CN-122004790-A - Dementia disease identification method and system based on space-time frequency multidimensional feature fusion

CN122004790ACN 122004790 ACN122004790 ACN 122004790ACN-122004790-A

Abstract

The invention discloses a dementia disease identification method and system based on space-time frequency multidimensional feature fusion, and belongs to the technical field of intelligent medical treatment. The method comprises the steps of preprocessing an electroencephalogram signal of a target user, extracting space-time frequency multi-dimensional characteristics of the preprocessed electroencephalogram signal, and classifying based on the space-time frequency multi-dimensional characteristics to obtain a dementia disease recognition result. The invention can be used for high-precision and automatic identification of dementia diseases such as Alzheimer disease, frontotemporal dementia and the like.

Inventors

LIU JIE
GUO WEICHAO

Assignees

中国科学院软件研究所

Dates

Publication Date: 20260512
Application Date: 20260410
Priority Date: 20251029

Claims (10)

1. A dementia disease identification method based on space-time frequency multidimensional feature fusion, which is characterized by comprising the following steps: preprocessing an electroencephalogram signal of a target user; Extracting space-time frequency multidimensional features of the preprocessed electroencephalogram signals; Classifying based on the space-time frequency multidimensional features to obtain the dementia disease recognition result.
2. The method of claim 1, wherein preprocessing the electroencephalogram signal of the target user comprises: Cutting the electroencephalogram signals by using a sliding window, and constructing a Raw object based on a cutting result; Electrode positioning is carried out on the Raw object in combination with the electrode position information to obtain the Raw object with the electrode position information, wherein the Raw object with the electrode position information is marked with bad leads through interactive visualization, and interpolation restoration is carried out on the bad leads by utilizing surrounding normal electrode data; Obtaining a re-referenced Raw object according to the average value of the electrodes of all the Raw objects with the electrode position information; Filtering the re-referenced Raw object; And removing artifacts from the filtered Raw object to obtain the preprocessed electroencephalogram signal.
3. The method of claim 2, wherein filtering the re-referenced Raw object comprises: Setting a band-pass filtering range; capturing the band-pass filtering range Frequency band(s), Frequency band(s), Frequency band(s), Frequency band and method for producing the same Electroencephalogram data in the frequency band is filtered by using a notch filter.
4. The method of claim 2, wherein the artifact removal is performed on the filtered Raw object to obtain a preprocessed electroencephalogram signal, comprising: Performing independent component analysis on the filtered Raw object to obtain a separation matrix; And multiplying the separation matrix by the filtered Raw object to obtain the preprocessed electroencephalogram signal.
5. The method of claim 1, wherein extracting the space-time-frequency multi-dimensional features of the preprocessed electroencephalogram signal comprises: Dividing an electroencephalogram signal into a plurality of brain region data sets based on a brain region where a guide channel is located, wherein the brain region data sets comprise a forehead region data set, a central region data set, a left temporal region data set, a right temporal region data set, a top region data set and a occipital region data set; extracting the electroencephalogram data of delta frequency band, theta frequency band, alpha frequency band, beta frequency band and gamma frequency band for the electroencephalogram signals in each brain region data set respectively; Extracting time domain features and frequency domain features of electroencephalogram data of each frequency band respectively, wherein the frequency domain features comprise spectrum energy and differential entropy, and the time domain features comprise zero crossing rate, peak-to-peak value and variance; dividing an electroencephalogram signal into a left hemisphere data set and a right hemisphere data set based on the brain position where the guide channel is located, and extracting hemispherical function asymmetry features based on the electroencephalogram signals in the left hemisphere data set and the right hemisphere data set, wherein the hemispherical function asymmetry features comprise a power asymmetry index and a complexity asymmetry index; Dividing an electroencephalogram signal into a forebrain region data set and a hindbrain region data set based on the brain position of the lead channel, and extracting functional gradient characteristics of the forebrain region and the hindbrain region based on the electroencephalogram signals in the forebrain region data set and the hindbrain region data set, wherein the functional gradient characteristics of the forebrain region and the hindbrain region comprise a forehead-occipital She Gonglv gradient ratio and a forehead-occipital lobe complexity gradient ratio; based on the time domain feature and the frequency domain feature of the brain electrical data of each frequency band in the brain region data set and the hemispherical function asymmetry feature and the front and back brain region function gradient feature, the space-time frequency multidimensional feature of the preprocessed brain electrical signal is obtained.
6. The method of claim 1, wherein classifying based on the time-space frequency multi-dimensional features to obtain a dementia-type disease recognition result comprises: The method comprises the steps of constructing a mixed classification model, wherein the mixed classification model comprises a CNN branch, an MLP branch, a splicing layer and an output layer, the CNN branch is used for extracting local space-time characteristics of an electroencephalogram signal, the MLP branch is used for extracting global abstract characteristics of the electroencephalogram signal, the splicing layer is used for fusing the local space-time characteristics with the global abstract characteristics, and the output layer is used for classifying fusion results; And (5) inputting the space-time frequency multidimensional features into the mixed classification model to obtain the dementia disease recognition result.
7. The method of claim 6, wherein the CNN branch comprises a plurality of convolution blocks and a global average pooling layer, and the plurality of convolution blocks form a pyramid structure with increasing feature depth, wherein each convolution block is formed by sequentially connecting a convolution layer, a batch normalization layer, an activation function layer, a pooling layer and a Dropout layer, and wherein a first convolution block is followed by a lightweight channel attention module that generates channel weights based on global average pooling and shared MLP.
8. The method of claim 6, wherein the MLP branch comprises a feature attention module and a number of fully connected layers, the feature attention module is configured to weight importance of the space-time-frequency multi-dimensional features based on fully connected layers and Sigmoid activation generation feature weight matrix, the number of fully connected layers form a refined structure of feature compression, and each fully connected layer is followed by a batch normalization layer and a Dropout layer.
9. The method of claim 6, wherein the hybrid classification model is trained based on a multi-level synthetic regularization strategy, wherein the multi-level synthetic regularization strategy includes structural regularization, training process regularization, and data plane regularization.
10. A dementia disease recognition system based on space-time-frequency multidimensional feature fusion, the system comprising: the preprocessing module is used for preprocessing the electroencephalogram signals of the target user; the feature extraction module is used for extracting space-time frequency multidimensional features of the preprocessed electroencephalogram signals; and the disease identification module is used for classifying based on the space-time frequency multidimensional characteristics to obtain a dementia disease identification result.

Description

Dementia disease identification method and system based on space-time frequency multidimensional feature fusion Technical Field The invention belongs to the technical field of intelligent medical treatment, and particularly relates to a dementia disease identification method and system based on space-time frequency multidimensional feature fusion. Background With the increasing aging of the population, the incidence of neurodegenerative diseases, especially dementia diseases, rises year by year, with Alzheimer's Disease (AD) and frontotemporal dementia (Frontotemporal Dementia, FTD) being the most common two types, severely threatening the cognitive health and quality of life of the elderly population. Currently, clinical diagnosis of dementia is mainly dependent on neuropsychological scale assessment, cerebrospinal fluid biomarker detection, and imaging means such as Positron Emission Tomography (PET) or Magnetic Resonance Imaging (MRI). However, these methods generally suffer from high cost, complex operation, poor accessibility, or invasive nature, and are difficult to meet for large-scale screening and long-term dynamic monitoring. As a non-invasive neuroelectrophysiology detection technology, the electroencephalogram (Electroencephalography, EEG) has the advantages of millisecond time resolution, portability of equipment, low running cost and the like, and has wide application prospect in cognitive function evaluation and nervous system disease detection. The research shows that AD patients often show the characteristics of lower power of rear alpha rhythm, enhanced activity of slow waves (delta, theta frequency bands), abnormal connection of brain network functions and the like, while FTD patients often see disorder of electric activity of frontal lobe areas. In recent years, EEG analysis methods based on machine learning have been widely used for automatic recognition studies of dementia. In the prior art, modeling is generally performed by combining traditional spectrum analysis with classifiers such as a Support Vector Machine (SVM) or random forest, but the feature extraction aspect is limited to single dimension (such as only frequency domain or only time domain), and joint information of EEG signals in three dimensions of time, space and frequency can not be fully mined, so that feature expression capability is limited, and stability and generalization capability of classification performance are affected. In addition, most researches lack systematic data preprocessing flow, rely on manual intervention for key steps such as bad conducting interpolation, re-referencing, artifact removal and the like, have low automation degree and influence the repeatability of results. Despite the advances made in the identification of dementia based on brain electricity, many shortcomings remain. The method represented by the patent CN120277489A adopts wavelet transformation to generate a time-frequency diagram and combines PSO to carry out channel selection, so that the classification performance is improved, but the method relies on an optimization algorithm with complex calculation, the automation degree of pretreatment and feature extraction processes is low, the system does not fuse time domain dynamic features, and the feature expression dimension is single. While patent CN113558636a only depends on a single nonlinear index of permutation entropy, performs state analysis under music stimulation, and has the advantage of simple operation, but the characteristic information is not fully utilized, it is difficult to comprehensively describe the differences of complex brain electrical modes of different dementia types such as AD and FTD, and a complete end-to-end analysis flow is not constructed, and the accuracy analysis and system integration capability are lacking. In general, the prior art generally suffers from the following drawbacks: (1) The preprocessing flow of the electroencephalogram signals lacks standardization, and most methods do not effectively integrate a module for removing physiological artifacts such as the electrooculogram and the electrocardiograph, so that noise interference in an original signal is serious, signal-to-noise ratio is low, and the accuracy of subsequent feature extraction and classification is affected; (2) The feature extraction method is limited to a single dimension, for example, only a frequency domain power feature or a nonlinear dynamics index is adopted, and multi-dimensional information of an electroencephalogram signal in a time domain, a space domain (brain area division) and a frequency domain cannot be fully fused, so that the discrimination capability of the extracted feature is insufficient, and dynamic change of a brain functional state is difficult to comprehensively characterize; (3) In the classification application of Alzheimer's disease and frontotemporal dementia, the prior art lacks systematic evaluation of classification performance, does not provid