0076] S4: A non-tubing blood pressure prediction model using a multi-head attention mechanism is constructed, and the multi-modal feature matrix corresponding to the multi-modal physiological signals is used as the training input data of the non-tubing blood pressure prediction model, the reference blood pressure values corresponding to the multi-modal physiological signals are used as labels, and deep learning training is performed on the non-tubing blood pressure prediction model to obtain a trained non-tubing blood pressure prediction model;

[0077] S5: Obtain multi-modal physiological signals of a test subject to be measured, process the multi-modal physiological signals of the test subject to be measured according to steps S2 and S3 to obtain a corresponding multi-modal feature matrix, and then input the multi-modal feature matrix into the trained non-tubing blood pressure prediction model to obtain a non-tubing blood pressure prediction result predicted by the non-tubing blood pressure prediction model.

[0078] Figure 2The workflow of multi-modal physiological signal collection, multi-modal feature extraction and cuffless blood pressure prediction model training in the method of the present application is shown. It can be seen that the multi-modal physiological signals of the subject, such as thoracic impedance signal ICG, photoplethysmogram signal PPG, electrocardiogram signal ECG, etc., collected synchronously by the multi-modal hemodynamic monitoring device based on the method of the present application are respectively subjected to noise reduction processing, and multiple physiological feature parameters of the pre-processed thoracic impedance signal ICG, photoplethysmogram PPG and electrocardiogram signal ECG are extracted, and a multi-modal feature matrix for cuffless blood pressure prediction is constructed based on multiple groups of multiple extracted physiological feature parameters, which is used as the training input data of the cuffless blood pressure prediction model, and the individual determined blood pressure value is used as the label to train the cuffless blood pressure prediction model based on the multi-modal feature matrix, and the trained cuffless blood pressure prediction model can directly predict the cuffless blood pressure prediction result according to the ICG, PPG and ECG multi-modal physiological signals of the subject to be tested. This end-to-end prediction mechanism can effectively improve the generalization ability and monitoring accuracy of continuous cuffless blood pressure monitoring, meet the continuous, non-invasive and high-precision blood pressure monitoring requirements, break through the bottleneck of the traditional cuffless blood pressure monitoring device in terms of insufficient accuracy and poor individual applicability, and has more significant clinical application value.

[0079] The reason why the method of the present application fuses thoracic impedance signal ICG, photoplethysmogram signal PPG, electrocardiogram signal ECG and other multi-modal physiological signals for continuous monitoring of cuffless blood pressure is to comprehensively consider the correlation between the three and blood pressure in different dimensions to enhance the accuracy of continuous blood pressure monitoring.

[0080] The electrocardiogram ECG records the electrical activity of the heart through the body surface electrode, and the waveform characteristics are highly coupled with mechanical contraction. The P wave of the ECG signal corresponds to the systole of the atrium, the atrial muscle contracts, and the blood is squeezed into the ventricle through the atrioventricular valve, and the QRS wave of the ECG signal indicates the ventricular depolarization, corresponding to the systole of the ventricle. During the isovolumic contraction period, the ventricle starts to contract, the intraventricular pressure rises rapidly, and the atrioventricular valve is closed. At this time, the aortic valve has not yet opened, and the ventricular volume is constant; during the ejection period, the ventricular pressure exceeds the aortic / pulmonary artery pressure, the aortic valve opens, and the blood is rapidly ejected into the artery (the rapid ejection period accounts for 60% to 70% of the blood volume), and then the ejection velocity slows down into the slow ejection period. The blood volume ejected into the artery with each heart fluctuation is the stroke volume SV. The T wave of the ECG signal corresponds to the diastole of the ventricle. The electrocardiogram can indirectly reflect the mechanical process of cardiac ejection and indirectly correlate with blood pressure.

[0081] Photoplethysmography (PPG) captures peripheral vascular blood volume pulsation through optical sensors, and its waveform characteristics contain hemodynamic information. The main wave is formed by the pressure wave generated by ventricular ejection conducting to the peripheral blood vessels, and its rising slope is positively correlated with CO and arterial compliance. The dicrotic wave is a characteristic notch formed by the backward flow of blood when the aortic valve closes, marking the beginning of diastole. The PPG waveform morphology, such as the rising slope, peak width, and dicrotic wave characteristics, can reflect vascular elasticity, peripheral resistance, and hemodynamic status. Currently, the core principles of non-invasive blood pressure measurement combining electrocardiogram and photoplethysmography mainly include the physiological correlation between blood pressure and pulse wave transmission time (PTT), pulse wave arrival time (PAT), or pulse wave velocity (PWV).

[0082] Pulse transit time (PTT) is generally defined as the propagation time of the pulse wave between two anatomical sites. It reflects the conduction time of the pressure wave generated by cardiac contraction from the aorta to the peripheral blood vessels, and a large number of studies have shown that it is negatively correlated with blood pressure.

[0083] Impedance cardiography (ICG) is a non-invasive technique for assessing cardiac function by measuring changes in thoracic electrical impedance. Its core mechanism is based on the fact that human tissues are conductive media, with blood having a higher electrical conductivity due to its high electrolyte and water content, while tissues such as the lungs and fat have lower electrical conductivity. ICG applies high-frequency (50-100 kHz) and low-amplitude (1-5 mA) alternating current to the thoracic cavity and measures voltage changes to calculate impedance (Z). During systole, left ventricular ejection rapidly increases aortic blood volume, causing changes in thoracic blood distribution. Due to the high conductivity of blood, the overall thoracic impedance (Z0) temporarily decreases. During diastole, blood flows back to the heart, and thoracic impedance returns to baseline. This periodic change forms the ICG waveform, and the peak of its differential signal corresponds to the maximum change in aortic blood flow velocity. ICG reflects cardiac pumping function by capturing changes in thoracic impedance. It can accurately calculate the output of each cardiac pump and further derive parameters such as cardiac output (CO), cardiac index (CI), and systemic vascular resistance (SVR), which are closely related to blood pressure changes. Additionally, by analyzing the ICG signal, key parameters such as the pre-ejection period (PEP) and left ventricular ejection time (LVET) can be extracted. PEP combined with pulse wave transmission time (PTT) can further calculate pulse wave arrival time (PAT), providing important evidence for blood pressure analysis. Meanwhile, LVET, an important indicator of afterload, is usually significantly prolonged when aortic pressure rises due to hypertension. ICG technology effectively reflects cardiac pumping function by capturing changes in thoracic impedance and, combined with hemodynamic models, quantitatively evaluates parameters such as CO and SVR, thereby providing in-depth physiological analysis of blood pressure changes. This method has important value in dynamic monitoring, cardiovascular disease management, and personalized treatment, and integrating multi-modal data can further improve the accuracy of analysis.

[0084] On the basis of blood pressure prediction based on electrocardiogram signal and photoplethysmogram signal, further combined with impedance cardiogram signal ICG, the prediction performance can be improved. The chest impedance detection technology directly provides the key information of the central blood flow dynamics of the heart by measuring the changes of the chest impedance (Z0) and its first-order differential (dZ / dt): the maximum value of its differential signal (dZ / dt max) is highly related to the aortic blood flow acceleration and myocardial contractility, and the area under the dZ / dt curve is closely related to the stroke volume (SV). This makes up for the shortcomings of ECG (only providing the timing of cardiac electrical excitation) and PPG (mainly reflecting the volume fluctuation of the peripheral microvascular bed with the change of arterial pressure) in quantifying the dynamic changes of the heart pumping function, thereby more accurately capturing the direct impact of cardiac contractility and stroke volume on blood pressure (especially systolic pressure). More importantly, the clear and distinguishable B point in the ICG waveform (usually corresponding to the time when the aortic valve opens) provides a more accurate physiological starting reference point for pulse wave transmission time. The transmission time (PAT_{ICG-PPG}) measured from the opening of the aortic valve (ICG B point) to the starting point of the peripheral pulse wave (PPG feature point) more purely reflects the transmission speed of the pulse wave in the arterial system, which is directly related to the hardness and tension state of the arterial blood vessel wall. This improves the correlation between pulse wave transmission time and blood pressure, and effectively eliminates the interference of individual variation factors such as ventricular electromechanical delay and isovolumic contraction period (PEP) that are unrelated to vascular state when using only the ECG R wave (ventricular electrical excitation starting point) to PPG starting point transmission time (PAT_{ECG-PPG}). In addition, the time interval between B point (aortic valve opening) and X point (aortic valve closing) in the ICG waveform directly provides the measurement of left ventricular ejection time (LVET), which is a key component of calculating cardiac output (CO=SV*HR) and provides indispensable information for accurately quantifying the effective pumping time of each heartbeat and perfecting the evaluation of systemic hemodynamics. In summary, the rich central hemodynamic parameters provided by ICG (including PEP, LVET, CO, etc.) are combined with the accurate cardiac cycle timing provided by ECG and the peripheral pulse waveform information provided by PPG, which realizes more comprehensive and more essential quantification of the core physiological mechanisms affecting blood pressure, such as myocardial contractility, ventricular ejection efficiency, arterial vascular characteristics, and systemic hemodynamic state. This multi-modal fusion not only enhances the physiological basis of the model, but also realizes cross-validation and information complementation through the temporal and spatial correlation and physiological constraint relationship between signals (such as: ICG B point should be slightly later than ECG R wave and earlier than PPG rising branch starting point), significantly enhancing the resistance to common noise sources such as motion artifacts, low perfusion, and respiratory interference, thereby achieving higher accuracy and stronger robustness in continuous non-invasive blood pressure prediction in static and dynamic scenarios.

[0085] In combination with the above analysis, the present application will be further described in detail as follows. As shown in Figure 3

[0086] S1: Obtain the multi-modal physiological signals of the test subject as a sample, and the reference blood pressure value corresponding to each group of multi-modal physiological signals; the multi-modal physiological signals include the thoracic impedance signal ICG, the photoplethysmogram signal PPG and the electrocardiogram signal ECG synchronously collected for the test subject.

[0087] In this step, in the specific implementation, a plurality of test subjects of different ages, different genders and different body types can be obtained as sample objects, and their basic physical characteristics, multi-modal physiological signals composed of thoracic impedance signals ICG, photoplethysmogram signals PPG and electrocardiogram signals ECG, and corresponding reference blood pressure values can be obtained, which are obtained by blood flow hemodynamic monitoring equipment based on thoracic impedance method. These devices are obtained by synchronous acquisition; wherein the preferred physiological parameter sampling rate is 200Hz to ensure data acquisition accuracy.

[0088] At the same time, the basic physical characteristics of the test subject can also be collected, including the height, weight, age and gender characteristics of the test subject.

[0089] Step S2: After time sequence cycle alignment processing according to the synchronous acquisition time sequence information of the multi-modal physiological signals, the multi-modal physiological signals are preprocessed and signal quality evaluated and screened, and then the screened multi-modal physiological signals are subjected to corresponding time sequence cycle feature point positioning processing.

[0090] In this step, in the specific implementation, the time sequence cycle alignment of the thoracic impedance signal ICG, the photoplethysmogram signal PPG and the electrocardiogram signal ECG can be aligned according to the synchronous acquisition time sequence information as the basis.

[0091] In the specific implementation, the preprocessing of the multi-modal physiological signals specifically includes:

[0092] S2101: The multi-modal physiological signals are respectively subjected to denoising processing by using wavelet decomposition method to improve the signal-to-noise ratio.

[0093] ​In the application, the wavelet decomposition method can be used to denoise various physiological signals, and the "sym4" wavelet basis is preferably used to decompose the ICG, PPG and ECG signals by 10 layers. After different fixed hard threshold denoising of different signals according to their main information central frequency bands, the wavelet reconstruction is performed to remove the high-frequency noise and baseline drift. After denoising, it is recommended that the frequency band of the ECG signal is 0.114-12.9 Hz, the frequency band of the ICG signal is 0.381-12.9 Hz, and the frequency band of the PPG signal is 0.381-25.8 Hz.

[0094] S2102: Standardizing the denoised multi-modal physiological signals to unify the dimension. In this way, the dimension influence can be eliminated.

[0095] In specific implementation, the signal quality evaluation and screening of the multi-modal physiological signals specifically includes:

[0096] S2201: Data cleaning is performed on the preprocessed multi-modal physiological signals to screen out defective signals with signal peak values or valley values exceeding the preset threshold range, invalid signals with mutual conflicts or unmatched time sequence periods between the multi-modal physiological signals of the same subject as a sample, and different repeated signals with completely identical data, and the remaining multi-modal physiological signals are determined as usable signals.

[0097] S2202: The multi-modal physiological signals determined as usable are screened again. For the multi-modal physiological signals of the same subject as a sample, the ICG, PPG and ECG signals are all determined as usable and the time sequence periods are aligned, the multi-modal physiological signals are retained, otherwise they are screened out. Thus, the multi-modal physiological signals retained after final screening are obtained.

[0098] In specific implementation, the feature point positioning processing of the corresponding time sequence periods for the screened multi-modal physiological signals specifically includes:

[0099] S2301: For the ICG, PPG and ECG signals of the same subject as a sample, the feature points of each cardiac cycle in the ECG signal are positioned first, including the P-wave peak point, Q-point, R-wave peak point, S-point, T-wave peak point and T-wave end point in the ECG signal. As shown in Figure 4 .

[0100] The specific way of locating the feature points of the electrocardiosignal ECG in each cardiac cycle is as follows: in a single cardiac cycle of the electrocardiosignal ECG, the R-wave peak point is located by peak value detection through a sliding window with the starting position of the cardiac cycle as the starting point, and the Q point is located by backtracking from the R-wave peak point to the first local minimum point, and the S point is located by forward searching from the R-wave peak point to the first local maximum point; then, the P-wave peak point and the T-wave peak point are located by first derivative zero-crossing point detection of the waveform before the S point, and finally the T-wave end point is located by identifying the inflection point at the end of the T wave. In this way, the feature points in each cardiac cycle of the electrocardiosignal ECG are located respectively.

[0101] S2302: In the signal cycle waveform of the thoracic impedance signal ICG corresponding to each cardiac cycle of the electrocardiosignal ECG, the feature points of the thoracic impedance signal ICG are located respectively, including the aortic valve opening point B, the wave peak point C and the aortic valve closing point X in each signal cycle waveform of the thoracic impedance signal ICG. As shown in Figure 4 .

[0102] The specific processing method of this step is as follows: first, determine the signal cycle of the thoracic impedance signal ICG corresponding to each cardiac cycle of the electrocardiosignal ECG according to the aligned time sequence cycle; then, in a single signal cycle of the thoracic impedance signal ICG, locate the wave peak point C by peak value detection through a sliding window with the starting position of the signal cycle as the starting point; then, locate the aortic valve opening point B by backtracking from the wave peak point C to the first local maximum point, and locate the aortic valve closing point X by forward searching from the wave peak point C to the lowest valley point in the signal cycle. In this way, the feature points of each signal cycle of the thoracic impedance signal ICG are located respectively.

[0103] S2303: In the signal cycle waveform of the photoplethysmogram signal PPG corresponding to each cardiac cycle of the electrocardiosignal ECG, the feature points of the photoplethysmogram signal PPG are located respectively, including the pulse wave starting point b, the systolic peak point c, the re-pulse wave starting point f, the re-pulse wave peak point g, the maximum slope point a of the ascending branch and the minimum slope point e of the descending branch in each signal cycle waveform of the photoplethysmogram signal PPG. As shown in Figure 4 .

[0104] The step is specifically processed as follows: first, according to the aligned time sequence period, the signal period of the photoplethysmogram PPG corresponding to each cardiac cycle of the electrocardiogram ECG is determined; then, in a single signal period of the photoplethysmogram PPG, the signal period starting position is taken as a starting point, the minimum point in the signal period is located by sliding window valley value detection to locate the pulse wave starting point b, and the maximum point in the signal period is located by sliding window peak value detection to locate the systolic peak point c; then, the maximum slope corresponding position of the rising branch between the pulse wave starting point b and the systolic peak point c is located by calculating the first derivative extreme value to locate the rising branch slope maximum point a, and the minimum slope corresponding position of the falling branch after the systolic peak point c is located to locate the falling branch slope minimum point e; thereafter, the first derivative zero-crossing point calculation detection is performed on the waveform region after the falling branch slope minimum point e, to determine whether there is a local area maximum point of the first derivative zero-crossing, if there is, the local area maximum point is determined as the located double pulse wave peak point g, if there is no local area maximum point of the first derivative zero-crossing, the local maximum point of the second reciprocal zero-crossing is further calculated to locate as the double pulse wave peak point g; finally, the local area minimum point between the falling branch slope minimum point e and the double pulse wave peak point g is located as the double pulse wave starting point f. In this way, the feature points of each signal period of the photoplethysmogram PPG are located respectively.

[0105] S2304: For the multi-modal physiological signals of each different subject as a sample after screening, the feature point positioning processing of the corresponding time sequence period is performed according to steps S2301-S2303 respectively.

[0106] Step S3: According to the correlation change relationship between each group of multi-modal physiological signals and the reference blood pressure values corresponding thereto, the related features reflecting pulse conduction and the related features reflecting hemodynamic state in the multi-modal physiological signals corresponding to the corresponding time sequence period are mined, and a multi-modal feature matrix corresponding to each group of multi-modal physiological signals is constructed respectively.

[0107] Among them, the related features reflecting pulse conduction in the multi-modal physiological signals corresponding to the corresponding time sequence period include one or more of the following:

[0108] a pulse wave propagation time PTTa from the R-wave peak point to a maximum point a of a rising branch slope of the photoplethysmographic signal PPG, a pulse wave propagation time PTTb from the R-wave peak point to a pulse wave starting point b of the photoplethysmographic signal PPG, a pulse wave propagation time PTTc from the R-wave peak point to a systolic peak point c of the photoplethysmographic signal PPG, a time span from the pulse wave starting point b to the systolic peak point c, a time span from the pulse wave starting point b to a re-pulse wave starting point f, a time span from the maximum point a of the rising branch slope of the photoplethysmographic signal PPG to a minimum point e of a falling branch slope, a time span from the systolic peak point c of the photoplethysmographic signal PPG to a pulse wave starting point b of a next signal cycle, a time span from the re-pulse wave starting point f of the photoplethysmographic signal PPG to an end point of a current signal cycle, a time span from a start point of the photoplethysmographic signal PPG to the end point of the current signal cycle, a time span of 1 / 6 of a maximum amplitude in a waveform of the photoplethysmographic signal PPG, a time span of 2 / 6 of the maximum amplitude in the waveform of the photoplethysmographic signal PPG, a time span of 3 / 6 of the maximum amplitude in the waveform of the photoplethysmographic signal PPG, a time span of 4 / 6 of the maximum amplitude in the waveform of the photoplethysmographic signal PPG, a time span of 5 / 6 of the maximum amplitude in the waveform of the photoplethysmographic signal PPG, a rising branch slope of the photoplethysmographic signal PPG, a falling branch slope of the photoplethysmographic signal PPG, an area under a curve of the photoplethysmographic signal PPG between the pulse wave starting point b and the end point of the current signal cycle in a same cardiac cycle, an area under the curve of the photoplethysmographic signal PPG between the maximum point a of the rising branch slope and the end point of the current signal cycle in the same cardiac cycle, an area under the curve of the photoplethysmographic signal PPG between the systolic peak point c and the end point of the current signal cycle in the same cardiac cycle, a ratio of an amplitude of the maximum point a of the rising branch slope to an amplitude of the systolic peak point c in the same cardiac cycle, a ratio of an amplitude of the minimum point e of the falling branch slope to the amplitude of the systolic peak point c in the same cardiac cycle, a ratio of an amplitude of the re-pulse wave starting point f to the amplitude of the systolic peak point c in the same cardiac cycle.

[0109] The relevant features reflecting the hemodynamic state in the multi-modal physiological signal corresponding to the timing cycle include one or more of the following:

[0110] a heart rate HR, a pre-ejection period PET, a left ventricular ejection time LVET, a thoracic fluid content TFC, a systolic time ratio STR, a stroke volume SV, a cardiac output CO, a body surface area BSA, a stroke volume index SVI, a cardiac index CI, an arterial compliance AC, an acceleration index of myocardial contraction ACI, a velocity index VI, a time span Qdz of a Q point of an electrocardiogram ECG to a C point of a peak of an impedance cardiogram ICG.

[0111] Wherein:

[0112] Heart rate HR refers to the number of heart fluctuations in a minute, i.e. the number of cardiac cycles contained in a minute.

[0113] Pre-ejection time PET is the time span between the Q point of the ECG signal and the aortic valve opening point B of the ICG signal in the same heartbeat cycle.

[0114] Ejection time LVET is the time span between the aortic valve opening point B and the aortic valve closing point X of the ICG signal in the same heartbeat cycle.

[0115] Thoracic fluid level TFC is the reciprocal of the ICG value of the thoracic impedance signal.

[0116] Systolic time ratio STR is the ratio of pre-ejection time PET to ejection time LVET, i.e.:

[0117]

[0118] Stroke volume SV refers to the blood volume in the process of each contraction of the heart, and the calculation formula is:

[0119]

[0120] Wherein, is the peak point amplitude of the ICG signal, and Z0 is the basic impedance value of the human body.

[0121] Cardiac output CO refers to the total amount of blood ejected by the heart per minute, and is defined as:

[0122] CO = SV x HR.

[0123] Body surface area BSA is the total area of the skin surface of the human body, and is defined as:

[0124] BSA (m 2 ) x 0.007184 x body weight 0.425 (kg) x height 0.725 (cm)

[0125] Stroke volume index SVI is an index standardized by the body surface area BSA of the stroke volume SV, which is used to eliminate the influence of individual body size difference on the evaluation of heart function, and is an important parameter for clinically evaluating the pumping efficiency of the heart, and is defined as:

[0126]

[0127] Cardiac index CI represents the cardiac output per unit of body surface area, i.e. the cardiac output per square meter of body surface area per minute, and is defined as:

[0128]

[0129] The acceleration index of myocardial contraction ACI is defined as:

[0130]

[0131] The velocity index VI is defined as:

[0132]

[0133] After obtaining the related features reflecting the pulse conduction and the related features reflecting the hemodynamic state of each group of multi-modal physiological signals through the deep mining process described above, the multi-modal feature set formed by these related features constitutes a multi-modal feature matrix corresponding to the group of multi-modal physiological signals.

[0134] In addition, in the case where the basic physical features of the test object are obtained, the basic physical features of the same test object are also added to the multi-modal feature matrix as set elements when constructing the multi-modal feature matrix corresponding to the multi-modal physiological signals. This enables the training of the cuffless blood pressure prediction model to also take into account the height, weight, age, gender features and other basic physical features of the test object, further improving the specificity and accuracy of the trained cuffless blood pressure prediction model in performing blood pressure prediction on different test objects.

[0135] S4: Construct a cuffless blood pressure prediction model using a multi-head attention mechanism, and use the multi-modal feature matrix corresponding to the multi-modal physiological signals as the training input data of the cuffless blood pressure prediction model, use the reference blood pressure value corresponding to the multi-modal physiological signals as the label, perform deep learning training on the cuffless blood pressure prediction model, and obtain the trained cuffless blood pressure prediction model.

[0136] The cuffless blood pressure prediction model constructed by the present application adopts a deep learning architecture, the core of which is a multi-head self-attention mechanism, which can automatically learn the complex dependency relationships between each multi-modal related feature in the multi-modal feature matrix corresponding to the multi-modal physiological signals, and establish a correlation with the cuffless blood pressure monitoring result, for subsequent accurate prediction of the cuffless blood pressure monitoring result. Specifically,

[0137] The cuffless blood pressure prediction model includes a first normalization layer, a multi-head self-attention network layer, a second normalization layer and a multi-layer perception network layer connected in sequence. The following will explain each network layer of the cuffless blood pressure prediction model.

[0138] (1) The first normalization layer is used to linearly transform the input multi-modal feature matrix into corresponding query vectors, key vectors and value vectors:

[0139] Qi = x i W Q , K i = x i W K , V i = x i W V ;

[0140] wherein, x i ∈ R M×N represents an arbitrary i-th multi-modal feature matrix, M represents the data dimension of the single item related feature contained in the multi-modal feature matrix, N represents the total number of related features contained in the multi-modal feature matrix; Q i , K i , V i are the query vector, key vector, and value vector corresponding to the i-th multi-modal feature matrix x i , respectively; are the linear conversion matrices of the query, key, and value corresponding to the multi-head self-attention mechanism, respectively; d k is the dimension of the attention head in the multi-head self-attention network layer.

[0141] (2) The multi-head self-attention network layer is used to extract the self-attention feature vector of the multi-modal feature matrix:

[0142]

[0143] wherein, H i represents the self-attention feature vector corresponding to the i-th multi-modal feature matrix x i ; represents the output of the l-th attention head in the multi-head self-attention network layer, l = 1, 2, …, L, L represents the number of attention heads in the multi-head self-attention network layer; Concat(·) is a concatenation function; W O is the output weight matrix of the multi-head self-attention network layer; Attention(·) represents the attention mechanism operation; Softmax(·) represents the normalized exponential function operation.

[0144] In specific applications, the number of attention heads in the multi-head self-attention network layer can be selected to be 4-6, which is the best to balance the feature extraction performance and the operation performance.

[0145] (3) The second layer normalization layer takes the addition of the self-attention feature vector and the multi-modal feature matrix as the input, and is used for linear transformation processing of the feature data dimension to obtain the attention superposition feature vector corresponding to the multi-modal feature matrix:

[0146] y i = LN(H i +x i);

[0147] wherein, y i represents the i-th multimodal feature matrix x i corresponding attention superimposed feature vector; LN(·) represents the linear transformation of the second layer normalization layer.

[0148] (4) The multi-layer perception network layer is used for fusion processing of the attention superimposed feature vector of the multimodal feature matrix, to obtain a prediction output of the cuffless blood pressure prediction result, the prediction result including predicted values of systolic blood pressure SBP, diastolic blood pressure DBP and mean arterial pressure MAP; the multi-layer perception network layer includes a plurality of hidden layers and an output layer, the output layer including 3 output neurons corresponding to output predicted values of systolic blood pressure SBP, diastolic blood pressure DBP and mean arterial pressure MAP, and the processing process is represented as:

[0149]

[0150] wherein, respectively represent the attention superimposed feature vector y i as input, the output of the 1st hidden layer, the (j-1)th hidden layer, the jth hidden layer and the Jth hidden layer in the multi-layer perception network layer, j=2, 3, …, J, J representing the number of hidden layers in the multi-layer perception network layer; W (1) , W (j) respectively represent the hidden layer weight of the 1st hidden layer and the jth hidden layer in the multi-layer perception network layer; b (1) , b (j) respectively represent the hidden layer bias of the 1st hidden layer and the jth hidden layer in the multi-layer perception network layer; σ(·) represents the ReLU activation function operation; O i,m represents the output of the mth output neuron of the output layer for the attention superimposed feature vector y i , m=1, 2, 3, O i,1 , O i,2 , O i,3 respectively correspond to the predicted values of systolic blood pressure SBP, diastolic blood pressure DBP and mean arterial pressure MAP. respectively represent the output layer weight and the output layer bias of the mth output neuron in the output layer.

[0151] In specific applications, the number of hidden layers in the multi-layer perception network layer can be selected as 3-5, which is the best to balance the recognition performance and operation performance.

[0152] In the training process of the cuffless blood pressure prediction model, the multi-modal feature matrix data corresponding to all modal physiological signals as training samples can be randomly divided into a training set and a test set in a ratio of 8:2, the reference blood pressure values corresponding to the multi-modal physiological signals are used as labels, the mean absolute error (MAE) is preferably used as the loss function, the learning rate is set to 0.001, the batch size is set to 32, the number of training rounds is set to 1000 rounds, the early stopping mechanism is used for model training (the training is stopped when the loss function of the validation set does not improve for 10 consecutive training rounds), and overfitting is prevented.

[0153] S5: Obtain the multi-modal physiological signals of the test subject, process them to obtain the corresponding multi-modal feature matrix, and input them into the trained cuffless blood pressure prediction model to obtain the cuffless blood pressure prediction result output by the cuffless blood pressure prediction model.

[0154] The specific process of processing the multi-modal physiological signals of the test subject to obtain the corresponding multi-modal feature matrix is consistent with the specific process of steps S2 and S3, and the purpose is to obtain the multi-modal feature matrix of the test subject in the same data format, which is input to the trained cuffless blood pressure prediction model for prediction. Therefore, the specific processing process will not be repeated.

[0155] In summary, compared with the prior art, the present application has the following technical advantages:

[0156] 1. In the cuffless blood pressure continuous monitoring method based on ICG, PPG and ECG multi-modal physiological signals of the present application, the chest impedance signal ICG, the photoplethysmogram PPG, the electrocardiogram ECG and other multi-modal physiological signals synchronously collected by the multi-modal hemodynamic monitoring device are used to participate in the continuous monitoring of the cuffless blood pressure. The synchronous collection mechanism ensures the strict alignment of multi-dimensional physiological data in the time sequence cycle dimension, providing a basis for the time sequence cycle consistency of subsequent feature fusion. In the preprocessing and signal quality assessment and screening stage, targeted noise reduction algorithms and screening mechanisms are used to significantly improve the signal-to-noise ratio of the original signal and effectively eliminate possible baseline drift, signal error conflicts, repeated signals and other disturbance factors in the multi-modal physiological signals, providing a basis for constructing high-quality data for feature extraction.

[0157] 2、In the method, the multi-modal signals screened according to the pretreatment and signal quality evaluation are subjected to further multi-dimensional feature extraction, covering hemodynamic parameters such as pre-ejection time and left ventricular ejection time of ICG, vascular elasticity indexes such as pulse wave transmission time and amplitude variation coefficient of PPG, and electrophysiological characteristics such as R-wave interval and QRS wave group form of ECG, a multi-modal feature matrix is constructed, different modalities and different dimensions of physiological characteristics are structured and fused, a composite feature space capable of comprehensively reflecting the state of the cardiovascular system is formed, and the limitations of a single feature in representing the blood pressure state are effectively overcome.

[0158] 3、In the method, a cuffless blood pressure prediction model adopting a multi-head attention mechanism is constructed based on the multi-modal feature matrix, the cuffless blood pressure prediction model is subjected to deep learning training with reference blood pressure values as labels, precise mapping from multi-modal physiological characteristics to blood pressure values is realized, and in the prediction stage, the ICG, PPG and ECG multi-modal physiological signals of the test subject to be measured can be directly used to realize real-time and continuous cuffless blood pressure value prediction. This end-to-end prediction mechanism can effectively improve the generalization ability and monitoring accuracy of cuffless blood pressure continuous monitoring, meet the requirements of continuous, non-invasive and high-precision blood pressure monitoring, break through the bottleneck of insufficient accuracy and poor individual applicability of traditional cuffless blood pressure monitoring equipment, and has more significant clinical application value.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit the technical solutions, and those of ordinary skill in the art should understand that those who modify or equivalently replace the technical solutions of the present application without departing from the purpose and scope of the technical solutions should be covered in the scope of the claims of the present application.

Claims

1. A cuffless continuous blood pressure monitoring method based on ICG, PPG, and ECG multimodal physiological signals, characterized in that, Includes the following steps: S1: Acquire the multimodal physiological signals of the subjects as samples, and the reference blood pressure value corresponding to each group of multimodal physiological signals; the multimodal physiological signals include chest impedance signal ICG, photoplethysmography (PPG) signal and electrocardiogram (ECG) signal synchronously acquired for the subjects; S2: After performing time-series period alignment processing based on the synchronous acquisition time-series information of the multimodal physiological signals, the multimodal physiological signals are preprocessed and signal quality is evaluated and screened. Then, the selected multimodal physiological signals are subjected to feature point localization processing for the corresponding time-series period. S3: Based on the correlation between each group of multimodal physiological signals and their respective corresponding reference blood pressure values, mine the relevant features reflecting pulse conduction and hemodynamic state in the multimodal physiological signals of the corresponding time period, and construct the multimodal feature matrix corresponding to each group of multimodal physiological signals respectively. S4: Construct a cuffless blood pressure prediction model using a multi-head attention mechanism, and use the multi-modal feature matrix corresponding to the multi-modal physiological signal as the training input data of the cuffless blood pressure prediction model. Use the reference blood pressure value corresponding to the multi-modal physiological signal as the label to perform deep learning training on the cuffless blood pressure prediction model to obtain the trained cuffless blood pressure prediction model. S5: Obtain the multimodal physiological signals of the subject to be tested. After processing the multimodal physiological signals of the subject to be tested according to steps S2 and S3 to obtain the corresponding multimodal feature matrix, input it into the trained cuffless blood pressure prediction model to obtain the cuffless blood pressure prediction result predicted and output by the cuffless blood pressure prediction model.

2. The cuffless continuous blood pressure monitoring method based on ICG, PPG, and ECG multimodal physiological signals according to claim 1, characterized in that, Step S2, the preprocessing of multimodal physiological signals specifically includes: S2101: Wavelet decomposition was used to denoise the multimodal physiological signals to improve the signal-to-noise ratio. After denoising, the retained frequency bands for ECG, ICG, ICG, and PPG were 0.381–12.9 Hz. S2102: Standardize the denoised multimodal physiological signals to unify their dimensions.

3. The cuffless continuous blood pressure monitoring method based on ICG, PPG, and ECG multimodal physiological signals according to claim 2, characterized in that, Step S2, specifically the signal quality assessment and screening of multimodal physiological signals, includes: S2201: Perform data cleaning on the preprocessed multimodal physiological signals to remove defective signals with signal peaks or valleys exceeding the preset threshold range, invalid signals with conflicts or mismatched time periods between multimodal physiological signals of the same subject as a sample, and different repeated signals with completely identical data. The remaining multimodal physiological signals are determined to be usable signals. S2202: The usable multimodal physiological signals are screened again. Multimodal physiological signals that are all usable and have time-cycle aligned for the same subject as a sample, such as chest impedance signal ICG, photoplethysmography pulse wave signal PPG and electrocardiogram signal ECG, are retained; otherwise, they are screened out. Thus, the final retained multimodal physiological signals are obtained.

4. The cuffless continuous blood pressure monitoring method based on ICG, PPG, and ECG multimodal physiological signals according to claim 3, characterized in that, In step S2, the feature point localization processing of the selected multimodal physiological signals for corresponding time periods specifically includes: S2301: For the same subject as a sample, the chest impedance signal ICG, photoplethysmography (PPG) and electrocardiogram (ECG) are used to locate the characteristic points of each cardiac cycle in the ECG, including the P wave peak point, Q point, R wave peak point, S point, T wave peak point, and T wave end point of each cardiac cycle in the ECG. S2302: Within the signal period waveform of the chest impedance signal ICG corresponding to each cardiac cycle of the electrocardiogram (ECG), the characteristic points of the chest impedance signal ICG are located respectively, including the aortic valve opening point B, the peak point C, and the aortic valve closing point X within the signal period waveform of the chest impedance signal ICG. S2303: Within the signal period waveform of the photoplethysmography (PPG) signal corresponding to each cardiac cycle of the electrocardiogram (ECG) signal, the feature points of the PPG signal are located, including the pulse wave initiation point b, the systolic peak point c, the diabetic wave initiation point f, the diabetic wave peak point g, the point with the maximum slope of the ascending branch a, and the point with the minimum slope of the descending branch e. S2304: For the multimodal physiological signals of different subjects after screening, feature point localization processing of the corresponding time period is performed according to steps S2301 to S2303.

5. The cuffless continuous blood pressure monitoring method based on ICG, PPG, and ECG multimodal physiological signals according to claim 4, characterized in that, In step S2301, the specific method for locating feature points of the ECG signal within each cardiac cycle is as follows: In a single cardiac cycle of an ECG signal, the R-wave peak point is first located by using the starting position of the cardiac cycle and sliding window peak detection. Based on the R-wave peak point, the first local minimum point is located backward to the Q point, and the first local maximum point is located forward to the S point. Then, the P-wave peak point and T-wave peak point are located by using first derivative zero-crossing point detection on the forward waveform at the S point. Finally, the T-wave end point is located by identifying the forward inflection point of the T-wave. In this way, characteristic points in each cardiac cycle of the electrocardiogram (ECG) signal can be located.

6. The cuffless continuous blood pressure monitoring method based on ICG, PPG, and ECG multimodal physiological signals according to claim 4, characterized in that, In step S2302, the specific method for locating feature points of the chest impedance signal ICG within the signal period waveform corresponding to each cardiac cycle of the electrocardiogram (ECG) is as follows: Based on the aligned timing cycles, determine the signal period of the chest impedance signal ICG corresponding to each cardiac cycle of the electrocardiogram (ECG). Then, within a single signal cycle of the chest impedance signal ICG, starting from the beginning of the signal cycle, the peak point C of the main wave of the chest impedance signal ICG is located by detecting the peak value through a sliding window. Then, based on the position of the peak point C, the first local maximum point is found backward to locate the aortic valve opening point B, and based on the position of the peak point C, the lowest valley point within the signal cycle is found forward to locate the aortic valve closure point X. In this way, the characteristic points of each signal cycle of the chest impedance signal ICG are located respectively.

7. The cuffless continuous blood pressure monitoring method based on ICG, PPG, and ECG multimodal physiological signals according to claim 4, characterized in that, In step S2303, the specific method for locating feature points of the photoplethysmography (PPG) signal within the signal period waveform corresponding to each cardiac cycle of the electrocardiogram (ECG) signal is as follows: Based on the aligned timing period, determine the signal period of the photoplethysmography (PPG) signal corresponding to each cardiac cycle of the electrocardiogram (ECG). Then, within a single signal cycle of the photoplethysmography (PPG) signal, starting from the beginning of the signal cycle, the minimum point within the signal cycle is detected using a sliding window trough to locate the pulse wave origin point b, and the maximum point within the signal cycle is detected using a sliding window peak to locate the systolic peak point c. Next, by calculating the extreme values ​​of the first derivative, the maximum slope of the rising branch between the pulse wave origin point b and the systolic peak point c is detected to locate the point a with the maximum rising branch slope, and the minimum slope of the falling branch after the systolic peak point c is detected to locate the point a with the minimum falling branch slope. At point e; then, in the waveform region after the minimum slope point e of the descending branch, the first derivative zero-crossing point is calculated and detected to determine whether there is a local maximum point where the first derivative crosses zero. If there is, the local maximum point is determined as the pinpointed diphtheria wave peak point g. If there is no local maximum point where the first derivative crosses zero, the calculation is further performed to find a local maximum point where the second reciprocal crosses zero, which is pinpointed as the diphtheria wave peak point g. Finally, between the minimum slope point e of the descending branch and the diphtheria wave peak point g, a local minimum point is found and pinpointed as the diphtheria wave starting point f. In this manner, the characteristic points of each signal cycle of the photoplethysmography (PPG) signal are located respectively.

8. The cuffless continuous blood pressure monitoring method based on ICG, PPG, and ECG multimodal physiological signals according to claim 4, characterized in that, In step S3, the relevant features reflecting pulse conduction in the multimodal physiological signals corresponding to the time period include one or more of the following: Pulse arrival time (PAT), pulse wave propagation time (PTTa) from the R-wave peak to the point a of maximum slope of the rising branch of the PPG signal (PPG), pulse wave propagation time (PTTb) from the R-wave peak to the pulse wave initiation point (b) of the PPG signal (PPG), pulse wave propagation time (PTTc) from the R-wave peak to the systolic peak point (c) of the PPG signal (PPG), time span from pulse wave initiation point (b) to systolic peak point (c), time span from pulse wave initiation point (b) to dicrotic initiation point (f), and PPG signal (PPG). The time span from the point of maximum slope a on the rising branch to the point of minimum slope e on the falling branch; the time span from the peak point c of the PPG contraction phase to the start point b of the next signal cycle; the time span from the start point f of the PPG diabetic wave to the end point of the current signal cycle; the time span from the start point to the end point of the current signal cycle; the time span when the amplitude of the PPG waveform is 1 / 6 of its maximum amplitude; the time span when the amplitude of the PPG waveform is at its maximum. The time span of 2 / 6, the time span in which the amplitude of the PPG waveform is at its maximum, the time span in which the amplitude of the PPG waveform is at its maximum, the time span in which the amplitude of the PPG waveform is at its maximum, the time span in which the amplitude of the PPG waveform is at its maximum, the slope of the rising branch of the PPG waveform, the slope of the falling branch of the PPG waveform, the area under the PPG curve between the pulse wave initiation point b and the current signal cycle end point within the same cardiac cycle, and the same... The area under the photoplethysmography (PPG) curve of the pulse wave signal between the point a with the maximum slope of the rising limb of the pulse wave within a cardiac cycle and the end of the current signal cycle; the area under the PPG curve of the pulse wave signal between the peak point c of the systolic phase within the same cardiac cycle and the end of the current signal cycle; the ratio of the amplitude of the point a with the maximum slope of the rising limb within the same cardiac cycle to the amplitude of the peak point c of the systolic phase; the ratio of the amplitude of the point e with the minimum slope of the falling limb within the same cardiac cycle to the amplitude of the peak point c of the systolic phase; and the ratio of the amplitude of the dicrotic wave initiation point f to the amplitude of the peak point c of the systolic phase within the same cardiac cycle. The relevant features reflecting hemodynamic state in multimodal physiological signals corresponding to time periods include one or more of the following: Heart rate (HR), pre-ejection time (PET), ejection time (LVET), pleural fluid level (TFC), systolic time ratio (STR), stroke volume (SV), cardiac output (CO), body surface area (BSA), stroke index (SVI), cardiac index (CI), arterial compliance (AC), myocardial contraction acceleration index (ACI), rate index (VI), and time span (Qdz) from the Q point of the ECG signal to the peak point C of the ICG signal. The set of relevant features reflecting pulse conduction and hemodynamic state corresponding to each group of multimodal physiological signals constitutes the multimodal feature matrix corresponding to that group of multimodal physiological signals.

9. The cuffless continuous blood pressure monitoring method based on ICG, PPG, and ECG multimodal physiological signals according to claim 8, characterized in that, When acquiring the multimodal physiological signals of the subjects, the basic physical characteristics of the subjects are also acquired; the basic physical characteristics include the subjects' height, weight, age and gender characteristics. When constructing the multimodal feature matrix corresponding to multimodal physiological signals, the basic physical characteristics of the same subject are also added to the multimodal feature matrix as set elements.

10. The cuffless continuous blood pressure monitoring method based on ICG, PPG, and ECG multimodal physiological signals according to claim 1, characterized in that, In step S4, the cuffless blood pressure prediction model includes a first normalization layer, a multi-head self-attention network layer, a second normalization layer, and a multilayer perceptron network layer, which are cascaded in sequence. The first normalization layer is used to linearly transform the input multimodal feature matrix into the corresponding query vector, key vector, and value vector: Q i =x i W Q ,K i =x i W K ,V i =x i W V ; Where, x i ∈R M×N Let represent any i-th multimodal feature matrix, M represent the data dimension of each relevant feature in the multimodal feature matrix, and N represent the total number of relevant features in the multimodal feature matrix; Q i K i V i Let x be the i-th multimodal feature matrix. i The corresponding query vector, key vector, and value vector; These are the linear transformation matrices for query, key, and value in a multi-head self-attention mechanism, respectively; d k The dimension of the attention head in a multi-head self-attention network layer; The multi-head self-attention network layer is used to extract the self-attention feature vector of the multimodal feature matrix: Among them, H i Let x represent the i-th multimodal feature matrix. i The corresponding self-attention feature vector; This represents the output of the l-th attention head in a multi-head self-attention network layer, where l = 1, 2, ..., L, and L represents the number of attention heads in the multi-head self-attention network layer; Concat(·) is the concatenation function; W O represents the output weight matrix of the multi-head self-attention network layer; Attention(·) represents the attention mechanism operation; Softmax(·) represents the normalized exponential function operation; The second normalization layer takes the sum of the self-attention feature vector and the multimodal feature matrix as input, and performs linear transformation processing on the feature data dimension to obtain the attention superimposed feature vector corresponding to the multimodal feature matrix: y i =LN(H i +x i ); Among them, y i Let x represent the i-th multimodal feature matrix. i The corresponding attention-stacked feature vector; LN(·) represents the linear transformation of the second normalization layer; The multilayer perceptron network layer is used to fuse the attention-superimposed feature vectors of the multimodal feature matrix to obtain the predicted blood pressure without cuffs. The predicted results include the predicted values ​​of systolic blood pressure (SBP), diastolic blood pressure (DBP), and mean blood pressure (MAP). The multilayer perceptron network layer includes multiple hidden layers and one output layer. The output layer includes three output neurons, which output the predicted values ​​of systolic blood pressure (SBP), diastolic blood pressure (DBP), and mean blood pressure (MAP) respectively. The processing procedure is as follows: in, These represent the attention-stacked feature vectors y and y, respectively. i The input is the output of the 1st, (j-1st), jth, and Jth hidden layers in the multilayer perceptron network, where j = 2, 3, ..., J, and J represents the number of hidden layers in the multilayer perceptron network; W (1) W (j) b represents the hidden layer weights of the 1st and jth hidden layers in a multilayer perceptron network, respectively; (1) b (j) These represent the hidden layer biases of the 1st and jth hidden layers in a multilayer perceptron network, respectively; σ(·) represents the ReLU activation function operation; O i,m This indicates that the output layer performs attention on the feature vector y. i The output of the m-th output neuron, m = 1, 2, 3, O i,1 O i,2 O i,3 These correspond to the predicted values ​​of systolic blood pressure (SBP), diastolic blood pressure (DBP), and mean systolic blood pressure (MAP), respectively. and represent the output layer weight and output layer bias of the m-th output neuron in the output layer, respectively.

Images