CN118177827A - Myocardial infarction positioning method based on electrocardiograph vector diagram - Google Patents
Myocardial infarction positioning method based on electrocardiograph vector diagram Download PDFInfo
- Publication number
- CN118177827A CN118177827A CN202410480147.9A CN202410480147A CN118177827A CN 118177827 A CN118177827 A CN 118177827A CN 202410480147 A CN202410480147 A CN 202410480147A CN 118177827 A CN118177827 A CN 118177827A
- Authority
- CN
- China
- Prior art keywords
- local
- feature
- features
- myocardial infarction
- point
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 239000013598 vector Substances 0.000 title claims abstract description 80
- 208000010125 myocardial infarction Diseases 0.000 title claims abstract description 59
- 238000000034 method Methods 0.000 title claims abstract description 40
- 238000010586 diagram Methods 0.000 title claims abstract description 32
- 230000007246 mechanism Effects 0.000 claims abstract description 22
- 230000008569 process Effects 0.000 claims abstract description 4
- 230000006870 function Effects 0.000 claims description 27
- 238000005070 sampling Methods 0.000 claims description 18
- 238000000605 extraction Methods 0.000 claims description 11
- 230000002776 aggregation Effects 0.000 claims description 9
- 238000004220 aggregation Methods 0.000 claims description 9
- 238000012549 training Methods 0.000 claims description 9
- 230000004807 localization Effects 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 6
- 238000001914 filtration Methods 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 6
- 230000000747 cardiac effect Effects 0.000 claims description 5
- 230000004927 fusion Effects 0.000 claims description 5
- 238000004422 calculation algorithm Methods 0.000 claims description 3
- 238000010276 construction Methods 0.000 claims description 3
- 238000011156 evaluation Methods 0.000 claims description 3
- 230000000750 progressive effect Effects 0.000 claims description 3
- 238000005516 engineering process Methods 0.000 abstract description 3
- 238000003745 diagnosis Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 239000011159 matrix material Substances 0.000 description 4
- 230000007547 defect Effects 0.000 description 3
- 238000002790 cross-validation Methods 0.000 description 2
- 230000001186 cumulative effect Effects 0.000 description 2
- 230000002427 irreversible effect Effects 0.000 description 2
- 108010028984 3-isopropylmalate dehydratase Proteins 0.000 description 1
- 208000013875 Heart injury Diseases 0.000 description 1
- 206010061216 Infarction Diseases 0.000 description 1
- 208000009785 Inferior Wall Myocardial Infarction Diseases 0.000 description 1
- 206010000891 acute myocardial infarction Diseases 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000036770 blood supply Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000030833 cell death Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 210000004351 coronary vessel Anatomy 0.000 description 1
- 230000034994 death Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000003111 delayed effect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 229940079593 drug Drugs 0.000 description 1
- 239000003814 drug Substances 0.000 description 1
- 239000012634 fragment Substances 0.000 description 1
- PCHJSUWPFVWCPO-UHFFFAOYSA-N gold Chemical compound [Au] PCHJSUWPFVWCPO-UHFFFAOYSA-N 0.000 description 1
- 230000007574 infarction Effects 0.000 description 1
- 201000009526 inferior myocardial infarction Diseases 0.000 description 1
- 150000002500 ions Chemical class 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000002107 myocardial effect Effects 0.000 description 1
- 208000037891 myocardial injury Diseases 0.000 description 1
- 238000005312 nonlinear dynamic Methods 0.000 description 1
- 230000037361 pathway Effects 0.000 description 1
- 201000009661 posterior myocardial infarction Diseases 0.000 description 1
- 230000004213 regulation of atrial cardiomyocyte membrane depolarization Effects 0.000 description 1
- 230000034225 regulation of ventricular cardiomyocyte membrane depolarization Effects 0.000 description 1
- 230000013577 regulation of ventricular cardiomyocyte membrane repolarization Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000010187 selection method Methods 0.000 description 1
- 230000035939 shock Effects 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 230000033772 system development Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
- A61B5/339—Displays specially adapted therefor
- A61B5/341—Vectorcardiography [VCG]
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
- A61B5/346—Analysis of electrocardiograms
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
- A61B5/346—Analysis of electrocardiograms
- A61B5/349—Detecting specific parameters of the electrocardiograph cycle
- A61B5/355—Detecting T-waves
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
- A61B5/346—Analysis of electrocardiograms
- A61B5/349—Detecting specific parameters of the electrocardiograph cycle
- A61B5/366—Detecting abnormal QRS complex, e.g. widening
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
- A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Cardiology (AREA)
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Biophysics (AREA)
- Public Health (AREA)
- Veterinary Medicine (AREA)
- General Health & Medical Sciences (AREA)
- Animal Behavior & Ethology (AREA)
- Surgery (AREA)
- Molecular Biology (AREA)
- Medical Informatics (AREA)
- Pathology (AREA)
- Biomedical Technology (AREA)
- Heart & Thoracic Surgery (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Fuzzy Systems (AREA)
- Signal Processing (AREA)
- Psychiatry (AREA)
- Physiology (AREA)
- Mathematical Physics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
- Measuring And Recording Apparatus For Diagnosis (AREA)
Abstract
The invention provides a myocardial infarction positioning method based on an electrocardiograph vector diagram, which utilizes a PointNet ++ multi-scale grouping (MSG) module to acquire global space-time characteristics of 3D-VCG, then identifies important point sets through a self-attention mechanism, processes the point sets through PointNet ++ to acquire local space-time characteristics, then uses a multi-head attention mechanism of a transducer to re-associate and fuse local and global characteristics, reduces the dimension of a generated trainable list through a forward feedback network, and generates class probability through a softmax layer so as to realize accurate positioning of myocardial infarction. The method of the invention has excellent performance on 11 MI positioning tasks, and the overall accuracy reaches 99.98%, which is obviously superior to the existing myocardial infarction positioning technology; the local-space-time characteristics of VCG in 3D space are effectively utilized, and the precise positioning of MI is realized.
Description
Technical Field
The invention relates to the technical field of medical signal processing, in particular to a myocardial infarction positioning method based on an electrocardiograph vector diagram.
Background
Myocardial Infarction (MI), also known as acute myocardial infarction, is caused by insufficient blood supply in the coronary artery region of the heart, and delayed diagnosis and treatment may lead to extensive myocardial cell death and irreversible heart injury, with higher disability rate and mortality rate, once the disease occurs, the patient needs to be sent to a hospital for rescue immediately, such as irreversible myocardial injury, shock and death phenomena occur when the patient is not cured in time.
Studies have shown that accurate identification and intervention within a critical "prime time" can significantly reduce infarct size and reduce mortality from myocardial infarction. As the electrical activity characteristics of the affected heart area may be significantly altered upon myocardial infarction. In this regard, the localization of myocardial infarction is often performed in clinical practice using a target 12-lead Electrocardiogram (ECG) and a vector electrocardiogram (VCG). Among them, the standard 12-lead electrocardiogram provides cardiac electrical activity patterns from 12 perspectives, and is considered as one of the gold standards for myocardial infarction diagnosis and localization.
However, the diagnostic capabilities of ECG are to be improved, especially for severe posterior and inferior myocardial infarction, due to the redundancy of lead information and the inability to provide sagittal plane electrocardiographic activity information. While electrocardiographic vector graphics (VCG) capture features of cardiac electrical activity from different directions and angles by recording signals from multiple electrodes or sensors.
With the development of new generation information technology, and the expression of the electrocardiographic characteristics of the electrocardiographic vector diagram is more comprehensive. Electrocardiogram vector graphics (VCG) has become a new research trend in the field of intelligent auxiliary diagnosis of myocardial infarction, improves the accuracy of myocardial infarction positioning through intelligent and automatic computer system development, and respectively represents atrial depolarization, ventricular depolarization and ventricular repolarization for diagnosis and treatment, medication of medical professionals and vector loops of each electrocardiograph vector graphics including a P loop, a QRST loop and a T loop. Myocardial infarction disrupts the flow of charged ions in the conductive pathways within the heart, resulting in changes in the local and spatiotemporal characteristics of the electrocardiographic vector map.
In this regard, some studies have directed to automatic extraction of local and spatiotemporal features of electrocardiographic vector graphics (VCG), and in conventional techniques researchers have focused on spatiotemporal feature extraction related to electrocardiographic vector graphics morphology, angle, energy and non-linear dynamics.
However, in the traditional method, the manual characteristic selection method which depends on priori expert knowledge has defects in the positioning accuracy and model generalization capability of myocardial infarction detection, so that the performance of the common myocardial infarction ward position positioning is poor;
In the existing large mainstream method, local or space-time characteristics are directly extracted from the original one-dimensional time sequence signals of the electrocardiograph vector diagram, or classification is carried out by combining the two characteristics, but the characteristic extraction process is complicated, the interrelationship of the original one-dimensional time sequence signals of the electrocardiograph vector diagram in space is ignored, model parameters are large, and the method is not suitable for the deployment of edge-end equipment; in another major method, local or space-time characteristics are obtained by using two-dimensional projection of an electrocardiograph vector diagram, accuracy of myocardial infarction positioning is improved by calculating correlation of original one-dimensional time sequence signals of the electrocardiograph vector diagram in a two-dimensional plane, however, at most, three visual projection images cannot fully represent complete characteristic representation of the electrocardiograph vector diagram in space, and the method can only analyze single beat of the electrocardiograph vector diagram, and cannot effectively analyze electrocardiograph vector diagram segments.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a myocardial infarction positioning method based on an electrocardiograph vector diagram, which realizes accurate and automatic MI positioning through deep learning analysis of a 3D vector electrocardiograph (3D-VCG) and overcomes the defects of the prior art.
The technical scheme of the invention is as follows: a myocardial infarction positioning method based on an electrocardiograph vector diagram comprises the following steps:
S1), acquisition and preprocessing of 3D vector electrocardiographic data
Collecting a target 3D-VCG signal, and preprocessing the target 3D-VCG signal through downsampling and filtering;
s2), global space-time feature extraction
Acquiring global space-time characteristics from a preprocessed target 3D-VCG signal point cloud space by using a PointNet ++ multi-scale grouping (MSG) module, calculating a central point from an input 3D-VCG point cloud by using a far point sampling algorithm, and realizing comprehensive representation of VCG global characteristics by using characteristic aggregation;
S3) extracting local features, determining key points by using a self-attention mechanism of a transducer, and extracting local space-time features within Euclidean distance of the key points through PointNet ++;
S4), feature fusion and MI localization
Local and global features are re-associated and fused by using a multi-head attention mechanism of a transducer, the dimension of the generated trainable list is reduced through a forward feedback network, and class probabilities are generated through a softmax layer, so that accurate positioning of myocardial infarction is realized.
Preferably, in step S1), the 3D-VCG data set of the cardiac patient and the healthy population provided by PHYSIKALISCH-TECHNISCHE BUNDESANSTALT (PTB) is used as the target 3D-VCG signal, the 3D-VCG data set covers 11 most common myocardial infarction categories, and the original sampling frequency of the data is set to 1000Hz.
Preferably, in step S1), the frequency of the downsampling is 512Hz, the filtering operation uses a second order Butterworth high-pass filter, the cut-off frequency is set to 0.5Hz to eliminate baseline drift, and then the data is segmented using a2 second non-overlapping window.
Preferably, in step S2), the PointNet ++ multi-scale grouping (MSG) module includes a plurality of Set Abstraction (SA) layers and a Feature Propagation (FP) layer, and the PointNet ++ uses a series of Set Abstraction (SA) layers to hierarchically extract local features.
Preferably, in step S2), the method specifically includes the following steps:
S21), sampling
The PointNet ++ multi-scale grouping (MSG) module selects a group of center points { c 1,c2,...,cm } from the preprocessed target 3D-VCG signal point cloud through a far-most point sampling strategy (FPS); the combination of these center points represents the global features of the point cloud; after sampling, each center point becomes the center of a group of local points, and the local points are used for subsequent feature extraction;
S22, grouping
For each center point c j, adjacent points are grouped according to a specified euclidean distance r and a maximum number of neighbors k, expressed as follows:
Pcj＝Group(P,cj,r)；
Wherein P cj represents a point group taking c j as a central point and is used for local feature calculation, and P represents an acquired 3D-VCG signal point cloud dataset;
S23), for each point group P cj, processing by using a small PointNet network to obtain a spatiotemporal feature f local,cj of the local point set:
flocal,cj＝PoinNet(Pcj)；
s24), feature aggregation
Splicing the space-time features f local,cj of all the local point sets through a multi-scale grouping (MSG) module to generate a global feature descriptor containing hierarchical local features; this global feature descriptor contains local feature information extracted from multiple scales; the expression of the global feature descriptor is as follows:
S25), feature upsampling
To combine local features with global features and restore to the resolution of the original point cloud, pointNet ++ uses a Feature Propagation (FP) layer; the Feature Propagation (FP) layer fuses features of different levels by upsampling functions.
Preferably, in step S24), for a given layer F up＝upsample(Flayerl,Flayerl-1), where the function upsample contains nearest neighbor interpolation or learning weights to fuse features from different layers, pointNet ++ provides a rich set of features, encapsulating local geometry and global context, after passing through multiple SA and FP layers; this hierarchical feature set is then processed by a series of shared MLPs, producing a refined global descriptor G final:
Gfinal＝MLP(Fup)。
Preferably, in step S2), the center point sampling number of the PointNet ++ multi-scale packet (MSG) module is set to 64, the euclidean distance r of the multi-scale packet (MSG) module is set to 0.1, 0.2, and 0.4 in sequence, and the number of points in each euclidean distance r is 16, 32, and 128, respectively.
Preferably, in step S3), the method specifically includes the following steps:
S31), for a given 3D-VCG signal point cloud dataset p= { P 1,p2,...,pN }, each point cloud P i has 3-dimensional features, and projecting the points through a progressive feed-forward network into a potential feature space, resulting in N potential feature vectors { platent 1,platent2,...,platentN }, and each feature vector platent i having dm-dimensional features;
s32), feeding the potential feature vectors obtained through the feed forward network into the self-attention mechanism architecture of the transducer, calculating the similarity and association degree between the query (Q) of each obtained potential feature and all keys (K) using the attention scoring function in the self-attention mechanism architecture of the transducer, and identifying key points according to the acquaintance score:
score(Q,K)＝σ(Q,KT)；
A(Q,K,V)＝score(Q,K)V；
Wherein score (Q, K) is a scoring function, σ (Q, K T) represents the dot product result between the Query (Query) and the Key (Key) through an activation function σ; the activation function σ employs a softmax function that converts the result of the dot product into a probability distribution such that the value of each element is between 0 and 1 and the sum of all elements is 1; q, K, V represent three Key parameters of Query (Query), key (Key) and Value (Value) of the self-attention mechanism architecture of the transducer respectively;
S33), aggregation of local features
Selecting M point clouds with highest score as key points, executing Euclidean distance for each selected key point p i j, finding points in the neighborhood of the selected key points within a specific Euclidean distance r, and forming a local point set by the neighborhood points and the key points; then processing the local point set through another feedforward network to obtain an aggregate local feature g j;
s34), construction of local feature vectors
Original features of each selected keypointScore/>The aggregated local features g j combine to form a local feature vector/>Wherein the local feature vector/>The method comprises the steps of including position information of points, importance scores and geometric features of local neighborhoods;
s35), arrangement of output local feature set
All local feature vectorsCombine into a local feature set F L and according to score/>The descending order of (3) ensures that the arrangement of the feature sets is not changed, and finally the ordered local feature set F L is output.
Preferably, in step S4), the method specifically includes the following steps:
S41), re-associating and fusing local and global features through a multi-head attention mechanism of a transducer, and inputting the fused features into a classifier SoftMax layer for model training;
S42), in the training process, optimizing the model parameters by using a cross entropy loss function L, and in the evaluation stage, measuring the performance of the model by using accuracy, recall and F1 score indexes, wherein the cross entropy loss function L is as follows:
Wherein Z represents the total number of samples in the training data set, J represents the number of MI categories, and r ic represents the true category of the ith sample; p ic represents the predicted class of the ith sample, and predicts whether each point cloud instance belongs to a particular class of MI, or is a healthy heart, by the trained model.
The beneficial effects of the invention are as follows:
1. The method of the invention has excellent performance on 11MI positioning tasks, and the overall accuracy reaches 99.98%, which is obviously superior to the existing myocardial infarction positioning technology; local-space-time characteristics of VCG in 3D space are effectively utilized, and precise positioning of MI is realized;
2. According to the invention, local and space-time characteristics of the electrocardiograph vector diagram are directly obtained in a three-dimensional space through the point cloud model, so that various myocardial infarction positioning is realized;
3. according to the invention, on the fusion of local and space-time characteristics of an electrocardiograph vector diagram, a transducer architecture is introduced, and the characteristic fusion is carried out through high-order dependency learning of the local and space-time characteristics;
4. The invention takes the electrocardiograph vector diagram heart beat as input, thus inputting the electrocardiograph vector diagram heart beat into a model for myocardial infarction positioning, and the input data can be the electrocardiograph vector diagram heart beat or the electrocardiograph vector diagram fragment, thereby increasing the mobility of the proposed method on various data sets.
Drawings
FIG. 1 is a schematic flow chart of a method according to an embodiment of the invention;
FIG. 2 is a flow chart of an embodiment of the present invention;
fig. 3 is a structural frame diagram of an embodiment of the present invention.
Detailed Description
The following is a further description of embodiments of the invention, taken in conjunction with the accompanying drawings:
as shown in fig. 1-3, the present embodiment provides a myocardial infarction positioning method based on an electrocardiograph vector diagram, which includes the following steps:
S1), acquisition and preprocessing of 3D vector electrocardiographic data
Collecting a target 3D-VCG signal, and preprocessing the target 3D-VCG signal through downsampling and filtering;
s2), global space-time feature extraction
Acquiring global space-time characteristics from a 3D-VCG point cloud space by using a PointNet ++ multi-scale grouping (MSG) module, calculating a central point from an input 3D-VCG point cloud by using a far point sampling algorithm, and realizing comprehensive representation of the VCG global characteristics by using characteristic aggregation;
S3) extracting local features, determining key points by using a self-attention mechanism of a transducer, and extracting local space-time features within Euclidean distance of the key points through PointNet ++;
S4), feature fusion and MI localization
Local and global features are re-associated and fused by using a multi-head attention mechanism of a transducer, the dimension of the generated trainable list is reduced through a forward feedback network, and class probabilities are generated through a classifier softmax layer, so that accurate positioning of myocardial infarction is realized.
As a preferred embodiment, in step S1), the 3D-VCG data set of the cardiac patient and the healthy population provided by PHYSIKALISCH-TECHNISCHE BUNDESANSTALT (PTB) is used as the collected target 3D-VCG signal, the original sampling frequency of the data is 1000Hz, and 426 records are extracted from the PTB in this embodiment, wherein 346 records are from 127 patients clinically diagnosed as myocardial infarction, the 11 most common myocardial infarction categories are covered, and in addition, 80 VCG records are obtained from 52 healthy individuals (called HC).
Preferably, in step S1), the frequency of the downsampling is 512Hz, the filtering is performed by using a second order Butterworth high-pass filter, the cut-off frequency is set to 0.5Hz, so as to eliminate baseline drift, and then the data is segmented by using a2 second non-overlapping window.
Preferably, in step S2), the PointNet ++ multi-scale grouping (MSG) module includes a plurality of Set Abstraction (SA) layers and a Feature Propagation (FP) layer, and the PointNet ++ uses a series of Set Abstraction (SA) layers to hierarchically extract local features.
As a preferred embodiment, step S2) specifically includes the following steps:
S21), sampling
The PointNet ++ multi-scale grouping (MSG) module selects a group of center points { c 1,c2,...,cm } from the point cloud through the furthest point sampling strategy (FPS); the combination of these center points represents the global features of the point cloud; after sampling, each center point becomes the center of a group of local points, and the local points are used for subsequent feature extraction;
S22, grouping
For each center point c j, adjacent points are grouped according to the specified Euclidean distance r and the maximum neighbor number k, and the expression is:
Pcj＝Group(P,cj,r)；
Wherein P cj represents a point set of the central point c j, and P represents an acquired 3D-VCG signal point cloud data set;
S23), for each point group P cj, processing by using a small PointNet network to obtain a spatiotemporal feature f local,cj of the local point set:
flocal,cj＝PoinNet(Pcj)；
s24), feature aggregation
By stitching the spatiotemporal features f local,cj of all local point sets, the multi-scale grouping (MSG) module generates a global feature descriptor containing hierarchical local featuresThis global feature descriptor/>Including local feature information extracted from multiple scales; wherein the global feature descriptor/>The expression of (2) is:
S25), feature upsampling
To combine local features with global features and restore to the resolution of the original point cloud, pointNet ++ uses a Feature Propagation (FP) layer; the Feature Propagation (FP) layer fuses features of different levels by upsampling functions.
Preferably as this embodiment, in step S24), for a given layerWherein the function upsample contains nearest neighbor interpolation or learning weights to fuse features from different layers, after passing through multiple SA and FP layers, pointNet ++ provides a rich set of features encapsulating local geometry and global context; this hierarchical feature set is then processed by a series of shared MLPs, producing a refined global descriptor G final:
Gfinal＝MLP(Fup)。
In step S2), the center point sampling number of the PointNet ++ multi-scale packet (MSG) module is set to 64, the euclidean distance r of the multi-scale packet (MSG) module is set to 0.1, 0.2, and 0.4 in sequence, and the number of points in each euclidean distance r is 16, 32, and 128, respectively.
As a preferred embodiment, step S3) specifically includes the following steps:
S31), for a given 3D-VCG signal point cloud dataset p= { P 1,p2,...,pN }, each point cloud P i has 3-dimensional features, and projecting the points through a progressive feed-forward network into a potential feature space, resulting in N potential feature vectors { platent 1,platent2,...,platentN }, and each feature vector platent i having dm-dimensional features;
s32), feeding the potential feature vectors obtained through the feed forward network into the self-attention mechanism architecture of the transducer, calculating the similarity and association degree between the query (Q) of each obtained potential feature and all keys (K) using the attention scoring function in the self-attention mechanism architecture of the transducer, and identifying key points according to the acquaintance score:
score(Q,K)＝σ(Q,KT)；
A(Q,K,V)＝score(Q,K)V；
Wherein score (Q, K) is a scoring function, σ (Q, K T) represents the dot product result between the Query (Query) and the Key (Key) through an activation function σ; the activation function σ employs a softmax function that converts the result of the dot product into a probability distribution such that the value of each element is between 0 and 1 and the sum of all elements is 1; q, K, V represent three Key parameters of Query (Query), key (Key) and Value (Value) of the self-attention mechanism architecture of the transducer respectively; the query vector Q is a vector for matching with the key vector K. In the self-attention mechanism, a query vector represents a location or sequence element that is currently being processed, which it wants to "query" for information at other locations. The query vector computes a similarity score by performing a dot product operation with all key vectors, which helps determine the importance of each element in the sequence to the current position; the key vector K is a representation of each element in the sequence that is not directly modified by the query vector. The role of the key vector is to provide matching objects for the query in order to calculate an attention score; in the self-attention mechanism, the key vector forms a database and the query vector finds the most relevant information by comparing with each key in this database. The value vector V is the actual content representation of each element in the sequence. Once the similarity scores between the query vector and the key vector are calculated, these scores will be used to weight the value vector. This means that for each position its output feature will be a weighted sum of the value vectors of all positions in the sequence, the weights being determined by the similarity scores between the query and the corresponding keys;
S33), aggregation of local features
Selecting M point clouds with highest score as key points, executing Euclidean distance for each selected key point p i j, finding points in the neighborhood of the selected key points within a specific Euclidean distance r, and forming a local point set by the neighborhood points and the key points; then processing the local point set through another feedforward network to obtain an aggregate local feature g j;
s34), construction of local feature vectors
Original features of each selected keypointScore/>The aggregated local features g j combine to form a local feature vector/>Wherein the local feature vector/>The method comprises the steps of including position information of points, importance scores and geometric features of local neighborhoods;
s35), arrangement of output local feature set
All local feature vectorsCombine into a local feature set F L and according to score/>The descending order of (3) ensures that the arrangement of the feature sets is not changed, and finally the ordered local feature set F L is output.
As a preferred embodiment, step S4) specifically includes the following steps:
S41), re-associating and fusing local and global features through a multi-head attention mechanism of a transducer, and inputting the fused features into a classifier SoftMax layer for model training;
S42), in the training process, optimizing model parameters by using a cross entropy loss function L, and in the evaluation stage, measuring the performance of the model by using accuracy, recall and F1 score indexes, wherein the cross entropy loss function L is as follows;
Wherein Z represents the total number of samples in the training data set, J represents the number of MI categories, and r ic represents the true category of the ith sample; p ic represents the predicted class of the ith sample, and predicts whether each point cloud instance belongs to a particular class of MI, or is a healthy heart, by the trained model.
The present embodiment performs t-SNE statistical analysis on local spatiotemporal features derived from the last FC layer covering the 3DVCG tensors of HC and 11 MI types.
In this embodiment, various indexes are cross-validated 5 times on the reference PTB database, and a cumulative confusion matrix under five-fold cross-validation is given in table 1, where the horizontal axis in the matrix represents the actual VCG segment class, and the vertical axis represents the prediction of the model. In the confusion matrix, diagonal values reflect the correctly identified instances of each MI type. Cross verification shows that only 61 of 24,933 3D-VCG test point clouds are classified in error, so that the recognition accuracy of 99.76% is realized. It is noted that the accuracy of identification of the categories HC, IPLMI, IPMI, PMI and PLMI is near perfect in this embodiment. Most errors are found in IMI and ALMI classifications. As shown in table 2, from the classification, this example shows an accuracy of 99.76%, a specificity of 99.97%, a precision of 99.72%, a recall of 99.77%, and an f1 score of 0.9975. The results demonstrate that the present embodiment can be proficiently utilized for accurate MI localization using 3D-VCG data.
The method of this embodiment is superior to existing VCG-based MI positioning methods in the reference PTB database. And acquiring the space-time characteristics of the 3D-VCG point set by utilizing the multi-scale characteristic acquisition capability of the MSG in Pointnet ++. For local feature extraction, key points are determined from the scores of interest, pointnet ++ calculates local features within Euclidean distance of these key points. The multi-head attention model of the transducer architecture integrates local and space-time features, captures the full-scope details of the 3D-VCG, and pinpoints 11 MI types and HC. In addition, the average recognition accuracy and f1 score of this embodiment are 99.98% and 0.9999%, respectively, ensuring that the recognition error for each MI type is less than 1%.
TABLE 1 cumulative confusion matrix under five-fold cross-validation
TABLE 2 Performance of eleven classes of myocardial infarction localization
The foregoing embodiments and description have been provided merely to illustrate the principles and best modes of carrying out the invention, and various changes and modifications can be made therein without departing from the spirit and scope of the invention as defined in the appended claims.
Claims (10)
1. The myocardial infarction positioning method based on the electrocardiograph vector diagram is characterized by comprising the following steps of:
S1), acquisition and preprocessing of 3D vector electrocardiographic data
Collecting a target 3D-VCG signal, and preprocessing the collected target 3D-VCG signal through downsampling and filtering;
s2), global space-time feature extraction
Acquiring global space-time characteristics from a preprocessed target 3D-VCG point cloud space by using a PointNet ++ multi-scale grouping (MSG) module, calculating a central point from the input 3D-VCG point cloud by using a farthest point sampling algorithm, and realizing comprehensive representation of the VCG global characteristics by using characteristic aggregation;
S3), local feature extraction
Determining key points by using a self-attention mechanism of a transducer, and extracting local space-time features within Euclidean distance of the key points through PointNet ++;
S4), feature fusion and MI localization
Local and global features are re-associated and fused by using a multi-head attention mechanism of a transducer, the dimension of the generated trainable list is reduced through a forward feedback network, and class probabilities are generated through a softmax layer, so that accurate positioning of myocardial infarction is realized.
2. The myocardial infarction positioning method based on an electrocardiographic vector diagram as set forth in claim 1, wherein: in step S1), the collected target 3D-VCG signals adopt a 3D-VCG data set of cardiac patients and healthy people provided by PHYSIKALISCH-TECHNISCHE BUNDESANSTALT (PTB), wherein the 3D-VCG data set covers 11 most common myocardial infarction categories, and the original sampling frequency of the data is set to be 1000Hz.
3. The myocardial infarction positioning method based on an electrocardiographic vector diagram as set forth in claim 1, wherein: in step S1), the downsampling frequency is 512Hz, the filtering operation adopts a second order Butterworth high-pass filter, the cut-off frequency is set to 0.5Hz to eliminate baseline drift, and then the data is segmented by adopting a 2 second non-overlapping window.
4. The myocardial infarction positioning method based on an electrocardiographic vector diagram as set forth in claim 1, wherein: in step S2), the PointNet ++ multi-scale packet (MSG) module includes a plurality of Set Abstraction (SA) layers and a Feature Propagation (FP) layer, and the PointNet ++ employs the plurality of Set Abstraction (SA) layers to extract local features in a layered manner.
5. The myocardial infarction positioning method based on an electrocardiographic vector diagram as set forth in claim 1, wherein: the step S2) specifically comprises the following steps:
S21), sampling
The PointNet ++ multi-scale grouping (MSG) module selects a group of center points { c 1,c2,...,cm } from the preprocessed 3D-VCG signal point cloud through a far-most point sampling strategy (FPS); the combination of these center points represents the global features of the point cloud; after sampling, each center point becomes the center of a group of local points, and the local points are used for subsequent feature extraction;
S22, grouping
For each center point c j, the neighboring points are grouped according to the specified Euclidean distance r and the maximum number of neighbors k:
Pcj＝Group(P,cj,r)；
Wherein P cj represents a point group taking c j as a central point, and P represents an acquired 3D-VCG signal point cloud dataset;
S23), for each point group P cj, processing by using a small PointNet network to obtain a spatiotemporal feature f local,cj of the local point set:
flocal,cj＝PoinNet(Pcj)；
s24), feature aggregation
Splicing the space-time features f local,cj of all local point sets through a multi-scale grouping (MSG) module to generate a global feature descriptor containing hierarchical local featuresThis global feature descriptor/>Including local feature information extracted from a plurality of scales; wherein the global feature descriptor/>The expression of (2) is:
S25), feature upsampling
In order to combine local features with global features and restore to the resolution of the original point cloud, the PointNet ++ Feature Propagation (FP) layer fuses features of different levels by upsampling functions.
6. The myocardial infarction positioning method based on an electrocardiographic vector diagram as set forth in claim 5, wherein: step S24), for a given layerWherein the function upsample contains nearest neighbor interpolation or learning weights to fuse features from different layers, after passing through multiple SA and FP layers, pointNet ++ provides a rich set of features, encapsulating local geometry and global context; this hierarchical feature set is then processed by a series of shared MLPs, producing a refined global descriptor G final:
Gfinal＝MLP(Fup)。
7. The myocardial infarction positioning method based on an electrocardiographic vector diagram as set forth in claim 5, wherein: in step S2), the center point sampling number of the PointNet ++ multi-scale packet (MSG) module is set to 64, the euclidean distance r of the multi-scale packet (MSG) module is set to 0.1, 0.2 and 0.4 in sequence, and the number of points in each euclidean distance r is respectively 16, 32 and 128.
8. The myocardial infarction positioning method based on an electrocardiographic vector diagram as set forth in claim 1, wherein: in step S3), the method specifically includes the following steps:
S31), for a given 3D-VCG signal point cloud dataset p= { P 1,p2,...,pN }, each point cloud P i has 3-dimensional features, projecting the point clouds into a potential feature space through a progressive feed-forward network, resulting in N potential feature vectors { platent 1,platent2,...,platentN }, and each feature vector platent i has dm-dimensional features;
s32), feeding the potential feature vectors obtained through the feed forward network into the self-attention mechanism architecture of the transducer, calculating the similarity and association degree between the query (Q) of each obtained potential feature and all keys (K) using the attention scoring function in the self-attention mechanism architecture of the transducer, and identifying key points according to the acquaintance score:
score(Q,K)＝σ(Q,KT)；
A(Q,K,V)＝score(Q,K)V；
Wherein score (Q, K) is a scoring function, σ (Q, K T) represents the dot product result between the Query (Query) and the Key (Key) through an activation function σ; the activation function σ employs a softmax function that converts the result of the dot product into a probability distribution such that the value of each element is between 0 and 1 and the sum of all elements is 1; q, K, V represent three Key parameters of Query (Query), key (Key) and Value (Value) of the self-attention mechanism architecture of the transducer respectively;
S33), aggregation of local features
Selecting M point clouds with highest score as key points, and selecting each key pointExecuting Euclidean distance, and finding points in the neighborhood of the Euclidean distance within a specific Euclidean distance r, wherein the neighborhood points and the key points form a local point set; then processing the local point set through another feedforward network to obtain an aggregate local feature g j;
s34), construction of local feature vectors
Original features of each selected keypointScore/>The aggregated local features g j combine to form a local feature vector/>Wherein the local feature vector/>The method comprises the steps of including position information of point cloud, importance scores and geometric features of local neighborhood;
s35), arrangement of output local feature set
All local feature vectorsCombine into a local feature set F L and according to score/>The descending order of (3) ensures that the arrangement of the feature sets is not changed, and finally the ordered local feature set F L is output.
9. The myocardial infarction positioning method based on an electrocardiographic vector diagram as set forth in claim 5, wherein: in step S4), the method specifically includes the following steps:
S41), re-associating and fusing local and global features through a multi-head attention mechanism of a transducer, and inputting the fused features into a classifier SoftMax layer of the model for model training.
10. The myocardial infarction positioning method based on an electrocardiographic vector diagram as set forth in claim 9, wherein: in the training process, the cross entropy loss function L is used for optimizing model parameters, and in the evaluation stage, the accuracy, recall and F1 score index are used for measuring the performance of the model, wherein the expression of the cross entropy loss function L is as follows:
Wherein Z represents the total number of samples in the training data set, J represents the number of MI categories, and r ic represents the true category of the ith sample; p ic represents the predicted class of the ith sample, and whether each point cloud instance belongs to a specific class of myocardial infarction or is a healthy heart is predicted by the trained model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410480147.9A CN118177827B (en) | 2024-04-22 | 2024-04-22 | Myocardial infarction positioning method based on electrocardiograph vector diagram |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410480147.9A CN118177827B (en) | 2024-04-22 | 2024-04-22 | Myocardial infarction positioning method based on electrocardiograph vector diagram |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN118177827A true CN118177827A (en) | 2024-06-14 |
| CN118177827B CN118177827B (en) | 2024-10-18 |
Family
ID=91409185
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410480147.9A Active CN118177827B (en) | 2024-04-22 | 2024-04-22 | Myocardial infarction positioning method based on electrocardiograph vector diagram |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN118177827B (en) |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108596142A (en) * | 2018-05-09 | 2018-09-28 | 吉林大学 | A kind of cardioelectric characteristic extracting process based on PCANet |
| CN109846471A (en) * | 2019-01-30 | 2019-06-07 | 郑州大学 | A kind of myocardial infarction detection method based on BiGRU deep neural network |
| CN110141220A (en) * | 2019-06-20 | 2019-08-20 | 鲁东大学 | Myocardial infarction automatic testing method based on multi-modal fusion neural network |
| CN115470828A (en) * | 2022-09-23 | 2022-12-13 | 华东师范大学 | Multi-lead electrocardiogram classification and identification method based on convolution and self-attention mechanism |
| CN116350231A (en) * | 2023-03-13 | 2023-06-30 | 浙江大学 | Semi-supervised electrocardiogram myocardial infarction positioning method based on hierarchical different granularity class constraint |
| KR20230133230A (en) * | 2022-03-10 | 2023-09-19 | 주식회사 메디컬에이아이 | Method, program and apparatus for contrasitive pre-training of neural network model based on electrocardiogram |
| CN117078956A (en) * | 2023-09-04 | 2023-11-17 | 吉林大学 | Point cloud classification segmentation network based on point cloud multi-scale parallel feature extraction and attention mechanism |
-
2024
- 2024-04-22 CN CN202410480147.9A patent/CN118177827B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108596142A (en) * | 2018-05-09 | 2018-09-28 | 吉林大学 | A kind of cardioelectric characteristic extracting process based on PCANet |
| CN109846471A (en) * | 2019-01-30 | 2019-06-07 | 郑州大学 | A kind of myocardial infarction detection method based on BiGRU deep neural network |
| CN110141220A (en) * | 2019-06-20 | 2019-08-20 | 鲁东大学 | Myocardial infarction automatic testing method based on multi-modal fusion neural network |
| KR20230133230A (en) * | 2022-03-10 | 2023-09-19 | 주식회사 메디컬에이아이 | Method, program and apparatus for contrasitive pre-training of neural network model based on electrocardiogram |
| CN115470828A (en) * | 2022-09-23 | 2022-12-13 | 华东师范大学 | Multi-lead electrocardiogram classification and identification method based on convolution and self-attention mechanism |
| CN116350231A (en) * | 2023-03-13 | 2023-06-30 | 浙江大学 | Semi-supervised electrocardiogram myocardial infarction positioning method based on hierarchical different granularity class constraint |
| CN117078956A (en) * | 2023-09-04 | 2023-11-17 | 吉林大学 | Point cloud classification segmentation network based on point cloud multi-scale parallel feature extraction and attention mechanism |
Non-Patent Citations (1)
| Title |
|---|
| XIAOYU LI: "Bat: Beat-aligned transformer for electrocardiogram classification[C]", 《2021 IEEE INTERNATIONAL CONFERENCE ON DATA MINING (ICDM)》, 31 December 2021 (2021-12-31), pages 320 - 329, XP034044813, DOI: 10.1109/ICDM51629.2021.00043 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN118177827B (en) | 2024-10-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| de Albuquerque et al. | Robust automated cardiac arrhythmia detection in ECG beat signals | |
| Chen et al. | Smart heart monitoring: Early prediction of heart problems through predictive analysis of ECG signals | |
| Xia et al. | Generative adversarial network with transformer generator for boosting ECG classification | |
| Bhagyalakshmi et al. | GB-SVNN: Genetic BAT assisted support vector neural network for arrhythmia classification using ECG signals | |
| Khazaee | Heart beat classification using particle swarm optimization | |
| Goswami et al. | Electrocardiogram signal classification using VGGNet: a neural network based classification model | |
| KR102632272B1 (en) | Classification method of atrial fibrillation and congestive heart failure using a convolutional artificial neural network | |
| CN113057585A (en) | Cognitive disorder detection model and training method thereof | |
| JP2023544242A (en) | Electrocardiogram processing system for detecting and/or predicting cardiac events | |
| Yang et al. | Proegan-ms: A progressive growing generative adversarial networks for electrocardiogram generation | |
| CN111261278A (en) | Heart disease detection method based on deep learning model of three-dimensional image | |
| Al-Huseiny et al. | Diagnosis of arrhythmia based on ECG analysis using CNN | |
| US20230181082A1 (en) | System and methods for electrocardiogram beat similarity analysis | |
| Bian et al. | Identifying electrocardiogram abnormalities using a handcrafted-rule-enhanced neural network | |
| Asadi et al. | Cardiac arrhythmia recognition with robust discrete wavelet-based and geometrical feature extraction via classifiers of SVM and MLP-BP and PNN neural networks | |
| Golrizkhatami et al. | Multi-scale features for heartbeat classification using directed acyclic graph CNN | |
| Imtiaz et al. | Cross-database and cross-channel electrocardiogram arrhythmia heartbeat classification based on unsupervised domain adaptation | |
| CN110522446A (en) | A kind of electroencephalogramsignal signal analysis method that accuracy high practicability is strong | |
| CN118177827B (en) | Myocardial infarction positioning method based on electrocardiograph vector diagram | |
| CN114431873A (en) | Method for constructing electrocardiogram heart rate multi-type prediction model based on graph convolution | |
| CN114983341A (en) | Multi-modal feature fusion based multi-classification prediction system for Alzheimer's disease | |
| Nasim et al. | An Evolutionary-Neural Mechanism for Arrhythmia Classification With Optimum Features Using Single-Lead Electrocardiogram | |
| Zhang et al. | Multi-scale and attention based ResNet for heartbeat classification | |
| Imtiaz et al. | Cross-Database and Cross-Channel ECG Arrhythmia Heartbeat Classification Based on Unsupervised Domain Adaptation | |
| Jayakrishnan et al. | Computational approach for heart disease prediction using machine learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |