CN-121971105-A - Atrial fibrillation anomaly detection method and system based on multi-sensing time-frequency pseudo anomaly

CN121971105ACN 121971105 ACN121971105 ACN 121971105ACN-121971105-A

Abstract

The invention discloses an atrial fibrillation anomaly detection method and system based on multi-sensing time-frequency pseudo anomalies, the method comprises the steps of obtaining an atrial electrophysiological signal sequence through an HDS high-density sensing atrial fibrillation analysis platform, constructing a multi-mode time-frequency tensor model after performing time-frequency conversion on the signal sequence, extracting a time-frequency feature subset related to atrial fibrillation anomalies, dynamically distributing feature weights to optimize features by utilizing an extrusion excitation time-frequency pseudo-selection model, constructing a double-domain hypersphere feature space based on the optimized features, calculating pseudo anomaly distance values through a double-domain hypersphere pseudo-anomaly distance calculation algorithm, dividing a suspected atrial fibrillation anomaly feature sample set by combining a preset threshold value, performing feature fusion on the suspected anomaly sample set, extracting common anomaly features and generating detection results. The system corresponds to six units, the comprehensive feature extraction and the abnormality recognition accuracy are greatly improved, the requirements of clinic on atrial fibrillation detection are met, and reliable technical support is provided for early atrial fibrillation diagnosis.

Inventors

LUO TAO
LIN TAO
Xiang Zhongzheng
FAN LINGJIE
GAO XIAOLI
LI WENYAO

Assignees

四川大学

Dates

Publication Date: 20260505
Application Date: 20260104

Claims (10)

1. S1, acquiring an atrial electrophysiological signal sequence through an HDS high-density sensing atrial fibrillation analysis platform, wherein the sequence comprises time domain signal data acquired by a plurality of sensing channels, and performing time-frequency domain conversion processing on the time domain signal data to obtain a multichannel time-frequency signal matrix; the method comprises the steps of S2, inputting a multi-channel time-frequency signal matrix into a multi-mode time-frequency tensor optimization algorithm, constructing a multi-mode time-frequency tensor model, conducting hierarchical optimization on mode dimensions in the tensor model, extracting time-frequency characteristic components in each mode through tensor decomposition operation, screening out a time-frequency characteristic subset related to atrial fibrillation, S3, inputting the time-frequency characteristic subset into an extrusion excitation time-frequency pseudo-selection model, conducting dynamic distribution on characteristic weights of different frequency segments in the time-frequency characteristic subset, enhancing characteristic capacity of atrial fibrillation abnormal characteristics, and outputting an optimized time-frequency characteristic matrix, S4, adopting a double-domain hyper-sphere pseudo-anomaly distance calculation algorithm to construct a double-domain hyper-sphere characteristic space, calculating pseudo-anomaly distance values of each time-frequency characteristic vector in the double-domain hyper-sphere characteristic space, S5, setting a pseudo-anomaly distance threshold, comparing the pseudo-anomaly distance values with the threshold, dividing a characteristic sample set of suspected atrial fibrillation abnormal and a normal characteristic set, S6, conducting characteristic fusion on the characteristic set of the atrial fibrillation abnormal characteristic set, and conducting characteristic fusion on the characteristic set of the suspected atrial fibrillation abnormal characteristic set.
2. The method for detecting abnormal atrial fibrillation based on multi-sensing time-frequency pseudo-anomalies according to claim 1, wherein the expression of the multi-modal time-frequency tensor optimization algorithm is: , is a multi-modal time-frequency tensor, For the number of sensing channels, For the number of time-domain sampling points, As the number of frequency bins to be counted, Is the first The first channel Sample point number The original time-frequency signal values of the frequency bins, In order for the parameters to be regularized, The matrix is constrained for atrial fibrillation anomaly characteristics, For the atrial fibrillation signal covariance matrix, In order to calculate the tensor product, Is a matrix trace operation.
3. The multi-sensor time-frequency pseudo-anomaly based atrial fibrillation anomaly detection method according to claim 1, wherein the expression of the two-domain hyper-sphere pseudo-anomaly distance calculation algorithm is: , Is a feature vector Is a pseudo-outlier distance value of (c), For the atrial fibrillation anomaly characteristic mean vector, Is a two-domain hyper-sphere weight matrix, For the abnormal characteristic variance of atrial fibrillation, Is the two-domain fusion coefficient of the two-domain fusion coefficient, In the form of a matrix determinant operation, Is an atrial fibrillation signal covariance matrix.
4. The multi-sensor time-frequency pseudo-anomaly based atrial fibrillation anomaly detection method of claim 1, characterized in that the expression of the extrusion excitation time-frequency pseudo-selection model is: , In order to squeeze the excited time-frequency characteristic value, Is the characteristic value of the original time-frequency, In order to be a frequency of the light, As the sampling points in the time domain, The function is activated for Sigmoid, For the full-connection layer weight matrix, For the function to be activated by the ReLU, The time-frequency characteristic mean value under the same frequency segment.
5. The method for detecting atrial fibrillation abnormalities based on multi-sensing time-frequency pseudo-abnormalities according to claim 1, characterized in that in said step S2, the multi-modal time-frequency tensor optimization algorithm further satisfies: , In the form of a channel mode factor matrix, Is a matrix of time-domain modal factors, As a matrix of frequency modal factors, The dimensions of each modal factor matrix are respectively, As a core tensor element, the formula is used to further refine feature extraction in the tensor decomposition process.
6. The method for detecting atrial fibrillation abnormalities based on multi-sensing time-frequency pseudo-abnormalities according to claim 1, characterized in that in said step S5, the determination expression of the pseudo-abnormality distance threshold is: , For the pseudo-anomaly distance threshold value, For the number of samples in the normal feature sample set, Is the first of the normal characteristic sample set The number of feature vectors is chosen to be the same, For the threshold adjustment coefficient, the equation determines the threshold by the pseudo-abnormal distance statistic characteristics of the normal sample.
7. The atrial fibrillation anomaly detection method based on multi-sensing time-frequency pseudo anomaly is characterized by comprising the following steps of S31, dividing an input time-frequency feature subset into a plurality of sub-frequency segments according to typical frequency distribution of atrial fibrillation anomaly signals, wherein each sub-frequency segment corresponds to a group of time-frequency feature data, S32, inputting the time-frequency feature data of each sub-frequency segment into an extrusion module of an extrusion excitation time-frequency pseudo-selection model, carrying out global average pooling operation on the feature data of each sub-frequency segment to obtain feature statistics values of each sub-frequency segment, S33, inputting the feature statistics values into an excitation module of the model, carrying out nonlinear transformation on the feature statistics values through a two-layer full-connection network to generate feature weight coefficients of each sub-frequency segment, S34, multiplying the feature weight coefficients with the time-frequency feature data of the corresponding sub-frequency segment element by element to obtain optimized time-frequency feature data of each sub-frequency segment, and integrating the optimized time-frequency feature data of all sub-frequency segments.
8. The atrial fibrillation anomaly detection method based on multi-sensing time-frequency pseudo anomaly is characterized by comprising the following substeps of S41, extracting all time-frequency characteristic vectors based on an optimized time-frequency characteristic matrix, calculating characteristic mean vectors and covariance matrices of atrial fibrillation anomaly samples, constructing core parameters of a double-domain hypersphere characteristic space based on the characteristic mean vectors and the covariance matrices, S42, determining inner-domain radiuses and outer-domain radiuses of the double-domain hypersphere characteristic space, setting the inner-domain radiuses based on the distribution range of normal sample characteristics, setting the outer-domain radiuses based on the distribution range of normal sample characteristics, S43, calculating Euclidean distances from the normal sample characteristics to the characteristic mean vectors of the normal sample characteristics, calculating the Mahalanobis distances between the normal sample characteristic mean vectors and the time-frequency characteristic vectors, and S44, fusing the Euclidean distances and the Mahalanobis distances according to a preset proportion, and obtaining pseudo anomaly distance values of the time-frequency characteristic vectors in the double-domain hypersphere characteristic space.
9. The atrial fibrillation anomaly detection method based on multi-sensing time-frequency pseudo anomalies according to claim 1 is characterized by comprising the following substeps of carrying out feature dimension normalization processing on each sample feature vector in a feature sample set of suspected atrial fibrillation anomalies to enable each dimensional feature to be in the same numerical range, carrying out fusion on normalized sample feature vectors by adopting a feature fusion algorithm, screening feature dimensions with higher correlation by calculating correlation among the sample feature vectors, carrying out weighted summation on the screened feature dimensions to generate a common abnormal feature vector of a sample set, wherein the vector comprises calibration feature information of suspected atrial fibrillation anomalies, and carrying out matching on the common abnormal feature vector and a preset atrial fibrillation anomaly feature template, and generating an atrial fibrillation anomaly detection result according to a matching result, wherein the type of the detected anomaly and the corresponding feature expression are clear.
10. An atrial fibrillation anomaly detection system based on multi-sensing time-frequency pseudo anomaly is characterized by being applied to the atrial fibrillation anomaly detection method based on multi-sensing time-frequency pseudo anomaly as claimed in claim 1, and comprises an HDS high-density sensing signal acquisition unit, a multi-mode time-frequency tensor optimization processing unit and an atrial electrophysiological signal acquisition unit, wherein the HDS high-density sensing signal acquisition unit is used for acquiring an atrial electrophysiological signal sequence and is connected with the multi-mode time-frequency tensor optimization processing unit and transmitting the acquired signal sequence to the multi-mode time-frequency tensor optimization processing unit; the multi-mode time-frequency tensor optimizing processing unit is used for performing time-frequency conversion and tensor optimizing processing on a received signal sequence, extracting a time-frequency characteristic subset, connecting the time-frequency characteristic subset with an extrusion excitation time-frequency pseudo-selecting unit, transmitting the time-frequency characteristic subset to the extrusion excitation time-frequency pseudo-selecting unit, the extrusion excitation time-frequency pseudo-selecting unit is used for performing characteristic weight distribution and optimization on the time-frequency characteristic subset, outputting an optimized time-frequency characteristic matrix, connecting the unit with a two-domain hyper-sphere pseudo-abnormal distance calculating unit, transmitting the optimized time-frequency characteristic matrix to the two-domain hyper-sphere pseudo-abnormal distance calculating unit, constructing a two-domain hyper-sphere characteristic space and calculating a pseudo-abnormal distance value, connecting the two-domain hyper-sphere abnormal distance calculating unit with an abnormal sample screening unit, transmitting the pseudo-abnormal distance value to the abnormal sample screening unit, setting a threshold value and dividing a suspected abnormal sample set, connecting the suspected abnormal sample set with a characteristic fusion and detection result generating unit, transmitting the suspected abnormal sample set to a characteristic fusion and detection result generating unit, a characteristic fusion and detection result generating unit for carrying out characteristic fusion on the suspected abnormal sample set and extracting characteristic and detecting result, the unit receives the suspected abnormal sample set transmitted by the abnormal sample screening unit, and outputs an atrial fibrillation abnormal detection result after finishing processing.

Description

Atrial fibrillation anomaly detection method and system based on multi-sensing time-frequency pseudo anomaly Technical Field The invention relates to the technical field of atrial fibrillation anomaly detection, in particular to an atrial fibrillation anomaly detection method and system based on multi-sensor time-frequency pseudo anomalies. Background Atrial fibrillation is a clinically common arrhythmia type, and early accurate detection of the atrial fibrillation has important significance for preventing thromboembolism, heart failure and other complications. In current clinical practice, atrial fibrillation detection is mostly dependent on conventional electrocardiograph, dynamic electrocardiograph and other devices, but the devices are easily interfered by the outside in the signal acquisition process, and the sensing mode of a single channel or a few channels is difficult to comprehensively capture the fine change of the electrophysiological activity of the atrium. With the development of high-density sensing technology, the HDS high-density sensing atrial fibrillation analysis platform is gradually applied to atrial electrophysiological signal acquisition, can acquire multichannel time domain signal data, and provides a basis for subsequent time-frequency domain feature extraction. However, how to effectively screen features related to atrial fibrillation abnormality from multichannel time-frequency signals, and construct an accurate abnormality detection model, is still a key problem to be broken through in the current atrial fibrillation detection technology development, and an atrial fibrillation abnormality detection method and system based on multi-sensing time-frequency pseudo-abnormality are proposed under the background. Existing atrial fibrillation detection techniques suffer from two significant drawbacks. On the one hand, in the feature extraction link, the existing method mostly adopts a single-mode time-frequency analysis means, does not perform tensor-level optimization processing on multi-channel time-frequency signals, and is difficult to fully mine the associated features of different sensing channels and different time-frequency dimensions, so that the extracted features have insufficient characterization capability on atrial fibrillation abnormality and can not accurately reflect the abnormal change rule of atrial electrophysiological activity. On the other hand, in the abnormality recognition link, in the prior art, normal samples and abnormal samples are divided based on a single distance calculation mode, a double-domain feature space is not constructed to perform pseudo-abnormality distance calculation, and the screening of suspected abnormal samples lacks an accurate threshold determination mechanism, so that the situation that the normal samples are misjudged to be abnormal or the abnormal samples are missed to be judged easily occurs, the reliability of atrial fibrillation abnormal detection is reduced, and the severe requirement of clinic on detection accuracy is difficult to meet. Disclosure of Invention In order to overcome the defects and shortcomings in the prior art, the invention provides an atrial fibrillation anomaly detection method and system based on multi-sensing time-frequency pseudo anomalies. The invention adopts the technical scheme that the atrial fibrillation anomaly detection method based on multi-sensing time-frequency pseudo anomaly comprises the following steps that S1, an atrial electrophysiological signal sequence is acquired through an HDS high-density sensing atrial fibrillation analysis platform, the sequence comprises time domain signal data acquired by a plurality of sensing channels, and the time domain signal data is subjected to time-frequency domain conversion processing to obtain a multi-channel time-frequency signal matrix; the method comprises the steps of S2, inputting a multi-channel time-frequency signal matrix into a multi-mode time-frequency tensor optimization algorithm, constructing a multi-mode time-frequency tensor model, conducting hierarchical optimization on mode dimensions in the tensor model, extracting time-frequency characteristic components in each mode through tensor decomposition operation, screening out a time-frequency characteristic subset related to atrial fibrillation, S3, inputting the time-frequency characteristic subset into an extrusion excitation time-frequency pseudo-selection model, conducting dynamic distribution on characteristic weights of different frequency segments in the time-frequency characteristic subset, enhancing characteristic capacity of atrial fibrillation abnormal characteristics, and outputting an optimized time-frequency characteristic matrix, S4, adopting a double-domain hyper-sphere pseudo-anomaly distance calculation algorithm to construct a double-domain hyper-sphere characteristic space, calculating pseudo-anomaly distance values of each time-frequency characteristic vector in the doubl