On the use of saliency maps and convolutional neural networks for improved Alzheimer's disease assessment. (31st December 2021)

• Record Type: Journal Article
• Authors: Lopes, Marilia Karla Soares; Cassani, Raymundo; Falk, Tiago H
• Is Part Of: Alzheimer's & dementia. Volume 17(2021)Supplement 6
• Page Start: n/a
• Page End: n/a
• Publication Date: 2021-12-31
• Subjects: Alzheimer's disease -- Periodicals; Alzheimer Disease -- Periodicals; Dementia -- Periodicals; Démence; Maladie d'Alzheimer; Périodique électronique (Descripteur de forme); Ressource Internet (Descripteur de forme); 616.83
• DOI: 10.1002/alz.054710
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: Alzheimer's disease (AD) is a progressive neurodegenerative disease accounting for 60–80∖% of dementia cases worldwide[2]. Early diagnosis can decrease the severity of the disorder in addition to improving the quality of life of patients. Biomarkers based on electroencephalography (EEG) have emerged as a promising tool in the study of AD, with the advantage of being non‐invasive, less expensive, and potentially portable when compared to other biomarkers. One such biomarker has been the so‐called "modulation‐spectral‐patch‐feature" proposed in [1]. Such features were found based on visual inspection of modulation spectra from healthy‐controls and age‐matched AD patients. Over the last few years, however, innovations in machine learning, in particular in deep‐neural‐networks, have revolutionized many biomedical image and signal processing applications[4, 5, 6]. In this paper, we explore their use in building better biomarkers for AD assessment. In particular, we explore the use of saliency‐maps(SM)[3] obtained from classification using convolutional‐neural‐networks(CNN) to extract optimal feature patches in a data‐driven sense. Method: The study collected data from fifty‐four participants, including 20‐healthy‐controls, 19‐Mild‐Cognitive‐Impairment, and 15‐moderate‐to‐severe‐AD, all age‐matched. Twenty‐channel EEG signals were acquired during eyes‐closed‐resting‐state sessions of eight minutes. The EEG recordings were pre‐processed to remove artifacts. The model performance was evaluated using 5‐fold‐cross‐validation(CV). The average SM was extracted from the validation‐set in order to highlight the relevant modulation "patches" in a data‐driven manner. The CNN model has two convolutional‐layers and three fully‐connected‐layers. Result: The tested CNN resulted in the following figures‐of‐merit over the 5‐fold‐CV: 89.4%+/‐2.3% for accuracy and 89.3%+/‐2.3% for f1. From the average SM of each fold, it can be seen that theta‐modulated‐by‐delta and beta‐modulated‐by‐theta were important modulation regions . Others regions such as beta‐modulated‐by‐alpha, alpha‐modulated‐by‐delta, and gamma‐modulated‐by‐beta were also important. While some of these regions match those found visually in[1], other provide more insights into AD assessment. Conclusion: In this paper, we show the efficacy of EEG modulation‐spectrograms coupled CNNs saliency‐maps to extract data specific features useful for Alzheimer's disease diagnosis. In particular, several main regions of interest were identified, thus building on previous work that relied on visual inspection. As future work, we will also explore the importance of each EEG channel, thus potentially leading to a more low‐cost and portable solution.