^{1}

^{1}

^{2}

^{3}

^{3}

^{3}

^{1}

^{2}

^{3}

There is not a specific Alzheimer's disease (AD) diagnostic test. AD diagnosis relies on clinical history, neuropsychological, and laboratory tests, neuroimaging and electroencephalography. Therefore, new approaches are necessary to enable earlier and more accurate diagnosis and to measure treatment results. Quantitative EEG (qEEG) can be used as a diagnostic tool in selected cases. The aim of this study was to answer if distinct electrode montages have different sensitivity when differentiating controls from AD patients. We analyzed EEG spectral peaks (delta, theta, alpha, beta, and gamma bands), and we compared references (Biauricular, Longitudinal bipolar, Crossed bipolar, Counterpart bipolar, and Cz reference). Support Vector Machines and Logistic Regression classifiers showed Counterpart bipolar montage as the most sensitive electrode combination. Our results suggest that Counterpart bipolar montage is the best choice to study EEG spectral peaks of controls versus AD.

Alzheimer’s disease (AD) diagnosis is based upon clinical history, neuropsychological and laboratory tests, neuroimaging, and electroencephalography (EEG). New approaches are necessary to earlier and more accurate diagnosis [

EEG visual analysis can be a helpful diagnostic test in AD [

Since the first quantitative EEG (qEEG) studies by Lehmann [

Despite the knowledge grounded in this field during the last decades, there are lots of unanswered questions that hinder qEEG consolidation as an AD diagnostic tool. Our objective was to study if distinct electrode montages have different sensitivity when differentiating controls from AD patients.

The dataset was composed of electroencephalograms (EEGs) recorded from two groups aged from 60 to 80 years: (S1) 12 normal subjects and (S2) 22 probable AD patients (NINCDS-ADRDA criteria) [

The EEGs were recorded with 12 bits resolution, band pass of 1–50 Hz, and sampling rate of 200 Hz. A

A 512-point Hamming Fast Fourier Transform (FFT) algorithm was used to process the epochs analysis. The windows were 2.5 seconds long with 90% of overlap between successive windows [

Feature extraction is the method used to mining some characteristics of a particular signal epoch producing data that can represent events [

Biauricular reference (Bar): Fp1-A1, Fp2-A2, F7-A1, F8-A2, F3-A1, F4-A2, C3-A1, C-A2, T3-A1, T4-A2, P3-A1, P4-A2, O1-A1, O2-A2;

Longitudinal Bipolar (Lbp): Fp1-F3, F3-C3, C3-P3, P3-O1, O1-T5, T5-T3, T3-F7, F7-Fp1, Fp2-F4, F4-C4, C4-P4, P4-O2, O2-T6, T6-T4, T4-F9, F8-Fp2;

Crossed Bipolar (Bcr): Fp1-Fp2, F7-F3, F3-Fz, Fz-F4, F4-F8, T3-C3, C3-Cz, Cz-C4, C4-T4, T5-P3, P3-Pz, Pz-P4, P4-T6, O1-O2;

Counterpart bipolar (Bco): F7-F8, F3-F4, T3-T4, C3-C4, P3-P4, T5-T6, O1-O2;

Cz reference (Czr): Fp1-Cz, Fp2-Cz, F3-Cz, F4-Cz, F7-Cz, F8-Cz, T3-Cz, T4-Cz, C3-Cz, C4-Cz, T5-Cz, T6-Cz, P3-Cz, P4-Cz, O1-Cz, O2-Cz.

Each of these electrode montages (Figure

Spectral peaks montage maps. Lines correspond to subtractions used to calculate spectral peaks. From left to right, top to bottom: Counterpart Bipolar (Bco), Longitudinal Bipolar (Lbp), Crossed Bipolar (Bcr), Biauricular reference (Bar), and Cz reference (Czr).

The EEG dataset was composed of 1360 epochs (40 epochs of 34 subjects). The analysis was based on the leave-one-subject-out process: 1320 epochs were used for training and 40 epochs from one subject for testing. It means that, each time, the classifier was trained with epochs from all individuals except the one going to be tested. This procedure was performed to test the classifiers discriminative capacity to work with data diverse from that presented in the training period. The leave

SVMs constitute a supervised Machine Learning (ML) technique based on the Statistical Learning Theory [

In this experiment, the Weka tool [

Logistic regression is part of a category of statistical models called generalized linear models. LR is a classification tool frequently used to help diagnosis [

Table

Accuracy, sensitivity, and specificity rates for each montage. Best results in bold and worst results in italic.

72,72 ± 36,80 | 84,09 ± 27,52 | 51,88 ± 43,39 | 66,03 ± 35,76 | 48,75 ± 38,04 | ||

79,41 | 58,33 | 50,00 | ||||

48,96 ± 39,32 | 76,59 ± 33,23 | |||||

67,65 | 77,27 | 50,00 | ||||

70,07 ± 36,81 | 85,57 ± 23,12 | 66,32 ± 32,50 | 76,14 ± 28,05 | 48,33 ± 33,50 | ||

76,47 | 95,45 | 67,65 | 81,82 | |||

70,22 ± 37,70 | 81,36 ± 31,15 | 49,79 ± 41,33 | 71,62 ± 28,37 | 80,45 ± 24,64 | 55,42 ± 28,52 | |

70,59 | 50,00 | 73,53 | 86,36 | 50,00 |

The second line of Table

It is important to note that high standard deviation (SD) is a methodological consequence of the leave-one-subject-out test. If an individual had bad epochs accuracy, the group mean was low and the SD high. Bco was the montage with lower SD, consequently, indicating less variability in number of correct diagnosis.

Table

Number of patients with epoch accuracy rates equal to 100%, exceeding or equal to 75%, less than or equal to 50%, and equal to 0% for each test. Best results in bold and worst results in italic.

Support Vector Machines | Logistic Regression | |||||||

Bipolar Counterpart | 15 | |||||||

Longitudinal Bipolar | 21 | 7 | 2 | 7 | 20 | 1 | ||

Crossed Bipolar | 20 | 8 | 11 | |||||

Biauricular Reference | 14 | 8 | 2 | 8 | 11 | |||

Cz Reference | 15 | 20 | 10 | 20 | 9 |

SVMs tests presented Lbp as the montage with maximum epoch accuracy (16 subjects with 100% accuracy), followed by Bcp e Czr (15 cases each). Bco was the montage with higher number of cases with accuracy greater than or equal to 75%, less cases with accuracy less than or equal to 50%, and without cases of 0% correct classification.

The LR tests ratified Bco as having the highest number of 100% accuracy results, the highest number of cases with accuracy greater than or equal to 75%, less cases with accuracy less than 50%, and no cases of 0% correct classification (in this last case similar to Bar and Czr, both with zero cases).

This study suggested that Bco was the more trustworthy montage because of his higher rates of 100% epoch accuracy and absence of 0% cases to both classifiers. Consequently, other parameters could be tested based on LR. The odds ratio values (ODDR) could be analyzed from the ratio AD/controls (Table

Odds ratio to Bipolar Counterpart LR test. In bold the significant ones (>1).

delta | theta | alpha | beta | gamma | |

F3-F4 | 0,371 ± 0,063 | 0,969 ± 0,230 | 0,913 ± 0,052 | ||

F7-F8 | 0,728 ± 0,221 | 0,659 ± 0,045 | |||

C3-C4 | 0,693 ± 0,176 | 0,734 ± 0,191 | 0,975 ± 0,047 | ||

T3-T4 | 0,753 ± 0,182 | 0,177 ± 0,068 | 0,263 ± 0,059 | 0,836 ± 0,087 | |

T5-T6 | 0,277 ± 0,075 | 0,104 ± 0,029 | 0,875 ± 0,080 | 0,649 ± 0,034 | |

P3-P4 | 0,574 ± 0,107 | 0,511 ± 0,091 | 0,019 ± 0,008 | 0,962 ± 0,118 | |

O1-O2 | 0,402 ± 0,126 | 0,767 ± 0,177 | 0,761 ± 0,072 | 0,792 ± 0,036 |

Among these ODDR features, the electrodes F3-F4, F7-F8, C3-C4, and T5-T6 presented values of ODDR > 1 to theta band; the electrodes F7-F8 and O1-O2 presented values of ODDR > 1 to delta band; F3-F4, F7-F8 and C3-C4 presented values of ODDR > 1 to beta band, and T3-T4 and P3-P4 presented values of ODDR > 1 to gamma band.

EEGs of mild DA have higher theta activity and low beta activity [

The analysis of the number of electrodes related to each montage demonstrates that the montages with higher number of signals were Lbp and Czr with 16 signals each, followed by Bar and Bcr with 14 signals. The montage with lowest number of signals was Bco (7 signals). We can say that Bco is also the more compact (less electrodes), consequently, less expensive in terms of processing time.

To sum, our results are in accordance with the literature that suggests that the spectral peak is an efficient tool in AD diagnosis [

Although more tests are needed to confirm the generalization power of our classifiers, we propose that spectral peak calculation using different montages of electrodes have an influence on the classification results (differentiation) of normal subjects and patients with AD. Our future goal is to generalize the results obtained increasing the number of probands.