CN113749630B

CN113749630B - Blood pressure monitoring system and method based on ECG signal and PPG signal

Info

Publication number: CN113749630B
Application number: CN202111086954.5A
Authority: CN
Inventors: 舒琳; 何家裕; 何岸; 徐向民; 刘涛; 曲若文
Original assignee: South China University of Technology SCUT; DO Technology Co ltd
Current assignee: South China University of Technology SCUT; DO Technology Co ltd
Priority date: 2021-09-16
Filing date: 2021-09-16
Publication date: 2022-06-07
Anticipated expiration: 2041-09-16
Also published as: CN113749630A

Abstract

The invention relates to the field of wearable equipment, in particular to a blood pressure monitoring system and a blood pressure monitoring method based on an ECG signal and a PPG signal, which comprises a physiological signal acquisition module, a data transmission module, a data preprocessing module, a feature extraction module, an abnormal value detection module, a feature selection module and a blood pressure regression prediction module; through the physiological signal collection module of deployment in intelligent wrist-watch, synchronous collection wearer's ECG signal and PPG signal, data preprocessing module carries out the preliminary treatment to physiological signal data, the characteristic of ECG signal and PPG signal is drawed to the characteristic extraction module, data abnormal value detection module carries out abnormal value detection and rejects relevant sample according to the physiological signal characteristic of drawing, the physiological signal characteristic relevant with blood pressure is selected out to the characteristic selection module, blood pressure regression prediction module carries out blood pressure prediction to the physiological signal characteristic of drawing, make blood pressure prediction model more adapt to every user. The invention improves the portability of the blood pressure monitoring equipment and can ensure the blood pressure monitoring precision.

Description

Blood pressure monitoring system and method based on ECG signal and PPG signal

Technical Field

The invention relates to the field of wearable equipment, in particular to a blood pressure monitoring system and method based on an ECG signal and a PPG signal.

Background

Hypertension has become a major factor which puzzles people's health at present, and is one of the most common, most common and representative chronic diseases in China, and the incidence of hypertension in China accounts for 3% -10% of the total number of people, so the method has great significance for monitoring and preventing hypertension in time. With the vigorous development of wearable medical equipment, a more convenient and accurate self-health monitoring mode is provided for people. Through wearable equipment, can monitor the person's of wearing physiological signal at any time, like electrocardiosignal, pulse signal, skin signal of telecommunication etc to obtain the person's of wearing health state, convenience of customers carries out timely prevention and treatment.

At present, the main method for measuring blood pressure is to measure blood pressure by an oscillometric method, pressurize a cuff to block blood flow of brachial artery, measure small pressure pulses transmitted by an arm when pressure is slowly reduced, identify the small pulses transmitted from the arm to the cuff by an instrument, and perform differentiation to obtain a blood pressure value.

Disclosure of Invention

In order to solve the technical problems in the prior art, the invention provides a blood pressure monitoring system and a blood pressure monitoring method based on an ECG signal and a PPG signal, wherein the ECG signal and the PPG signal of a user are synchronously acquired and subjected to processing such as feature extraction, physiological signal features related to blood pressure are selected, a personal blood pressure prediction model is trained by combining a migration learning algorithm with an established database, and the blood pressure prediction is carried out on the selected physiological signal features, so that the blood pressure prediction is more adaptive to each user, the portability of blood pressure monitoring equipment is improved, and the accuracy of blood pressure monitoring can be ensured.

The invention discloses a blood pressure monitoring system based on an ECG signal and a PPG signal, which comprises a physiological signal acquisition module, a data transmission module, a data preprocessing module, a feature extraction module, an abnormal value detection module, a feature selection module and a blood pressure regression prediction module, wherein the physiological signal acquisition module is used for acquiring a physiological signal;

the physiological signal acquisition module is used for synchronously acquiring an ECG signal and a PPG signal of a user;

the data transmission module is used for packaging the ECG signal and the PPG signal into a data packet and transmitting the data packet to the mobile terminal of the mobile phone;

the data preprocessing module is used for preprocessing the acquired ECG signal and the PPG signal, and comprises time alignment, data segmentation, signal denoising, normalization processing, peak detection and data rejection;

the characteristic extraction module is used for extracting the characteristics of the ECG signal and the PPG signal;

the abnormal value detection module is used for detecting and eliminating abnormal samples distributed in the characteristic space to obtain pure sample data;

the characteristic selection module is used for screening out physiological signal characteristics related to blood pressure through a genetic algorithm;

the blood pressure regression prediction model module is used for performing blood pressure prediction model training on each user by using a transfer learning algorithm in combination with a physiological signal blood pressure database, performing blood pressure regression prediction on the selected physiological signal characteristics in combination with a linear regression algorithm, and outputting predicted values of systolic pressure and diastolic pressure; and adjusting parameters of the blood pressure prediction model so that the blood pressure prediction model is adapted to each user.

The monitoring method is realized on the basis of the blood pressure monitoring system based on the ECG signal and the PPG signal, and comprises the following steps:

step 1, physiological signal data acquisition, namely synchronously acquiring an ECG signal and a PPG signal of a user;

step 2, data transmission, namely packaging the ECG signal and the PPG signal into a data packet and transmitting the data packet to a mobile terminal of the mobile phone, and uploading the data packet to a server by the mobile terminal of the mobile phone;

step 3, data preprocessing, namely preprocessing the ECG signal and the PPG signal, wherein the preprocessing comprises data segmentation, time alignment, signal denoising, normalization, peak detection and data rejection;

step 4, feature extraction, namely extracting features of an ECG signal and a PPG signal from the physiological signal data after data preprocessing;

step 5, detecting abnormal data values, namely detecting abnormal data values through an isolated forest algorithm according to the extracted physiological signal characteristics, and removing abnormal samples;

step 6, selecting characteristics, namely selecting physiological signal characteristics related to blood pressure;

step 7, blood pressure regression prediction, namely training a blood pressure prediction model for each user by using a transfer learning Tradaboost algorithm in combination with a physiological signal blood pressure database, performing blood pressure regression prediction on the selected physiological signal characteristics in combination with a linear regression algorithm, and outputting predicted values of systolic pressure and diastolic pressure; and adjusting parameters of the blood pressure prediction model so that the blood pressure prediction model is adapted to each user.

Compared with the prior art, the invention has the following advantages and beneficial effects:

the physiological signal acquisition module arranged on the intelligent watch or bracelet device is used for synchronously acquiring the ECG signal and the PPG signal of a wearer (namely a user), so that the intelligent watch or bracelet device is simple to use and convenient to carry, and can be used for monitoring the blood pressure at any time; preprocessing the physiological signal data through a data preprocessing module; the feature extraction module extracts features of an ECG signal and a PPG signal, the data abnormal value detection module detects abnormal values according to the extracted features of the physiological signals and eliminates related samples, the feature selection module selects the features of the physiological signals related to blood pressure, the blood pressure regression prediction module trains a personal blood pressure prediction model of a user by combining a migration learning algorithm with an established database, and the blood pressure prediction is carried out on the extracted features of the physiological signals, so that the blood pressure prediction model is more adaptive to each user, and the blood pressure monitoring precision is ensured.

Drawings

FIG. 1 is an overall system block diagram in an embodiment of the invention;

FIG. 2 is a flowchart of an overall data preprocessing module according to an embodiment of the present invention;

FIG. 3 is a graph of a segment of an ECG signal cut into 5 seconds in an embodiment of the present invention;

FIG. 4 is a graph of a PPG signal fragment cut into 5 seconds according to an embodiment of the present invention;

fig. 5 is a schematic diagram of extracting features of a PPG raw signal according to an embodiment of the present invention;

fig. 6 is a schematic diagram of PAT feature extraction of ECG and PPG peak time differences in an embodiment of the invention;

FIG. 7 is an effect diagram of an isolated forest algorithm in an embodiment of the present invention;

FIG. 8 is a flow chart of creating a model for predicting a blood pressure of a person using the Tradaboost algorithm in an embodiment of the present invention.

Detailed Description

The technical solutions of the present invention will be described in detail with reference to the accompanying drawings and examples, and it is obvious that the described examples are some, but not all, examples of the present invention, and the embodiments of the present invention are not limited thereto. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Examples

As shown in fig. 1, a blood pressure monitoring system based on an ECG signal and a PPG signal includes a physiological signal acquisition module, a data transmission module, a data preprocessing module, a feature extraction module, an abnormal value detection module, a feature selection module, and a blood pressure regression prediction module. The physiological signal acquisition module arranged on the watch or the bracelet device is used for synchronously acquiring an ElectroCardioGram (ElectroCardiogram) ECG signal and a PhotoPlethysmoGraphy (PhotoPlethysmogry) PPG signal of limb leads of a wearer, and the physiological signal acquisition module is simple to use, convenient to carry and capable of monitoring blood pressure at any time; preprocessing the physiological signal data through a data preprocessing module; the characteristic extraction module extracts the characteristics of an ECG signal and a PPG signal, the data abnormal value detection module detects abnormal values according to the extracted physiological signal characteristics and eliminates related abnormal samples, the characteristic selection module selects physiological signal characteristics related to blood pressure, the blood pressure regression prediction module trains a personal blood pressure prediction model of a user by combining a migration learning algorithm with an established database, and the blood pressure prediction is carried out on the extracted physiological signal characteristics, so that the blood pressure prediction model is more adaptive to each user, and the precision of blood pressure monitoring equipment is ensured.

The physiological signal acquisition module comprises an ECG sensor and a PPG sensor and is used for synchronously acquiring two physiological signal data of an ECG signal and a PPG signal to obtain the ECG signal and the PPG signal with the sampling frequency of 250 Hz. Preferably, the ECG sensor and the PPG sensor are integrated on one smart watch or bracelet device, the ECG sensor employs three single-lead limb electrodes, two of which are integrated at the bottom of the smart watch or bracelet device, and the other electrode is integrated at the right side of the smart watch or bracelet device; what PPG sensor adopted is photoelectric sensor, integrate in intelligent wrist-watch or bracelet device bottom, be located the central authorities of ECG sensor bottom bipolar electrode on intelligent wrist-watch or the bracelet device, wear intelligent wrist-watch or bracelet device in left wrist department, when opening and gathering physiological signal data, right hand forefinger dabs intelligent wrist-watch or bracelet device right side electrode, 250Hz timer interrupt is opened to intelligent wrist-watch or bracelet device's master control procedure, through signal amplification circuit, ECG signal and PPG signal in the reading sensor such as analog-to-digital conversion circuit, synchronous collection sampling frequency is 250Hz ECG signal and PPG signal.

And the data transmission module is used for transmitting the physiological signal data ECG signal and the PPG signal from the sensor end to the mobile phone end through Bluetooth communication, and the mobile phone end uploads the physiological signal data to the server through a network. In this embodiment, the bluetooth module disposed at the smart watch end is connected with the bluetooth of the mobile phone mobile end, updates the time information on the smart watch end synchronously, and then packs and transmits the acquired physiological signal data to the mobile phone mobile end. The data of one second is taken as one frame in the data packet, each frame of data carries time information on the watch, and then the mobile terminal of the mobile phone uploads the data to the server through the network, so that the calculation amount of the watch terminal is reduced, and the calculation speed of the whole blood pressure prediction model can be guaranteed.

And the data preprocessing module is used for preprocessing the acquired ECG signals and the PPG signals such as time alignment, data segmentation, signal denoising, normalization processing, peak detection, data elimination operation and the like, and eliminating physiological signal data segments which do not meet the conditions. As shown in fig. 2, in the overall process of the data preprocessing module, the data transmission module transmits the physiological signal data packet to the mobile terminal of the mobile phone, the physiological signal data packet takes 1 second as one frame, the frame header is the current time information, and the data portion is the acquired ECG signal and the PPG signal. The two physiological signals are aligned by time information in the physiological signal data packet, and then data slicing is performed at intervals of T seconds, preferably T is 5. Denoising the ECG signal, firstly designing a median filter of 200ms and 600ms, filtering QRS complex, P wave and T wave of the electrocardiosignal to obtain an ECG signal baseline, and subtracting the original signal from the baseline to obtain the electrocardiosignal with the baseline drift eliminated; then designing a 50Hz digital wave trap to remove power frequency noise interference; and finally, designing a low-pass filter with the cut-off frequency of 45Hz to filter the influence of the electromyographic noise to obtain a pure electrocardiosignal. PPG signal denoising firstly decomposes noise detail components by using wavelet transformation, then carries out signal reconstruction by wavelet inverse transformation, thereby eliminating baseline drift, and then designs a 50Hz digital wave trap to remove power frequency noise interference and a band-pass filter with cut-off frequency of 0.5Hz-4Hz to filter high-frequency noise. The ECG signal and the PPG signal after filtering are normalized, then peak detection is carried out, and the amplitude threshold value and the time interval threshold value of the peak detection are set, so that the maximum value is prevented from being identified by mistake. And data are removed according to the peak detection result, because the wave crests of the ECG signal and the PPG signal which are simultaneously acquired in a normal resting state are alternately appeared, the number of the wave crests of the ECG signal and the PPG signal and the positions of the wave crests are checked whether to alternately appear in a physiological signal segment of 5 seconds, and the physiological signal data segment which does not meet the condition is removed.

The feature extraction module is used for extracting features of the ECG signal and the PPG signal, the extracted features comprise time-frequency domain features of the ECG signal, time-frequency domain features of the PPG signal and time-frequency domain connection of the two physiological signals, and multi-feature fusion is carried out to comprehensively extract physiological information related to blood pressure. The time domain characteristics of the ECG signal comprise the variance, the mean, the maximum value, the minimum value, the arrangement entropy and the sample entropy of a detail signal and an approximate signal obtained after extracting first-order wavelet transform of the ECG signal, and the frequency domain characteristics comprise frequency band energy and frequency band energy ratio; the time domain features of the PPG signal comprise heart rate features for extracting the PPG raw signal, Pearson correlation coefficient between pulses, PPG raw signal, first derivative signal, maximum amplitude, minimum amplitude, period, minimum amplitude/maximum amplitude of second derivative signal, systolic phase time, diastolic phase time, systolic phase time/diastolic phase time, systolic phase time + diastolic phase time of the signal when the pulse amplitude is 10%, 25%, 33%, 50%, 66%, 75%, 100% of the signal, systolic phase signal area, diastolic phase signal area, systolic phase signal area or diastolic phase signal area, systolic phase signal area and diastolic phase signal area; the variance, the mean value, the maximum value, the minimum value, the arrangement entropy, the sample entropy, the frequency band energy and the frequency band energy ratio of detail signals and approximate signals obtained after the PPG signals are subjected to first-order, second-order and third-order wavelet transformation; and the fundamental frequency, fundamental amplitude, second harmonic frequency, second harmonic amplitude, third harmonic frequency, and third harmonic amplitude of the PPG raw signal. The correlation in time of the two physiological signals refers to the time difference between the synchronously acquired ECG and PPG peaks, called the pulse arrival time.

Specifically, as shown in fig. 3 and 4, the data preprocessing module divides the physiological signal into physiological signal data segments of 5 seconds, and extracts the following characteristic parameters of the physiological signal data segments:

(1) the ECG time-frequency domain characteristic parameters comprise: the detail signal (cD) obtained by the wavelet transform, the minimum value (E _ Min _ cD1, E _ Min _ cD2, E _ Min _ cD3, E _ Min _ cA3) of the approximation signal (cA), the maximum value (E _ Max _ cD1, E _ Max _ cD2, E _ Max _ cD3, E _ Max _ cA3), the variance (E _ Var _ cD1, E _ Var _ cD2, E _ Var _ cD3, E _ Var _ cA3), the Mean value (E _ Mean _ cD1, E _ Mean _ sad _ cD1, E _ sac _ cD1, E _ Mean _ cA 1), the band Energy (E _ Energy _ cD 4, E _ Energy _ ce _ cD1, E _ Energy _ E _ cA 1, E _ Energy _ guard _, E _ SampEn _2_ cD2, E _ SampEn _2_ cD3, E _ SampEn _2_ cA3), permutation entropy (E _ PE _ cD1, E _ PE _ cD2, E _ PE _ cD3, E _ PE _ cA 3);

(2) as shown in fig. 5, the PPG time-frequency domain characteristic parameters include: pearson correlation coefficient (pearson) between pulses, maximum amplitude (Ih), minimum amplitude (Il) of pulses, pulse signal period (T _ pi), minimum/maximum amplitude (Il/Ih), pulse amplitude of 10%, 25%, 33%, 50%, 66%, 75%, systolic time at 100% (St10, St25, St33, St50, St66, St75, St), diastolic time (Dt10, Dt25, Dt33, Dt50, Dt66, Dt75, Dt), ratio of systolic and diastolic periods (St10/Dt10, St25/Dt25, St25/Dt 4, St 25/25, St25/Dt25, St/25, St25+ 25A 25+ 25, St25+ 25A 25, St25+ 25, St25+ 25, St25+ D25, St25+ 25, St 25/D25, St 25/D4, St 25/D25, St 25A 25, St 25/D25, St # 25, St 25/D4, St25, St # 25/D25, St # 25, and D25, St # 25/D25, St # 25, St #, A _ St), diastolic area (a _ Dt10, a _ Dt25, a _ Dt33, a _ Dt50, a _ Dt66, a _ Dt75, a _ Dt), ratio of systolic and diastolic area (a _ St10/a _ Dt10, a _ St25/a _ Dt25, a _ St25+ a _ St25, a _ St25+ a _ Dt25, a _ St25+ a _ St25, a _ St25+ a _ Dt25, a _ sto _ Dt 25); first and second derivative peak values (Ih _1d, Ih _2d), troughs/peaks (Il _1d/Ih _1d, Il _2d/Ih _2d), rise times (St _1d, St _2d), fall times (Dt _1d, Dt _2d), start-to-trough times (St _1d + Dt _1d, St _2d + Dt _2d), fall times/rise times (Dt _1d/St _1d, Dt _2d/St _2d), rise areas (a _ St _1d, a _ St _2d), fall areas (a _ Dt _1d, a _ Dt _2d), sum of rise and fall areas (a _ St _1d + a _ Dt _1d, a _ Dt _2d), rise and fall areas (a _ Dt _1d + a _ Dt _1d, a _ Dt _2d), and rise/fall areas (a _1d + a _ Dt _1d, a _ Dt _ d, a _ Dt _2d), and fall areas (a _1d, a _ d/d) A _ St _2d/a _ Dt _2d), fundamental frequency (f _ base), fundamental amplitude (sp _ mag _ base), second harmonic frequency (f _2), second harmonic amplitude (sp _ mag _2), third harmonic frequency (f _3), third harmonic amplitude (sp _ mag _3), Heart Rate (HR), detail signal (cD) obtained by wavelet transformation, and minimum value (P _ Min _ cD1, P _ Min _ cD2, P _ Min _ cD3, P _ Min _ cA3), maximum value (P _ Max _ cD1, P _ Max _ cD2, P _ Max _ cD3, P _ Max _ cD 3), variance (P _ Var _ cD1, P _ Var _ cD2, P _ Var _ cD3, P _ Var _ cA _ 27), average value (en _ cD _ 22), end _ cD 22, etc., variance (P _ Var _ cD 388, P _ cD3, etc.), average _ cD 22, etc. m _ cD3, m _ cD 54, m _ cD _ 12, m _ cD _ c _ 12, m _ cD _ c _ cD _ c _ 22, m _ cD _3, m _ cD _ c _3, m _ cD _ 12, m _ cD _ c _ cD _ 12, m _ cD _ P _ cD _ c _ P _ c _ cD _ 22, m _ cD _ 22, m _ cD _ 9, m _ cD _ P _ cD _ 22, m _ cD _ 22, m _ cD _ P _ cD _ 22, m _ cD _ P _ cD _ 22, m _ cD _ 9, m _ cD _ 22, m _ cD _ 24, m _ cD _ 9, m _ cD _ 24 _ cD _ 22, m _ cD _, Band Energy fraction (P _ Ratio _ Energy _ cD1, P _ Ratio _ Energy _ cD2, P _ Ratio _ Energy _ cD3, P _ Ratio _ Energy _ cA3), m-1 sample entropy (P _ sample _1_ cD1, P _ sample _1_ cD2, P _ sample _1_ cD3, P _ sample _1_ cA3), m-2 sample entropy (P _ sample _ en _2_ cD1, P _ sample _2_ cD2, P _ sample _2_ cD3, P _ sample _2_ cD 36 3), permutation entropy (P _ PE _ cD1, P _ PE _ c _ cD2, P _ sample _ PE _ cA _ 3626, P _ PE _ c _ cA _ c _ cD 3);

preferably, the correlation matrix between pulses of the PPG signal over a certain time period is obtained by calculating the pearson correlation coefficient between pulses, reflecting the degree of change of the PPG signal pulses in a short time. Extraction of n pulse periods S from a 5 second PPG signal₁、S₂、S₃…S_nThe calculation formula from pulse to pulse is as follows:

wherein S is_i、S_jRepresenting two different pulse signal periods separated from a 5 second PPG signal segment, the correlation matrix between pulses is as follows:

and calculating the mean value, the variance and the minimum value of all elements in the correlation matrix, wherein the mean value reflects the average level of the correlation between the pulses, the variance reflects the variation degree of the correlation between the pulses, and the minimum value reflects the maximum degree of the variation between the pulses.

(3) As shown in fig. 6, the characteristic parameter of the correlation in time between the ECG signal and the PPG signal is the time difference (PAT) between the ECG signal peak and the PPG signal peak.

And the abnormal value detection module is used for detecting and eliminating abnormal samples which are distributed in a sparse area and have high distance density in the characteristic space, so that the eliminated pure sample data can represent the blood pressure information of the current user better. Abnormal value detection is carried out on the samples in the feature space through an isolated forest algorithm, sample points which are distributed in a sparse region and are far away from a high-density group in the feature space are considered as abnormal samples, the abnormal samples are isolated, and the abnormal samples are removed. The acquired 4-minute physiological signal data is cut into physiological signal data fragments of 5 seconds, then filtering is carried out, unqualified samples are removed according to the positions of wave crests, the sizes of the wave crests and the number of the wave crests, strong noise and motion artifacts which are difficult to filter still exist in the residual sample data fragments, the characteristics extracted from the samples cannot represent the blood pressure condition of the current user, therefore, the segmented sample points are mapped to the high-dimensional feature space through the isolated forest algorithm, and cutting different subspaces according to each randomly selected dimension characteristic until only one sample point exists in the subspaces, calculating the abnormal score s of each sample point, setting an abnormal sample proportion value, preferably setting the abnormal sample proportion value to be 0.2, and finally rejecting abnormal samples which are distributed in a sparse area and are far away from a high-density group in the characteristic space. As shown in fig. 7, the effect graph of the isolated forest algorithm in the embodiment.

Preferably, the specific steps of carrying out outlier detection on the sample in the feature space by using the isolated forest algorithm comprise:

step 1: randomly selecting X sample points from the rest sample data fragments and mapping the X sample points to a high-dimensional feature space;

step 2: randomly selecting one-dimensional features, generating a hyperplane by setting a random threshold, and segmenting sample data in a feature space into two subspaces, namely generating two branches of an isolated tree;

and step 3: continuously repeating the step 2 in the cut subspace until only one sample point or the number of branches of the isolated tree in the cut subspace reaches the upper limit;

and 4, step 4: generating a plurality of isolated trees, integrating the result of each tree, and calculating the abnormal score s of each residual sample data segment, wherein the calculation formula is as follows:

where n (X) is the number of nodes in each tree and C (X) is the average path of X samples.

And 5: and setting an abnormal value score threshold, and when the abnormal score of the sample exceeds the threshold, the sample is considered as an abnormal sample and is removed from the residual sample data fragments.

The feature selection module is used for searching the global optimal feature subset through a genetic algorithm, screening out physiological signal features related to blood pressure and eliminating redundant features, reducing the calculated amount of a blood pressure regression model and improving the blood pressure regression prediction precision, and mainly comprises the following steps: initializing the population quantity; setting a fitness function; randomly selecting a specific number of individuals from the population; exchanging chromosomes among individuals to form new filial generation individuals; and (5) carrying out mutation on the chromosomes of the offspring individuals according to a certain probability.

The genetic algorithm is developed based on a skearn-genetic 0.2 toolkit, and the set parameters are as follows: man _ position ═ 100, cross _ proba ═ 0.5, rotation _ proba ═ 0.3, n _ generations ═ 100; the number of selected features is 5 to 40; the use of 5-fold cross-validation methods to avoid selected features has limitations and specificity.

The blood pressure regression prediction model module is used for performing personalized model training on each user by using a migration learning algorithm Tradaboost in combination with an established physiological signal blood pressure database, performing blood pressure regression prediction on the selected physiological signal characteristics in combination with a linear regression algorithm, and outputting predicted values of systolic pressure (SBP) and diastolic pressure (DBP) corresponding to the physiological signal characteristics; by adjusting the model parameters, the blood pressure prediction model is more adaptive to each user.

As shown in fig. 8, the specific steps of performing personalized model training for each user by using the Tradaboost transfer learning algorithm include:

step 1, establishing a physiological signal blood pressure database of a group, collecting data in a sound insulation test room, recruiting subjects, and collecting real blood pressure and corresponding 4-minute physiological signals of the subjects in three time periods of 8:00-10:00, 14:00-16:00 and 18:00-20:00 in five consecutive days as source domain data;

step 2, acquiring real blood pressure and corresponding 4-minute physiological signals in three time intervals of 8:00-10:00, 14:00-16:00 and 18:00-20:00 in one day as target domain data by using an electronic sphygmomanometer and a physiological signal acquisition module;

step 3, updating corresponding sample weights by calculating errors of a predicted value and a true value of the blood pressure in the source domain data in the training process by using a Tradaboost transfer learning algorithm, wherein the closer the predicted value is to the true value, the greater the corresponding weight is; conversely, the smaller the weight; by continuously reducing the weight of sample data contradictory to the target domain data in the source domain data, data which are not matched with the target domain in the source domain are filtered, so that the model is more suitable for each user.

Preferably, the specific training step of adjusting the model parameters includes:

(1) initializing the weight of a single sample, wherein the calculation formula is as follows:

where i represents a single sample, n represents the number of source domain samples, and m represents the number of target domain samples, then all sample weight vectors are:

(2) obtaining the blood pressure predicted value y of each physiological signal sample by using a regression algorithm_pUsing the true blood pressure value y of the physiological signal sample_iCalculating an error

The formula is as follows:

wherein, i isForm individual samples, D_tThe true blood pressure value and the predicted blood pressure value y of the representative sample_pMaximum absolute error of (d);

(3) and updating the sample weight, wherein the formula is as follows:

where i represents a single sample, n represents the number of source domain samples, m represents the number of target domain samples,

error of true blood pressure value and predicted blood pressure value, Z_tIn order to be a normalization constant, the method comprises the following steps of,

n is the set iteration number;

(4) the iteration number N is 10, the steps (2) and (3) are repeated, the sample weight is continuously updated, and the more the predicted value is close to the true value, the larger the corresponding weight is; conversely, the smaller the weight; by continuously reducing the weight of sample data contradictory to the target domain data in the source domain, data unmatched to the target domain in the source domain are filtered, and the trained model is more suitable for each user.

When a user uses the blood pressure monitoring system for blood pressure monitoring for the first time, real blood pressure data measured in three time periods of 8:00-10:00, 14:00-16:00 and 18:00-20:00 and physiological signal data of 4 minutes in one day need to be provided for personal model calibration, and when the calibrated model is used for blood pressure prediction, only 5 seconds of physiological signals simultaneously acquired by a watch need to be acquired, so that the current blood pressure value of the user can be predicted. The invention improves the portability of the blood pressure monitoring equipment and can ensure the blood pressure monitoring precision.

Based on the same inventive concept, the present embodiment further provides a blood pressure monitoring method based on an ECG signal and a PPG signal, including the following steps:

step 2, data transmission, namely packaging and transmitting the ECG signal and the PPG signal to a mobile terminal of the mobile phone, and uploading the data to a server by the mobile terminal of the mobile phone;

step 3, data preprocessing, namely preprocessing acquired physiological signal data such as data segmentation, time alignment, signal denoising, normalization, peak detection, data rejection and the like;

step 5, data abnormal value detection, namely performing abnormal value detection according to the extracted physiological signal characteristics through an isolated forest algorithm, removing relevant abnormal samples and removing data segments which do not meet conditions;

The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents thereof, and all such changes, modifications, substitutions, combinations, and simplifications are intended to be included in the scope of the present invention.

Claims

1. A blood pressure monitoring system based on an ECG signal and a PPG signal, comprising: the physiological signal collection module, the data transmission module, the data preprocessing module, the feature extraction module, the abnormal value detection module, the feature selection module and the blood pressure regression prediction module;

the data transmission module is used for packaging the ECG signal and the PPG signal into a data packet and transmitting the data packet to a mobile terminal of the mobile phone;

the data preprocessing module is used for performing time alignment, data segmentation, signal denoising, normalization processing, peak detection and data rejection on the ECG signal and the PPG signal;

the characteristic extraction module is used for extracting the characteristics of the ECG signal and the PPG signal; the features extracted by the feature extraction module from the ECG signal and the PPG signal include time-frequency domain features of the ECG signal, time-frequency domain features of the PPG signal, and a relation between the ECG signal and the PPG signal in time;

the relation of the ECG signal and the PPG signal in time refers to the time difference between the synchronously acquired ECG signal and the PPG signal peak;

obtaining a correlation matrix between pulses of the PPG signal in a certain time period by calculating a Pearson correlation coefficient between the pulses;

the blood pressure regression prediction model module is used for performing blood pressure regression prediction on the selected physiological signal characteristics through a linear regression algorithm by combining a physiological signal blood pressure database, constructing a blood pressure prediction model, performing blood pressure prediction model training on each user by using a transfer learning algorithm, and outputting predicted values of systolic pressure and diastolic pressure; and adjusting parameters of the blood pressure prediction model so that the blood pressure prediction model is adapted to each user.

2. The blood pressure monitoring system according to claim 1, wherein the physiological signal acquisition module comprises an ECG sensor and a PPG sensor which are deployed on the smart watch or bracelet device, the ECG sensor adopts three single-lead limb electrodes, two electrodes are integrated at the bottom of the smart watch or bracelet device, and the other electrode is integrated at the right side of the smart watch or bracelet device; the PPG sensor adopts a photoelectric sensor and is integrated at the bottom of an intelligent watch or a bracelet device.

3. The blood pressure monitoring system of claim 1, wherein the data preprocessing module is specifically configured to:

the method comprises the steps of aligning an ECG signal and a PPG signal according to time information in a data packet, carrying out data segmentation of T seconds length on the ECG signal and the PPG signal, filtering ECG signal noise through a median filter, a digital wave trap and a low-pass filter, filtering PPG signal noise through wavelet transformation, the digital wave trap and a band-pass filter, carrying out normalization and peak detection on the ECG signal and the PPG signal fragment, and screening the ECG signal and the PPG signal fragment according to the number and the position of peaks.

4. Blood pressure monitoring system according to claim 1, characterized in that n pulse periods S are extracted from a T seconds PPG signal₁、S₂、S₃···S_nThe pearson correlation coefficient between pulses is calculated as follows:

wherein S is_i、S_jRepresenting two different pulse signal periods separated from a T-second PPG signal segment, the correlation matrix between pulses is as follows:

5. the blood pressure monitoring system according to claim 1, wherein the abnormal value detection module is configured to perform abnormal value detection on the sample in the feature space by using an isolated forest algorithm, wherein sample points distributed in a sparse region and relatively far from a high-density group in the feature space are considered as abnormal samples, isolate the abnormal samples and reject the abnormal samples;

the abnormal value detection module specifically comprises the following steps of:

wherein n (X) is the number of nodes in each tree, and c (X) is the average path of X samples;

6. The blood pressure monitoring system of claim 1, wherein the genetic algorithm used by the feature selection module has the following parameter settings: man _ position ═ 100, cross _ proba ═ 0.5, rotation _ proba ═ 0.3, n _ generations ═ 100; the number of selected features is 5 to 40; the use of 5-fold cross-validation methods to avoid selected features has limitations and specificity.

7. The blood pressure monitoring system of claim 1, wherein the specific step of the blood pressure regression prediction model module performing personalized model training for each user using a transfer learning algorithm comprises:

step 1, establishing a group physiological signal blood pressure database, collecting data in a sound insulation test room, and collecting real blood pressure of a subject in multiple periods of continuous multiple days and corresponding physiological signals of preset time as source domain data;

step 2, collecting real blood pressure of a plurality of time intervals in a day and corresponding physiological signals at preset time by using an electronic sphygmomanometer and a physiological signal collecting module as target domain data;

8. The blood pressure monitoring system of claim 7, wherein step 3 further comprises:

The formula is as follows:

wherein i represents a single sample, D_tRepresenting the maximum absolute error of the real blood pressure value and the blood pressure predicted value of the sample;

(3) and updating the sample weight, wherein the formula is as follows:

n is the set iteration number;

(4) iterating for multiple times, repeating the steps (2) and (3), continuously updating the sample weight, wherein the closer the predicted value is to the true value, the larger the corresponding weight is; otherwise, the smaller the corresponding weight is; by continuously reducing the weight of sample data contradictory to the target domain data in the source domain, data unmatched to the target domain in the source domain are filtered, and the trained model is more suitable for each user.