Atrial fibrillation identification method and system based on unbalanced PPG signal

Abstract

The invention discloses an atrial fibrillation recognition method and system based on an unbalanced PPG signal, and the method comprises the steps: obtaining a human pulse signal, carrying out the data preprocessing and complementary set empirical mode decomposition, and obtaining an improved marginal spectrum feature with a representation amplitude-frequency relation; an optimization mechanism based on Newton-Raphson parameters is introduced, the improved marginal spectrum features with the characterization amplitude-frequency relation are recognized and classified through an error sample category weight adaptive adjustment algorithm, and an atrial fibrillation recognition result is obtained; and carrying out visualization processing on the atrial fibrillation identification result to realize atrial fibrillation monitoring. According to the method, the weights of the samples with prediction errors can be dynamically adjusted according to category distribution, the classification boundary is improved, and the atrial fibrillation monitoring precision is improved. The atrial fibrillation recognition method and system based on the unbalanced PPG signal can be widely applied to the technical field of atrial fibrillation recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of atrial fibrillation recognition, and in particular to a method and system for atrial fibrillation recognition based on unbalanced PPG signals. Background Art

[0002] Atrial fibrillation (AF) is a common persistent or paroxysmal cardiac arrhythmia characterized by irregular atrium contractions and disordered electrical activity. Photoplethysmography (PPG) is a non-invasive biomedical signal that uses photoelectric sensors to monitor blood flow and heartbeat. When the heart beats, blood flow causes the skin's blood vessels to dilate and constrict, altering blood flow beneath the skin and, in turn, affecting the intensity and waveform of the PPG signal. With the increasing popularity of wearable devices, PPG technology has also been incorporated into smart bracelets and smartwatches, collecting data such as heart rate, blood pressure, cardiac output, and blood oxygen saturation for health monitoring and fitness tracking. This technology can be used for continuous heart rate monitoring and early cardiovascular disease screening. Combining deep learning and artificial intelligence technologies can enable more accurate and personalized atrial fibrillation monitoring and management, providing better medical services and health management for patients. By analyzing the waveform characteristics of the PPG signal, it is possible to distinguish AF from normal sinus rhythm. However, the low prevalence of atrial fibrillation (AF) leads to a widespread class imbalance in PPG signal datasets used to detect AF, with normal heart rhythm data far outnumbering AF data. This data distribution not only causes the model to favor the normal class during training, ignoring AF samples, but also can affect the generalization ability of the classification model, leading to an increased missed diagnosis rate for AF.

[0003] However, related technologies are time-consuming and inefficient in diagnosing atrial fibrillation. The models used are relatively simple and may suffer from problems such as data underfitting and poor generalization ability. They are unable to perform predictions and classifications well. When training with data with unbalanced categories, the model decision boundary is prone to bias toward the majority class and ignore minority class samples. In addition, a large amount of manual adjustment of parameters and model parameters is required, which is time-consuming and labor-intensive, and is also subject to human subjectivity, which may lead to a decrease in recognition rate. Summary of the Invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide a method and system for identifying atrial fibrillation based on unbalanced PPG signals, which can dynamically adjust the weights of samples with prediction errors according to category distribution, improve classification boundaries, and improve the accuracy of atrial fibrillation monitoring.

[0005] The first technical solution adopted by the present invention is: a method for identifying atrial fibrillation based on an unbalanced PPG signal, comprising the following steps:

[0006] The human pulse signal is acquired and data preprocessed and complementary set empirical mode decomposition is performed to obtain an improved marginal spectrum feature that represents the amplitude-frequency relationship;

[0007] A Newton-Raphson parameter optimization mechanism was introduced, and an adaptive error sample weight adjustment algorithm was used to identify and classify the improved marginal spectrum features that represent the amplitude-frequency relationship, thereby obtaining atrial fibrillation recognition results.

[0008] The atrial fibrillation recognition results are visualized to realize atrial fibrillation monitoring.

[0009] Furthermore, the step of acquiring a human pulse signal and performing data preprocessing and complementary set empirical mode decomposition to obtain an improved marginal spectrum feature having a representation of an amplitude-frequency relationship specifically includes:

[0010] Collect human pulse signals through a photoelectric capacitance pulse sensor;

[0011] De-noising, filtering and artifact removal are performed on the human pulse signal in sequence to obtain a pre-processed human pulse signal;

[0012] The pre-processed human pulse signal is decomposed by complementary set empirical mode decomposition algorithm to obtain several intrinsic mode functions;

[0013] The Hilbert-Huang transform is performed on several intrinsic mode functions to obtain improved marginal spectrum features that characterize the amplitude-frequency relationship.

[0014] Furthermore, the step of introducing a Newton-Raphson parameter optimization mechanism and identifying and classifying the improved marginal spectrum features representing the amplitude-frequency relationship through an error sample category weight adaptive adjustment algorithm to obtain an atrial fibrillation identification result specifically includes:

[0015] Based on the deep extreme learning machine, L1 regularization is introduced to construct a sparse regularized DELM weak classifier, which includes an input layer, a hidden layer and an output layer;

[0016] Based on the improved marginal spectrum feature that characterizes the amplitude-frequency relationship, the sparse regularized DELM weak classifier is pre-trained to obtain the trained sparse regularized DELM weak classifier;

[0017] The improved marginal spectrum features that represent the amplitude-frequency relationship are weighted to construct sample data with initialized sample weights;

[0018] Based on the Newton-Raphson parameter optimization mechanism, the sample data with initialized sample weights are input into the trained sparse regularized DELM weak classifier to adaptively adjust the weights of the error sample categories to obtain a strong classifier.

[0019] The improved marginal spectrum features that represent the amplitude-frequency relationship are identified and classified by a strong classifier to obtain the atrial fibrillation recognition results.

[0020] Furthermore, the step of pre-training the sparse regularized DELM weak classifier based on the improved marginal spectrum feature having the characteristic amplitude-frequency relationship to obtain the trained sparse regularized DELM weak classifier specifically includes:

[0021] Generate the weight matrix from the input layer to the hidden layer of the sparse regularized DELM weak classifier through random normal distribution, and initialize the bias term from the input layer to the hidden layer to obtain the initialized sparse regularized DELM weak classifier;

[0022] The improved marginal spectrum features that characterize the amplitude-frequency relationship are input into the initialized sparse regularized DELM weak classifier for feature mapping to obtain the mapped improved marginal spectrum features;

[0023] The mapped improved marginal spectrum features are trained through a linear regression model to obtain the output layer weights of the sparse regularized DELM weak classifier;

[0024] The output layer weights are optimized by the least square method to obtain the optimized output layer weights;

[0025] The optimized output layer weights and the mapped improved marginal spectrum features are constrained by L1 regularization to obtain the trained sparse regularized DELM weak classifier.

[0026] Furthermore, the step of inputting the sample data with initialized sample weights into the trained sparse regularized DELM weak classifier based on the Newton-Raphson parameter optimization mechanism to adaptively adjust the weights of the error sample categories to obtain a strong classifier specifically includes:

[0027] The sample data with initialized sample weights are input into the trained sparse regularized DELM weak classifier for recognition, and the prediction results of this round of weak classifier are obtained;

[0028] Get the global error rate based on the prediction results of this round of weak classifiers;

[0029] Determine the weak classification weight according to the global error rate;

[0030] The Newton-Raphson parameter optimization mechanism is introduced to update the sample weights of the trained sparse regularized DELM weak classifier according to the weak classification weights to obtain the updated sparse regularized DELM weak classifier.

[0031] The sample data with initialized sample weights are cyclically input into the updated sparse regularized DELM weak classifier to perform the sample weight update step until the preset number of iterations is met to obtain a strong classifier.

[0032] Furthermore, the expression for obtaining the global error rate based on the prediction results of the current round of weak classifiers is specifically as follows:

[0033]

[0034] In the above formula, ε t represents the global error rate, i represents the i-th sample data, h t (X i ) is the prediction result of the weak classifier of the i-th sample in the t-th round, Y i represents the true label of the i-th sample, D t (i) represents sample data with initialized sample weights.

[0035] Furthermore, the expression for determining the weak classification weight according to the global error rate is specifically as follows:

[0036]

[0037] In the above formula, represents the weak classification weight, ε t Represents the global error rate.

[0038] Furthermore, the expression for performing sample weight update processing on the trained sparse regularized DELM weak classifier according to the weak classification weight is specifically as follows:

[0039]

[0040]

[0041] In the above formula, D t+1 (i) 1 indicates the weight update of the predicted correct sample, D t+1 (i) 2 indicates the prediction error of the majority class healthy sample update, D t+1 (i) 3 indicates the update of the predicted minority atrial fibrillation sample, D t (i) represents sample data with initialized sample weights, represents the weak classification weight, α represents the value used to adjust the attention to FN, β represents the value used to adjust the attention to FP, γ represents the value used to adjust the amplitude of sample update, FN represents missed diagnosis samples, FP represents misdiagnosed samples, TN represents healthy samples, and TP represents atrial fibrillation samples.

[0042] Furthermore, the expression of the strong classifier is specifically as follows:

[0043]

[0044] In the above formula, H(x) represents a strong classifier, T represents the total number of iterations, and t represents the tth round of iteration. represents the weak classification weight, h t (x) represents the prediction result of the weak classifier in this round.

[0045] The second technical solution adopted by the present invention is: an atrial fibrillation recognition system based on unbalanced PPG signals, comprising:

[0046] The first module is used to obtain human pulse signals and perform data preprocessing and complementary set empirical mode decomposition to obtain improved marginal spectrum features that characterize the amplitude-frequency relationship;

[0047] The second module is used to introduce a Newton-Raphson parameter optimization mechanism and use an adaptive error sample weight adjustment algorithm to identify and classify the improved marginal spectrum features that represent the amplitude-frequency relationship to obtain atrial fibrillation recognition results.

[0048] The third module is used to visualize the atrial fibrillation recognition results to achieve atrial fibrillation monitoring.

[0049] The beneficial effects of the method and system of the present invention are as follows: the present invention obtains human pulse signals and performs data preprocessing and complementary set empirical mode decomposition to obtain improved marginal spectrum features with characterization of amplitude-frequency relationship, further introduces a Newton-Raphson parameter optimization mechanism, and uses an error sample category weight adaptive adjustment algorithm to identify and classify the improved marginal spectrum features with characterization of amplitude-frequency relationship to obtain atrial fibrillation recognition results. By introducing a Newton-Raphson parameter optimization mechanism and taking the model's atrial fibrillation recognition accuracy as input, the optimal hyperparameter combination of the model can be determined quickly and accurately. By using an error sample category weight adaptive adjustment algorithm, the weights of the predicted errors are dynamically adjusted according to the category distribution, with special attention paid to the weight improvement of minority atrial fibrillation samples, improving the classification boundary, and improving the accuracy of atrial fibrillation monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flowchart of the steps of a method for identifying atrial fibrillation based on an unbalanced PPG signal according to the present invention;

[0051] Figure 2 This is a structural block diagram of an atrial fibrillation recognition system based on unbalanced PPG signals according to the present invention;

[0052] Figure 3 1 is a schematic diagram of a pulse signal acquisition process according to a specific embodiment of the present invention;

[0053] Figure 4 1 is a schematic diagram of a marginal spectrum feature extraction process according to a specific embodiment of the present invention;

[0054] Figure 5 1 is a flow chart of an algorithm for adaptively adjusting the weight of error sample categories provided by a specific embodiment of the present invention;

[0055] Figure 6 It is a schematic diagram of a process of Newton-Raphson parameter optimization provided by a specific embodiment of the present invention. DETAILED DESCRIPTION

[0056] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are provided for ease of description only and do not limit the order of the steps. The order of execution of the steps in the embodiments can be adaptively adjusted based on the understanding of those skilled in the art.

[0057] Reference Figure 1 The present invention provides a method for identifying atrial fibrillation based on an unbalanced PPG signal, the method comprising the following steps:

[0058] S100, acquiring a human pulse signal and performing data preprocessing and complementary set empirical mode decomposition to obtain an improved marginal spectrum feature that characterizes an amplitude-frequency relationship;

[0059] S110, collecting human pulse signals through a photoelectric capacitance pulse sensor;

[0060] S120, performing denoising, filtering, and artifact removal processing on the human pulse signal in sequence to obtain a preprocessed human pulse signal;

[0061] In this embodiment, if Figure 3 As shown in the figure, a photoplethysmography sensor is used to collect PPG signals, and then the signals are preprocessed by denoising, filtering, and removing artifacts. The preprocessed PPG signal data is then transmitted to the pulse signal processing module for signal processing.

[0062] S130, decomposing the preprocessed human pulse signal using a complementary set empirical mode decomposition algorithm to obtain a number of intrinsic mode functions;

[0063] S140. Performing Hilbert-Huang transform on several intrinsic mode functions to obtain improved marginal spectrum features that characterize the amplitude-frequency relationship.

[0064] In this embodiment, the collected PPG signal is decomposed by complementary ensemble empirical mode decomposition (CEEMD), and the marginal spectrum features are obtained by Hilbert-Huang (HHT) transform. The feature extraction process is as follows: Figure 4As shown in Figure 2, the marginal spectrum features are the amplitude and frequency characteristics of the PPG signal. It should be noted that the preprocessed PPG signal is first decomposed using the complementary ensemble empirical mode decomposition (CEEMD). Using the CEEMD algorithm, the original signal is adaptively decomposed into several intrinsic mode functions (IMFs), denoted as IMF1, IMF2, IMF3, and so on to IMFn. Each IMF component represents a frequency component at a different time scale in the signal and satisfies the physical meaning of instantaneous frequency. A Hilbert-Huang transform is then performed on each decomposed IMF component. Specifically, the Hilbert transform is performed on each IMF component to calculate its instantaneous frequency and instantaneous amplitude to generate an analytical signal. The instantaneous frequency and amplitude of each IMF component are integrated in the frequency domain to obtain the marginal spectrum of the signal. The marginal spectrum reflects the energy distribution characteristics of the PPG signal at different frequencies and contains joint characteristics of amplitude and frequency. The marginal spectrum features of all IMF components are extracted and integrated into a feature set that characterizes the PPG signal.

[0065] S200, introducing a Newton-Raphson parameter optimization mechanism, and identifying and classifying the improved marginal spectrum features that characterize the amplitude-frequency relationship through an error sample category weight adaptive adjustment algorithm to obtain an atrial fibrillation recognition result;

[0066] First of all, it should be noted that the Adaboost algorithm can enhance a weak learner whose prediction accuracy is only slightly higher than random guessing into a strong learner with high prediction accuracy. The characteristic of Adaboost is to increase the weight of the predicted error class samples, so that they will receive more attention in the next round of iteration, allowing the model to focus on difficult-to-classify samples and improve the overall classification accuracy. However, when processing PPG data with imbalanced categories, traditional Adaboost does not distinguish between the majority class (healthy samples) and minority class (atrial fibrillation samples) samples that are predicted incorrectly. Even if the weight of the erroneous minority class samples is increased, due to the low proportion of minority class samples in the overall data set, the erroneous minority class samples after the weight increase still do not receive enough attention from the model, resulting in the classification model decision boundary being biased towards the majority class, and ultimately resulting in a serious problem of a large number of missed diagnoses. In response to the demand for rapid and accurate detection of atrial fibrillation based on category-imbalanced PPG signals, an integrated boosting algorithm (Weight-Boost) based on the Adaboost algorithm with adaptive adjustment of the weights of the error sample categories is proposed. The Weight-Boost algorithm process is as follows: Figure 5 shown.

[0067] S210. Based on a deep extreme learning machine, L1 regularization is introduced to construct a sparse regularized DELM weak classifier, wherein the sparse regularized DELM weak classifier includes an input layer, a hidden layer, and an output layer;

[0068] In this embodiment, the Deep Extreme Learning Machine (DELM) is a fast and efficient multi-hidden layer feedforward neural network characterized by the random generation of weights and biases for the input and hidden layers. This does not require updating each round, and only the weights of the output layer are trained. Due to the randomness of the weights and biases, the high-dimensional features generated by the hidden layer often contain a large number of linearly correlated or nearly correlated components. Some of the generated features do not contribute to the final classification task, but instead increase model complexity and training time. L1 regularization is introduced to increase sparsity by summing the absolute values ​​of the weight parameters, causing some weights to approach zero, indirectly removing invalid features and greatly shortening classification time.

[0069] S220, pre-training the sparse regularized DELM weak classifier based on the improved marginal spectrum feature that characterizes the amplitude-frequency relationship to obtain a trained sparse regularized DELM weak classifier;

[0070] Specifically, a weight matrix from the input layer to the hidden layer of the sparse regularized DELM weak classifier is generated by random normal distribution, and the bias term from the input layer to the hidden layer is initialized to obtain the initialized sparse regularized DELM weak classifier; the improved marginal spectrum features with characterization of the amplitude-frequency relationship are input into the initialized sparse regularized DELM weak classifier for feature mapping to obtain the mapped improved marginal spectrum features; the mapped improved marginal spectrum features are trained by a linear regression model to obtain the output layer weights of the sparse regularized DELM weak classifier; the output layer weights are optimized by the least squares method to obtain the optimized output layer weights; the optimized output layer weights and the mapped improved marginal spectrum features are constrained by L1 regularization to obtain the trained sparse regularized DELM weak classifier.

[0071] In this embodiment, the data is first input, with the normalized preprocessed marginal spectrum as input. Then, random weights and biases are assigned, using a random normal distribution to generate the weight matrix W from the input layer to the hidden layer, and the bias term b is randomly initialized.

[0072] Further feature mapping is performed, and the input data is mapped to a high-dimensional feature space using randomly initialized weights and biases. The hidden layer output H is generated through the nonlinear activation function sigmoid, where the formula for obtaining H is as follows:

[0073] H=δ(WX+b)

[0074] Among them, δ is the sigmoid activation function, W is the weight matrix, b is the bias term, and X is the input data.

[0075] Then, regression optimization is performed, and the mapped feature H is trained using a linear regression model to obtain the output layer weight β. The objective function is as follows:

[0076]

[0077] The objective function is to minimize the regression error. The optimal solution of β can be obtained by the least squares method as shown below:

[0078] β=YH T (H T H) -1

[0079] Among them, Y is the true label of X, H T is the transpose of the output matrix H.

[0080] Further sparse regularization is performed. During the regression optimization process, L1 regularization constraint β and hidden layer output H are applied, and its expression is:

[0081]

[0082] Among them, λ1 and λ2 control the degree of sparsification.

[0083] Then, feature selection is performed. Through the sparsification effect of L1 regularization, unimportant features are automatically eliminated to make the model simpler and improve the generalization ability of the model. Finally, the prediction results are output and the optimized regression weight β and feature map H are used to predict new data.

[0084] S230, assigning weights to the improved marginal spectrum features that represent the amplitude-frequency relationship, and constructing sample data with initialized sample weights;

[0085] S240, based on the Newton-Raphson parameter optimization mechanism, input the sample data with the initialized sample weights into the trained sparse regularized DELM weak classifier to perform adaptive adjustment on the weights of the error sample categories to obtain a strong classifier;

[0086] Specifically, the sample data with initialized sample weights are input into the trained sparse regularized DELM weak classifier for identification to obtain the prediction results of this round of weak classifiers; the global error rate is obtained according to the prediction results of this round of weak classifiers; the weak classification weights are determined according to the global error rate; the Newton-Raphson parameter optimization mechanism is introduced, and the sample weights of the trained sparse regularized DELM weak classifier are updated according to the weak classification weights to obtain an updated sparse regularized DELM weak classifier; the sample data with initialized sample weights are cyclically input into the updated sparse regularized DELM weak classifier to perform the sample weight updating step until the preset number of iterations is met to obtain a strong classifier.

[0087] In this embodiment, the sample weight D is initialized. Assuming there are N training sample data, each training sample is initially assigned the same weight 1 / N, which is expressed as follows:

[0088]

[0089] Where N is the total number of samples, i=1,…,N.

[0090] Further training of the sparse regularized DELM weak classifier, the samples with initialized sample weights are used as input to train the sparse regularized DELM weak classifier.

[0091] Then calculate the global error rate, and calculate the global error rate ε according to the actual prediction effect of this round t , the calculation formula is as follows:

[0092]

[0093] Among them, h t (X i ) is the prediction result h of the i-th sample in the t-th round weak classifier t (x i )∈{-1,1}. Y i is the true label Y of the i-th sample i ∈{-1,1}.

[0094] And calculate the weak classifier weights, calculate the weak classification weights according to the current global error rate The calculation formula is as follows:

[0095]

[0096] In the above formula, represents the weak classification weight, ε t Represents the global error rate.

[0097] Finally, the sample weights are updated. The weights of correctly predicted samples are reduced in the next round, and the weights of incorrectly predicted samples are updated according to their category distribution. The updated weights are normalized, including:

[0098] The weight of the predicted correct sample is updated as follows:

[0099]

[0100] The update of the predicted majority class (healthy) sample is as follows:

[0101]

[0102] The update of the incorrectly predicted minority class (atrial fibrillation) sample is as follows:

[0103]

[0104] In the above formula, D t+1 (i) 1 indicates the weight update of the predicted correct sample, D t+1 (i) 2 indicates the prediction error of the majority class healthy sample update, D t+1 (i) 3 indicates the update of the predicted minority atrial fibrillation sample, D t (i) represents sample data with initialized sample weights, represents the weak classification weight, α represents the value used to adjust the attention to FN, β represents the value used to adjust the attention to FP, γ represents the value used to adjust the amplitude of sample update, FN represents missed diagnosis samples, FP represents misdiagnosed samples, TN represents healthy samples, and TP represents atrial fibrillation samples.

[0105] Then a strong classifier is formed. If the iteration conditions are met, the weak classifiers are weighted according to the weights to form the final strong classifier. The strong classifier formula is as follows:

[0106]

[0107] In the above formula, H(x) represents a strong classifier, T represents the total number of iterations, and t represents the tth round of iteration. represents the weak classification weight, h t (x) represents the prediction result of the weak classifier in this round.

[0108] It should also be noted that in order to solve the problem of frequent manual adjustment of the three hyperparameters α, β, and γ during the sample weight update process, the NRBO parameter optimization module is introduced. The model's recognition accuracy of atrial fibrillation is used as input, and the NRSR search rule and TAO avoidance of local optimal strategy are used to calculate the optimal hyperparameter combination. The NRBO parameter optimization network structure is as follows: Figure 6 As shown in the figure, the NRSR search rule combines the fast convergence and regularization of the Newton method to ensure stable and efficient parameter updates by dynamically adjusting the step size and learning rate. The TAO strategy uses mechanisms such as multi-point initialization, random perturbations, and furnace annealing to prevent the optimization process from falling into local optima and gradually guide the global search towards local convergence. This combination enables automated optimization within the NRBO module, reducing the complexity of manual hyperparameter tuning while improving model performance.

[0109] S250 , identifying and classifying the improved marginal spectrum features that characterize the amplitude-frequency relationship through a strong classifier to obtain an atrial fibrillation identification result.

[0110] S300: Visualize the atrial fibrillation recognition result to achieve atrial fibrillation monitoring.

[0111] In summary, the embodiment of the present invention collects the original human pulse signal, and after the collection is completed, it is pre-processed by filtering, denoising, removing artifacts, etc., and then transmitted to the feature processing module for feature extraction; the pre-processed signal is further decomposed by CEEMD to obtain multiple IMF components of different frequencies, and then the spectrum energy value of each IMF component is calculated, and then the Hilbert-Huang Transform (HHT) method based on the CEEMD decomposition method is used for feature extraction, and finally the improved Hilbert spectrum obtained is integrated on the time axis to obtain the improved marginal spectrum feature that characterizes the amplitude-frequency relationship; the Weight-Boost classification model is pre-trained The method is trained, and a parameter adaptive optimization method is selected to further optimize its atrial fibrillation detection performance. The validation set is then used to evaluate the performance of the model. The marginal spectrum features are input into the trained Weight-Boost model to identify and classify the feature data to obtain the diagnosis results. The diagnostic recognition rate is also calculated for subsequent system optimization. Finally, the model recognition rate is obtained for feedback optimization. According to the NRSR search rule and the working principle of TAO to avoid local optimality, the error sample weight of the diagnosis module is optimized and adjusted in a direction that is conducive to improving the diagnostic performance. The final results are displayed, and warnings and health recommendations are made based on the signal calculation results. The embodiment of the present invention proposes a weight adaptive boosting algorithm (NRBO-WeightBoost) based on Newton-Raphson (NRBO) and adaptive boosting (Adaboost). This method uses an improved DELM as the basic classifier, fully considers the influence of the current number of missed and misdiagnosed samples in the error sample weight update process, adaptively assigns corresponding update weights to error samples of different categories, and uses NRBO as the weight adaptive process to find the optimal hyperparameter combination to achieve the best atrial fibrillation diagnosis effect.

[0112] Therefore, compared with the prior art, the embodiments of the present invention have the following improvements:

[0113] 1) Fast feature extraction method: CEEMD decomposition combined with Hilbert-Huang transform is used to extract the marginal spectral features of PPG signals to improve the efficiency and accuracy of feature extraction.

[0114] 2) Propose sparse regularized DELM: introduce L1 regularization to enhance feature sparsity, optimize model performance and reduce computational complexity.

[0115] 3) Adaptive adjustment mechanism for the weight of incorrect sample categories: The weight of samples with prediction errors is dynamically adjusted according to the category distribution, with special attention paid to the weight increase of minority classes (atrial fibrillation samples) to improve the classification boundary.

[0116] 4) Diagnostic performance improvement method based on dynamic feedback optimization: Through the NRBO module, continuous feedback is used to optimize the model weight distribution and gradually improve the accuracy of atrial fibrillation detection.

[0117] 5) Comprehensively apply fast feature extraction methods, deep learning technology, category-based weight update and feedback optimization mechanism to build a comprehensive optimized diagnostic system to improve the overall performance and accuracy of AF detection.

[0118] Reference Figure 2 , an atrial fibrillation recognition system based on unbalanced PPG signals, comprising:

[0119] The first module 201 is used to obtain a human pulse signal and perform data preprocessing and complementary set empirical mode decomposition to obtain an improved marginal spectrum feature that represents the amplitude-frequency relationship;

[0120] The second module 202 is used to introduce a Newton-Raphson parameter optimization mechanism and identify and classify the improved marginal spectrum features that represent the amplitude-frequency relationship through an error sample category weight adaptive adjustment algorithm to obtain an atrial fibrillation recognition result;

[0121] The third module 203 is used to visualize the atrial fibrillation recognition result to achieve atrial fibrillation monitoring.

[0122] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0123] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for identifying atrial fibrillation based on unbalanced PPG signals, characterized in that: The following steps are involved: The human pulse signal is acquired and data preprocessed and complementary set empirical mode decomposition is performed to obtain an improved marginal spectrum feature that represents the amplitude-frequency relationship; A Newton-Raphson parameter optimization mechanism was introduced, and an adaptive error sample weight adjustment algorithm was used to identify and classify the improved marginal spectrum features that represent the amplitude-frequency relationship, thereby obtaining atrial fibrillation recognition results. The atrial fibrillation recognition results are visualized to realize atrial fibrillation monitoring.

2. The atrial fibrillation identification method based on unbalanced PPG signals according to claim 1, characterized in that: The step of acquiring a human pulse signal and performing data preprocessing and complementary set empirical mode decomposition to obtain an improved marginal spectrum feature having an amplitude-frequency relationship specifically includes: Collect human pulse signals through a photoelectric capacitance pulse sensor; De-noising, filtering and artifact removal are performed on the human pulse signal in sequence to obtain a pre-processed human pulse signal; The pre-processed human pulse signal is decomposed by complementary set empirical mode decomposition algorithm to obtain several intrinsic mode functions; The Hilbert-Huang transform is performed on several intrinsic mode functions to obtain improved marginal spectrum features that characterize the amplitude-frequency relationship.

3. The atrial fibrillation identification method based on unbalanced PPG signals according to claim 2, characterized in that: The step of introducing a Newton-Raphson parameter optimization mechanism and identifying and classifying the improved marginal spectrum features representing the amplitude-frequency relationship through an error sample category weight adaptive adjustment algorithm to obtain an atrial fibrillation recognition result specifically includes: Based on the deep extreme learning machine, L1 regularization is introduced to construct a sparse regularized DELM weak classifier, which includes an input layer, a hidden layer and an output layer; Based on the improved marginal spectrum feature that characterizes the amplitude-frequency relationship, the sparse regularized DELM weak classifier is pre-trained to obtain the trained sparse regularized DELM weak classifier; The improved marginal spectrum features that represent the amplitude-frequency relationship are weighted to construct sample data with initialized sample weights; Based on the Newton-Raphson parameter optimization mechanism, the sample data with initialized sample weights are input into the trained sparse regularized DELM weak classifier to adaptively adjust the weights of the error sample categories to obtain a strong classifier. The improved marginal spectrum features that represent the amplitude-frequency relationship are identified and classified by a strong classifier to obtain the atrial fibrillation recognition results.

4. The atrial fibrillation identification method based on unbalanced PPG signals according to claim 3, characterized in that: The step of pre-training the sparse regularized DELM weak classifier based on the improved marginal spectrum feature having the characteristic of the amplitude-frequency relationship to obtain the trained sparse regularized DELM weak classifier specifically includes: Generate the weight matrix from the input layer to the hidden layer of the sparse regularized DELM weak classifier through random normal distribution, and initialize the bias term from the input layer to the hidden layer to obtain the initialized sparse regularized DELM weak classifier; The improved marginal spectrum features that characterize the amplitude-frequency relationship are input into the initialized sparse regularized DELM weak classifier for feature mapping to obtain the mapped improved marginal spectrum features; The mapped improved marginal spectrum features are trained through a linear regression model to obtain the output layer weights of the sparse regularized DELM weak classifier; The output layer weights are optimized by the least square method to obtain the optimized output layer weights; The optimized output layer weights and the mapped improved marginal spectrum features are constrained by L1 regularization to obtain the trained sparse regularized DELM weak classifier.

5. The atrial fibrillation identification method based on unbalanced PPG signals according to claim 4, characterized in that: The step of inputting sample data with initialized sample weights into the trained sparse regularized DELM weak classifier based on the Newton-Raphson parameter optimization mechanism to adaptively adjust the weights of the error sample categories to obtain a strong classifier specifically includes: The sample data with initialized sample weights are input into the trained sparse regularized DELM weak classifier for recognition, and the prediction results of this round of weak classifier are obtained; Get the global error rate based on the prediction results of this round of weak classifiers; Determine the weak classification weight according to the global error rate; The Newton-Raphson parameter optimization mechanism is introduced to update the sample weights of the trained sparse regularized DELM weak classifier according to the weak classification weights to obtain the updated sparse regularized DELM weak classifier. The sample data with initialized sample weights are cyclically input into the updated sparse regularized DELM weak classifier to perform the sample weight update step until the preset number of iterations is met to obtain a strong classifier.

6. The atrial fibrillation identification method based on unbalanced PPG signals according to claim 5, characterized in that: The expression for obtaining the global error rate based on the prediction results of the current round of weak classifiers is specifically as follows: In the above formula, ε t represents the global error rate, i represents the i-th sample data, h t (X i ) is the prediction result of the weak classifier of the i-th sample in the t-th round, Y i represents the true label of the i-th sample, D t (i) represents sample data with initialized sample weights.

7. The atrial fibrillation identification method based on unbalanced PPG signals according to claim 6, characterized in that: The expression for determining the weak classification weight according to the global error rate is specifically as follows: In the above formula, represents the weak classification weight, ε t Represents the global error rate.

8. The atrial fibrillation identification method based on unbalanced PPG signals according to claim 7, characterized in that: The expression for performing sample weight update processing on the trained sparse regularized DELM weak classifier according to the weak classification weight is specifically as follows: In the above formula, D t+1 (i) 1 indicates the weight update of the predicted correct sample, D t+1 (i) 2 indicates the prediction error of the majority class healthy sample update, D t+1 (i) 3 indicates the update of the predicted minority atrial fibrillation sample, D t (i) represents sample data with initialized sample weights, represents the weak classification weight, α represents the value used to adjust the attention to FN, β represents the value used to adjust the attention to FP, γ represents the value used to adjust the amplitude of sample update, FN represents missed diagnosis samples, FP represents misdiagnosed samples, TN represents healthy samples, and TP represents atrial fibrillation samples.

9. The atrial fibrillation identification method based on unbalanced PPG signals according to claim 8, characterized in that: The expression of the strong classifier is specifically as follows: In the above formula, H(x) represents a strong classifier, T represents the total number of iterations, and t represents the tth round of iteration. represents the weak classification weight, h t (x) represents the prediction result of the weak classifier in this round.

10. An atrial fibrillation recognition system based on unbalanced PPG signals, characterized in that: Includes the following modules: The first module is used to obtain human pulse signals and perform data preprocessing and complementary set empirical mode decomposition to obtain improved marginal spectrum features that characterize the amplitude-frequency relationship; The second module is used to introduce a Newton-Raphson parameter optimization mechanism and use an adaptive error sample weight adjustment algorithm to identify and classify the improved marginal spectrum features that represent the amplitude-frequency relationship to obtain atrial fibrillation recognition results. The third module is used to visualize the atrial fibrillation recognition results to achieve atrial fibrillation monitoring.

Images