Intelligent supervision method for wearable electrocardiogram monitoring equipment

Abstract

The invention discloses an intelligent supervision method for wearable electrocardiogram monitoring equipment. The intelligent supervision method comprises the following steps that S1, multi-channel electrocardiogram signals of a monitored person are continuously collected through twelve-channel electrocardiogram monitoring equipment based on a flexible circuit; s2, performing fusion preprocessing on the multi-channel electrocardiosignal to obtain a high-quality fused electrocardiosignal; s3, on the basis of the fused electrocardiosignal and the preprocessed signals of part of channels, extracting a multi-dimensional electrocardiosignal feature vector with inter-channel relevance; s4, inputting the multi-dimensional electrocardio feature vector into a trained abnormal electrocardio event recognition model, and performing intelligent analysis and classification to recognize a specific abnormal electrocardio mode; and S5, according to the type and severity of the identified abnormal electrocardiogram mode, dynamically generating and executing a graded intelligent supervision response strategy. According to the invention, more intelligent supervision can be realized.

Description

Technical Field

[0001] This invention relates to the field of electrocardiogram (ECG) monitoring, and more specifically to an intelligent monitoring method for wearable ECG monitoring devices. Background Technology

[0002] With the increasing demand for cardiovascular health management, wearable electrocardiogram (ECG) monitoring devices are gradually becoming important tools for daily health monitoring and cardiovascular disease prevention and control. Compared with traditional fixed ECG monitoring devices in hospitals (such as 12-lead ECG machines), these devices achieve portable and lightweight designs by relying on flexible circuits and miniature sensors. They can continuously collect ECG signals by fitting snugly against the skin, breaking the limitation of short-term monitoring only in medical settings. They can meet the long-term follow-up monitoring needs of cardiovascular disease patients and provide early warnings of potential ECG abnormalities for healthy individuals, making them widely used in home health management and chronic disease prevention and control. However, with the upgrading of device functions, multi-channel acquisition has become the mainstream trend. Multi-channel signals are easily affected by human activity, environmental interference, and differences in device hardware, resulting in inconsistent quality of the raw ECG signals collected, which poses challenges for subsequent signal analysis and abnormality identification.

[0003] Against this backdrop, the regulatory needs for wearable ECG monitoring devices are becoming increasingly prominent. Regulation not only refers to real-time analysis of the data collected by the devices, but also needs to cover the entire process, including signal quality optimization, accurate identification of abnormal ECGs, and risk response. On the one hand, without an effective signal preprocessing mechanism, noise and baseline drift in single-channel signals will directly affect data reliability. Simple multi-channel averaging fusion cannot fully utilize high-quality channel information, which can easily lead to deviations in subsequent feature extraction. On the other hand, existing regulatory methods mostly rely on single-dimensional feature identification to identify anomalies, ignoring the correlation of electrical activity between different channels. Moreover, the parameters of the anomaly identification model are mostly fixed values, which cannot adapt to changes in ECG signals under different monitoring scenarios, and are prone to false alarms or missed alarms. At the same time, after long-term use, the consistency of data collection across channels is prone to decline due to hardware aging, and there is a lack of dynamic graded risk response strategies. This results in either over-alarming for low-risk anomalies or insufficient response to high-risk anomalies. The accuracy, adaptability, and stability of the existing regulatory system are difficult to meet the actual application needs of multi-channel wearable ECG monitoring devices. Therefore, an intelligent regulatory method for wearable ECG monitoring devices is proposed. Summary of the Invention

[0004] The present invention solves the above-mentioned technical problems through the following technical solution. The present invention includes the following steps: S1: Continuously acquire multi-channel electrocardiogram (ECG) signals of the monitored person using a twelve-channel ECG monitoring device based on flexible circuits; S2: Perform fusion preprocessing on the multi-channel ECG signals to obtain high-quality fused ECG signals; S3: Based on the fused ECG signal and the preprocessed signal of some channels, extract a multidimensional ECG feature vector with inter-channel correlation; S4: Input the multidimensional ECG feature vector into a trained abnormal ECG event recognition model for intelligent analysis and classification to identify specific abnormal ECG patterns. S5: Based on the type and severity of the identified abnormal ECG patterns, dynamically generate and execute a graded intelligent monitoring response strategy; S6: Periodically perform the equipment self-calibration process by applying standard test signals to the equipment and calculating calibration parameters based on the response signals of each channel to maintain the consistency of the twelve-channel signal acquisition.

[0005] Furthermore, the fusion preprocessing in step S2 includes: S21: Perform wavelet transform-based denoising processing on each single-channel ECG signal to obtain the preliminary purified signal for each channel; S22: Perform inter-channel signal quality assessment on the preliminary purified signals of each channel. The signal quality assessment results are obtained by weighting the signal-to-noise ratio and baseline drift of each channel signal. S23: Based on the signal quality assessment results, construct a channel weight set, wherein channels with high signal quality are assigned higher weights; S24: The preliminary purified signals from each channel are weighted and fused according to their corresponding channel weights to generate a fused ECG signal; the calculation process of the weighted fusion is expressed as follows: ; Where is the fused signal at time t. Let be the initial purification signal of the i-th channel at time t. Let be the weight of the i-th channel, and satisfy . .

[0006] Furthermore, the extraction process of the multidimensional ECG feature vector with inter-channel correlation extracted in step S3 includes: S31: Extract global time-domain features and frequency-domain features from the fused electrocardiogram signal; S32: Extract channel-specific features from the preliminary purification signals of at least two specific channels, the channel-specific features being used to reflect the electrical activity characteristics of different cardiac sites; S33: Calculate the mutual information value between the preliminary purified signals of the at least two specific channels to quantify the correlation between the channels. The mutual information value is obtained by calculating the relationship between the joint probability distribution of the two channel signals and their respective marginal probability distributions. The calculation formula is as follows: ; in, This represents the mutual information value between channel X and channel Y. Let p(x) and p(y) be the joint probability distribution, and p(x) and p(y) be the marginal probability distributions. S34: The global time-domain features, frequency-domain features, channel-specific features, and mutual information values ​​are combined to form the multidimensional ECG feature vector.

[0007] Furthermore, the abnormal ECG event recognition model in step S4 is an improved support vector machine model. The improvement lies in introducing the inter-channel mutual information value from the multi-dimensional ECG feature vector as a model parameter adjustment factor. The decision function used by this support vector machine model during classification is constructed based on statistical learning theory, and the expression of the decision function is: ; in, The input is a multidimensional ECG feature vector. Let be the feature vector of the i-th training sample, yi be the corresponding sample label, αi be the Lagrange multiplier, and b be the bias term. Let be the kernel function; wherein, the kernel function is a radial basis function, and its expression is: ,parameter The value is dynamically adjusted by the average value of the inter-channel mutual information values, and the adjustment relationship is as follows: ,in As the baseline parameter, It is the average value of all mutual information values ​​in the multidimensional electrocardiogram feature vector.

[0008] Furthermore, the dynamic generation of the hierarchical intelligent regulatory response strategy in step S5 includes: S51: Determine the type of abnormal event and the initial risk score based on the output of the abnormal ECG event recognition model; S52: The initial risk score is corrected by combining the real-time activity status of the monitored personnel to obtain a comprehensive risk score; the correction is achieved through an activity status compensation function, which utilizes activity intensity data from the inertial measurement unit; S53: Compare the comprehensive risk score with multiple preset threshold intervals to determine the corresponding alarm level; S54: Execute a response action corresponding to the alarm level, the response action including different combinations of data logging, local prompting, remote alarm and sending a distress signal containing abnormal ECG fragments and location information.

[0009] Furthermore, in step S52, the activity state compensation function is used to quantitatively assess the potential impact of physical activity on electrocardiogram signals and risk scores. Its core lies in performing linear compensation on the initial risk score based on activity intensity. Specifically, the comprehensive risk score is obtained through the following calculation process: ; Wherein, is the revised overall risk score. The initial risk score is determined in step S51, A is the normalized real-time activity intensity index obtained by analyzing and processing the data from the inertial measurement unit, and β is a pre-set compensation coefficient greater than zero, used to adjust the degree of influence of the activity status on the final risk assessment.

[0010] Furthermore, after executing the remote alarm in step S5, a feedback learning step is also included: S55: Receive feedback information from the monitoring center or medical staff regarding this alarm, including alarm confirmation, false alarm labeling, or severity correction; S56: Using the feedback information and the corresponding multidimensional ECG feature vector and real-time activity status data, the abnormal ECG event recognition model is incrementally updated to optimize the model's future classification performance.

[0011] Furthermore, the device self-calibration process in step S6 specifically includes: S61: Apply a standard test signal to the twelve-channel ECG monitoring device; S62: Acquire the response signals of each channel to the standard test signal; S63: Calculate the gain error and phase error between the response signal of each channel and the standard test signal; S64: Based on the gain error and phase error, generate a set of calibration parameters, and apply these parameters in subsequent signal acquisition to correct the original signal in real time to maintain the consistency of the twelve-channel signal.

[0012] Compared with existing technologies, this invention has the following advantages: The intelligent monitoring method for wearable ECG monitoring devices continuously acquires signals through a twelve-channel ECG monitoring device based on flexible circuits. It improves ECG signal quality through preprocessing methods such as wavelet transform denoising and channel quality weighted fusion. Then, it extracts multi-dimensional ECG feature vectors containing global time-domain / frequency-domain features, channel-specific features, and inter-channel mutual information values. An improved support vector machine model with dynamically adjusted parameters based on mutual information values ​​enhances the accuracy of abnormal ECG pattern recognition. Subsequently, it dynamically generates a graded monitoring response strategy based on the abnormality type, severity, and real-time activity status of the monitored individual, and optimizes model performance through feedback learning to achieve precise intelligent monitoring. Simultaneously, it periodically performs a device self-calibration process to calibrate the gain and phase error of each channel to maintain consistency in multi-channel signal acquisition and ensure the reliability of long-term monitoring data. Overall, this method improves the intelligence, precision, and stability of wearable ECG monitoring device monitoring, providing efficient and reliable assurance for ECG monitoring and making the system more worthy of widespread adoption. Attached Figure Description

[0013] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0014] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0015] like Figure 1 As shown, this embodiment provides a technical solution: an intelligent monitoring method for wearable electrocardiogram (ECG) monitoring devices, comprising the following steps: S1: Continuously acquire multi-channel electrocardiogram (ECG) signals of the monitored person using a twelve-channel ECG monitoring device based on flexible circuits; S2: Perform fusion preprocessing on the multi-channel ECG signals to obtain high-quality fused ECG signals; S3: Based on the fused ECG signal and the preprocessed signal of some channels, extract a multidimensional ECG feature vector with inter-channel correlation; S4: Input the multidimensional ECG feature vector into a trained abnormal ECG event recognition model for intelligent analysis and classification to identify specific abnormal ECG patterns. S5: Based on the type and severity of the identified abnormal ECG patterns, dynamically generate and execute a graded intelligent monitoring response strategy; S6: Periodically perform the equipment self-calibration process by applying standard test signals to the equipment and calculating calibration parameters based on the response signals of each channel to maintain the consistency of the twelve-channel signal acquisition.

[0016] Furthermore, the fusion preprocessing in step S2 includes: S21: Perform wavelet transform-based denoising processing on each single-channel ECG signal to obtain the preliminary purified signal for each channel; S22: Perform inter-channel signal quality assessment on the preliminary purified signals of each channel. The signal quality assessment results are obtained by weighting the signal-to-noise ratio and baseline drift of each channel signal. S23: Based on the signal quality assessment results, construct a channel weight set, wherein channels with high signal quality are assigned higher weights; S24: The preliminary purified signals from each channel are weighted and fused according to their corresponding channel weights to generate a fused ECG signal; the calculation process of the weighted fusion is expressed as follows: ; Where is the fused signal at time t. Let be the initial purification signal of the i-th channel at time t. Let be the weight of the i-th channel, and satisfy . ; By employing a process that progresses from single-channel denoising based on wavelet transform to channel quality assessment combining signal-to-noise ratio and baseline drift, to weighting based on quality, and finally to multi-channel weighted fusion, this approach can effectively filter out interference such as electromyography and power frequency signals in single-channel ECG signals to improve signal purity. Furthermore, it can quantitatively assess the effectiveness of each channel signal, prioritizing the retention of high-quality channel information while mitigating interference from low-quality channels. Ultimately, this generates a high-quality fused ECG signal, providing a reliable data foundation for subsequent multi-dimensional ECG feature extraction and abnormal ECG identification. This approach avoids analytical errors caused by single-channel signal defects and improves the accuracy and stability of the overall regulatory method.

[0017] When monitored individuals perform daily household chores (such as sweeping), the ECG signal from channel 4 of the wearable device is mixed with electromyographic interference (EMG) at frequencies of 15-40Hz due to arm muscle activity. The QRS complex in the original signal is covered by clutter, with an interference amplitude of 0.6mV. Denoising was performed using a three-level wavelet decomposition based on the db6 wavelet basis: the high-frequency coefficients (corresponding to EMG interference) after decomposition were processed using a soft thresholding method (threshold set at 0.2mV), while the low-frequency coefficients (corresponding to the valid ECG signal) were retained. A preliminary purified signal was then obtained through wavelet reconstruction. After processing, the EMG interference amplitude was reduced to below 0.08mV, and key waveforms such as the QRS complex (peak value 1.2mV) and P wave (peak value 0.3mV) were clearly distinguishable. The signal-to-noise ratio improved from 18dB before denoising to 32dB, meeting the signal purity requirements for subsequent feature extraction.

[0018] For example, when assessing the quality of the initial purified signals from 12 channels, the signal-to-noise ratio (SNR) is weighted at 0.6, and the baseline drift (BD) is weighted at 0.4 (the smaller the baseline drift, the higher the score; both are out of 1). The assessment formula is: ; in, (Preset ideal signal-to-noise ratio) (Preset maximum acceptable drift).

[0019] If channel 6 , Channel 11 , Substitute into the formula to calculate: ; ; The quantitative scores clearly show that the signal quality of channel 6 (0.834) is far superior to that of channel 11 (0.337), providing an objective basis for subsequent weight allocation.

[0020] Based on the above quality assessment scores, a channel weight set is constructed (satisfying) ), of which channel 6 is due to High was given (Higher than average weight) ), Channel 11 because Low given (Below average weight), the remaining channel weights are allocated according to the score as follows: w1=0.14, w2=0.15, w3=0.16, w4=0.12, w5=0.10, w7=0.09, w8=0.06, w9=0.05, w10=0.04, w12=0.03. At t=2.5s, the initial purified signal values ​​of each channel are: S1=0.9mV, S2=1.0mV, S3=1.2mV, S4=0.8mV, S5=0.95mV, S6=1.1mV, S7=0.85mV, S8=0.75mV, S9=0.7mV, S10=0.65mV, S11=0.6mV, S12=0.8mV. Substituting into the weighted fusion formula: ; The fused signal retains key ECG information from high-quality channels such as channels 2, 3, and 6, while reducing interference from low-quality channels such as channel 11. Compared to using only the channel 11 signal (0.6mV, with significant drift), the waveform stability of the fused signal is improved by 45%. Compared to the unweighted average signal ((0.9+1.0+…+0.8) / 12≈0.82mV), the identification of key ECG features (such as QRS peak value) of the fused signal is improved by 20%, laying the foundation for subsequent extraction of accurate multidimensional ECG features.

[0021] The extraction process of the multidimensional ECG feature vector with inter-channel correlation extracted in step S3 includes: S31: Extract global time-domain features and frequency-domain features from the fused electrocardiogram signal; S32: Extract channel-specific features from the preliminary purification signals of at least two specific channels, the channel-specific features being used to reflect the electrical activity characteristics of different cardiac sites; S33: Calculate the mutual information value between the preliminary purified signals of the at least two specific channels to quantify the correlation between the channels. The mutual information value is obtained by calculating the relationship between the joint probability distribution of the two channel signals and their respective marginal probability distributions. The calculation formula is as follows: ; in, This represents the mutual information value between channel X and channel Y. Let p(x) and p(y) be the joint probability distribution, and p(x) and p(y) be the marginal probability distributions. S34: The global time-domain features, frequency-domain features, channel-specific features, and mutual information values ​​are combined to form the multidimensional ECG feature vector; By extracting and fusing global time-domain / frequency-domain features, channel-specific features of specific channels, and mutual information values ​​between channels to form a multidimensional ECG feature vector, it is possible to capture the overall waveform patterns (such as heart rate and rhythm stability) and frequency distribution characteristics (such as the proportion of high and low frequency energy in ECG signals) from a global perspective. It can also reflect the differences in local electrical activity in different parts of the heart (such as atria and ventricles) through specific channel features, and quantify the signal correlation between channels through mutual information values ​​to reflect the conduction coordination of cardiac electrical activity. This avoids the limitations of single-dimensional features (such as only global features) in not being able to cover local abnormalities or channel correlation abnormalities, and provides more comprehensive and accurate feature inputs for subsequent abnormal ECG event recognition models, thereby improving the model's recognition accuracy and specificity for specific abnormal ECG patterns (such as myocardial ischemia and arrhythmias).

[0022] Taking the fused electrocardiogram signal of a middle-aged male subject at rest as an example, from the perspective of global time domain features, three core features were extracted: RR interval (the time interval between the peak values ​​of two adjacent QRS complexes), QRS width (the time from the start to the end of a QRS complex), and P wave amplitude (the maximum voltage value of the P wave). The average RR interval was calculated as follows: (Corresponding to a heart rate of approximately 65 beats per minute), average QRS wave width (Normal range <0.12s), average P-wave amplitude ; From the perspective of global frequency domain characteristics, a Fourier transform was performed on the fused ECG signal to divide it into a low-frequency band (0.04-0.15Hz, reflecting sympathetic nerve activity) and a high-frequency band (0.15-0.4Hz, reflecting vagal nerve activity), and the low-frequency power was calculated. High-frequency power Low frequency / high frequency ratio These characteristics can reflect the heart rate rhythm, cardiac conduction velocity, and autonomic nervous system regulation state of the monitored person at rest, providing a basis for identifying global abnormalities such as "arrhythmia" and "conduction block".

[0023] Based on the same monitored individual, two representative channels were selected: precordial lead V1 (corresponding to the right ventricular region, reflecting right ventricular electrical activity) and limb lead II (corresponding to the atrium-to-ventricle conduction path, reflecting atrioventricular conduction characteristics). Channel-specific features were extracted from the preliminary cleansing signals of the two channels: for lead V1, the S-wave depth (the minimum voltage value of the S-wave, reflecting right ventricular repolarization) was extracted and calculated. (Normal range -0.2 to -0.4 mV, indicating normal right ventricular repolarization); For lead II, the PR interval (the time from the onset of the P wave to the onset of the QRS complex, reflecting the atrioventricular conduction time) is extracted and calculated. (Normal range 0.12-0.20s, indicating normal atrioventricular conduction). If the S wave depth in lead V1 changes to 0.6mV (outside the normal range) during subsequent monitoring, combined with the specific characteristics of this channel, it can be preliminarily determined that there may be abnormal electrical activity in the right ventricle (such as a tendency for right ventricular hypertrophy). However, relying solely on global characteristics is insufficient to accurately locate such local abnormalities.

[0024] Taking the initial cleansing signals from lead V1 (channel X) and lead II (channel Y) as an example, calculate the mutual information value between the two channels. First, the two-channel signals are discretized, dividing the signal amplitude into 5 intervals: ([-0.5,-0.2),[-0.2,0.1),[0.1,0.4),[0.4,0.7),[0.7,1.0]), the joint probability distribution p(x,y) is obtained statistically: Some key values ​​are as follows: , , , , The remaining joint probabilities sum to 0.34, and the marginal probability distribution p(x): , , , , ), , , , , Substitute into the mutual information calculation formula: ; Substitute the above probability values ​​into the calculation (example of key terms expanded): ; ), and finally obtain .

[0025] This value indicates a strong correlation between the V1 and II lead signals, reflecting good coordination of electrical activity in the right ventricle and atrioventricular conduction pathway. If the monitored person experiences myocardial ischemia, the correlation between the two channels will weaken, and the mutual information value may drop below 0.3 bits. This feature can help identify abnormal patterns such as "local myocardial electrical activity conduction abnormalities," while single global or channel-specific features are difficult to capture such correlation abnormalities.

[0026] The abnormal ECG event recognition model in step S4 is an improved support vector machine (SVM) model. The improvement lies in introducing the inter-channel mutual information value from the multi-dimensional ECG feature vector as a model parameter adjustment factor. The decision function used by this SVM model during classification is constructed based on statistical learning theory, and the expression of the decision function is: ; in, The input is a multidimensional ECG feature vector. Let be the feature vector of the i-th training sample, yi be the corresponding sample label, αi be the Lagrange multiplier, and b be the bias term. Let be the kernel function; wherein, the kernel function is a radial basis function, and its expression is: ,parameter The value is dynamically adjusted by the average value of the inter-channel mutual information values, and the adjustment relationship is as follows: ,in As the baseline parameter, The average value of all mutual information values ​​in the multidimensional electrocardiogram feature vector; The abnormal ECG event recognition model is designed as an improved support vector machine (SVM) model that incorporates inter-channel mutual information values ​​as parameter adjustment factors. The parameter γ of the radial basis kernel function is dynamically adjusted by the average mutual information value, enabling the model to adaptively optimize the classification decision boundary based on the correlation between ECG signal channels. This retains the strong generalization ability of traditional SVM based on statistical learning theory, while solving the problem of poor adaptability of fixed-parameter SVM to differences in ECG signal channel correlation. At the same time, by combining multi-dimensional ECG feature vectors (including global, channel-specific, and mutual information features), the recognition accuracy of specific abnormal ECG patterns (such as myocardial ischemia and conduction block) that depend on abnormal channel correlation is significantly improved, reducing false alarms and false negatives caused by fixed parameters, and making the model classification more consistent with the actual physiological characteristics of ECG signals.

[0027] As in the scenario described above involving a middle-aged male subject, in a particular monitoring session, his multidimensional electrocardiogram feature vector contained mutual information values ​​between three channels: V1 and II leads, V2 and III leads, and aVR and aVL leads. (V1-II) (V2-III) (aVR-aVL).

[0028] First, calculate the average mutual information using the formula: .

[0029] Model preset baseline parameters According to the adjustment relationship The dynamic parameters were calculated. At this point, the kernel function has a strong adaptability to the classification boundary of normal electrocardiogram signals and can accurately distinguish between normal signals and obvious abnormal signals.

[0030] If the monitored individual subsequently develops mild myocardial ischemia, the coordination of electrical activity in different parts of the heart decreases, the correlation between channels weakens, and the three sets of mutual information values ​​change. , , Recalculate the average mutual information: Corresponding dynamic parameters .

[0031] Downregulating γ allows the kernel function sensitivity to adapt to changes in signal correlation, avoiding the loss of mild ischemic abnormalities due to parameter mismatch with signal characteristics when γ is fixed (e.g., always using 0.075).

[0032] Based on a mild myocardial ischemia scenario, five key support vector samples were taken after the model training was completed, including the feature vector of sample 1. (In order: global temporal domain feature value, channel-specific feature value, mutual information value), tags (1 indicates normal, -1 indicates abnormal), Lagrange multipliers ; Feature vector of sample 2 ,Label Lagrange multipliers ; Feature vector of sample 3 ,Label Lagrange multipliers ; Feature vector of sample 4 ,Label Lagrange multipliers ; Feature vector of sample 5 ,Label Lagrange multipliers And the model's preset bias term b = -0.2.

[0033] Current monitoring multidimensional ECG feature vector to be classified First, the radial basis function kernel is used to calculate the kernel function value between each support vector and the current vector. The kernel function formula is as follows: ,in (Dynamic parameters in the aforementioned ischemic scenario).

[0034] Taking sample 3 as an example, first calculate the vector norm: Substituting this into the kernel function, we get: .

[0035] Similarly, calculate the kernel function values ​​for other samples: , , .

[0036] Then substitute into the decision function formula First, calculate the summation term: After adding the bias term, we get -0.5 + (-0.2) = -0.7, and finally... The model successfully classified the current vector as an anomaly, accurately identifying mild myocardial ischemia. However, using a fixed γ=0.075 could lead to deviations in kernel function calculations, with the summation term potentially approaching 0, causing the decision function to misclassify the data as normal. This demonstrates the advantage of the improved SVM in enhancing anomaly recognition accuracy through dynamic parameter adaptation.

[0037] The dynamic generation of the hierarchical intelligent regulatory response strategy in step S5 includes: S51: Determine the type of abnormal event and the initial risk score based on the output of the abnormal ECG event recognition model; S52: The initial risk score is corrected by combining the real-time activity status of the monitored personnel to obtain a comprehensive risk score; the correction is achieved through an activity status compensation function, which utilizes activity intensity data from the inertial measurement unit; S53: Compare the comprehensive risk score with multiple preset threshold intervals to determine the corresponding alarm level; S54: Execute a response action corresponding to the alarm level, the response action including different combinations of data logging, local prompting, remote alarm and sending a distress signal containing abnormal ECG fragments and location information; By combining the determination of the abnormality type and initial risk classification with real-time activity status to correct the risk classification, comparing the alarm level with the threshold, and then executing the corresponding response action, the dynamic hierarchical supervision process solves the problem of assessing risk based solely on the abnormality type while ignoring the activity status (such as misjudging physiological ECG fluctuations caused by exercise as pathological abnormalities), thus achieving accuracy in risk assessment. Furthermore, by matching different risk levels with hierarchical response actions, it avoids over-supervision of low-risk cases (such as triggering alarms when no remote alarm is needed) or under-supervision of high-risk cases (such as only providing local prompts when emergency intervention is required). At the same time, the response actions include abnormal ECG segments and location information, providing key data for subsequent processing, significantly improving the practicality, targeting, and safety of supervision.

[0038] In a scenario involving a middle-aged male being monitored, after analyzing his ECG signal, the abnormal ECG event recognition model outputs the abnormality type as "occasional premature ventricular contractions" (premature ventricular contraction, which may be physiological or pathological, requiring judgment based on the specific scenario). Based on the model's preset risk scoring rules (initial risk score for occasional premature ventricular contractions...), the model then assigns a risk score... Frequent premature ventricular contractions Premature ventricular contractions combined with ST segment depression ), determine the initial risk score for this test. .

[0039] The device collects data such as the number of steps and acceleration of the monitored person through an inertial measurement unit (IMU), analyzes the data to determine that the person is currently running, and sets the real-time activity intensity index A=0.8 after normalization (preset normalization standards: resting state A=0, slow walking A=0.3, running A=0.8, vigorous exercise A=1), while also setting a compensation coefficient. (Clinically validated to balance the impact of activity intensity on the risk of electrocardiographic abnormalities).

[0040] Substitute into the comprehensive risk score calculation formula: , can be obtained .

[0041] If the person being monitored is at rest (A=0) at this time, then By correcting for activity status, it effectively distinguishes between "occasional premature beats during exercise (high physiological probability, risk not significantly increased)" and "occasional premature beats at rest (pathological possibility should be noted, risk remains at the initial value)," avoiding misjudgment of risk due to ignoring activity status.

[0042] The threshold range corresponding to the preset alarm level (low risk: Medium risk: High risk: ), this time The risk level is classified as medium, and the corresponding medium-risk response will be executed: The device will locally vibrate and display a text message to the monitored person stating, "Occasional premature ventricular contractions (PVCs) detected (medium risk). It is recommended to immediately reduce exercise intensity, maintain steady breathing, and continue monitoring." Simultaneously, the device will automatically record the abnormal ECG segment (including the time of occurrence and PVC waveform data) and real-time location information (GPS-located to "XX City XX Park Plastic Running Track"), without triggering a remote alarm (to avoid consuming medical resources). If, during subsequent monitoring, the person experiences occasional premature ventricular contractions again at rest (…), A=0, If the risk is low, only "data recording + slight vibration alert" will be executed; if "frequent premature ventricular contractions" occur ( Even if resting A=0, In cases of high risk, the device immediately triggers a remote alarm, sending a distress signal to the monitoring center / linked medical staff. The signal includes an abnormal ECG segment (a complete waveform of nearly 10 seconds), the real-time location (accurate to within 10 meters), and the name, age, and medical history of the monitored person. This ensures timely intervention in high-risk situations and demonstrates the precise adaptability of the tiered response system.

[0043] In step S52, the activity state compensation function is used to quantitatively assess the potential impact of physical activity on electrocardiogram signals and risk scores. Its core lies in performing linear compensation on the initial risk score based on activity intensity. Specifically, the comprehensive risk score is obtained through the following calculation process: ; Wherein, is the revised overall risk score. The initial risk score is determined in step S51, A is the normalized real-time activity intensity index obtained by analyzing and processing the data of the inertial measurement unit, and β is a pre-set compensation coefficient greater than zero, which is used to adjust the degree of influence of the activity status on the final risk assessment. By constructing a linear activity state compensation function based on activity intensity, the potential impact of physical activity on electrocardiogram signals and risk scores can be quantitatively assessed. The initial risk score is corrected using a clear linear calculation method, avoiding risk assessment bias caused by ignoring activity states (such as physiological electrocardiogram fluctuations caused by exercise and pathological abnormalities at rest). This allows the comprehensive risk score to better reflect the actual physiological scenarios of the monitored individuals, providing a more accurate risk basis for subsequent tiered regulatory response strategies, reducing false alarms or missed alarms caused by activity interference, and further improving the objectivity and relevance of risk assessment.

[0044] In a scenario involving a middle-aged male being monitored, the abnormal electrocardiogram event identification model has determined the abnormality type to be "occasional premature ventricular contractions," with an initial risk score. (In the preset scoring rules, the baseline risk score for occasional premature ventricular contractions is fixed at 30).

[0045] The device collects the person's real-time motion data through an inertial measurement unit (IMU): when he is running, the IMU records a step frequency of 150 steps / minute and a peak vertical acceleration of 0.7g. After normalization (preset resting state activity intensity A=0, vigorous exercise A=1), the real-time activity intensity index during running is obtained as A1=0.8. When he was in a seated resting state, the IMU recorded a cadence of 0 and an acceleration of 0, and the normalized activity intensity index A2=0; when he was walking slowly (cadence of 70 steps / minute, acceleration of 0.2g), the normalized activity intensity index A3=0.3.

[0046] Meanwhile, based on clinical electrocardiogram monitoring data, a preset compensation coefficient was established. (This value can balance the impact of activity intensity on risk assessment, avoiding overestimating or underestimating the role of the activity.)

[0047] Based on the comprehensive risk score calculation formula Calculate the results under different activity states: Running status (A1=0.8): Substituting into the formula yields... .

[0048] The overall risk score is higher than the initial score at this point because physical activity during running may cause physiological fluctuations in the electrocardiogram signal. The risk score needs to be corrected to indicate that "attention should be paid to the potential impact of the current activity on electrocardiogram abnormalities, and it is recommended to adjust the exercise intensity appropriately." Resting state (A2=0): Substituting into the formula yields .

[0049] If there is no activity interference at rest and the overall risk score is consistent with the initial score, and occasional premature ventricular contractions still occur at this time, there is a greater possibility of pathological tendencies, and monitoring should be carried out according to the level corresponding to the initial risk. Slow walking state (A3=0.3): Substituting into the formula, we get... .

[0050] Walking is less strenuous than running, so the adjustment range for the overall risk score is smaller. This reflects the slight impact of the activity without excessively raising the risk level, which is consistent with actual physiological logic.

[0051] If this activity state compensation function is not used, the risk scores for all activity states will be calculated as follows: This approach can lead to two misjudgments: First, the risk level of occasional premature ventricular contractions (PVCs) during running (high physiological probability) may be the same as that of occasional PVCs at rest (high pathological probability). This could result in a failure to promptly address physiological fluctuations during running or to focus on abnormalities at rest. Second, the impact of minor activities during slow walking may not be quantified, making it impossible to accurately distinguish between "activity-related physiological abnormalities" and "non-activity pathological abnormalities." However, by using a linear compensation function, the differences in comprehensive risk scores across different activity states are clearly defined (42, 30, 34.5), which can directly correspond to different regulatory response intensities (e.g., triggering local alerts for medium-risk activities during running, recording only low-risk activities at rest, and increasing monitoring frequency for low-risk activities during slow walking), significantly improving the accuracy of risk assessment.

[0052] After executing the remote alarm in step S5, a feedback learning step is also included: S55: Receive feedback information from the monitoring center or medical staff regarding this alarm, including alarm confirmation, false alarm labeling, or severity correction; S56: Using the feedback information and the corresponding multidimensional ECG feature vector and real-time activity status data, the abnormal ECG event recognition model is incrementally updated to optimize the model's future classification performance. By employing a feedback learning process following a remote alarm, the professional feedback information from the monitoring center or medical personnel (alarm confirmation, false alarm marking, severity correction) is combined with corresponding ECG characteristics and activity status data to incrementally update the abnormal ECG event recognition model. This forms a closed loop of "recognition to alarm to feedback to optimization," which not only solves the potential limitations of the initial training data but also allows the model to continuously adapt to changes in the individual ECG characteristics of the monitored individuals (such as long-term physiological state adjustments and the progression of underlying diseases). This reduces false alarms and missed alarms in similar scenarios in the future, improves the long-term stability and personalized adaptability of the model's classification performance, and ensures that the intelligent monitoring method is continuously optimized as it is used.

[0053] The device self-calibration process in step S6 specifically includes: S61: Apply a standard test signal to the twelve-channel ECG monitoring device; S62: Acquire the response signals of each channel to the standard test signal; S63: Calculate the gain error and phase error between the response signal of each channel and the standard test signal; S64: Based on the gain error and phase error, generate a set of calibration parameters, and apply these parameters in subsequent signal acquisition to correct the original signal in real time to maintain the consistency of the twelve-channel signal.

[0054] The above process can accurately quantify the hardware deviations (such as gain drift and phase shift) of each channel of the twelve-channel ECG monitoring device, and eliminate such deviations through real-time correction. This effectively maintains the consistency of multi-channel signal acquisition, avoids differences in signal accuracy between channels caused by hardware aging due to long-term use of the equipment and environmental interference (such as temperature changes), ensures the reliability and accuracy of ECG data in long-term monitoring, and provides a signal foundation of uniform accuracy for subsequent multi-channel signal fusion, feature extraction and anomaly identification, further improving the stability of the entire intelligent monitoring method and the credibility of the analysis results.

[0055] If the scenario of using a twelve-channel ECG monitoring device for middle-aged male subjects continues, the device performs a self-calibration process according to a preset cycle (e.g., once a week). First, a standard test signal is simultaneously applied to all twelve channels of the device. This signal is a sinusoidal signal simulating a normal ECG waveform, with the following parameters set: frequency... (Corresponding to a heart rate of 60 beats / min, simulating the baseline rhythm of an electrocardiogram signal), standard amplitude (Typical amplitude range of analog ECG signal), standard phase .

[0056] Subsequently, the response signals of each channel to the standard signal were acquired, and key parameters were recorded: Channel 1 response amplitude. Phase ; Channel 3 response amplitude Phase ; Channel 8 response amplitude Phase (The response parameters of the other channels are within the above range), and it can be seen that there are obvious amplitude and phase deviations in different channels due to hardware differences.

[0057] Based on the calibration logic of this case, the gain error and phase error of each channel are calculated respectively.

[0058] The formula for calculating gain error is as follows: ( (where i is the amplitude of the response of the i-th channel); the formula for calculating the phase error is: ( (where i is the phase of the response of the i-th channel). Substitute the channel data above to calculate: Channel 1: ; Channel 3: ; Channel 8: .

[0059] The calculation results show that channel 8 has the largest gain error (-8%) and phase error (-12°). If it is not calibrated, the amplitude of the actual ECG signal it acquires will be underestimated by 8% and the phase will be delayed by 12°, causing the signal of this channel to be out of sync with the signals of other channels, affecting the accuracy of subsequent multi-channel fusion.

[0060] Based on the above error calculation results, calibration parameters and gain calibration coefficients for each channel are generated. (Used to correct amplitude deviation), phase calibration coefficient (Used to correct phase deviation). The calibration parameters for each channel are as follows: Channel 1 , ; Channel 3 , ; Channel 8 , .

[0061] In subsequent signal acquisition, the device applies calibration parameters to the raw signals of each channel in real time: if the raw amplitude of an actual ECG signal acquired by channel 8 is... The phase is -15°, and the amplitude after calibration is (Close to the true amplitude), the phase after calibration is -15° + 12° = -3° (synchronized with the phase of other channels).

[0062] After calibration, the response of each channel to the standard signal is tested again: Channel 1 response amplitude Phase 0.2°; Channel 3 response amplitude Phase -0.3°; Channel 8 response amplitude Phase 0.8°, gain error of all channels controlled within Within, the phase error is controlled within Within this timeframe, the consistency of multi-channel signals is significantly improved. During subsequent fusion preprocessing based on the calibrated signals, channel 8 no longer suffers from signal deviations that negatively impact fusion quality. The QRS complex identification of the fused ECG signal is enhanced, providing a more accurate signal foundation for subsequent anomaly detection.

[0063] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0064] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0065] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for intelligent monitoring of a wearable electrocardiogram (ECG) monitoring device, characterized in that, Includes the following steps: S1: Continuously acquire multi-channel electrocardiogram (ECG) signals of the monitored person using a twelve-channel ECG monitoring device based on flexible circuits; S2: Perform fusion preprocessing on multi-channel ECG signals to obtain high-quality fused ECG signals; S3: Based on the fused ECG signal and the preprocessed signal of some channels, extract multidimensional ECG feature vectors with inter-channel correlation; S4: Input the multidimensional ECG feature vector into a trained abnormal ECG event recognition model for intelligent analysis and classification to identify specific abnormal ECG patterns. S5: Based on the type and severity of the identified abnormal ECG patterns, dynamically generate and execute a graded intelligent monitoring response strategy; S6: Periodically perform the equipment self-calibration process by applying standard test signals to the equipment and calculating calibration parameters based on the response signals of each channel to maintain the consistency of the twelve-channel signal acquisition.

2. The intelligent monitoring method for a wearable electrocardiogram monitoring device according to claim 1, characterized in that: The fusion preprocessing in step S2 includes: S21: Perform wavelet transform-based denoising processing on each single-channel ECG signal to obtain the preliminary purified signal for each channel; S22: Perform inter-channel signal quality assessment on the preliminary purified signals of each channel. The signal quality assessment results are obtained by weighting the signal-to-noise ratio and baseline drift of each channel signal. S23: Based on the signal quality assessment results, construct a channel weight set, in which channels with high signal quality are assigned higher weights; S24: The preliminary purified signals from each channel are weighted and fused according to their corresponding channel weights to finally generate a fused ECG signal.

3. The intelligent monitoring method for a wearable electrocardiogram monitoring device according to claim 1, characterized in that: The extraction process of the multidimensional ECG feature vector with inter-channel correlation extracted in step S3 includes: S31: Extract global time-domain and frequency-domain features from the fused ECG signal; S32: Extract channel-specific features from the preliminary purification signals of at least two specific channels. These channel-specific features are used to reflect the electrical activity characteristics of different cardiac sites. S33: Calculate the mutual information value between the preliminary clean signals of at least two specific channels to quantify the correlation between the channels. The mutual information value is obtained by calculating the relationship between the joint probability distribution of the two channel signals and their respective marginal probability distributions. S34: Combine global time-domain features, frequency-domain features, channel-specific features, and mutual information values ​​to form a multidimensional ECG feature vector.

4. The intelligent monitoring method for a wearable electrocardiogram monitoring device according to claim 1, characterized in that: The abnormal ECG event recognition model in step S4 is an improved support vector machine model. The improvement lies in introducing the inter-channel mutual information value in the multi-dimensional ECG feature vector as a model parameter adjustment factor. The decision function used by this support vector machine model in classification is constructed based on statistical learning theory.

5. A smart monitoring method for a wearable electrocardiogram (ECG) monitoring device according to claim 4, characterized in that: The dynamic generation of the hierarchical intelligent regulatory response strategy in step S5 includes: S51: Based on the output of the abnormal ECG event identification model, determine the type of abnormal event and the initial risk score; S52: The initial risk score is corrected by combining the real-time activity status of the monitored personnel to obtain a comprehensive risk score; the correction is achieved through an activity status compensation function, which utilizes activity intensity data from the inertial measurement unit; S53: Compare the comprehensive risk score with multiple preset threshold ranges to determine the corresponding alarm level; S54: Execute response actions corresponding to the alarm level. Response actions include different combinations of actions, from data logging and local prompting to remote alarms and sending distress signals containing abnormal ECG fragments and location information.

6. The intelligent monitoring method for a wearable electrocardiogram monitoring device according to claim 1, characterized in that: In step S52, the activity state compensation function is used to quantitatively assess the potential impact of physical activity on electrocardiogram signals and risk scores. Its core is to perform linear compensation on the initial risk score based on activity intensity.

7. The intelligent monitoring method for a wearable electrocardiogram monitoring device according to claim 6, characterized in that: After executing the remote alarm in step S5, a feedback learning step is also included: S55: Receive feedback from the monitoring center or medical staff regarding this alert. The feedback includes alert confirmation, false alarm flags, or severity corrections. S56: Using feedback information and corresponding multidimensional ECG feature vectors and real-time activity status data, the abnormal ECG event recognition model is incrementally updated to optimize the model's future classification performance.

8. The intelligent monitoring method for a wearable electrocardiogram monitoring device according to claim 1, characterized in that: The device self-calibration process in step S6 specifically includes: S61: Apply a standard test signal to the twelve-channel ECG monitoring device; S62: Acquire the response signals of each channel to the standard test signal; S63: Calculate the gain error and phase error between the response signal of each channel and the standard test signal; S64: Based on gain error and phase error, a set of calibration parameters is generated, and these parameters are applied in real time to correct the original signal in subsequent signal acquisition to maintain the consistency of the twelve-channel signal.

Images