A MCI classification algorithm based on cerebellar multi-band feature fusion

Abstract

Alzheimer's disease is a common disease in the elderly with prominent memory loss, reduced cognitive function, and emotional and personality changes. It has a high incidence and mortality rate for the elderly. Early mild cognitive impairment (MCI) is the early stage of Alzheimer's disease development and the best treatment effect. In order to overcome the deficiencies in the prior art, the present invention proposes an MCI classification algorithm based on cerebellar multi-band feature fusion for the classification of early Alzheimer's disease population, and based on the analysis of cerebellar data, provides a new mode of cerebellar data preprocessing and feature extraction.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing, and in particular relates to an EEG data processing and feature extraction method, which is suitable for auxiliary diagnosis and evaluation of clinical MCI patients. Background Art

[0002] Alzheimer's disease is a common disease in the elderly, in which patients show prominent memory loss, reduced cognitive function, and emotional and personality changes. It has a high incidence and mortality rate in the elderly. MCI is the early stage of Alzheimer's disease development and the best treatment period. The diagnosis of this disease initially relied on invasive biochemical tests. With the development of science and technology, the means of obtaining brain information have also become more abundant. Nuclear magnetic resonance technology and EEG atlas recording technology have gradually become one of the main bases for the diagnosis of MCI. EEG signal acquisition and analysis is one of the auxiliary means of MCI disease diagnosis.

[0003] Compared with MRI technology, EEG can provide feedback on the subject's brain physiological activities with high temporal resolution, and is one of the key biomarkers for determining abnormal brain physiological electrical activities in MCI patients. According to previous studies, the frontal lobe, parietal lobe, and hippocampus are the main brain regions with lesions in MCI patients, and the diagnosis, intervention, and treatment of MCI diseases are also mainly carried out in these brain regions. However, with the development of brain science in recent years, the electrical activity of the cerebellum has gradually entered people's field of vision, and significant abnormalities have occurred during the development of MCI diseases.

[0004] Although the current EEG analysis of the frontal lobe, parietal lobe and hippocampus can cover the abnormal brain electrical activity of most patients, the abnormal brain activity of some patients is still masked by the human organ with the strongest specific performance. The abnormal electrical activity of the cerebellum in the brain activity of MCI patients has brought new ideas for the diagnosis, intervention and treatment of the disease. The cerebellum, as a brain tissue distributed outside the cerebral cortex, is located in the cranial fossa at the posterior and lower part of the brain. Compared with the traditional EEG channel data distributed in the cerebral cortex, the physical depth of the electrical activity is greater. In addition, there is abundant muscle tissue near the cranial fossa, which will cause the collected cerebellar EEG activities to be mixed with multiple components. Moreover, as an independent tissue, the electrical activity also has different characteristic distributions in the time-frequency domain from the EEG data of other channels.

[0005] We collected EEG signals with cerebellar electrical channel distribution from most healthy adults, the elderly, and patients with MCI. Analysis found that cerebellar electrical activity is mainly distributed in the medium and high frequency bands or even higher frequency bands, and is discretely distributed in different frequency bands with certain frequency band heterogeneity. Moreover, the distribution between different subjects has certain biological specificity. Therefore, for the feature analysis and application of cerebellar data, it is necessary to introduce multi-band information of subject-specific weight distribution. Summary of the invention

[0006] In order to overcome the deficiencies in the prior art, the present invention constructs a classification method for multi-band information fusion of cerebellar electroencephalogram (EEG) for the classification of early Alzheimer's disease population. Based on the analysis of cerebellar data, a new mode of preprocessing and feature extraction of cerebellar data is provided. The method mainly solves: (1) The current lack of effectiveness of cerebellar EEG information in the early mild cognitive impairment (MCI) patient population, and improves the accuracy of the classification problem between MCI and healthy people; (2) The high classification effect provides a solution to the problem that the EEG acquisition operation steps are complicated and time-consuming, which is unfriendly to patients; (3) The effectiveness of cerebellar EEG classification brings new ideas for the diagnosis, intervention and treatment of MCI diseases. The method specifically includes:

[0007] Step 1: Construct cerebellar EEG data in epochs:

[0008] Resting-state EEG data were collected from MCI patients and healthy control subjects who met the biochemical indicators;

[0009] The acquired resting-state EEG data were preprocessed using the Matlab toolkit eeglab;

[0010] The electromyographic data of the posterior fossa of MCI patients and healthy control subjects were collected using electromyographic acquisition equipment;

[0011] The preprocessed EEG data were subjected to component stripping using the FastICA algorithm;

[0012] The distance between the stripped EEG data and the electromyographic data is calculated using the Euclidean distance, and the components within the threshold are removed to obtain the cerebellum coverage channel EEG data Z;

[0013] Step 2: Construct multi-band weight fusion features

[0014] Construct bandpass filters for the three larger frequency bands of cerebellar electrical activity: intermediate frequency, hyper-gamma, and ultra-high frequency;

[0015] The obtained cerebellum coverage channel EEG data Z is input into three filters to obtain filtered data, and the filtered data is divided into multiple epoch segments at set time intervals; the energy proportions of the three data in the data Z are calculated respectively;

[0016] The maximum difference among the three frequency band components is calculated using the common space mode to obtain its eigenvalue;

[0017] According to the proportion of the three components in the data Z, they are allocated and fused in a weighted manner to obtain the multi-band fusion features of weighted allocation;

[0018] Step 3: Feed the features into classifier training:

[0019] The multi-band fusion features obtained in step 2 are fed into the open source support vector machine model in segments with epoch length L, and the classifier is trained by k-fold cross validation;

[0020] Set the classifier to output the classification results of the data in a non-binary form;

[0021] The epoch segmentation results of each subject are averaged to obtain the evaluation probability result of his / her MCI.

[0022] The beneficial effects of the present invention are:

[0023] First, the present invention successfully constructs cerebellum data into a biomarker with high classification significance, obtains a high classification rate in the experiment, and can provide more dimensional disease diagnosis assistance and reference;

[0024] Second, the present invention provides new features for EEG diagnosis, makes up for the shortcomings of traditional EEG diagnosis, and improves the overall classification rate of EEG in MCI diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is the model result diagram of the present invention. DETAILED DESCRIPTION

[0026] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that the following description is only used to explain the present invention and is not used to limit the present invention.

[0027] The present invention proposes an MCI classification algorithm based on cerebellum multi-band feature fusion, comprising the following steps:

[0028] (1) Comprehensively evaluate the biochemical indicators of clinical subjects and accurately select MCI patients and healthy subjects, which is the data basis for high-quality feature extraction and fusion;

[0029] (2) Electrodes for collecting oculoculograms were placed in the posterior cranial fossa of the subjects, and EEG data containing cerebellar bioelectric information were collected from the two groups of subjects at rest;

[0030] (3) EEG data preprocessing: Based on the Matlab toolkit eeglab, both types of data were preprocessed with the same standards: low-frequency filtering below 0.1, whole-brain average re-referencing, and channel data inspection;

[0031] (4) Collecting electromyographic information of the posterior cranial fossa from two groups of subjects, the MCI group and the healthy control group, and preprocessing the electromyographic information;

[0032] (5) The FastICA algorithm is used to perform source component stripping on the preprocessed EEG data and electromyographic data. The stripped data is recorded as X(x1,x2,x3...x n ) and Y t , X is the source component in the EEG data, Y t is the myoelectric component at time t;

[0033] (6) Use the Euclidean distance to calculate the distance between the n source components in X and the myoelectric source component, that is, C = (c1, c2, c3...c n ),in i∈(1,n), t represents the data sampling time, x it Represents the source information of the i-th source component in the EEG at time t;

[0034] (7) According to the threshold result, the component close to the electromyography is set to zero, and the cerebellum coverage channel data is selected, which is recorded as Z;

[0035] (8) construct three second-order Butterworth bandpass filters: intermediate frequency (30 Hz-70 Hz), super gamma (80 Hz-150 Hz), and ultra-high frequency (160 Hz-320 Hz);

[0036] (9) Send Z to three bandpass filters to obtain filtered data z1, z2, z3. For z1, z2, z3, epoch segmented data z1', z'2, z'3 are obtained at 500ms intervals. The data dimension is 2*500*(L*N), where L is the epoch length of the test data ((t-20) / 500), and N is the number of subjects in the data. Calculate the energy proportion of the three data in data Z respectively.

[0037] (10) The data z1', z'2, z'3 are recorded as N1 and N2 according to the MCI patients and the healthy control group, respectively, and the maximum difference between the two types of data is calculated using the common space model to obtain its feature F (f1, f2, f3);

[0038] (11) The feature F is fused in the form of the ratio of the total effective feature output to obtain the feature And concatenate them in columns to get the fused feature F";

[0039] (12) The obtained feature F" is sent into the open source support vector machine classifier in segments with epoch length L for k-fold cross-validation training;

[0040] (13) Set the classifier to output the classification result A (a1, a2, a3...a N·L ), represents whether the classifier believes that the data provider is an MCI patient, a i =0 means the data owner is HC, a i =1 means the data owner is MCI, N represents the subject number, and i ranges from 1 to N;

[0041] (14) The average of each subject’s epoch results is used to obtain the estimated probability of MCI S (s1, s2, s3…s N ), s i Represents the estimated probability that the i-th subject is MCI, and i ranges from 1 to N.

[0042] See also Figure 1 In the figure, the points corresponding to the star-shaped points 1, 2, 3, and 4 are the detection and verification results of the classifier corresponding to the features of the frontal lobe, parietal lobe, hippocampus, and cerebellum proposed in the present invention on MCI patients. The accuracy of the classifier corresponding to the cerebellum feature in classifying data is 83.3%, which is better. Among the circular points 1, 2, 3, and 4, 1, 2, and 3 are the results of the frontal lobe, parietal lobe, and hippocampus features after adding the cerebellum features and sending them to the classifier for training, and 4 represents the result after all the features are sent to the classifier. Figure 1 It can be seen that the fusion of cerebellum features and other features can improve the effect of the classifier. It shows that cerebellum features are effective and non-repetitive features, which has new significance for the diagnosis, intervention and treatment of MCI.

[0043] The above content has been introduced in detail to the present invention, but the description of the specific implementation mode is only used to explain the method of the present invention and its core idea, so as to facilitate the technical personnel in this technical field to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation mode. For ordinary technical personnel in this technical field, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. An MCI classification algorithm based on cerebellar multi-band feature fusion, characterized by: The following steps are involved: Step 1: Construct cerebellar EEG data in epochs: Resting-state EEG data were collected from MCI patients and healthy control subjects who met the biochemical indicators; The acquired resting-state EEG data were preprocessed using the Matlab toolkit eeglab; The electromyographic data of the posterior fossa of MCI patients and healthy control subjects were collected using electromyographic acquisition equipment; The preprocessed EEG data were subjected to component stripping using the FastICA algorithm; The distance between the stripped EEG data and the electromyographic data is calculated using the Euclidean distance, and the components within the threshold are removed to obtain the cerebellum coverage channel EEG data Z; Step 2: Construct multi-band weight fusion features Construct bandpass filters for the three larger frequency bands of cerebellar electrical activity: intermediate frequency, hyper-gamma, and ultra-high frequency; The obtained cerebellum coverage channel EEG data Z is input into three filters to obtain filtered data, and the filtered data is divided into multiple epoch segments at set time intervals, and the energy proportions of the three data in data Z are calculated respectively; The maximum difference among the three frequency band components is calculated using the common space mode to obtain its eigenvalue; According to the energy proportion of the three components in the data Z, they are allocated and fused in a weighted manner to obtain the multi-band fusion features of weighted allocation; Step 3: Send the features to the classifier training to get the evaluation results The multi-band fusion features obtained in step 2 are fed into the open source support vector machine model in segments with epoch length L, and the classifier is trained by k-fold cross validation; Set the classifier to output the classification results of the data in a non-binary form; The epoch segmentation results of each subject are averaged to obtain the evaluation probability result of his / her MCI.

Images