Primate brain pattern-based automated Alzheimer's disease detection model using EEG signals

Electroencephalography (EEG) may detect early changes in Alzheimer's disease (AD), a debilitating progressive neurodegenerative disease. We have developed an automated AD detection model using a novel directed graph for local texture feature extraction with EEG signals. The proposed graph was created from a topological map of the macroscopic connectome, i.e., neuronal pathways linking anatomo-functional brain segments involved in visual object recognition and motor response in the primate brain. This primate brain pattern (PBP)-based model was tested on a public AD EEG signal dataset. The dataset comprised 16-channel EEG signal recordings of 12 AD patients and 11 healthy controls. While PBP could generate 448 low-level features per one-dimensional EEG signal, combining it with tunable q-factor wavelet transform created a multilevel feature extractor (which mimicked deep models) to generate 8,512 (= 448 × 19) features per signal input. Iterative neighborhood component analysis was used to choose the most discriminative features (the number of optimal features varied among the individual EEG channels) to feed to a weighted k-nearest neighbor (KNN) classifier for binary classification into AD vs. healthy using both leave-one subject-out (LOSO) and tenfold cross-validations. Iterative majority voting was used to compute subject-level general performance results from the individual channel classification outputs. Channel-wise, as well as subject-level general results demonstrated exemplary performance. In addition, the model attained 100% and 92.01% accuracy for AD vs. healthy classification using the KNN classifier with tenfold and LOSO cross-validations, respectively. Our developed multilevel PBP-based model extracted discriminative features from EEG signals and paved the way for further development of models inspired by the brain connectome.

Data and relevant material are available from the Datashare website: http://datashare.is.ed.ac.uk/handle/10283/2783 (Smith et al. 2017).

The authors state that this work has not received any funding.

Authors and Affiliations

1. Department of Digital Forensics Engineering, College of Technology, Firat University, Elazig, Turkey

   Sengul Dogan & Turker Tuncer

2. Department of Computer Engineering, College of Engineering, Ardahan University, Ardahan, Turkey

   Mehmet Baygin

3. Vocational School of Technical Sciences, Firat University, Elazig, 23119, Turkey

   Burak Tasci

4. School of Science and Technology, Singapore University of Social Sciences, 463 Clementi Road, Singapore, 599494, Singapore

   Hui Wen Loh

5. School of Business (Information System), University of Southern Queensland, Toowoomba, QLD, 4350, Australia

   Prabal D. Barua

6. Faculty of Engineering and Information Technology, University of Technology Sydney, Sydney, NSW, 2007, Australia

   Prabal D. Barua

7. Department of Cardiology, National Heart Centre Singapore, Singapore 169609, Singapore

   Ru-San Tan

8. Duke-NUS Medical School, Singapore 169857, Singapore

   Ru-San Tan

9. Department of Electronics and Computer Engineering, Ngee Ann Polytechnic, Singapore, 599489, Singapore

   U. Rajendra Acharya

10. Department of Biomedical Engineering, School of Science and Technology, SUSS University, Singapore, Singapore

    U. Rajendra Acharya

11. Department of Biomedical Informatics and Medical Engineering, AsiaUniversity, Taichung, Taiwan

    U. Rajendra Acharya

Contributions

All authors contributed equally to the study.

Corresponding author

Correspondence to Sengul Dogan.

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Appendix

Tunable Q-factor wavelet transform (TQWT)

As a representative wavelet transform expressing wavelets in the frequency domain, discrete-time signals of finite length are analyzed using radix-2 FFTs (Selesnick 2011). Compared to traditional Q-fixed wavelet transforms, TQWT has the excellent feature of estimating the Q factor easily and continuously adjusting it depending on the different oscillations of different signals. Moreover, successful error feature generation using TQWT-based parsing relies on appropriate TQWT parameters such as parsing level, Q factor, and redundancy (Kong et al. 2018). As noted, TQWT takes three parameters (Q: oscillatory value, r: redundancy, J: number of level). The first parameter is Q and it assigns number of oscillations. If Q is 1, there is no oscillations.

TQWT contains two filter banks and these two-channel filter banks consist of a low-pass and a high-pass channel. The low-pass channel has a low-pass filter (LPF) followed by a low-pass scaling factor (LCF). The high-pass channel has a high-pass filter (HPF) followed by a high-pass scaling factor (HCF). The ratio of the center frequency (CF) of each subband to the bandwidth is equal to the Q factor used to implement the TQWT. The following expressions can be used to calculate the MF and bandwidth of a subband (Selesnick 2011):

Equations 1 and 2 define the effect of scaling factors on the center frequency and bandwidth is observed. Herein j represents the subband number where 1 ≤ J ≤ J + 1 fs is the sampling frequency. The Q factor is controlled by R, and its connection is given in Eqs. 3 and 4 (Zhang et al. 2001).

Iterative neighborhood component analysis (INCA)

In order to explain INCA better, we first gave information about NCA in this section. Neighborhood component analysis (NCA) is a technique for reducing dimensions and selecting features. In machine learning applications, it is crucial to measure features. In classification studies, NCA is widely used as one of the most successful learning algorithms. Using NCA, classification operations are carried out by learning the projections of the vectors that optimize the nearest neighbor classifier's accuracy criteria. As another option, the NCA can select a linear projection that optimizes the performance of the nearest neighbor classifier within the projection area (Goldberger et al. 2004a).

NCA is a method that aims to learn the Mahalanobis distance to be used in classification and performs the classification using this distance. While the NCA method calculates Mahalanobis distance, it learns the projection matrix and prevents the inverse operation of the matrix.

This method uses the Mahalanobis distance formula described below to define the Mahalanobis distance.

In this equation, P is defined as the projection matrix that transforms the data. Thus, NCA moves from learning the Mahalanobis distance to learning the P matrix.

A data in the training data determines the class label by choosing another data as a neighbor. Data with transformed probability values are defined using softmax regression with euclidean distances.

Based on the stochastic selection rules, the probability of correctly classifying the data is calculated, and the set of data in the same class is displayed.

NCA then calculates the P matrix by maximizing the number of correctly classified points.

The f value that changes according to the P matrix gives rise to the following gradient rule.

Thus, a gradient-based optimizer is used.

INCA is an improved version of the NCA feature selector and it is an iterative version of the NCA. The main objective of the INCA feature selector is to choose the optimized number of features. To choose optimal feature vector, we used a loop and loss function. The steps of the INCA are given below.

Step 1 Calculate the qualified indexes by deploying the NCA algorithm.

Step 2 Define a loop range to decrease time complexity.

Step 3 Choose feature vectors iteratively by using the created loop.

Step 4 Apply the used loss function to the selected feature vectors and create a loss array.

Step 5 Select the feature vector with minimum loss value.

kNN

The KNN algorithm is one of the more common algorithms used in general and one of the most widely known and widely used algorithms. Given unlabeled test data, kNN finds the closest k in the data set and then assigns the most appropriate label (Guo et al. 2003).

KNN performs object classification according to the closest training examples. A majority vote of its neighbors classifies an object.

Two situations should be considered in the kNN algorithm. The first of these is the correct choice of k, which affects performance. When k values are large, it may ignore small but important patterns. The other case is to calculate the distance between test samples and neighbors (Zhang 2016). The most popular measure of distance in distance functions are euclidean, manhattan, and Minkowski.

X1, X2, …, Xn and Y1, Y2, …, Yn represent the feature vector where n is the dimension of the feature space. The mathematical representation of the Euclidean distance is given in Eq. 11.

The mathematical representation of the manhattan distance, which calculates the difference of two data points in absolute value, is given in Eq. 12.

The mathematical representation of the Minkowski distance, where p ∈ (0, ∞) for a constant p, is given in Eq. 13.

Mode-based majority voting

Majority voting relies on the principle of normalization, which is derived from the sum of the probabilities given by the classifiers. As a result of this classification, the highest probability class combination result is obtained within the normalized result (Suen and Lam 2000).

Springer Nature or its licensor holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions