Portable intelligent electrocardiogram monitoring method and system

Abstract

The invention discloses a portable intelligent electrocardiogram monitoring method and system, and the method comprises the steps: collecting a chest lead electrocardiogram signal through a chest paste, collecting a physiological signal through a bracelet, and storing the collected signals as a health signal; power line interference and muscle artifacts in the health signals are eliminated through GF low-pass filtering, and baseline drift in the health signals is eliminated through IIR high-pass filtering; a DisCo-FFS feature selection method is adopted to capture time sequence features of the electrocardiosignals and reduce data complexity, and an input data set is constructed; constructing a FilterDF-form diagnosis model, and introducing a DOA dream optimization algorithm to optimize the FilterDF-form diagnosis model; performing diagnosis by using the DOA-Filter DF-form model, and outputting a diagnosis result of the type of the heart disease; a diagnosis result is sent to a user client through the central processing module, and corresponding professional suggestions are provided; the wearable comfort and the use convenience are improved, and the comprehensiveness and the accuracy of heart health monitoring are remarkably enhanced through synchronous acquisition of multi-mode signals.

Description

Technical Field

[0001] The present invention belongs to the technical field of electrocardiogram monitoring, and particularly relates to a portable intelligent electrocardiogram monitoring method and system. Background Art

[0002] Cardiovascular diseases have become one of the main threats to residents' health. As the core means for the diagnosis and management of cardiovascular diseases, electrocardiogram monitoring has become increasingly important, which has promoted the rapid development of medical-grade and consumer-grade electrocardiogram monitoring technologies.

[0003] Currently, medical-grade electrocardiogram monitoring mainly relies on professional equipment such as 12-lead electrocardiograph machines. High-precision electrocardiogram signals are collected through multiple electrodes, and manual or algorithm-assisted diagnosis is carried out based on clinical standards. Such equipment usually uses technologies such as filtering, baseline correction, and feature extraction to process signals to ensure compliance with medical diagnosis requirements.

[0004] However, although medical-grade equipment has high precision, it is large in volume and complex in operation, and cannot meet the needs of daily continuous monitoring. Moreover, existing algorithms are mostly optimized for static scenarios, and the signal quality drops sharply under the user's movement state, restricting the practical value. As a result, the existing technology has limitations in scene adaptability and diagnostic ability. Summary of the Invention

[0005] Object of the Invention: The object of the present invention is to provide a portable intelligent electrocardiogram monitoring method that is exquisitely designed, lightweight, easy to carry and wear, and can improve the diagnostic ability. On the other hand, a portable intelligent electrocardiogram monitoring system is provided.

[0006] Technical Solution: The electrocardiogram monitoring method of the present invention includes the following steps:

[0007] (1) Collect electrocardiogram signals by attaching a portable chest patch to the human body in accordance with the chest lead method, and collect physiological signals through a smart bracelet. Store the electrocardiogram signals and physiological signals as healthy signals, effectively realizing synchronous acquisition of multi-channel electrocardiogram and physiological data, and providing a comprehensive data basis for subsequent analysis;

[0008] (2) Use GF low-pass filtering to eliminate power line interference and muscle artifacts in the healthy signals, and then use IIR high-pass filtering to eliminate baseline drift in the healthy signals, ensuring that the healthy signals are purer, improving the signal quality, and providing reliable data input for subsequent processing and diagnosis;

[0009] (3) Adopt the DisCo-FFS feature selection method to process the filtered electrocardiogram signals, capture the temporal characteristics of the electrocardiogram signals and reduce the data complexity, construct an input data set, enhance the interpretability of the electrocardiogram signals, reduce redundant information at the same time, and improve the efficiency of the subsequent model;

[0010] (4) Construct a FilterDF-former diagnostic model, and update the FilterDF-former diagnostic model by introducing the DOA dream optimization algorithm to form an optimized DOA-FilterDF-former model, improving the diagnostic accuracy so that the model can more accurately predict and analyze heart diseases;

[0011] (5) Process the input data set using the DOA-FilterDF-former model and output the classification results of health signals, enhancing the reliability and practicality of the intelligent diagnostic system;

[0012] (6) Send the classification results of the health signals to the user client and output preset prompt information associated with the classification results, facilitating the user to obtain health information in a timely manner and take corresponding measures, and improving the user's health management level.

[0013] Preferably, the health signal acquisition process described in step 1 includes:

[0014] (11) Collect six chest lead electrocardiogram signals E1, E2, E3, E4, E5, and E6 through an intelligent chest patch. The intelligent chest patch is provided with V1-V6 electrode areas, corresponding to the positions of the right margin of the sternum at the 4th intercostal space, the left margin of the sternum at the 4th intercostal space, the midpoint of the connection line between V2 and V4, the intersection of the left midclavicular line and the 5th intercostal space, the same level as V4 on the left anterior axillary line, and the same level as V4 on the left midaxillary line respectively;

[0015] (12) Collect physiological signals through an intelligent bracelet. The intelligent bracelet is worn on the user's wrist, and the physiological signals include pulse E7, body temperature E8, and blood pressure E9 signals;

[0016] (13) Store the E1-E6 electrocardiogram signals and E7-E9 physiological signals as a health signal data set E = {E1, E2, E3, E4, E5, E6, E7, E8, E9} for subsequent analysis and processing.

[0017] The above health signal acquisition process realizes the synchronous acquisition of multi-channel electrocardiogram signals and physiological signals through the collaborative work of the intelligent chest patch and the intelligent bracelet. The accurate acquisition and storage of these signals provide comprehensive and reliable data support for subsequent data analysis and diagnosis, helping to improve the accuracy of health management and disease warning.

[0018] Preferably, the high- and low-pass filtering process described in step 2 is as follows:

[0019] (21) Perform segmented processing on the health signal data, divide it into 10 segments of equal-length signal data, and each segment is further divided into 9 signal sequences;

[0020] (22) Apply a Gaussian low-pass filter to each signal sequence to remove high-frequency noise in the signal, reduce power line interference and muscle artifacts, and calculate the filtered output through the following formula:

[0021]

[0022] Where y(n) is the output signal; b k is the coefficient of the filter, σ is the standard deviation of the Gaussian function, n is the order of the filter, and k is the number of delay steps; the cut-off frequency of the Gaussian low-pass filter is set to 35 Hz to 45 Hz, the passband ripple is 1 dB, and the stopband attenuation is 80 dB;

[0023] (23) Apply an IIR high-pass filter to the signal sequence of each segment after low-pass filtering to remove baseline drift in the signal, and calculate the filtered output through the following formula:

[0024]

[0025] Where a i is the feedback coefficient; m is the order of the feedback part; i is the number of delay steps; the cut-off frequency of the IIR high-pass filter is set to 0 Hz to 0.05 Hz, the passband ripple is 1 dB, and the stopband attenuation is 80 dB;

[0026] (24) Perform time synchronization processing on the signal sequences after low- and high-pass filtering, including start time alignment and sampling rate adjustment, and arrange and combine them in the original segment order. Finally, merge them to form the processed healthy signal dataset E' = {E'1, E'2, E'3, E'4, E'5, E'6, E'7, E'8, E'9}.

[0027] The above high- and low-pass filtering processes effectively remove noise and interference in the healthy signal through precise filtering techniques, including high-frequency noise, power line interference, muscle artifacts, and baseline drift, thus ensuring the clarity and accuracy of the signal; the combination of the Gaussian low-pass filter and the IIR high-pass filter, combined with strict frequency and attenuation settings, ensures the smoothness and stability of the signal; finally, through time synchronization and sampling rate adjustment, the processed healthy signal dataset is obtained, providing high-quality and less noisy signal data for subsequent analysis and diagnosis.

[0028] Preferably, the DisCo-FFS feature selection process described in step 3 is as follows:

[0029] (31) Use the processed electrocardiogram signal E'1 as the input to establish an initial feature set F 0 = {E′1};

[0030] (32) Use F 0Train a lightweight classifier to obtain the classifier output y for all sample events pred ;

[0031] (33) Screen the classifier output y pred Construct a confusion set E″ from samples with values between 0.3 and 0.7, where:

[0032] E″ = {x ∈ E′ ∣ 0.3 < y pred (x) < 0.7}

[0033] In the formula, E″ is a subset of the healthy signal dataset;

[0034] (34) Calculate the candidate correlation scores, including for each candidate feature E′ i (i = 2, 3,..., 9), calculate its distance correlation with the reference label y ref ; for each event x ∈ E″, construct the feature vector X = {(E′1(x), E′ i (x)) ∣ x ∈ E″}; the corresponding reference label is Y = {y ref (x) ∣ x ∈ E″}; use the following formula to calculate the distance correlation dCor between X and Y:

[0035]

[0036] In the formula, the distance covariance dCov 2 (X, Y) is calculated by the formula:

[0037] dCov 2 (X, Y) = E v [||X - X′|| ||Y - Y′||] + E v [||X - X′||] E v [||Y - Y′||] - 2E v [||X - X′||||Y - Y″||]

[0038] In the formula, X', Y', X″, Y″ represent independent and identically distributed random vectors of X and Y, ||·|| represents the Euclidean norm, and E v represents the expected value;

[0039] (35) Select the feature E′ with the maximum dCor value i , add it to the selected feature set F 0 to form a new feature set F 1 ;

[0040] (36) Repeat steps (32) to (35) until the performance metrics of the classifier saturate, stop the feature selection process, and obtain the final input feature set F k={E′1, E′2,..., E′ n}, n = k, where k is the number of finally selected features.

[0041] The above DisCo-FFS feature selection process accurately filters out the features with the maximum correlation with the reference label by combining a lightweight classifier and the distance correlation score; through an iterative process, the feature set is gradually constructed and optimized to ensure that the selected features make a significant contribution to the improvement of the classifier performance; finally, by setting the stopping criterion, the optimal input feature set is obtained, effectively improving the performance and accuracy of the subsequent model, while reducing the interference of redundant features and ensuring the efficiency and accuracy of feature selection.

[0042] Preferably, the process of constructing the FilterDF-former diagnostic model described in step 4 is as follows:

[0043] (41) Standardize the input feature set F k ={E′1, E′2,..., E′ n} so that its mean is 0 and variance is 1:

[0044]

[0045] where μ i and σ i are the mean and standard deviation of the i-th health signal respectively;

[0046] (42) Construct each health signal E′ i into a time series vector X i ={E i,1 , E i,2 ,..., E i,T}, where T is the number of time steps;

[0047] (43) Perform DTW (Dynamic Time Warping), including calculating the similarity between the time feature vectors of different health signals: d DTW (X i , X j ) = DTW(X i , X j ); where X i and X j are the time series vectors of two health signals, and d is the similarity; select the time feature vector of the health signal with the highest similarity as the neighbor according to the DTW distance to construct the time correlation graph A t ; calculate the cosine similarity of the node embeddings and select the most similar node as the neighbor to form the learned spatio-temporal correlation graph A st ;

[0048] (44) Use the spatio-temporal correlation graph Ast Extended to the global spatio-temporal graph A g , enabling nodes at different positions and time slices to interact freely:

[0049]

[0050] In the formula, A is the adjacency matrix of the topological structure graph of the health signal, and A st is the adjacency matrix of the spatio-temporal correlation graph; NT represents the dimension;

[0051] (46) Perform data embedding processing. First, perform spatio-temporal hybrid embedding. The sequence learned through the random walk strategy fuses the topological information of the health signal and the temporal correlation of the temporal similarity graph to form the spatio-temporal hybrid embedding E st ; then perform temporal embedding. Use the fixed coding method to perform position coding on the time series information, associate the sequence elements with points on the sine or cosine curve, and generate the position embedding E t :

[0052]

[0053] In the formula, pos represents the time step, d model represents the total dimension of the embedding vector; d represents the dimension index of the embedding vector;

[0054] Next, perform spatial embedding. Extract information based on the topological structure graph of the health signal and the temporal series similarity of the nodes, calculate the topological graph, and obtain the normalized graph Laplacian matrix through matrix decomposition. Its corresponding eigenvector is used as the spatial embedding E s ; finally, add the spatio-temporal hybrid embedding E st , the temporal embedding E t and the spatial embedding E s to get the final data embedding E data = E s + E t + E st ;

[0055] (46) Encoder processing. Introduce the self-attention mechanism to capture the dependencies between different positions and time steps in the input data:

[0056]

[0057] In the formula, Q, K, and V are the query, key, and value matrices respectively. QK T is the dot product operation, calculating the attention weights of Q on V, is used to scale the dot product to prevent the problem of gradient disappearance, and softmax() is the transfer function;

[0058] Finally, use residual connection and layer normalization to output the feature Efinal = LayerNorm(E data + Attention(Q, K, V));

[0059] (47) Decoder processing, taking the feature E after encoder processing final and the original data E data as the input of the decoder. After the decoder's processing, the final diagnosis result Y diag = Linear(E final + E data ), and finally outputs the heart disease type Y = [Y1, Y2, Y3, Y4], where: Y1 is normal, Y2 is myocardial infarction, Y3 is myocardial hypertrophy, and Y4 is arrhythmia.

[0060] By fusing the topological information, temporal similarity, and spatial features of healthy signals, the model effectively captures the spatio-temporal dependence relationships between signals; the encoder introduces a self-attention mechanism to accurately capture the complex relationships between different time steps and positions; the decoder then decodes based on the features of the encoder and finally outputs the heart disease type. This technology effectively improves the diagnostic accuracy and reliability of heart diseases.

[0061] Preferably, the introduction of the DOA dream optimization algorithm to update the FilterDF-former diagnostic model in step 4 includes optimizing the initialization parameters W and the L2 regularization parameter λ of the model. The specific process is as follows:

[0062] Initialize the parameter vector, combining the initialization parameters and regularization parameters of the FilterDF-former diagnostic model into a parameter vector X = (W, λ);

[0063] Generate a random population, where each individual corresponds to a group of parameter vectors Xi, where i = 1, 2,..., N, and N is the population size, and calculate the individual fitness;

[0064] In the exploration stage, divide the population into 5 groups according to the difference in memory ability, and update the individuals in each group as follows:

[0065] For the q-th group (q = 1, 2, 3, 4, 5), perform the following process:

[0066] Memory strategy, reset the position of each individual in the group to the position of the best individual in the group:

[0067]

[0068] In the formula, is the position of the best individual in the q-th group at iteration t;

[0069] Forgetting and replenishment strategy. The individual updates the position information in the forgetting dimension. Combining the global and local search functions, the update formula is as follows:

[0070]

[0071] In the formula, α and β are step factors that control the amplitude of the update, and rand is a random number between 0 and 1;

[0072] Dream sharing strategy. The individual randomly obtains the position information of other individuals in the forgetting dimension to enhance the escape ability from the local optimal solution. The update formula is as follows:

[0073]

[0074] In the formula, m is a natural number randomly selected within the range of [1, N], and m ≠ i;

[0075] In the exploitation stage, grouping is no longer performed. All individuals display the best individual before in each iteration and update the position in the forgetting dimension. The following process is executed:

[0076] Memory strategy. Reset the positions of all individuals to the position of the global best individual:

[0077]

[0078] In the formula, is the position of the optimal individual in the whole population at iteration t;

[0079] Forgetting and replenishment strategy. The individual updates the position information in the forgetting dimension. The update formula is as follows:

[0080]

[0081] In the formula, γ and δ are step factors;

[0082] Convergence judgment. Judge whether the preset maximum number of iterations T max is reached; if so, the algorithm terminates; otherwise, continue the iteration;

[0083] After the algorithm converges, select the individual with the largest fitness value from the population as the optimal solution, that is, find the optimal combination of initialization parameters and regularization parameters (W', λ').

[0084] By jointly encoding the model parameters (W) and the regularization parameter (λ) as a parameter vector, and utilizing the grouped memory strategy, forgetting-supplementation mechanism, and dream sharing strategy, multi-stage intelligent optimization is achieved: in the exploration stage, global exploration and local development are balanced through 5 groups of differential memory resets and random dimension updates; in the development stage, the population is guided to converge by the globally optimal individual, and the dynamic step size factors (α / β / γ / δ) are combined to enhance the parameter fine-tuning accuracy; finally, the algorithm can efficiently escape from the local optimum, automatically obtain the optimal parameter combination (W', λ'), significantly improve the generalization ability of the FilterDF-former model and the heart disease classification accuracy, and at the same time, overfitting can be suppressed through L2 regularization optimization.

[0085] Preferably, the model diagnosis process described in step 5 is as follows:

[0086] (51) Divide the healthy signal dataset into a training set, a validation set, and a test set according to a preset ratio;

[0087] (52) Use the divided training set to train the improved DOA-FilterDF-former model, enabling the model to learn the patterns and relationships in the data;

[0088] (53) Use the validation set to validate the model, adjust the model parameters, and complete the training;

[0089] (54) Use the trained model to process the test set to obtain the healthy signal classification result.

[0090] By training the model using the training set, it can learn the patterns and relationships in the data; using the validation set to validate the model and adjusting the model parameters according to the validation results to ensure its better adaptation to the data; using the trained model to diagnose the test set to obtain an accurate diagnosis result of the heart disease type; this process effectively improves the generalization ability and diagnostic accuracy of the model.

[0091] Preferably, the process described in step 6 is as follows:

[0092] (61) Integrate the healthy signal classification results through the central processing module to form a complete result report;

[0093] (62) Send the result report to the user client by wireless or wired means, and the user client includes mobile phones, tablets;

[0094] (63) According to the healthy signal classification result, the system automatically generates corresponding preset prompt information, including medical advice and lifestyle adjustment advice.

[0095] This process not only improves the automation of diagnosis and recommendations, but also enhances users' awareness of health management and self-care ability, provides comprehensive health support services, and further optimizes the medical service experience.

[0096] The electrocardiogram monitoring system described in the present invention includes:

[0097] A signal acquisition module, including a portable chest patch and a smart bracelet. The portable chest patch is used to adhere to the human body in accordance with the chest lead method to collect electrocardiogram signals, and the smart bracelet is used to collect physiological signals and store the electrocardiogram signals and physiological signals as health signals;

[0098] A signal preprocessing module, including a GF low-pass filter unit and an IIR high-pass filter unit. The GF low-pass filter unit is used to eliminate power line interference and muscle artifacts in the health signals; the IIR high-pass filter unit is used to eliminate baseline drift in the health signals;

[0099] A feature extraction module, which is used to process the filtered electrocardiogram signals by using the DisCo-FFS feature selection method, capture the temporal features of the electrocardiogram signals and reduce the data complexity, and construct an input data set;

[0100] A model diagnosis module, which is used to construct a FilterDF-former diagnosis model and update the FilterDF-former diagnosis model by introducing a DOA dream optimization algorithm to form an optimized DOA-FilterDF-former model;

[0101] A result output module, which is used to process the input data set by using the DOA-FilterDF-former model and output the classification result of the health signals;

[0102] A central processing module, which is used to send the classification result of the health signals to the user client through the central processing module and output a preset prompt message associated with the classification result.

[0103] A computer-readable storage medium, on which a computer program is stored. The computer program is characterized in that when the computer program is executed by a processor, it implements the portable intelligent electrocardiogram monitoring method described above.

[0104] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: 1. It not only improves the wearing comfort and usage convenience, but also significantly enhances the comprehensiveness and accuracy of cardiac health monitoring through the synchronous acquisition of multi-modal signals; 2. Through the cascaded processing of low-pass and high-pass filtering, the quality of the input data for the subsequent diagnostic model is significantly optimized; 3. The DisCo-FFS feature selection method improves the model performance and reduces the risk of overfitting, with high computational efficiency and strong compatibility; 4. The FilterDF-former model alleviates the complex spatio-temporal dependence relationship between data and improves the model's understanding and processing ability of data; 5. The dream optimization algorithm (DOA) avoids local optimal solutions and enhances the global search ability, thereby improving the optimization efficiency and accuracy. Brief Description of the Drawings

[0105] Figure 1 is a flowchart of the present invention;

[0106] Figure 2 is a schematic diagram of the DisCo-FFS feature selection method of the present invention;

[0107] Figure 3 is a schematic diagram of the dream optimization algorithm (DOA) of the present invention;

[0108] Figure 4 is a schematic diagram of the overall framework of the present invention. Detailed Embodiment

[0109] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0110] As Figure 1 shown, the present invention includes the following steps:

[0111] Step 1: Collect electrocardiogram signals by attaching a portable chest patch to the human body in the chest lead manner, and collect physiological signals by a smart bracelet, and store the electrocardiogram signals and physiological signals as health signals.

[0112] Step 1.1: Wear a smart chest patch. The areas marked V1-V6 on the chest patch correspond to the positions of the right margin of the sternum at the 4th intercostal space, the left margin of the sternum at the 4th intercostal space, the midpoint between V2 and V4, the intersection of the left midclavicular line and the 5th intercostal space, the left anterior axillary line at the same level as V4, and the left midaxillary line at the same level as V4. Ensure that the monitoring electrodes are accurately placed and collect the electrocardiogram signals (E1, E2, E3, E4, E5, E6) of the chest leads in the above 6 areas;

[0113] Step 1.2: Wear the smart bracelet on the wrist and collect the pulse (E7), body temperature (E8), and blood pressure physiological signal (E9);

[0114] Step 1.3: Store the collected electrocardiogram signals and physiological signals as healthy signals (E = {E1, E2, E3, E4, E5, E6, E7, E8, E9}) for subsequent data analysis and cardiac health monitoring.

[0115] Step 2: Use a GF low-pass filter to eliminate power line interference and muscle artifacts in the healthy signals, and then use an IIR high-pass filter to eliminate baseline drift in the healthy signals.

[0116] Step 2.1: Slice the extracted waveform data so that the original signal data is processed into 10 segments of signal data with a fixed length, and each segment is divided into 9 sequences;

[0117] Step 2.2: Select a Gaussian low-pass filter, set the cut-off frequency range to 35 Hz to 45 Hz, the passband ripple to 1 dB, and the stopband attenuation to 80 dB;

[0118] Step 2.3: Apply the designed low-pass filter to each signal sequence segment, perform filtering calculations on each signal point, remove high-frequency noise in the signal, reduce power line interference and muscle artifacts, and obtain new filtered signal values. The filtering calculation formula is as follows:

[0119]

[0120] In the formula, y(n) is the output signal, b k is the coefficient of the filter, σ is the standard deviation of the Gaussian function, n is the order of the filter; k is the number of delay steps.

[0121] Step 2.4: Select an IIR high-pass filter, set the cut-off frequency range to 0 Hz to 0.05 Hz, the passband ripple to 1 dB, and the stopband attenuation to 80 dB;

[0122] Step 2.5: Apply the designed high-pass filter to each segmented signal sequence after low-pass filtering, remove baseline drift in the signal, perform filtering calculations on each signal point, and obtain new filtered signal values. The filtering calculation formula is as follows:

[0123]

[0124] In the formula, a i is the feedback coefficient; m is the order of the feedback part; i is the number of delay steps.

[0125] Step 2.6: Synchronize and align each sequence signal processed by low- and high-pass filtering on the time axis.

[0126] Check and adjust the starting time and sampling rate of the signals to make them consistent. Then, arrange and combine them in the original segmented order to merge into a complete processed healthy signal dataset E' = {E'1, E'2, E'3, E'4, E'5, E'6, E'7, E'8, E'9}.

[0127] Step 3: As Figure 2 shown, use the DisCo-FFS feature selection method to process the filtered electrocardiogram signals, capture the temporal features of the electrocardiogram signals and reduce the data complexity, and construct the input dataset.

[0128] Step 3.1: Initialize the feature set, start from a group of initial features, denoted as F 0 = {E′1};

[0129] Step 3.2: Train the classifier, use F 0 to train a lightweight classifier, obtain, and get the classifier output y of all sample events pred ;

[0130] Step 3.3: Define the confusion set, select a subset E” from the entire dataset, screen the samples that the classifier is most “confused” about, and select the classification output y pred for events between 0.3 and 0.7:

[0131] E″ = {x ∈ E′ ∣ 0.3 < y pred (x) < 0.7}

[0132] In the formula, E” is a subset in the dataset.

[0133] Step 3.4: Calculate the candidate correlation score, for each candidate feature E′ i (i = 2, 3,..., 9), calculate its distance correlation with the reference label y ref For each event x ∈ E″, construct the feature vector:

[0134] X = {(E′1(x), E′ i (x)) ∣ x ∈ E″}

[0135] The corresponding reference label is:

[0136] Y = {y ref (x) ∣ x ∈ E″}

[0137] Use the following formula to calculate the distance correlation dCor between X and Y:

[0138]

[0139] In the formula, the calculation formula for the distance covariance dCov2(X, Y) is:

[0140]

[0141] Wherein, X', Y', X", Y" represent random vectors that are independently and identically distributed as X and Y, ||·|| represents the Euclidean norm, and E v represents the expected value.

[0142] Step 3.5: Select the best feature, and select the feature E′ with the highest distance correlation score i , and add it to the selected feature set F 0 to form a new feature set F 1 .

[0143] Step 3.6: Retrain the classifier using the updated feature set F 1 to obtain a new classification output y pred ; repeat the process of selecting the confusion set, calculating the correlation score, and adding the best feature until the performance metrics of the classifier saturate, and stop the feature selection process.

[0144] The set after the final feature selection is represented as F k ={E′1, E′2,..., E′ n}, n = k, where k is the number of finally selected features.

[0145] Step 4: Construct a FilterDF-former diagnostic model, and update the FilterDF-former diagnostic model by introducing the DOA dream optimization algorithm, as Figure 3 shown, to form an optimized DOA-FilterDF-former model.

[0146] Step 4.1: Standardize the data F k ={E′1, E′2,..., E′ n} so that its mean is 0 and its variance is 1;

[0147]

[0148] Wherein, μ i and σ i are respectively the mean and standard deviation of the i-th healthy signal.

[0149] Step 4.2: Construct each healthy signal E′ i as a time series vector X i ={E i,1 , E i,2 ,..., E i,T}, where T is the number of time steps.

[0150] Step 4.3: Perform Dynamic Time Warping (DTW) to calculate the similarity between the time feature vectors of different health signals:

[0151] d DTW (X i ,X j ) = DTW(X i ,X j )

[0152] where X i and X j are the time series vectors of two health signals, and d is the similarity.

[0153] According to the DTW distance, select the most similar nodes as neighbors to construct the time-related graph A t , calculate the cosine similarity of the node embeddings, and select the most similar nodes as neighbors to form the learned spatio-temporal association graph A st .

[0154] Step 4.4: Expand the spatio-temporal association graph A st into the global spatio-temporal graph A g to enable nodes at different positions and time slices to interact freely:

[0155]

[0156] where A is the adjacency matrix of the topological structure graph of the health signal, A st is the adjacency matrix of the spatio-temporal association graph, and NT represents the dimension.

[0157] Step 4.5: Perform data embedding. First, perform spatio-temporal hybrid embedding. The sequence learned through the random walk strategy is used to fuse the topological information of the health signal and the time correlation of the time similarity graph to form the spatio-temporal hybrid embedding E st ;

[0158] Then perform time embedding. Positional encoding is performed on the time series information. Using a fixed encoding method, the sequence elements are associated with points on a sine or cosine curve to generate the positional embedding E t :

[0159]

[0160] where pos represents the time step, d model represents the total dimension of the embedding vector, and d represents the dimension index of the embedding vector;

[0161] Next, perform spatial embedding. Based on the topological structure graph of the health signal and the time series similarity of the nodes, information is extracted, the topological graph is calculated, and the normalized graph Laplacian matrix is obtained through matrix decomposition. Its corresponding eigenvector is used as the spatial embedding E s;

[0162] Finally, the spatio-temporal hybrid embedding E st , the temporal embedding E t and the spatial embedding E s are added together to obtain the final data embedding E data :

[0163] E data = E s + E t + E ST .

[0164] Step 4.6: Encoder part, introducing the self-attention mechanism to capture the dependencies between different positions and time steps in the input data:

[0165]

[0166] where Q, K, and V are the query, key, and value matrices respectively; QK T is the dot product operation, calculating the attention weights of Q over V; is used to scale the dot product to prevent the vanishing gradient problem; softmax() is the transfer function.

[0167] Finally, the residual connection and layer normalization are used to output the features:

[0168] E final = LayerNorm(E data + Attention(Q, K, V))

[0169] Step 4.7: Decoder part, taking the features E final processed by the encoder and the original data E data as the input of the decoder, and obtaining the final diagnosis result after the decoder's processing:

[0170] Y diag = Linear(E final + E data )

[0171] Finally, the heart disease types Y = [Y1, Y2, Y3, Y4] are output, where: Y1 is normal, Y2 is myocardial infarction, Y3 is myocardial hypertrophy, and Y4 is arrhythmia.

[0172] Step 4.8: Initialize the parameter vector, combining the initialization parameters and regularization parameters of the FilterDF-former diagnostic model into a parameter vector X = (W, λ).

[0173] Step 4.9: Population initialization, generating a random population, where each individual corresponds to a set of parameter vectors X i, where \(i = 1, 2, \ldots, N\) and \(N\) is the population size, calculate the individual fitness.

[0174] Step 4.10: Enter the exploration stage. Divide the population into 5 groups according to the difference in memory ability, and update the individuals in each group as follows:

[0175] For the \(q\)-th group (\(q = 1, 2, 3, 4, 5\)), perform the following process:

[0176] Memory strategy, reset the position of the individuals in each group to the position of the best individual in that group:

[0177]

[0178] In the formula, is the position of the best individual in the \(q\)-th group at iteration \(t\).

[0179] Forgetting and replenishment strategy, the individual updates the position information in the forgetting dimension, combining the global and local search functions. The update formula is:

[0180]

[0181] In the formula, \(\alpha\) and \(\beta\) are step factors that control the amplitude of the update; rand is a random number between 0 and 1.

[0182] Dream sharing strategy, the individual randomly obtains the position information of other individuals in the forgetting dimension to enhance the escape ability from the local optimal solution. The update formula is:

[0183]

[0184] In the formula, \(m\) is a natural number randomly selected in the range \([1, N]\) and \(m\neq i\).

[0185] Step 4.11: Enter the exploitation stage. In the exploitation stage, no grouping is performed. All individuals display the best individual before each iteration and update the position in the forgetting dimension. Perform the following process:

[0186] Memory strategy, reset the position of all individuals to the position of the global best individual:

[0187]

[0188] In the formula, is the position of the optimal individual in the whole at iteration \(t\).

[0189] Forgetting and replenishment strategy, the individual updates the position information in the forgetting dimension. The update formula is:

[0190]

[0191] Where γ and δ are step factors.

[0192] Step 4.12: Convergence judgment to determine whether the maximum number of iterations T is reached max . If it is reached, the algorithm terminates; otherwise, continue the iteration.

[0193] After the algorithm converges, select the individual with the largest fitness value from the population as the optimal solution, that is, the found optimal combination of initialization parameters and regularization parameters ((W', λ')

[0194] Step 5: Use the DOA-FilterDF-former model for diagnosis and output the diagnosis result of the heart disease type.

[0195] Step 5.1: Divide the healthy signal dataset into a training set, a validation set, and a test set according to a preset ratio;

[0196] Step 5.2: Use the divided training set to train the improved DOA-FilterDF-former model and let the model learn the patterns and relationships in the data;

[0197] Step 5.3: Use the validation set to validate the model and adjust the model parameters to complete the training;

[0198] Step 5.4: Use the trained model to diagnose the test set and obtain the diagnosis result of the heart disease type.

[0199] Step 6: Send the diagnosis result to the user client through the central processing module and provide corresponding professional suggestions.

[0200] Step 6.1: The central processing module integrates the diagnosis results output by the decoder to form a complete diagnosis report.

[0201] Step 6.2: Send the diagnosis report to the user client wirelessly or wired, such as mobile phones, tablets, etc.

[0202] Step 6.3: According to the diagnosis result, the system automatically generates corresponding professional suggestions, such as medical advice, lifestyle adjustment advice, etc., to help users better manage their health.

[0203] As Figure 4 shown, the electrocardiogram monitoring system corresponding to the electrocardiogram monitoring method includes the following modules:

[0204] The signal acquisition module includes a portable chest patch and a smart bracelet. The portable chest patch is used to collect electrocardiogram signals by attaching to the human body in the chest lead mode, and the smart bracelet is used to collect physiological signals around the body and store the electrocardiogram signals and physiological signals as healthy signals;

[0205] The signal preprocessing module includes a GF low-pass filtering unit and an IIR high-pass filtering unit. The GF low-pass filtering unit is used to eliminate power line interference and muscle artifacts in the health signal; the IIR high-pass filtering unit is used to eliminate baseline drift in the health signal.

[0206] The feature extraction module is used to process the filtered electrocardiogram signal by using the DisCo-FFS feature selection method, capture the temporal features of the electrocardiogram signal and reduce the data complexity, and construct an input data set.

[0207] The model diagnosis module is used to construct a FilterDF-former diagnosis model by introducing the DOA dream optimization algorithm.

Claims

1. A portable intelligent electrocardiogram monitoring method, characterized in that, It includes the following steps: (1) Collect electrocardiogram signals by attaching a portable chest patch to the human body in accordance with the chest lead method, and collect physiological signals by a smart bracelet. Store the electrocardiogram signals and physiological signals as health signals; (2) Use a GF low-pass filter to eliminate power line interference and muscle artifacts in the health signals, and then use an IIR high-pass filter to eliminate baseline drift in the health signals; (3) Adopt the DisCo-FFS feature selection method to capture the temporal features of the filtered health signals and reduce data complexity, and construct an input data set; (4) Construct a FilterDF-former classification model, and update the FilterDF-former classification model by introducing a DOA dream optimization algorithm to form an optimized DOA-FilterDF-former model; (5) Use the DOA-FilterDF-former model to process the input data set and output the classification result of the health signals; (6) Send the classification result of the health signals to the user client and output a preset prompt message associated with the classification result.

2. The electrocardiogram monitoring method according to claim 1, wherein, The health signal acquisition process described in step 1 includes: (11) Collect 6 chest lead electrocardiogram signals E1, E2, E3, E4, E5, and E6 through a smart chest patch. The smart chest patch is provided with V1-V6 electrode regions, corresponding to the positions of the 4th intercostal space on the right margin of the sternum, the 4th intercostal space on the left margin of the sternum, the midpoint of the connection line between V2 and V4, the intersection of the left midclavicular line and the 5th intercostal space, the same level as V4 on the left anterior axillary line, and the same level as V4 on the left midaxillary line respectively; (12) Collect physiological signals through a smart bracelet. The smart bracelet is worn on the user's wrist. The physiological signals include pulse E7, body temperature E8, and blood pressure E9 signals; (13) Store the E1-E6 electrocardiogram signals and E7-E9 physiological signals as a health signal data set E = {E1, E2, E3, E4, E5, E6, E7, E8, E9} for subsequent analysis and processing.

3. The electrocardiogram monitoring method according to claim 1, wherein, The high- and low-pass filtering processing process described in step 2 is as follows: (21) Perform segmented processing on the health signal data, divide it into 10 segments of equal-length signal data, and each segment is further divided into 9 signal sequences; (22) Apply a Gaussian low-pass filter to each signal sequence to remove high-frequency noise in the signal, reduce power line interference and muscle artifacts, and calculate the filtered output through the following formula: where y(n) is the output signal; b k is the coefficient of the filter, σ is the standard deviation of the Gaussian function, n is the order of the filter, and k is the number of delay steps; the cut-off frequency of the Gaussian low-pass filter is set to 35 Hz to 45 Hz, the passband ripple is 1 dB, and the stopband attenuation is 80 dB; (23) Apply an IIR high-pass filter to the signal sequences of each segment after low-pass filtering to remove baseline drift in the signal, and calculate the filtered output through the following formula: where a i is the feedback coefficient; m is the order of the feedback part; i is the number of delayed steps; the cut-off frequency of the IIR high-pass filter is set to 0 Hz to 0.05 Hz, the passband ripple is 1 dB, and the stopband attenuation is 80 dB; (24) Perform time synchronization processing on the signal sequences after low- and high-pass filtering, including start time alignment and sampling rate adjustment, and arrange and combine them in the original segmented order, and finally merge them to form a processed health signal data set E' = {E'1, E'2, E'3, E'4, E'5, E'6, E'7, E'8, E'9}.

4. The electrocardiogram monitoring method according to claim 1, wherein The DisCo-FFS feature selection process described in step 3 is as follows: (31)Using the processed electrocardiogram signal E'1 as the input, an initial feature set F is established 0 ={E′1}; (32) Use F 0 Train a lightweight classifier to obtain the classifier output y of all sample events pred ; (33) Screening classifier output y pred Construct a confusion set E” for samples between 0.3 and 0.7, where: E″ = {x ∈ E′ | 0.3 < y pred (x) < 0.7} In the formula, E” is a subset in the health signal data set; (34) Calculate the candidate correlation score, including for each candidate feature E i ′ (i = 2, 3,..., 9), calculate its distance correlation with the reference label y ref ; for each event x ∈ E″, construct the feature vector X = {(E1′(x), E i ′(x)) | x ∈ E″}; the corresponding reference label is Y = {y ref (x) | x ∈ E″}; use the following formula to calculate the distance correlation dCor between X and Y: where the distance covariance dCov 2 (X, Y) is calculated as follows: dCov 2 (X,Y) = E v [||X - X′||||Y - Y′||] + E v [||X - X′||]E v [||Y - Y′||] - 2E v [||X - X′||||Y - Y″||] where X', Y', X'', Y'' are random vectors that are independently and identically distributed as X and Y, ||·|| represents the Euclidean norm, and E v represents the expected value; (35) Select the feature E with the maximum dCor value i ′, and add it to the selected feature set F 0 to form a new feature set F 1 ; (36) Repeat steps (32) to (35) until the performance metrics of the classifier saturate, stop the feature selection process, and obtain the final input feature set F k ={E1′, E′2,..., E′ n}, n = k, where k is the number of finally selected features.

5. The electrocardiogram monitoring method according to claim 1, characterized in that The process of constructing the FilterDF-former diagnostic model described in step 4 is as follows: (41)Normalize the input feature set F k ={E1′, E′2,..., E′ n} so that its mean is 0 and variance is 1: where μ i and σ i are the mean and standard deviation of the i-th health signal, respectively; (42)Construct each health signal E i ′ as a time series vector X i = {E i,1 , E i,2 ,..., E i,T}, where T is the number of time steps; (43) Perform DTW (Dynamic Time Warping), including calculating the similarity between the time feature vectors of different health signals: d DTW (X i , X j ) = DTW(X i , X j ); where X i and X j are the time series vectors of two health signals, and d is the similarity; select the time feature vector of the health signal with the highest similarity according to the DTW distance as the neighbor to construct the time correlation graph A t ; calculate the cosine similarity of the node embeddings, select the most similar nodes as neighbors to form the learned spatio-temporal correlation graph A st ; (44) Expand the spatio-temporal correlation graph A st into the global spatio-temporal graph A g so that nodes at different positions and time slices can interact freely: where A is the adjacency matrix of the topological structure diagram of the health signal, and A st is the adjacency matrix of the spatio-temporal correlation diagram; NT represents the dimension; (45)Perform data embedding processing. First, perform spatio-temporal hybrid embedding. The sequence learned through the random walk strategy fuses the topological information of the health signal and the temporal correlation of the temporal similarity graph to form the spatio-temporal hybrid embedding E st ; then perform temporal embedding. Use the fixed coding method to perform position coding on the time series information, associate the sequence elements with the points on the sine or cosine curve, and generate the position embedding E t : where pos represents the time step, d model represents the total dimension of the embedding vector; d represents the dimension index of the embedding vector; Next, perform spatial embedding. Extract information based on the topological structure diagram of the health signals and the time series similarity of the nodes, calculate the topological graph, and obtain the normalized graph Laplacian matrix through matrix decomposition. Its corresponding eigenvectors serve as the spatial embedding E of the nodes s ; Finally, add the spatio-temporal hybrid embedding E st , the time embedding E t and the spatial embedding E s to obtain the final data embedding E data = E s + E t + E st ; (46) Encoder processing, introducing the self-attention mechanism to capture the dependencies between different positions and time steps in the input data: where Q, K, and V are the query, key, and value matrices respectively, and QK T is a dot product operation that calculates the attention weights of Q over V, which is used to scale the dot product to prevent the vanishing gradient problem, and softmax() is the transfer function; Finally, use residual connection and layer normalization to output feature E final = LayerNorm(E data + Attention(Q, K, V)); (47)Decoder processing, taking the feature E after encoder processing final and the original data E data as the input of the decoder. After the decoder's processing, the final diagnosis result Y diag = Linear(E final + E data ). Finally, the heart disease type Y = [Y1, Y2, Y3, Y4] is output, where: Y1 is normal, Y2 is myocardial infarction, Y3 is myocardial hypertrophy, and Y4 is arrhythmia.

6. The electrocardiogram monitoring method according to claim 1, characterized in that, The introduction of the DOA dream optimization algorithm to update the FilterDF-former diagnostic model described in step 4 includes optimizing the model initialization parameter W and the L2 regularization parameter λ. The specific process is as follows: Initialize the parameter vector, combine the FilterDF-former diagnostic model initialization parameter and the regularization parameter into a parameter vector X = (W, λ); Generate a random population, each individual corresponding to a group of parameter vectors Xi, where i = 1, 2, …, N, and N is the population size, and calculate the individual fitness; In the exploration stage, divide the population into 5 groups according to the difference in memory ability, and update the individuals in each group as follows: For the q-th group (q = 1, 2, 3, 4, 5), perform the following process: Memory strategy, reset the position of the individuals in each group to the position of the best individual in the group: In the formula, is the best individual position of the q-th group at iteration t; Forgetting and replenishment strategy, the individual updates the position information in the forgetting dimension, combining the global and local search functions, and the update formula is: In the formula, α and β are step factors that control the update amplitude, and rand is a random number between 0 and 1; Dream sharing strategy, the individual randomly obtains the position information of other individuals in the forgetting dimension to enhance the escape ability from the local optimal solution, and the update formula is: In the formula, m is a natural number randomly selected in the range of [1, N], and m ≠ i; In the exploitation stage, no grouping is performed. All individuals show the best individual before each iteration and update the position in the forgetting dimension. Perform the following process: Memory strategy, reset the position of all individuals to the position of the global best individual: Wherein, is the position of the optimal individual in the whole at iteration t; Forgetting and replenishment strategy, the individual updates the position information in the forgetting dimension, and the update formula is: In the formula, γ and δ are step factors; Convergence judgment to determine whether the preset maximum number of iterations T is reached max ; if it is reached, the algorithm terminates; otherwise, continue the iteration; After the algorithm converges, select the individual with the largest fitness value from the population as the optimal solution, that is, the found optimal combination of initialization parameters and regularization parameters (W', λ').

7. The electrocardiogram monitoring method according to claim 1, characterized in that The processing process of the DOA-FilterDF-former model described in step 5 is as follows: (51) Divide the healthy signal dataset into a training set, a validation set, and a test set according to a preset ratio; (52) Use the divided training set to train the improved DOA-FilterDF-former model to make the model learn the patterns and relationships in the data; (53) Use the validation set to validate the model and adjust the model parameters to complete the training; (54) Use the trained model to process the test set to obtain the healthy signal classification result.

8. The electrocardiogram monitoring method according to claim 1, characterized in that, The process described in step 6 is as follows: (61) Integrate the healthy signal classification result through the central processing module to form a complete result report; (62) Send the result report to the user client by wireless or wired means, and the user client includes mobile phones and tablets; (63) According to the healthy signal classification result, the system automatically generates the corresponding preset prompt information, including medical advice and lifestyle adjustment advice.

9. A portable intelligent electrocardiogram monitoring system, characterized in that, Include: The signal acquisition module includes a portable chest patch and a smart bracelet. The portable chest patch is used to collect electrocardiogram signals by attaching to the human body in the chest lead mode. The smart bracelet is used to collect physiological signals and store the electrocardiogram signals and physiological signals as health signals; The signal preprocessing module includes a GF low-pass filter unit and an IIR high-pass filter unit. The GF low-pass filter unit is used to eliminate power line interference and muscle artifacts in the health signals. The IIR high-pass filter unit is used to eliminate baseline drift in the health signals; The feature extraction module is used to process the filtered electrocardiogram signals by using the DisCo-FFS feature selection method, capture the temporal features of the electrocardiogram signals and reduce the data complexity, and construct an input data set; The model diagnosis module is used to construct a FilterDF-former diagnosis model and update the FilterDF-former diagnosis model by introducing the DOA dream optimization algorithm to form an optimized DOA-FilterDF-former model; The result output module is used to process the input data set by using the DOA-FilterDF-former model and output the classification result of the health signals; The central processing module is used to send the classification result of the health signals to the user client through the central processing module and output a preset prompt message associated with the classification result.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the portable intelligent electrocardiogram monitoring method according to any one of claims 1 to 8.

Images