CN-122020422-A - Electrocardiosignal classification system and method based on multi-mode fusion

CN122020422ACN 122020422 ACN122020422 ACN 122020422ACN-122020422-A

Abstract

The invention discloses an electrocardiosignal classification system and method based on multi-mode fusion, and belongs to the technical field of artificial intelligence and medical diagnosis. The system comprises a time domain processing unit, a frequency domain processing unit, a biological information processing unit, a characteristic fusion unit and a classifier unit, wherein ultrafine granularity multi-label classification of arrhythmia is realized by fusing time domain and frequency domain characteristics of electrocardiosignals and biological information such as gender and age of patients. Aiming at the problems of high dimension, strong co-occurrence and extremely unbalanced categories of the electrocardiographic data label, the invention provides dynamic weight sampling, data enhancement and Focal Loss function based on category frequency adjustment, and the recognition capability of the model on rare arrhythmia is remarkably improved. On a high-difficulty data set containing 195-class labels, the system provided by the invention realizes classification performance superior to that of the existing standard, and has good clinical applicability and generalization capability.

Inventors

MA LONG
YUAN QIKUN
YU QIMEI
LUO YI
QIN JIEHUA
WANG MEI
CHEN MING

Assignees

中山市博爱医院(中山市妇幼保健院、中山市妇幼保健计划生育服务中心、中山市妇女儿童医院)

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (8)

1. An electrocardiosignal classification system based on multi-modal fusion, which is characterized by comprising: The time domain processing unit is used for extracting characteristics of the input multi-lead electrocardio time domain signals; the frequency domain processing unit is used for carrying out frequency domain transformation on the electrocardio time domain signals and extracting frequency domain characteristics; The biological information processing unit is used for receiving the static biological characteristics of the patient and performing characteristic mapping; The feature fusion unit is used for splicing and fusing the time domain features, the frequency domain features and the biological information features; And the classifier unit is used for classifying the multi-label arrhythmia based on the fused features.
2. The system for classifying electrocardiograph signals based on multi-modal fusion of claim 1 wherein the time domain processing unit and the frequency domain processing unit each comprise a convolution block, a residual block and a maximum pooling layer, wherein the first convolution block is followed by the maximum pooling layer, and the maximum pooling layer is adopted in the subsequent residual block.
3. The system for classifying electrocardiographic signals based on multi-modal fusion as claimed in claim 2, wherein the time domain processing unit includes 7 residual blocks, and the frequency domain processing unit includes 5 residual blocks.
4. The system for classifying electrocardiographic signals based on multi-modal fusion as claimed in claim 1, wherein the static biometric characteristics received by the biometric information processing unit include at least gender and age, and the characteristic transformation is performed through at least one linear layer.
5. The system for classifying electrocardiographic signals based on multi-modal fusion of claim 1, further comprising a data preprocessing module for weight-based resampling and data enhancement of electrocardiographic signals, wherein the sample weights are dynamically calculated according to rarity of tags contained in the sample weights.
6. The system for classifying cardiac signals based on multi-modal fusion as set forth in claim 5, wherein said data enhancement includes adding at least one of random noise, amplitude scaling and time offset.
7. The electrocardiosignal classification system based on multi-modal fusion as claimed in claim 1, wherein the classifier unit uses Sigmoid activation function to conduct multi-label probability prediction, and Focal Loss based on category frequency dynamic adjustment weight is adopted as a Loss function in the training process.
8. A multi-label classification method using the multi-modality fusion-based electrocardiographic signal classification system according to any one of claims 1 to 7, comprising the steps of: Acquiring multi-lead electrocardio time domain signals and biological information of a patient; respectively extracting time domain features and frequency domain features; extracting biological information characteristics; Fusing the three types of characteristics; and classifying the multi-label arrhythmia based on the fusion characteristics.

Description

Electrocardiosignal classification system and method based on multi-mode fusion Technical Field The invention belongs to the technical fields of artificial intelligence, digital medical treatment and signal processing, and particularly relates to an Electrocardiogram (ECG) automatic analysis and diagnosis system and method based on deep learning, which are particularly suitable for classifying ultrafine-granularity and multi-label arrhythmia of multi-lead electrocardiosignals. Background Cardiovascular disease is one of the leading causes of death worldwide, and arrhythmia is a common manifestation. Electrocardiography (ECG) is a non-invasive, convenient diagnostic tool, and is of great importance in arrhythmia screening and diagnosis. Traditional ECG analysis is highly dependent on physician experience, and has the problems of strong subjectivity, difficult classification of complex arrhythmias, and the like. In recent years, artificial intelligence technology based on deep learning brings breakthrough to ECG automatic analysis. The existing methods focus on single-label or small-label (usually not more than 50 classes) heart rhythm classification, and are difficult to deal with the common occurrence of multiple symptoms in clinical practice. In addition, the existing model is mostly input by adopting a single mode (such as only a time domain or only a frequency domain), and cannot fully utilize the multidimensional information of electrocardiosignals and the influence of individual differences (such as age and sex) of patients on electrocardiogram manifestations. When the ultra-high-dimensional (hundreds of classes) and extremely unbalanced fine-granularity diagnosis labels are oriented, the existing model is easy to be overfitted to high-frequency common classes, and the identification capability to low-frequency rare classes is insufficient, so that the clinical missed diagnosis risk is caused. Therefore, there is an urgent need for an automatic classification system for electrocardiographic signals, which can integrate multi-modal information, effectively process the problems of high co-occurrence of labels and extreme unbalance, so as to improve the diagnosis accuracy, and particularly the detection rate of rare arrhythmia. Disclosure of Invention Aiming at the defects existing in the prior art, the invention aims to provide an electrocardiosignal classification system and an electrocardiosignal classification method based on multi-mode fusion, which solve the problems in the background art. The invention is realized by the following technical scheme that the electrocardiosignal classification system based on multi-mode fusion comprises: The time domain processing unit is used for extracting characteristics of the input multi-lead electrocardio time domain signals; the frequency domain processing unit is used for carrying out frequency domain transformation on the electrocardio time domain signals and extracting frequency domain characteristics; The biological information processing unit is used for receiving the static biological characteristics of the patient and performing characteristic mapping; The feature fusion unit is used for splicing and fusing the time domain features, the frequency domain features and the biological information features; And the classifier unit is used for classifying the multi-label arrhythmia based on the fused features. As a preferred embodiment, the time domain processing unit and the frequency domain processing unit each comprise a convolution block, a residual block and a maximum pooling layer, wherein the first layer convolution block is followed by the maximum pooling layer, and the subsequent residual block adopts the maximum pooling layer. As a preferred embodiment, the time domain processing unit comprises 7 residual blocks and the frequency domain processing unit comprises 5 residual blocks. As a preferred embodiment, the static biometric features received by the biometric information processing unit include at least gender and age, and are transformed by at least one linear layer. As a preferred embodiment, the system further comprises a data preprocessing module for weight-based resampling and data enhancement of the electrocardiosignal, wherein the weight of the sample is dynamically calculated according to the rarity degree of the tag contained in the sample. As a preferred embodiment, the data enhancement includes adding at least one of random noise, amplitude scaling, and time offset. As a preferred embodiment, the classifier unit uses Sigmoid activation function to perform multi-label probability prediction, and uses Focal Loss with dynamically adjusted weights based on class frequency as a Loss function in the training process. A multi-label classification method of an electrocardiosignal classification system based on multi-mode fusion comprises the following steps: Acquiring multi-lead electrocardio time domain signals and biological information of a