Adaptive elimination method of motion artifact of electrocardiosignal and flexible electrocardio electrode device

Abstract

The application discloses a kind of electrocardiosignal motion artifact adaptive elimination method and flexible electrocardio electrode device, belong to biomedical signal processing and wearable medical equipment technical field.For the problem that all signal frames are processed by unified processing strategy in existing method, the noise reduction degree and waveform fidelity cannot be dynamically balanced, the application proposes adaptive hierarchical processing architecture: three characteristic indexes of baseline drift amplitude, electromyographic interference energy proportion and R wave detection confidence are extracted for each signal frame, and the signal frame is determined as mild, moderate or severe artifact frame;Mild frame uses linear filter denoising, moderate frame inputs lightweight convolutional coding and decoding network for end-to-end denoising, and severe frame is superimposed on the basis of network denoising adaptive filter correction based on acceleration reference signal.In 8km / h jogging scene, R wave detection accuracy can reach 88.2%, compared with about 12 percentage points of pure U-Net method, and mild frame processing speed is improved by about 5 times.

Description

Technical Field

[0001] This invention belongs to the field of biomedical signal processing and wearable medical device technology, specifically relating to an adaptive method for eliminating motion artifacts in dynamic electrocardiogram (ECG) monitoring scenarios, and a flexible ECG electrode device used in conjunction with this method. Background Technology

[0002] Ambulatory electrocardiogram (ECG) monitoring (Holter monitoring, exercise stress testing, wearable ECG patches, etc.) is an important tool for cardiovascular disease screening and chronic disease management. It can continuously record ECG signals while the subject is in a natural activity state, capturing paroxysmal arrhythmias, transient ST segment changes, and exercise-induced myocardial ischemia events that are difficult to detect on resting ECG. However, during daily activities or exercise, various physical factors combine to create motion artifacts, which seriously interfere with the quality of ECG signals.

[0003] The mechanisms of motion artifacts mainly fall into three categories: First, relative displacement between the skin and electrodes. During human movement, the skin surface deforms and shifts, causing disturbances in the electric double layer structure at the electrode-skin interface. This generates low-frequency potential fluctuations related to the movement frequency, with the spectrum mainly concentrated in the 0.1Hz~10Hz range, highly overlapping with the baseline frequency band of the electrocardiogram (ECG) signal, manifesting as baseline drift. Second, electromyographic crosstalk introduced by muscle contraction. Active contraction of skeletal muscles during exercise generates electromyographic signals, with a spectrum ranging from 20Hz to 500Hz. This significantly overlaps with the QRS complex frequency band (5Hz~40Hz) of the ECG signal. During high-intensity exercise, such as running or climbing stairs, the energy of electromyographic interference can exceed the energy of the ECG signal by several times, causing the QRS complex to be submerged in noise. Third, baseline drift caused by respiration. Respiratory movements cause the thoracic cavity to expand and contract periodically, changing the skin tension at the electrode attachment site and generating a low-frequency drift of approximately 0.15 Hz to 0.5 Hz. This drift overlaps with the low-frequency band of heart rate variability (HRV) analysis, interfering with the accurate measurement of the ST segment baseline.

[0004] The three factors mentioned above superimpose during exercise, forming a broadband, non-stationary composite noise that severely interferes with the quality of electrocardiogram (ECG) signals. This leads to a significant decrease in the discernibility of key waveforms such as the P wave and ST segment, and may even cause missed or false detections of the R wave, resulting in clinical misdiagnosis. Studies have shown that under 8km / h jogging conditions, the accuracy of R wave detection in undenoised raw ECG signals can be as low as below 40%, and the measurement error of ST segment shift can exceed 0.2mV, far exceeding the acceptable range for clinical diagnosis. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive method for eliminating motion artifacts in electrocardiogram (ECG) signals and a flexible ECG electrode device to solve the following technical problems existing in the prior art: Existing denoising methods use a uniform processing strategy for all signal frames, which cannot adaptively switch the processing strategy according to the actual intensity of motion artifacts. This results in over-filtering in the low artifact segment, causing distortion of subtle waveforms such as P waves and T waves, or insufficient denoising in the high artifact segment, leading to missed detection of R waves. The invention fails to achieve a dynamic balance between denoising strength and waveform fidelity.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] This invention provides an adaptive method for eliminating motion artifacts in electrocardiogram (ECG) signals, comprising the following steps: S1, continuously acquiring ECG signals from the body surface using flexible electrodes attached to the body surface, and simultaneously acquiring motion reference signals using a triaxial accelerometer integrated within the electrode module. The sampling frequency of the ECG signals is 100Hz~350Hz, preferably 250Hz; S2, dividing the acquired raw ECG signals into signal frames according to a time window. The time window length of the signal frame is 2~3 seconds, preferably 2 seconds. Artifact intensity features are extracted from each signal frame. The artifact intensity features include three indicators: baseline drift amplitude, electromyographic interference energy ratio, and R-wave detection confidence. The baseline drift amplitude is taken as the maximum absolute deviation between the signal frame after high-pass filtering at a cutoff frequency of 0.5Hz and the cubic spline fitted baseline. The electromyographic interference energy ratio is obtained by performing a fast Fourier transform on the signal frame and taking the frequency band of 20Hz~500Hz and 0.0... The ratio of the sum of squares of the amplitude spectrum within the full frequency band of 5Hz~150Hz is calculated. The R-wave detection confidence is obtained by detecting each candidate R-wave position using the Pan-Tompkins algorithm, and then taking the average of the ratio of the signal peak value at each candidate position to the root mean square of the signal within the frame. S3: The three feature indicators are compared with preset thresholds respectively, and the current signal frame is determined as a mild artifact frame, a moderate artifact frame, or a severe artifact frame according to the classification rules. S4: According to the determination result of step S3, the microcontroller integrated in the electrode module performs the corresponding denoising process. Linear filtering is performed to denoise the mild artifact frame, the moderate artifact frame is input into the trained deep neural network for denoising, and the severe artifact frame is superimposed with an adaptive filtering correction based on the acceleration reference signal after denoising by the deep neural network. S5: The processed signal frames are spliced ​​in chronological order, and the splicing artifacts in the inter-frame overlapping area are eliminated by weighted averaging using the Hanning window, and a continuous ECG signal is output.

[0008] This invention employs an adaptive architecture of "grading first, then processing," classifying signal frames into three levels—mild, moderate, and severe—based on real-time extracted artifact intensity features. Different processing strategies are applied to each level: mild artifact frames undergo linear filtering denoising using a 4th-order Butterworth filter, resulting in fast processing speed without introducing waveform distortion; moderate artifact frames are input to a trained encoder-decoder deep neural network for end-to-end denoising. This network uses a 3-layer encoder and 3-layer decoder structure, with convolutional kernel sizes of 15, 9, and 5, and channel numbers of 8, 16, and 16, respectively. Each layer includes batch normalization and ReLU activation, with a downsampling factor of 2. Skip connections are provided between corresponding layers of the encoder and decoder, and the skip connections adopt a channel-dimensional splicing method. Heavy artifact frames are superimposed with recursive least squares filtering correction based on acceleration reference signals after denoising by deep neural networks. The envelope of the three-axis signals of the accelerometer is used as the reference input, and the output signal of the deep network is used as the main input signal. The envelope of the three-axis signals is calculated by summing the absolute values ​​of the three-axis acceleration signals respectively. The filter order is 32, and the forgetting factor is 0.97~0.99.

[0009] Furthermore, the grading rule in step S3 is as follows: when the proportion of EMG interference energy is less than the first threshold, the baseline drift amplitude is less than the second threshold, and the R-wave detection confidence is greater than the third threshold, it is determined to be a mild artifact frame; when the proportion of EMG interference energy is greater than or equal to the first threshold and less than the fourth threshold, and the R-wave detection confidence is greater than or equal to the sixth threshold, or the baseline drift amplitude is greater than or equal to the second threshold and less than the fifth threshold, and the R-wave detection confidence is greater than or equal to the sixth threshold, it is determined to be a moderate artifact frame; when the proportion of EMG interference energy is greater than or equal to the fourth threshold, or the baseline drift amplitude is greater than or equal to the fifth threshold, or the R-wave detection confidence is less than the sixth threshold, it is determined to be a severe artifact frame. This grading rule comprehensively considers the interrelationship of the three artifact characteristic indicators and can accurately distinguish motion artifacts of different intensities.

[0010] Preferably, the first threshold is 0.20, the second threshold is 0.10mV, the third threshold is 3.5, the fourth threshold is 0.42, the fifth threshold is 0.32mV, and the sixth threshold is 2.0. This set of thresholds, calibrated using a grid search on the labeled dataset, achieves a classification accuracy of 89.4% on the labeled set, effectively distinguishing between light, moderate, and heavy artifact frames.

[0011] Furthermore, in step S4, the linear filtering denoising for frames with mild artifacts employs a high-pass filter with a cutoff frequency of 0.5Hz to remove baseline drift and a low-pass filter with a cutoff frequency of 45Hz to remove high-frequency electromyographic noise. Both filters are fourth-order Butterworth filters. This linear filtering method has a fast processing speed, with a single frame processing time not exceeding 8ms, and does not cause distortion to the ECG waveform, making it suitable for signal frames with low artifact intensity.

[0012] Furthermore, the total number of parameters in the deep neural network described in step S4 does not exceed 150K, and after 8-bit integer quantization compression, the number of parameters is approximately 112K. The deep neural network is executed by a microcontroller, and the single-frame inference time does not exceed 50ms. Through lightweight design and model quantization, this deep neural network can run in real time on a microcontroller with an ARM Cortex-M4 core (64MHz clock speed), meeting the real-time requirements of edge deployment.

[0013] Furthermore, the deep neural network is trained as follows: the training dataset consists of two parts. The first part selects records free of severe noise from the MIT-BIH arrhythmia database as clean signal sources, and superimposes motion artifacts, electromyographic interference, and baseline drift records from the MIT-BIH noise stress test database onto the clean signals at different signal-to-noise ratios to generate training samples. The second part uses actual measured data from subjects wearing flexible electrodes in three states: sitting, walking, and jogging, with synchronously acquired signals from a medical 12-lead electrocardiograph as labels. The training loss function is a weighted combination of mean squared error loss and waveform morphology loss, where waveform morphology loss is defined as the negative value of the Pearson correlation coefficient between the processed signal and the labeled signal within the QRS interval. Moderate artifact samples are assigned a mean squared error loss weight of 0.6 and a waveform morphology loss weight of 0.4, while severe artifact samples are assigned a mean squared error loss weight of 0.8 and a waveform morphology loss weight of 0.2. This training strategy enables the network to maintain the key features of the electrocardiogram waveform while denoising.

[0014] Furthermore, the signal frames described in step S2 are segmented by overlapping adjacent windows by 0.5 seconds, i.e., a step size of 1.5 seconds, with each frame corresponding to 500 sampling points. This overlapping segmentation method can effectively eliminate splicing artifacts through Hanning window weighted averaging during the splicing process in step S5.

[0015] This invention also provides a flexible electrocardiogram (ECG) electrode device for collecting surface ECG signals in conjunction with the above-described method. The device comprises the following layers stacked sequentially from the skin contact surface outwards: a medical pressure-sensitive adhesive layer with a thickness of 50-80 μm, having perforated windows corresponding to electrode positions; a porous conductive gel layer, using polyacrylamide-sodium alginate double-network hydrogel as a matrix, containing 0.8-1.5 wt% sodium chloride, 6-8 wt% glycerol, and 0.08-0.15 wt% N,N'-methylenebisacrylamide, with through-pores of 40-180 μm in diameter distributed within the gel layer, a porosity of 25%-45%, and a thickness of 300-500 μm. The porous conductive gel layer is connected to the medical pressure-sensitive adhesive layer... A pressure-sensitive adhesive layer is fixed to the skin surface; a silver-silver chloride composite conductive layer is formed on the upper surface of a flexible support substrate by screen printing, with a thickness of 8~15μm and a sheet resistance ≤15Ω / sq; the flexible support substrate is made of thermoplastic polyurethane, with a thickness of 120~200μm, and has through-hole breathable micropores; a flexible printed circuit layer integrates an ECG analog front-end chip, a triaxial accelerometer, a microcontroller, and a communication module, and is electrically connected to the silver-silver chloride composite conductive layer through conductive silver paste vias; a waterproof and breathable layer is an expanded polytetrafluoroethylene film with a thickness of 30~60μm and a moisture permeability ≥8000g / m² / 24h, covering the flexible printed circuit layer.

[0016] This flexible ECG electrode device employs a layered structure design to reduce skin-electrode contact resistance and its movement fluctuations from the source. The porous conductive gel layer utilizes a polyacrylamide-sodium alginate dual-network hydrogel system. The polyacrylamide first network provides mechanical strength, while the sodium alginate second network provides ionic conductivity. Glycerin acts as a moisturizer to maintain the gel's high water content and ionic conductivity. The perforated channels within the gel layer enhance breathability and allow sweat to escape, preventing skin discomfort caused by prolonged wear. The silver-silver chloride composite conductive layer is formed using a screen printing process. Flake-shaped silver powder in the conductive paste provides electronic conduction pathways, while silver chloride powder forms a stable electrochemical interface. The combined use of these two materials achieves low sheet resistance and a stable electrode potential.

[0017] Furthermore, the porous conductive gel layer is prepared by template dissolution using sodium chloride crystal particles as the pore-forming template, with a pore size of 60-150 μm. By controlling the pore size of the pore-forming agent, the pore size and porosity of the gel layer can be precisely controlled. A preferred particle size range of 80-120 μm yields a gel with a porosity of approximately 32%, balancing air permeability and mechanical integrity.

[0018] Furthermore, the conductive paste used in the silver-silver chloride composite conductive layer comprises, by weight: 60-70 parts of flake silver powder, 5-10 parts of silver chloride powder, 5-8 parts of ethyl cellulose, 15-20 parts of terpineol, and 2-4 parts of dibutyl phthalate. It is used after being milled by a three-roll mill until the fineness measured by a scraper fineness meter is ≤15μm. The average particle size of the flake silver powder is 2-5μm, the average particle size of the silver chloride powder is 1-3μm, ethyl cellulose is used as a film-forming agent, terpineol as a solvent, and dibutyl phthalate as a plasticizer. After mixing and grinding the components in the specified proportions, a conductive paste with good rheological properties is obtained, suitable for screen printing processes.

[0019] Preferably, the preparation process of the porous conductive gel layer is as follows: acrylamide monomer, sodium alginate, sodium chloride, glycerol, and N,N'-methylenebisacrylamide are dissolved in deionized water in a certain proportion, ammonium persulfate is added as an initiator and tetramethylethylenediamine as an accelerator, the prepolymer solution is injected into a mold pre-arranged with sodium chloride crystal particles, cured at 50~60℃ for 1~2 hours, and then soaked in deionized water for 12~24 hours to dissolve the sodium chloride template.

[0020] Preferably, the silver-silver chloride composite conductive layer is screen-printed using a 200-mesh polyester screen, with a printing pressure of 1.2 kg / cm², a squeegee angle of 65°, and a printing speed of 80 mm / s. After printing, it is sintered in an oven at 120°C for 25 minutes. These process parameters result in a conductive layer with uniform thickness and good adhesion. The measured average thickness of the conductive layer is 11.2 μm, the sheet resistance measured by the four-probe method is 9.8~13.5 Ω / sq, and the cross-cut adhesion test rating is 4B.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] 1. By extracting three feature indicators—baseline drift amplitude, EMG interference energy ratio, and R-wave detection confidence—signal frames are classified into mild, moderate, and severe artifact frames. Different processing strategies are applied to different levels, achieving the technical effect of adaptively switching processing methods based on the actual intensity of motion artifacts. In an 8km / h jogging scenario, the method of this invention achieves an output signal-to-noise ratio of 13.6±1.8dB and an R-wave detection accuracy of 88.2%, which are approximately 48% and 12 percentage points higher than the pure U-Net method (output signal-to-noise ratio 9.2±2.0dB, R-wave accuracy 75.9%), respectively. The improvement is even more significant compared to the adaptive filtering method (output signal-to-noise ratio 5.4±2.6dB, R-wave accuracy 52.8%), demonstrating the effectiveness of the adaptive hierarchical processing strategy.

[0023] 2. For frames with mild artifacts, a linear filter is used to process them, skipping the deep neural network. This improves the processing speed by about 5 times compared to deep network methods, with a single frame processing time of no more than 8ms. At the same time, it avoids the slight waveform distortion that may be introduced by deep networks. In a seated scenario, the P wave and T wave morphology fidelity score reaches 1.9 points (out of 2), showing good consistency with the reference electrocardiogram signal. This achieves zero waveform damage denoising in low artifact scenarios.

[0024] 3. For heavily artifact-laden frames, a recursive least-squares adaptive filtering correction based on acceleration reference signals is superimposed on the deep neural network denoising. This effectively "cancels" artifact components synchronized with the motion rhythm by utilizing the motion pattern information contained in the acceleration signals. Compared with the scheme without the graded strategy (SNR 10.5±1.9dB and R-wave accuracy 79.6% in jogging scenes), the method of this invention improves the SNR of jogging scenes by about 30% and the R-wave accuracy by about 11 percentage points, demonstrating the necessity of superimposing adaptive filtering correction on heavily artifact-laden frames.

[0025] 4. The matching flexible ECG electrode device adopts a porous double-network conductive gel layer. The gel layer contains through-channels with a pore size of 40-180 μm and a porosity of 25%-45%. Compared with traditional non-porous gel electrodes, the porous structure enhances breathability and allows sweat to escape. In 10 volunteers who wore the electrode continuously for 72 hours, 5 / 5 experienced no redness, itching, or rash, demonstrating good biocompatibility and wearing comfort. The measured average contact impedance of the electrode of this invention is 9.2±2.1 kΩ, lower than the 11.8±3.4 kΩ of the commercially available 3M Red Dot 2560 electrode. After walking at 4 km / h for 5 minutes, the impedance fluctuation is 6.3±2.8%, only about 28% of that of commercially available electrodes (22.5±7.1%), reducing contact impedance and its movement fluctuations from the source.

[0026] 5. The deep neural network adopts a lightweight design, with three layers each for the encoder and decoder. The convolutional kernel size is gradually reduced from 15 to 5, and the total number of network parameters does not exceed 150K. After 8-bit integer quantization, the number of parameters is about 112K. It can run in real time on a microcontroller with an ARM Cortex-M4 core (64MHz main frequency). The single-frame inference time is 38±6ms, which meets the real-time requirements of edge deployment. There is no need to upload data to the cloud for processing, which reduces system latency and communication power consumption, while protecting the privacy of users' physiological data. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0028] Figure 2 This is a schematic cross-sectional view of the layered structure of a flexible electrocardiogram electrode device.

[0029] Figure 3 This is a schematic diagram of the three-dimensional scatter point distribution for pseudo-intensity grading.

[0030] Figure 4 This is a bar chart comparing the output signal-to-noise ratio and R-wave detection accuracy of the four methods in Example 2 under different motion states. Detailed Implementation

[0031] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention. In the following embodiments, unless otherwise specified, the materials involved are all commercially available industrial-grade or medical-grade products, and the electronic components involved are all standard packaged devices.

[0032] Example 1

[0033] Fabrication and performance testing of flexible ECG electrodes

[0034] This embodiment details the material formulation, preparation process, and measured results of key performance indicators for each layer of the flexible ECG electrode.

[0035] 1. Formulation screening and preparation of porous conductive gel layers

[0036] The system is based on a polyacrylamide-sodium alginate dual-network hydrogel. A typical formulation is as follows: 2.5g acrylamide, 0.3g sodium alginate, 0.2g sodium chloride, 1.2g glycerol, 0.018g N,N'-methylenebisacrylamide, and 14mL deionized water. To prepare the solution, the above components are stirred at room temperature for 40 minutes on a magnetic stirrer until completely dissolved. Then, 80μL of 10wt% ammonium persulfate aqueous solution is added as an initiator, and 6μL of tetramethylethylenediamine is added as an accelerator. After rapid and homogenous mixing, the solution is poured into a polytetrafluoroethylene mold with an inner cavity of 30mm × 30mm × 0.4mm.

[0037] The pore-forming process employed a template dissolution method. Before injecting the prepolymer solution, a layer of analytical grade sodium chloride crystal particles was evenly spread at the bottom of the mold. After the prepolymer solution was injected, the mold was placed in a 55°C oven for curing for 90 minutes. After curing, the gel was removed from the mold and immersed in 100 mL of deionized water, changing the water every 6 hours for a total of 18 hours to fully dissolve the sodium chloride template. After removal, the surface moisture was absorbed with lint-free paper, and the gel was sealed and stored for later use. The resulting gel had a thickness of 380±45 μm, a porosity of 32±6%, and an equilibrium water content of 78±5%.

[0038] The skin contact impedance of the gel at 10 Hz was tested using a two-electrode method with an electrochemical workstation. Five healthy volunteers (3 males and 2 females, aged 23-41 years) were tested. The electrodes were applied to the inner side of the left forearm, and the impedance was measured after 5 minutes of inactivity to allow it to stabilize. The average of the five measurements was taken for each formulation. The test results showed that the average contact impedance of the electrode of this invention was 9.2 ± 2.1 kΩ, compared to 11.8 ± 3.4 kΩ for the commercially available 3M Red Dot 2560 electrode.

[0039] Regarding impedance fluctuations after volunteers walked at 4 km / h for 5 minutes, the impedance change of the electrode of this invention was 6.3 ± 2.8%, while that of the control electrode was 22.5 ± 7.1%. This indicates that the porous conductive gel layer effectively reduces contact impedance and its motion fluctuations.

[0040] 2. Preparation of silver-silver chloride composite conductive layer

[0041] The conductive paste formulation is as follows: 65 parts by weight of flake silver powder (average particle size 3.2 μm, purity ≥99.9%), 8 parts by weight of silver chloride powder (average particle size 2.0 μm, purity ≥99.5%), 7 parts by weight of ethyl cellulose (viscosity 45 mPa·s), 18 parts by weight of terpineol, and 3 parts by weight of dibutyl phthalate. Each component is ground three times on a three-roll mill to a fineness of 12 μm (measured by a squeegee fineness meter). A 200-mesh polyester screen is used to screen print conductive patterns onto a TPU substrate (150 μm thick, purchased from Bayer, Germany, brand name Desmopan 385E). The printing pressure is 1.2 kg / cm², the squeegee angle is 65°, and the printing speed is 80 mm / s.

[0042] After printing, the material was sintered in an oven at 120℃ for 25 minutes and then naturally cooled to room temperature. The measured average thickness of the conductive layer was 11.2μm, the sheet resistance was 9.8~13.5Ω / sq measured by the four-probe method, the adhesion was rated as 4B according to the cross-cut adhesion test, and the measured peeling area was ≤5%.

[0043] 3. Overall electrode assembly and comprehensive performance testing

[0044] according to Figure 2 The layer sequence shown is completed by stacking and assembling each layer, with the top layer covering an ePTFE waterproof and breathable membrane (purchased from Gore, USA, 45μm thick, with a measured moisture permeability of 9200g / m² / 24h). After assembly, the overall thickness is approximately 0.8~0.9mm, and the weight of a single electrode is approximately 1.6g. The following tests were performed on the assembled electrode:

[0045] (1) Tolerance to continuous wear: Five volunteers wore the device continuously for 72 hours, and the skin condition at the application site was checked daily during the period. After 72 hours, the proportion of no redness, swelling, itching or rash was 5 / 5, and only 1 case reported a slight itching sensation at the edge of the application, which subsided spontaneously within 2 hours after the electrode was removed.

[0046] (2) Signal acquisition quality: The electrodes were connected to the ADS1292R ECG simulation front end (Texas Instruments) and ECG signals were acquired at a sampling rate of 250Hz under sitting conditions. The signals were compared with those acquired simultaneously by a medical 12-lead ECG machine. The QRS wave detection consistency rate was 99.6%, the P wave and T wave morphology were identifiable, and the ST segment offset deviation was ≤0.05mV. The electrodes prepared in this embodiment can meet the basic requirements for clinical-grade ECG signal acquisition.

[0047] Example 2

[0048] Experimental verification of the adaptive motion artifact elimination method

[0049] This embodiment fully demonstrates the processing flow and performance verification results of the method of the present invention on actual data. The comparison method selected adaptive filtering with accelerometer signal as reference input and sym8 wavelet thresholding with 5-layer decomposition, and U-Net-based deep learning method with 3-layer encoder-decoder and no artifact classification mechanism, representing three representative denoising schemes.

[0050] 1. Test data collection

[0051] Eight healthy volunteers (5 males and 3 females, aged 22-38 years) were selected. Each volunteer had a flexible electrode as described in Example 1 attached to their chest at the V2 position, and a commercial triaxial accelerometer was fixed next to the electrode. Simultaneously, medical Ag / AgCl electrodes were attached to the volunteers' torsos at the standard 12-lead positions, and a wireless module was connected to a Nihon Kohden ECG-1350 electrocardiograph as a reference tag signal. The data collection scenarios and sequence were as follows: Segment 1: Sitting still, keeping the upper body stationary for 5 minutes; Segment 2: Walking on flat ground in a corridor at a speed of approximately 3.5 km / h for 5 minutes; Segment 3: Jogging slowly on an outdoor track at a speed of approximately 8 km / h for 3 minutes; Segment 4: After resting in a seated position for 2 minutes, three consecutive standing-up-sitting movements were performed, holding the position for approximately 10 seconds after each standing up before sitting down, to induce baseline drift caused by postural changes.

[0052] All signals were synchronously acquired by the ADS1292R front-end chip at a sampling rate of 250Hz and transmitted in real time to the laptop's storage via Bluetooth. The dataset was named "FECG-MA-1" and had a total effective acquisition time of approximately 120 minutes.

[0053] 2. Signal Preprocessing

[0054] The raw ECG signal is first filtered by a 0.05Hz high-pass filter to remove the DC component, and then by a 60Hz notch filter to eliminate power frequency interference. The signal is then divided into frames with a window length of 2 seconds, a step size of 1.5 seconds, and adjacent windows overlapping by 0.5 seconds. Each frame corresponds to 500 sampling points.

[0055] 3. Threshold calibration for artifact intensity grading

[0056] Before the formal testing, 800 frames of signal were randomly selected from the FECG-MA-1 dataset, including approximately 500 frames from a sitting period and approximately 300 frames from a jogging period. Two trained electrocardiogram technicians independently and manually labeled each frame as mild, moderate, or severe artifacts. 744 frames had consistent labeling (93.0% consistency rate), and the inconsistent frames were discussed and a consensus was reached. The labeled data was used to calibrate the classification thresholds, employing a grid search method to maximize classification accuracy.

[0057] The final calibration results are as follows: the first threshold (upper limit of mild EMG interference energy percentage) is 0.20, the second threshold (upper limit of mild baseline drift amplitude) is 0.10mV, the third threshold (lower limit of mild R-wave detection confidence) is 3.5, the fourth threshold (upper limit of moderate EMG interference energy percentage) is 0.42, the fifth threshold (upper limit of moderate baseline drift amplitude) is 0.32mV, and the sixth threshold (lower limit of moderate R-wave detection confidence) is 2.0. Under this set of thresholds, the classification accuracy on the labeled set is 89.4%. The main type of misclassification is that mild frames are classified as moderate frames (accounting for approximately 60% of misclassifications). In practical applications, this means that a stronger denoising strategy was used for some signal frames that could have been easily processed; this is a conservative error and will not cause loss of waveform information.

[0058] 4. Training of Deep Neural Networks

[0059] The core design difference between the deep neural network in this embodiment and the classic U-Net is that the encoder and decoder each have 3 layers instead of the usual 4-5 layers, and the convolutional kernel size is gradually reduced from 15 to 5. This design aims to compress the network depth while maintaining a sufficient receptive field. Figure 4 As shown, the encoder has a kernel size of 15 and 8 channels in the first layer, a kernel size of 9 and 16 channels in the second layer, and a kernel size of 5 and 16 channels in the third layer. Each layer includes batch normalization and ReLU activation, and the downsampling factor is 2. The decoder contains 3 corresponding transposed convolution upsampling modules. There are skip connections between the corresponding layers of the encoder and decoder. The skip connection method is to splice along the channel dimension.

[0060] The training set consists of two parts. The first part consists of 20 noise-free records (record numbers 100, 101, 103, 105, 106, 108, 112, 113, 114, 115, 116, 117, 119, 121, 122, 123, 200, 202, 205, 213) selected from the MIT-BIH arrhythmia database. Synthetic data with different signal-to-noise ratios (-8dB to 10dB, with a step size of approximately 3dB) are generated using a noise superposition method, resulting in approximately 12,000 training sample frames. The second part consists of all data collection segments (sitting + walking + jogging + postural changes) from 6 volunteers in the FECG-MA-1 dataset. Using reference electrocardiogram signals as labels, approximately 3,600 measured training sample frames are generated.

[0061] Training parameters: Adam optimizer, initial learning rate 1×10⁻⁶ -4 The model was trained for 25 epochs, with the weight decaying to 0.8 times the original value every 10 epochs. The batch size was 16. The loss function used was weighted mean squared error, where the pixel weights in the QRS intervals 60ms before and after the peak value of the reference label R-wave were set to three times that of the rest of the region. This achieved waveform fidelity without introducing an additional loss term. This simple weighting strategy proved to be more stable than an additional correlation coefficient loss term in practice. After model training, 8-bit integer quantization was performed. The quantized model had approximately 112K parameters. It was deployed to an nRF52840 microcontroller in TensorFlowLite for Microcontrollers format. In actual tests, the inference time for a single frame with 500 sampling points was 38±6ms.

[0062] 5. Noise Reduction Performance Comparison

[0063] The remaining 30 minutes of data from the two volunteers in the FECG-MA-1 dataset who did not participate in training were used as the test set. Four methods were applied: the method of this invention, adaptive filtering, wavelet threshold denoising, and U-Net denoising. Evaluation metrics included: SNR (Signal-to-Noise Ratio), defined as the SNR between the processed signal and the reference electrocardiogram signal, using the reference electrocardiogram signal as the standard signal, and defined as 10×log... 10 (Signal power / noise power); R-wave detection accuracy, i.e., the proportion of detected R-wave positions that deviate from the reference signal R-wave position within ±40ms; P-wave and T-wave morphological fidelity, evaluated in a single-blind manner by an attending cardiologist, and divided into three levels according to recognizability: clearly recognizable: 2 points, roughly recognizable: 1 point, unrecognizable: 0 points. The average score of all P-waves / T-waves is taken, and the value is the mean ± standard deviation of all frame test results of 2 subjects.

[0064] The performance of the four methods in different motion scenarios is shown in the table below:

[0065] Table 1 Comparison of test results under different motion scenarios

[0066] Notable data characteristics in the table above: First, in walking and jogging scenarios, the SNR improvement of the method in this invention compared to U-Net is significantly greater than that in the sitting scenario. This is because after artifact classification, targeted processing strategies are used for moderate and severe frames, while the pure U-Net method processes all frames in the same mode, resulting in excessive denoising at low-intensity artifacts. U-Net's P / T wave morphology score in the sitting scenario is 1.8, slightly lower than that of this invention, confirming this phenomenon. Second, in jogging scenarios, the standard deviation of R-wave accuracy for all methods increases significantly, reflecting the differences in movement patterns among different subjects. For example, different stride frequencies and landing impact forces have a significant impact on artifact intensity. Finally, the SNR standard deviation is the largest among all scenarios in the postural change scenario. This is because the speed of standing up and sitting down varies from person to person, and the amplitude and steepness of baseline drift vary considerably.

[0067] 6. Validation of the effectiveness of the tiered strategy

[0068] To further verify the contribution of the artifact classification strategy, the following comparative experiment was conducted: The deep network structure and parameters were kept identical, but the classification determination in step S3 was removed. That is, all frames uniformly followed the second processing strategy (deep network denoising) for moderate artifacts, and the adaptive correction of the superimposed acceleration reference signal for heavily affected frames was no longer applied. On the same jogging test set, the SNR after removing the classification decreased from 13.6±1.8dB to 10.5±1.9dB, and the R-wave accuracy decreased from 88.2% to 79.6%.

[0069] The reason for this is that during high-intensity exercise, relying solely on deep networks is insufficient to effectively separate large-amplitude electromyographic interference from QRS waves. In this case, the acceleration signal contains motion pattern information highly correlated with motion artifacts. Using this as a reference input for adaptive filtering can effectively "cancel out" the artifact component synchronized with the motion rhythm. For frames with mild artifacts, skipping deep networks and directly using classical filters not only reduces computational overhead and increases processing speed by approximately five times, but also avoids the subtle waveform distortions that deep networks might introduce. This result confirms the necessity of a hierarchical strategy in practical engineering.

[0070] Based on the above test results, the method of the present invention can automatically match a suitable denoising strategy according to the actual intensity of the artifact during the transition from resting to moving state, achieving a dynamic balance between denoising strength and waveform fidelity, which is superior to existing single-strategy denoising schemes.

[0071] The above description is merely a preferred embodiment of the present invention and does not limit the scope of patent protection of the present invention. Any equivalent structural or procedural modifications made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. An adaptive method for eliminating motion artifacts in electrocardiogram signals, characterized in that, Includes the following steps: S1. Electrocardiogram (ECG) signals are collected from the body surface using flexible electrodes attached to the body surface, and motion reference signals are collected simultaneously using a triaxial accelerometer integrated into the electrode module. The sampling frequency of the ECG signals is 100~350Hz. S2. Divide the original ECG signal into signal frames according to a time window, with the time window length of each signal frame being 2-3 seconds. Extract artifact intensity features for each signal frame. The artifact intensity features include: baseline drift amplitude, which is the maximum absolute deviation between the signal frame after high-pass filtering at a cutoff frequency of 0.5Hz and the baseline fitted by cubic splines; EMG interference energy ratio, which is the ratio of the sum of squares of the amplitude spectrum in the 20Hz-500Hz frequency band to the sum of the squares of the amplitude spectrum in the 0.05Hz-150Hz full frequency band after performing a fast Fourier transform on the signal frame; and R-wave detection confidence, which is the average of the ratio of the signal peak value at each candidate position to the root mean square value of the signal within the frame, obtained by detecting each candidate R-wave position using the Pan-Tompkins algorithm. S3. Compare the three feature indicators with the preset thresholds respectively, and determine the current signal frame as a mild artifact frame, a moderate artifact frame or a severe artifact frame according to the classification rules. S4. Based on the determination result of step S3, the microcontroller integrated in the electrode module performs the corresponding denoising process: linear filtering is performed to denoise the mild artifact frames, the medium artifact frames are input into the trained deep neural network for denoising, and the severe artifact frames are superimposed with adaptive filtering correction based on the acceleration reference signal after denoising by the deep neural network. S5. The processed signal frames are spliced ​​together in chronological order. The overlapping areas between frames are eliminated by weighted averaging using a Hanning window to remove splicing artifacts, and a continuous ECG signal is output.

2. The method according to claim 1, characterized in that, The grading rule mentioned in step S3 is as follows: When the proportion of electromyographic interference energy is less than the first threshold, the baseline drift amplitude is less than the second threshold, and the R-wave detection confidence is greater than the third threshold, it is judged as a mild artifact frame. A frame is classified as having moderate artifacts when one of the following conditions is met: (1) The proportion of electromyographic interference energy is greater than or equal to the first threshold and less than the fourth threshold, and the confidence level of R-wave detection is greater than or equal to the sixth threshold; (2) The baseline drift amplitude is greater than or equal to the second threshold and less than the fifth threshold, and the R-wave detection confidence is greater than or equal to the sixth threshold; A frame is classified as a severe artifact frame when the proportion of electromyographic interference energy is greater than or equal to the fourth threshold, the baseline drift amplitude is greater than or equal to the fifth threshold, or the R-wave detection confidence is less than the sixth threshold.

3. The method according to claim 2, characterized in that, The first threshold is 0.20, the second threshold is 0.10mV, the third threshold is 3.5, the fourth threshold is 0.42, the fifth threshold is 0.32mV, and the sixth threshold is 2.

0.

4. The method according to claim 1, characterized in that, The deep neural network described in step S4 adopts an encoder-decoder structure. The encoder contains three consecutive one-dimensional convolutional modules with kernel sizes of 15, 9, and 5, and channel numbers of 8, 16, and 16, respectively. Each layer includes batch normalization and ReLU activation, and the downsampling factor after each convolution is 2. The decoder contains three corresponding transposed convolutional upsampling modules. There are skip connections between corresponding layers of the encoder and decoder. The skip connections are constructed using a channel-dimensional concatenation method. The total number of network parameters does not exceed 150K. The deep neural network is executed by a microcontroller, and the single-frame inference time does not exceed 50ms.

5. The method according to claim 1, characterized in that, In step S4, the adaptive filtering correction of the heavily artifact frames adopts recursive least squares filtering, with the envelope of the three-axis signal of the accelerometer as the reference input and the output signal of the deep network as the main input signal. The envelope of the three-axis signal is calculated by summing the absolute values ​​of the three-axis acceleration signals respectively. The filter order is 32 and the forgetting factor is 0.97~0.

99.

6. A flexible electrocardiogram electrode device for acquiring surface electrocardiogram signals, characterized in that, It includes the following layers stacked sequentially from the skin-contact side outwards: The medical pressure-sensitive adhesive layer has a thickness of 50~80μm and has hollowed-out windows corresponding to the electrode positions. The porous conductive gel layer uses polyacrylamide-sodium alginate dual-network hydrogel as the matrix, containing 0.8~1.5wt% sodium chloride, 6~8wt% glycerol, and 0.08~0.15wt% N,N'-methylenebisacrylamide. The gel layer has through-pores with a pore size of 40~180μm, a porosity of 25%~45%, and a thickness of 300~500μm. The porous conductive gel layer is fixed to the skin surface by the medical pressure-sensitive adhesive layer. A silver-silver chloride composite conductive layer is formed on the upper surface of a flexible support substrate by screen printing, with a thickness of 8~15μm and a sheet resistance of ≤15Ω / sq. The flexible support substrate is made of thermoplastic polyurethane with a thickness of 120~200μm and has through-holes for ventilation. The flexible printed circuit layer integrates an electrocardiogram analog front-end chip, a triaxial accelerometer, a microcontroller, and a communication module, and is electrically connected to the silver-silver chloride composite conductive layer through conductive silver paste vias; The waterproof and breathable layer is an expanded polytetrafluoroethylene film with a thickness of 30~60μm and a moisture permeability of ≥8000g / m² / 24h, which covers the flexible printed circuit layer.

7. The flexible electrocardiogram electrode device according to claim 6, characterized in that, The porous conductive gel layer is prepared by template dissolution method using sodium chloride crystal particles as pore-forming templates, with a pore-forming agent particle size of 60~150μm.

8. The flexible electrocardiogram electrode device according to claim 6, characterized in that, The conductive paste used in the silver-silver chloride composite conductive layer comprises, by weight, 60-70 parts of flake silver powder, 5-10 parts of silver chloride powder, 5-8 parts of ethyl cellulose, 15-20 parts of terpineol, and 2-4 parts of dibutyl phthalate. It is used after being ground by three rollers until the fineness measured by a scraper fineness meter is ≤15μm.

Images