IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 17, NO. 6, NOVEMBER 2013 1039 Diagnosis of Early Alzheimer’s Disease Based on EEG Source Localization and a Standardized Realistic Head Model Haleh Aghajani, Edmond Zahedi, Senior Member, IEEE, Mahdi Jalili, Member, IEEE, Adib Keikhosravi, and Bijan Vosoughi Vahdat Abstract—In this paper, distributed electroencephalographic (EEG) sources in the brain have been mapped with the objec- tive of early diagnosis of Alzheimer’s disease (AD). To this end, records from a montage of a high-density EEG from 17 early AD patients and 17 matched healthy control subjects were considered. Subjects were in eyes-closed, resting-state condition. Cortical EEG sources were modeled by the standardized low-resolution brain electromagnetic tomography (sLORETA) method. Relative loga- rithmic power spectral density values were obtained in the four conventional frequency bands (alpha, beta, delta, and theta) and 12 cortical regions. Results show that in the left brain hemisphere, the theta band of AD subjects shows an increase in the power, whereas the alpha band shows a decreased activity (P -value <0.05). In the right brain hemisphere of AD subjects, a decreased activity is observed in all frequency bands. It was also noticed that the right temporal region shows a significant difference between the two groups in all frequency bands. Using a support vector machine, control and patient groups are discriminated with an accuracy of 84.4%, sensitivity 75.0%, and specificity of 93.7%. Index Terms—Alzheimer’s disease (AD), brain source localiza- tion, classification, electroencephalography (EEG), standardized low-resolution brain electromagnetic tomography (sLORETA). I. INTRODUCTION T HE number of people with Alzheimer’s disease (AD) is predicted to grow to 15 million by 2050 in the United States only [1] with other countries following a similar trend. Although scientists have not yet found a potent treatment for AD, medications are available to delay the manifestation of symptoms. Early diagnosis and taking these medications has proven to be helpful to postpone the symptoms [2]. Electroencephalography (EEG), a relatively inexpensive and noninvasive technique, has proven its potential in early diag- Manuscript received November 17, 2012; revised January 25, 2013 and February 28, 2013; accepted March 11, 2013. Date of publication March 19, 2013; date of current version November 12, 2013. H. Aghajani, A. Keikhosravi, and B. V. Vahdat are with the Department of Electrical Engineering, Sharif University of Technology, Tehran 11365-11155, Iran (e-mail: hl.aghajani@gmail.com; adib.k.bme@gmail.com; vahdat@sharif. edu). E. Zahedi is with the Department of Electrical Engineering, Sharif University of Technology, Tehran 11365-11155, Iran, and also with the Department of Elec- trical, Electronic, and Systems Engineering, National University of Malaysia (UKM), Bangi 43600, Malaysia (e-mail: zahedi@sharif.edu). M. Jalili is with the Department of Computer Engineering, Sharif University of Technology, Tehran 11365-11155, Iran (e-mail: mjalili@sharif.ir). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/JBHI.2013.2253326 nosis of AD [1], [2]. With this respect, investigated features include reduced coherence [1], [2], EEG slowing [2]–[4], re- duced complexity [2], [5], and chaoticity [6]. The EEG originates mostly from the neocortex [7] by deep- in brain sources and reaches the surface of the scalp [8]. From a distant recording electrode (at the scalp) and using inverse problem techniques [8], [9], the distribution of current sources and sinks can be approximated as dipoles field [10]. There are two major approaches to solve the inverse problem: parametric and nonparametric [11]. In parametric approaches, EEG sources are considered as a single (single-dipole fitting) or multidipoles (multiple-dipole fitting) [11], [12]. In nonpara- metric algorithms, it is assumed that dipoles with fixed lo- cation/direction are distributed all over the cortex or entire brain volume [11]. Most nonparametric methods, including low- resolution brain electromagnetic tomography (LORETA), ex- tract the electrical activity of neurons directly as current den- sity [13]. Standardized LORETA (sLORETA) estimates the cur- rent density obtained through the minimum norm solution [14], and the localization inference is based on standardized values of this estimate [15]. Results obtained through simulations [14] and analytical approaches [16], [17] have shown that under ideal conditions when the signal-to-noise ratio is high, sLORETA has no localization bias. The head model can vary from simple multilayer spherical models to complex ones reflecting the anatomy, using magnetic resonance (MR) and computerized tomography (CT) images. To solve the forward problem in a realistic head model, numerical methods such as boundary element method (BEM) and finite el- ement method (FEM) are required [11]. The advantage of FEM over BEM is that it does not assume homogeneity and isotropy within each region of the head [18]. However, an extensive list of parameters makes FEM a complicated and time-consuming technique. Furthermore, FEM requires in vivo tissue conductiv- ity and anisotropy values that are mostly unknown [19]. There- fore, BEM—which is simpler, faster, and more practical—has become popular [20]. Another major impediment toward using a true realistic head model is the requirement to have expen- sive, high-resolution three-dimensional (3-D) anatomic images of the brain in order to define the required boundaries of BEM for each subject [19]. Using a digital atlas may be a solution to this problem [19]. In [21], a spherical head model is considered and the brain sources related to different bands extracted for 38 patients with AD and 24 healthy individuals [21]. An accuracy of 84% was achieved between two groups using linear discriminant 2168-2194 © 2013 IEEE