Continuous ECG Waveform Reconstruction System and Method Using Smart Wristband Motion Sensor

Abstract

This invention proposes a continuous electrocardiogram (ECG) waveform reconstruction system and method using a smart wristband motion sensor, belonging to the field of mobile computing application technology. The system includes a smart wristband device with a built-in motion sensor and a processing unit. The motion sensor of the smart wristband device collects motion sensing signals from the user's wrist and sends the signals to the processing unit. The processing unit processes the motion sensing signals acquired by the motion sensor, extracts body surface vibration signals related to the heartbeat, segments the body surface vibration signals into single-heartbeat cycle segments, and finally reconstructs the ECG waveform corresponding to the heartbeat cycle. This invention, based on heartbeat-related body surface vibration signals, can accurately reconstruct ECG waveforms. The system is easy to use, can continuously reconstruct ECG waveforms, the measurement process is transparent to the user, and requires no user intervention.

Description

Technical Field

[0001] This invention relates to a continuous electrocardiogram (ECG) waveform reconstruction system, specifically to a continuous ECG waveform reconstruction system and method utilizing a smart wristband motion sensor, belonging to the field of mobile computing application technology. Background Technology

[0002] Continuous electrocardiogram (ECG) waveform measurement systems have shown great potential in areas such as driver fatigue warning, biometric recognition, and emotion recognition. Traditional ECG waveform measurement systems utilize electrodes and electrolyte gels to measure potential changes on the human body surface. However, due to their reliance on specialized equipment and operational complexity, these systems are difficult to widely apply in daily life. To achieve convenient home ECG waveform measurement, some systems integrate electrodes into household appliances and furniture (such as toilets and chairs). However, these methods can only be used when the user is in contact with the measuring device and cannot acquire continuous ECG waveforms for extended periods. Currently, some systems attempt to sew fabric electrodes into textiles (such as pajamas and belts) for ECG waveform measurement, but the high cost hinders the widespread deployment of such systems. In recent years, some new commercial smartwatches have also incorporated electrodes for measuring ECG waveforms, but these devices still require the user to perform specific actions (such as covering the electrodes with a finger) during use and cannot continuously acquire the wearer's ECG waveforms.

[0003] In addition, some systems reconstruct electrocardiogram (ECG) waveforms using signals related to the heartbeat cycle. Unlike directly measuring an ECG using changes in body surface potential, these systems do not require the user to touch electrodes. By utilizing the correlation between the acquired sensor signals and the heartbeat cycle, they can accurately reconstruct the corresponding ECG waveform. For example, wireless signals can be used to measure the minute chest vibrations caused by the heartbeat to predict the ECG waveform. However, the user needs to remain still during the acquisition process to avoid interference from body movement in sensing chest vibrations. Some systems also use vibration sensors deployed in the mattress to reconstruct the ECG waveform of a user during sleep; however, this only works when the user is lying flat, and the measurement is forced to stop when the user lies on their side or gets off the mattress. Furthermore, some systems use pulse oximeters worn on the fingertips to record pulse signals at the fingertips and predict ECG waveforms, but wearing the measuring device on the finger for extended periods can affect hand touch function and cause discomfort.

[0004] In summary, the existing system has various defects and shortcomings. Summary of the Invention

[0005] The purpose of this invention is to overcome the technical shortcomings of existing electrocardiogram (ECG) waveform measurement and reconstruction systems, such as high cost, inconvenience of use, inability to continuously acquire ECG waveforms, and the need for users to perform special actions. This invention creatively proposes a continuous ECG waveform reconstruction system and method using a smart wristband motion sensor.

[0006] A continuous electrocardiogram waveform reconstruction system utilizing a smart wristband motion sensor includes a smart wristband device with a built-in motion sensor and a processing unit.

[0007] The motion sensor in the smart wristband device is used to collect motion sensing signals from the user's wrist and send the signals to the processing unit.

[0008] The processing unit processes motion sensors to acquire human motion sensing signals, extracts body surface vibration signals related to heartbeat, segments the body surface vibration signals into single heartbeat cycle segments, and finally reconstructs the electrocardiogram waveform corresponding to the heartbeat cycle.

[0009] The above system is implemented as follows, including the following steps:

[0010] Step 1: Use the motion sensor (such as a gyroscope) of the smart wristband device to collect motion sensing signals from the target user's wrist.

[0011] Specifically, users wear smart wristband devices with built-in motion sensors, which continuously acquire motion sensing signals from the wrist.

[0012] Step 2: The processing unit extracts surface vibrations related to heartbeat from the motion sensing signal. The purpose is to extract surface vibrations caused by the continuous changes in the center of gravity of blood flow during the heartbeat cycle from the chaotic motion sensing signal.

[0013] Step 2.1: Use a bandpass filter to process the wrist motion sensing signal acquired in Step 1 to remove irrelevant noise.

[0014] Step 2.2: Based on the filtered motion sensing signal extracted in Step 2.1, further remove noise from the messy motion sensing signal using stationary wavelet transform, and extract the body surface vibration signal related to heartbeat.

[0015] Step 3: The processing unit segments the surface vibration signal related to heartbeat. The purpose is to segment the surface vibration signal into single-heartbeat cycle segments based on the specific waveform of the surface vibration signal related to heartbeat.

[0016] Step 3.1: Detect local maxima and minima in the body surface vibration signals related to heartbeat, and construct local triangles. Based on the characteristics of the local triangles, identify the peak points related to ventricular contraction.

[0017] Step 3.2: Based on the peak points related to ventricular contraction identified in Step 3.1, the body surface vibration signal related to heartbeat is segmented into segments of a single heartbeat cycle.

[0018] Step 4: Reconstruct the corresponding electrocardiogram (ECG) waveform using body surface vibration signal segments related to the heartbeat. The purpose is to establish the correspondence between body surface vibration signals and ECG waveforms. The corresponding ECG waveform is reconstructed using body surface vibration signals.

[0019] Step 4.1: Based on the fragments of body surface vibration signals extracted in Step 3.2, establish an encoder-decoder network model capable of reconstructing electrocardiogram waveforms.

[0020] Step 4.2: Using a generative adversarial network, train the encoder-decoder network model established in Step 4.2, and accurately reconstruct the corresponding electrocardiogram waveform based on the body surface vibration signals related to heartbeat.

[0021] Beneficial effects

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] 1. This invention utilizes a smart wristband device to collect wrist motion sensing information of the target user and accurately reconstructs the electrocardiogram waveform based on the body surface vibration signal related to heartbeat.

[0024] 2. This invention is easy to use, can continuously reconstruct ECG waveforms, and the measurement process is transparent to the user without requiring user intervention. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the present invention.

[0026] Figure 2 This is a schematic diagram of the three coordinate axes of the gyroscope in the smart wristband device according to an embodiment of the present invention, as well as the sensing signals and electrocardiogram waveforms acquired on the three coordinate axes.

[0027] Figure 3 This is a schematic diagram illustrating the extraction of motion signals related to heartbeat from body surface vibrations using stationary wavelet transform, as described in an embodiment of the present invention.

[0028] Figure 4 This is a schematic diagram illustrating the construction of local triangles and the features of local triangles in an embodiment of the present invention.

[0029] Figure 5 This is a schematic diagram of the encoder-decoder network model developed in an embodiment of the present invention.

[0030] Figure 6 This is a schematic diagram of the generative adversarial network structure constructed according to an embodiment of the present invention.

[0031] Figure 7 This is a prototype diagram used in an embodiment of the present invention.

[0032] Figure 8 This refers to the waveform reconstruction error of the reconstructed electrocardiogram waveform in the embodiments of the present invention.

[0033] Figure 9 The correlation coefficients for reconstructing electrocardiogram waveforms in embodiments of the present invention.

[0034] Figure 10 The waveform reconstruction error and correlation coefficient of this invention are given at different sampling frequencies in the embodiments of the present invention.

[0035] Figure 11 The waveform reconstruction error and correlation coefficient are shown in the embodiments of the present invention when the wristband device is worn in different positions on the wrist. Detailed Implementation

[0036] The principles and features of the present invention will be further described in detail below with reference to embodiments and accompanying drawings. The examples given are for illustrative purposes only and are not intended to limit the scope of the invention.

[0037] A continuous electrocardiogram waveform reconstruction system utilizing a smart wristband motion sensor includes a smart wristband device with a built-in motion sensor and a processing unit.

[0038] The motion sensor in the smart wristband device is used to collect motion sensing signals from the user's wrist and send the signals to the processing unit.

[0039] The processing unit processes motion sensors to acquire human motion sensing signals, extracts body surface vibration signals related to heartbeat, segments the body surface vibration signals into single heartbeat cycle segments, and finally reconstructs the electrocardiogram waveform corresponding to the heartbeat cycle.

[0040] Figure 1 A schematic diagram illustrating an embodiment of the present invention is shown. When blood flows periodically through blood vessels in conjunction with the heartbeat, the blood's center of gravity changes periodically. The user's body is affected by these changes in the blood's center of gravity, generating opposing forces, particularly rotational energy, which causes minute vibrations in the body. Therefore, motion sensors, especially gyroscopes, are used to collect body surface vibrations related to the heartbeat. By analyzing the body surface vibration signals related to the heartbeat, the corresponding electrocardiogram (ECG) signal is estimated, enabling continuous ECG measurement.

[0041] The above system is implemented as follows, including the following steps:

[0042] Step 1: Use the gyroscope of the smart wristband device to collect motion sensing signals from the target user's wrist.

[0043] Step 1.1: The user wears a smart wristband device with a built-in motion sensor. The motion sensor continuously acquires motion sensing signals from the wrist.

[0044] Specifically as follows:

[0045] Motion sensors, including a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer, are sensitive to minute movements. Analyzing the data from the three-axis gyroscope, the X, Y, and Z axes of the smart wristband device... Figure 2 As shown in (1), the sensing signals of the X, Y, and Z axes of the three-axis gyroscope are defined as G, G, and G, respectively. X G Y and G Z .like Figure 2 As shown in (2), the electrocardiogram and sensor signal G X G Y and G Z There is a clear correlation, among which the key inflection point of the electrocardiogram is related to G. X The key inflection points have a corresponding relationship; therefore, the X-axis sensing signal G from the three-axis gyroscope of the smart wristband device is used. X It serves as a motion sensing signal for the wrist.

[0046] Step 2: The processing unit extracts the body surface vibrations related to heartbeat from the motion sensing signals.

[0047] Step 2.1: Use a bandpass filter to process the wrist motion sensing signal acquired in Step 1.1 to initially remove irrelevant noise.

[0048] Since the body surface vibrations related to heartbeat collected by smart wristband devices are inevitably affected by the user's body movements, a bandpass filter is used to remove signals from other frequency bands to perform preliminary noise reduction on the motion sensing signals.

[0049] Since the body surface vibrations related to the human heartbeat are mainly distributed in the 5Hz-30Hz range, preferably, the present invention retains the portion of the motion sensing signal with a frequency range of 5Hz-30Hz. However, other settings within the [0Hz, 50Hz] range are also within the scope of the present invention.

[0050] Step 2.2: Based on the filtered motion sensing signal extracted in Step 2.1, further remove noise from the messy motion sensing signal using stationary wavelet transform, and extract the body surface vibration signal related to heartbeat.

[0051] Specifically as follows:

[0052] First, a stationary wavelet transform is applied to the motion sensing signal extracted in step 2.1. Let the mother wave be db4, and it is decomposed into J approximate components a1, a2, ..., a4 according to the frequency distribution range. J and J detail components d1, d2, ..., dJ containing high-frequency information J J is the order set during the stationary wavelet transform process. Preferably, the present invention selects a value of 6 for J, but other settings within the range of [2, 10] are also within the scope of the present invention.

[0053] like Figure 3 As shown in (1), the motion sensing signal is affected by body movement noise, making it impossible to observe obvious body surface vibrations related to heartbeat. Figure 3 As shown in (2), after applying the stationary wavelet transform, periodic vibrations can be observed in each detail component of the motion sensing signal.

[0054] Secondly, calculate d for each detail component. j short-term energy e j And detect short-term energy e j All local maximum amplitudes P j (k), k = 1, 2, ..., K, marks the time corresponding to the occurrence of a local maximum amplitude. Define a threshold. and It is the detail component d j short-term energy e j All local maximum amplitudes P j The mean and variance of (k). This is because motion sensing signals caused by most body motion noise have larger amplitudes compared to surface vibration signals related to heartbeat. Compare each P individually. j Amplitude and threshold of (k) if Then classify it as P j (k) represents body surface vibrations related to heartbeat.

[0055] Furthermore, a small portion of motion sensing signals caused by body movement noise and surface vibration signals related to heartbeat exhibit similar amplitudes. Given that motion sensing signals caused by body movement noise are non-periodic while heartbeat-related surface vibration signals are periodic, the time interval between adjacent local maxima is used to further process other unclassified P signals. j (k).

[0056] Define two adjacent local maxima P j (k-1) and P j The time interval for (k) to occur is I. j (k). Setting a threshold For e j The time interval between all adjacent local maxima in the middle is I j The average value of (k). If and Then classify it as P j (k) represents body surface vibrations related to heartbeat.

[0057] Finally, the local maximum amplitude P, which is classified as a body surface vibration related to heartbeat, is used. j (k) Separation dj Vibrations and noises on the body surface related to heartbeat.

[0058] Specifically, the maximum amplitude P of each body surface vibration related to heartbeat was examined. j The occurrence time of (k), the occurrence time of local minima searched forward and backward, the start and end times of body surface vibrations related to heartbeat were determined, and d was retained. j The corresponding data in the middle, other data are identified as noise, and d j The corresponding data is replaced with 0. An inverse stationary wavelet transform is performed to obtain the body surface vibration signal related to the heartbeat.

[0059] like Figure 3 As shown in (3), after applying the inverse stationary wavelet transform, the influence of body motion noise is eliminated in the motion sensing signal, and the body surface vibration signal related to heartbeat is extracted.

[0060] Step 3: The processing unit segments the body surface vibration signals related to heartbeat.

[0061] Step 3.1: Detect local maxima and local minima in the body surface vibration signals related to heartbeat, construct local triangles, and identify peaks related to ventricular contraction based on the characteristics of the local triangles.

[0062] Because surface vibration signals associated with heartbeats have dynamic characteristics (e.g., varying intervals between consecutive heartbeats, and different waveforms in each individual heartbeat cycle), clear features are lacking to define the boundaries of a single heartbeat cycle. Observation shows that surface vibration signals associated with heartbeats always exhibit a significant spike during ventricular contraction. By selecting the spike associated with ventricular contraction, the surface vibration signals associated with heartbeats can be segmented.

[0063] Specifically, all local maxima and local minima are searched in the body surface vibration signals related to heartbeat. A local triangle is constructed using each local maximum and its two adjacent local minima, and features are extracted from these local triangles. For example... Figure 4 As shown, the local maximum point in the local triangle is defined as the vertex, the adjacent left local minimum point is defined as the left base point, and the adjacent right local minimum point is defined as the right base point. The sum of the distances from the vertex to the left and right base points, a+b, the difference d between the ordinates of the vertex and the left base point, the difference e between the ordinates of the vertex and the right base point, and the angle f of the angle where the vertex is located in the local triangle are selected as features.

[0064] In addition, the difference c between the left and right base points, the angle g of the angle where the left base point is located in the local triangle, and the angle h of the angle where the right base point is located in the local triangle can be selected as supplementary features to help identify the peak points.

[0065] A random forest classifier is trained to identify whether vertices (i.e., local maxima) in a local triangle are spurs related to ventricular contraction. During identification, the selected features are input into the random forest classifier to predict the probability that a vertex belongs to a spur. If the probability is greater than a set value (e.g., 0.5), then the vertex is a spur related to ventricular contraction.

[0066] Step 3.2: Based on the peak points related to ventricular contraction identified in Step 3.1, the body surface vibration signal related to heartbeat is segmented into segments corresponding to a single heartbeat cycle.

[0067] Furthermore, in order to obtain a heartbeat-related surface vibration signal containing a single heartbeat cycle, the 250 milliseconds before the peak can be determined as the start point of a heartbeat cycle, and the 250 milliseconds before the next peak can be determined as the end point of this heartbeat cycle. However, other settings within the range of [100 milliseconds, 400 milliseconds] are also within the scope of this invention.

[0068] Step 4: Reconstruct the corresponding electrocardiogram waveform using body surface vibration signal segments related to heartbeat.

[0069] Step 4.1: Based on the fragments of body surface vibration signals extracted in Step 3.2, establish an encoder-decoder network model capable of reconstructing electrocardiogram waveforms.

[0070] Specifically, an encoder-decoder network model is developed based on a Long Short-Term Memory (LSTM) neural network, such as... Figure 5 As shown, segments of body surface vibration signals are input into the encoder-decoder network model. Due to the inconsistent length of the input segments, ranging from approximately 0.6 seconds to 1.1 seconds, linear interpolation is used to stretch the signal length to 1.2 seconds (corresponding to 120 sample data points at a sampling rate of 100Hz). The encoder uses a bidirectional long short-term memory neural network (BLSTM) to extract hidden information related to the heartbeat and integrates it into a contraction feature consisting of N samples (e.g., 30). An attention mechanism (applying the softmax function) is used to assign weights to the N samples in the contraction feature extracted by the encoder. The decoder consists of two unidirectional long short-term memory neural network layers, reconstructing the ECG waveform based on the contraction feature and its corresponding weights. Finally, the reconstructed ECG waveforms are all 1.2 seconds long, and inverse interpolation is used to adjust the length to the original segment length.

[0071] Step 4.2: Using a generative adversarial network, train the encoder-decoder network model established in Step 4.2 to accurately reconstruct the corresponding electrocardiogram waveform based on the body surface vibration signals related to heartbeat.

[0072] Specifically, to achieve accurate ECG waveform reconstruction anytime and anywhere, a training method for the encoder-decoder network model is established. To this end, a deep learning model based on generative adversarial networks is built to assist the training of the encoder-decoder network model offline.

[0073] like Figure 6 As shown, the established generative adversarial network model consists of a generator that reconstructs the ECG waveform and a discriminator that distinguishes the reconstructed ECG waveform from the real ECG waveform. The encoder-decoder network model established in step 4.1 serves as the generator, learning the complex mapping from heartbeat-related surface vibration signals to the ECG waveform. The discriminator takes the reconstructed ECG waveform output by the generator and its corresponding real ECG waveform as inputs, and stretches the input data length to 1.2 seconds using linear interpolation. Then, two bidirectional long short-term memory neural network layers, one fully connected layer, and one decision layer (using the softmax function) are used to distinguish whether the current input signal is the real ECG waveform or the reconstructed ECG waveform. The parameters of the generator and discriminator are trained alternately to reduce the difference between the generated data and the real data, enabling the generator model to accurately reconstruct the ECG waveform.

[0074] The reconstructed electrocardiogram waveform is denoted as E = {E1, E2, ..., E...} i ,…,E L The actual electrocardiogram waveform corresponding to E is denoted as A = {A1, A2, ..., A}. i ,…,A L}, E i A represents the amplitude of the electrocardiogram waveform signal. i Let L be the amplitude of the actual ECG signal and L be the length of the two waveforms. The loss function Loss is the waveform reconstruction error L of the reconstructed ECG waveform by the generator. e The discrimination error L between the discriminator and the real ECG waveform and the reconstructed ECG waveform a Together they form, Loss = L e +L a .set up:

[0075]

[0076] L a =log[1-P EA (2)

[0077] Among them, P EA This represents the proportion of reconstructed ECG waveforms that the discriminator incorrectly identifies as real ECG waveforms. During training, parameters in the adversarial network are iteratively generated until the loss converges.

[0078] Test and verification

[0079] To verify the beneficial effects of the present invention, a wristband-style prototype system was developed for testing. The prototype system is as follows: Figure 7 As shown, the prototype consists of an integrated motion sensor (capable of acquiring signals from a three-axis gyroscope) and an adjustable wristband.

[0080] A total of 20 healthy volunteers (10 men and 10 women, aged 20-33) were recruited to participate in data collection. During the data collection process, each volunteer wore a wristband-type prototype device and a medical electrocardiogram measuring device to record data for approximately 30 minutes under both static and dynamic conditions (swinging forearms, swinging upper arms, and walking) for analysis and training.

[0081] Waveform reconstruction error and correlation coefficient were used to evaluate system performance. The waveform reconstruction error was defined as the average ratio of the absolute value of the amplitude difference between the reconstructed ECG waveform and the corresponding real ECG waveform to the amplitude of the real ECG waveform, i.e., equation (1). A waveform reconstruction error close to 0 indicates that the system can accurately reconstruct the ECG waveform. The correlation coefficient was defined as:

[0082]

[0083] in The average value of the reconstructed ECG waveform E amplitude. The average amplitude of the actual electrocardiogram waveform A corresponding to E is denoted by E. A correlation coefficient close to 1 indicates that the system can accurately reconstruct the electrocardiogram waveform.

[0084] First, the overall performance of the invention was tested. A generative adversarial network was trained using data from 20 volunteers, enabling the generator to accurately reconstruct the corresponding electrocardiogram waveform from body surface vibration signals related to heartbeat. Figure 8 The waveform reconstruction error box plot is shown for four-fold cross-validation (75% of all experimental data was randomly selected for model training, and the remaining 25% was used for model testing). The boxes are drawn from the lower quartile to the upper quartile, with the horizontal line representing the median. The mean waveform reconstruction error for all 20 volunteers was 5.989%, and the standard deviation was 2.496%. Figure 9 The correlation coefficients for performing four-fold cross-validation are shown. The mean correlation coefficient for all 20 volunteers was 0.926, and the standard deviation was 0.030. This indicates that the present invention can accurately reconstruct electrocardiogram waveforms.

[0085] Then, the performance of the invention was tested at different sampling rates, demonstrating that the invention can achieve low waveform reconstruction error and high correlation coefficient at various sampling rates. All volunteers had motion sensing signals from their wrists collected at 60Hz, 100Hz, 150Hz, and 200Hz, respectively. Figure 10The paper demonstrates the waveform reconstruction error and correlation coefficient of ECG waveforms reconstructed from collected data under four sampling rates. As the sampling rate increases, the waveform reconstruction error decreases, while the correlation coefficient increases. In all cases, the waveform reconstruction error is less than 10%, and the correlation coefficient is greater than 0.8. Experiments confirm that this invention can accurately reconstruct ECG waveforms under various sampling rates. Since the motion sensors built into commercial smart wristband devices often support sampling rates of 60-200Hz, this invention can be applied to commercial smart wristband devices with different sampling rates.

[0086] Finally, the performance of the smart wristband device of this invention was tested under different wrist positions, proving that the invention can achieve high accuracy under different wearing positions. Wrist motion sensing signals were collected from all volunteers at positions 1 cm, 2 cm, and 3 cm above the ulnar styloid process (towards the elbow). Figure 11 The waveform reconstruction error and correlation coefficient were shown under different acquisition positions. In all three cases, the waveform reconstruction error was less than 10%, and the correlation coefficient was greater than 0.8. This confirms that the present invention can accurately reconstruct ECG waveforms under various wearing positions.

[0087] The specific examples described above are further explanations of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications and equivalent substitutions made within the principles and spirit of the present invention should be within the scope of protection of the present invention.

Claims

1. A continuous electrocardiogram waveform reconstruction system utilizing a smart wristband motion sensor, characterized in that, A smart wristband device comprising a built-in motion sensor and a processing unit; The implementation method of the system comprises the following steps: Step 1: using the motion sensor of the smart wristband device, collecting the motion sensing signal of the wrist of the target user; Step 2: the processing unit extracts the body surface vibration related to heartbeat in the motion sensing signal; Step 2.1: using a band-pass filter to process the wrist motion sensing signal collected in step 1, removing irrelevant noise and retaining the part of the motion sensing signal with a frequency range of 0Hz-50Hz; Step 2.2: based on the filtered motion sensing signal extracted in step 2.1, further removing noise in the chaotic motion sensing signal based on stationary wavelet transform, and extracting the body surface vibration signal related to heartbeat; Step 3: the processing unit segments the body surface vibration signal related to heartbeat; Step 3.1: detecting local maximum and local minimum points in the body surface vibration signal related to heartbeat, and constructing a local triangle; according to the characteristics of the local triangle, identifying the peak point related to ventricular contraction; Step 3.2: according to the peak point related to ventricular contraction identified in step 3.1, the body surface vibration signal related to heartbeat is segmented into single heartbeat period segments; Wherein, the X milliseconds before the peak point is the starting point of a heart cycle, and the X milliseconds before the next peak point is the end point of the heartbeat cycle, and X is in the range of 100 milliseconds to 400 milliseconds; Step 4: reconstructing the corresponding electrocardiogram waveform using the body surface vibration signal segment related to heartbeat; Step 4.1: developing an encoder-decoder network model based on a long short-term memory neural network, and inputting the body surface vibration signal segment into the encoder-decoder network model; Using linear interpolation method, the signal length is stretched to 1.2 seconds, the encoder extracts the hidden information related to heartbeat and integrates it into a contraction feature composed of X samples by using a bidirectional long short-term memory neural network; the X samples in the contraction feature extracted by the encoder are respectively assigned weights by using an attention mechanism; the decoder is composed of two unidirectional long short-term memory neural network layers, and reconstructs the electrocardiogram waveform according to the contraction feature and the weight corresponding to the contraction feature; finally, the reconstructed electrocardiogram waveform length is 1.2 seconds, and the length is adjusted to the original segment length by using inverse interpolation; Step 4.2: establishing a deep learning model based on a generative adversarial network to assist the training of the encoder-decoder network model in an offline manner; The established generative adversarial network model is composed of a generator for reconstructing electrocardiogram waveform and a discriminator for distinguishing the reconstructed electrocardiogram waveform from the real electrocardiogram waveform, wherein the encoder-decoder network model established in step 4.1 is used as the generator to learn the complex mapping of the heartbeat-related body surface vibration signal to the electrocardiogram waveform; the discriminator takes the reconstructed electrocardiogram waveform output by the generator and the corresponding real electrocardiogram waveform as inputs, respectively, and stretches the input data length to 1.2 seconds by linear interpolation; then, two bidirectional long short-term memory neural network layers, one fully connected layer and one decision layer are used to distinguish whether the current input signal is a real electrocardiogram waveform or a reconstructed electrocardiogram waveform; the parameters of the generator and the discriminator are trained alternately to reduce the difference between the generated data and the real data, so that the generator model can accurately reconstruct the electrocardiogram waveform; Let the reconstructed electrocardiogram waveform be denoted as }, Let the corresponding real electrocardiogram waveform be denoted as }, Let the amplitude of the electrocardiogram waveform be denoted as Let the amplitude of the real electrocardiogram signal be denoted as Let the length of the two waveforms be denoted as The waveform reconstruction error of the electrocardiogram waveform reconstructed by the generator And the discrimination error of the discriminator to distinguish the real electrocardiogram waveform from the reconstructed electrocardiogram waveform Together constitute, ; Let: wherein, is the proportion of reconstructions of electrocardiogram waveforms that the discriminator incorrectly identifies as real electrocardiogram waveforms; during the training process, the parameters in the generative adversarial network are iteratively generated until convergence.

2. The continuous electrocardiograph waveform reconstruction system using a smart wristband motion sensor of claim 1, wherein, The motion sensor comprises a three-axis accelerometer, a three-axis gyroscope and a three-axis magnetometer, the three-axis gyroscope data is analyzed, and the three-axis gyroscope built in the smart wristband device 、 、 Three coordinate axes, three-axis gyroscope 、 、 The sensing signals of the three axes are respectively defined as and ECG and sensing signals and have a significant correlation, wherein the key inflection points of the ECG have a corresponding relationship with The key inflection points, using the sensing signals of the axis of the three-axis gyroscope of the smart wristband device as the motion sensing signals of the wrist.

3. The continuous electrocardiograph waveform reconstruction system using a smart wristband motion sensor of claim 1, wherein, In step 2.1, the part of the motion sensing signal with a frequency range of 5-30 Hz is reserved.

4. The continuous electrocardiograph waveform reconstruction system using a smart wristband motion sensor of claim 1, wherein, In step 2.2, firstly, the stationary wavelet transform is applied to the motion sensing signal extracted in step 2.1; let the mother wave be , which is decomposed into approximate components containing low-frequency information and detail components containing high-frequency information , , according to the frequency distribution range, where is the order set in the stationary wavelet transform process, and the value range is [2, 10]. Second, calculate the short-time energy of each detail component , and detect all local maximum amplitudes in the short-time energy , , and mark the time corresponding to the local maximum amplitude.​​ Define threshold , and It's about the details. short-term energy All local maximum amplitudes The mean and variance; compare each Amplitude and threshold ,if Then classify Vibrations of the body surface related to heartbeat; Using the time interval between the occurrence of two adjacent local maximum amplitudes, further processing other unclassified ; define two adjacent local maximum amplitudes and The time interval between the occurrence of ; set the threshold value to The average of the time interval between the occurrence of all adjacent local maximum values in If and , classify as body surface vibration related to heartbeat; Finally, the local maximum amplitude of body surface vibrations classified as heartbeat-related was utilized. Separation Surface vibrations and noises related to heartbeat; examine the maximum amplitude of each surface vibration related to heartbeat. The occurrence time of the heartbeat is determined by searching for the occurrence time of local minima forward and backward, identifying the start and end times of surface vibrations related to the heartbeat, and preserving the occurrence time of the local minimum. The corresponding data is used to identify other data as noise. The corresponding data is replaced with 0; an inverse stationary wavelet transform is performed to obtain the body surface vibration signal related to the heartbeat.

5. The continuous electrocardiograph waveform reconstruction system using a smart wristband motion sensor of claim 4, wherein, Selecting the value 6.

6. The continuous electrocardiograph waveform reconstruction system using a smart wristband motion sensor of claim 1, wherein, In step 3.1, all local maximum and minimum points in the heartbeat-related body surface vibration signal are searched, a local triangle is constructed using each local maximum point and the two adjacent local minimum points on the left and right of the local maximum point, and features are extracted from the local triangle; The local maximum point in the local triangle is defined as the vertex, the adjacent left local minimum point is defined as the left bottom point, the adjacent right local minimum point is defined as the right bottom point, the sum of the distances a+b from the vertex to the left and right bottom points, the difference d between the vertical coordinates of the vertex and the left bottom point, the difference e between the vertical coordinates of the vertex and the right bottom point, and the angle f of the vertex in the local triangle are selected as features; In addition, the difference c between the horizontal coordinates of the left and right bottom points, the angle g of the left bottom point in the local triangle, and the angle h of the right bottom point in the local triangle are selected as supplementary features to assist in identifying the sharp peak point. A random forest classifier is trained to identify whether the vertex, i.e., the local maximum point, in the local triangle is a sharp peak point related to ventricular contraction; during identification, the selected features are input into the random forest classifier to predict the probability that the vertex belongs to a sharp peak point, and if the probability is greater than a set value, the vertex is a sharp peak point related to ventricular contraction.

7. The continuous electrocardiograph waveform reconstruction system using a smart wristband motion sensor of claim 6, wherein, In step 3.1, the difference c between the horizontal coordinates of the left and right bottom points, the angle g of the left bottom point in the local triangle, and the angle h of the right bottom point in the local triangle are selected as supplementary features to assist in identifying the sharp peak point.

8. The continuous electrocardiograph waveform reconstruction system using a smart wristband motion sensor of claim 1, wherein, In step 3.2, the 250 milliseconds before the sharp peak point is determined as the start point of a heart cycle, and the 250 milliseconds before the next sharp peak point is determined as the end point of the heart cycle.

Images