The development of an automated machine learning pipeline for the detection of Alzheimer's Disease

Abstract

Although Alzheimer's disease is the most prevalent form of dementia, there are no treatments capable of slowing disease progression. A lack of reliable disease endpoints and/or biomarkers contributes in part to the absence of effective therapies. Using machine learning to analyze EEG offers a possible solution to overcome many of the limitations of current diagnostic modalities. Here we develop a logistic regression model with an accuracy of 81% that addresses many of the shortcomings of previous works. To our knowledge, no other study has been able to solve the following problems simultaneously: (1) a lack of automation and unbiased removal of artifacts, (2) a dependence on a high level of expertise in data pre-processing and ML for non-automated processes, (3) the need for very large sample sizes and accurate EEG source localization using high density systems, (4) and a reliance on black box ML approaches such as deep neural nets with unexplainable feature selection. This study presents a proof-of-concept for an automated and scalable technology that could potentially be used to diagnose AD in clinical settings as an adjunct to conventional neuropsychological testing, thus enhancing efficiency, reproducibility, and practicality of AD diagnosis.

Figure 1

(A) Power spectral density (all channel montage obtained by averaging PSD for all channels in each subject) in healthy control subjects (HC, n = 23) and those with Alzheimer’s disease diagnosis (AD, n = 18) in the 3–35 Hz frequency range (shaded areas indicate SEM). (B) Mean power in the frequency bands delta (1–4 Hz), theta (5–9 Hz), alpha (10–13 Hz) and beta (14–32 Hz) based on data in panel A (*p < 0.05). Figure generated using Matlab R2021b Update 3 (https://www.mathworks.com/).

Figure 2

(A) Power spectral density in 14 individual EEG channels in HC (n = 23) and AD (n = 18) subjects in the delta, theta, alpha, and beta frequency bands. (B,C) Heat maps showing t-test values for individual channels in each band (B) and for each 1 Hz bin between 5 and 11 Hz (C), red hue indicates p < 0.05, and yellow rectangles indicate 4 features selected for machine learning based on the lowest p values. Figure generated using Matlab R2021b Update 3 (https://www.mathworks.com/).

Figure 3

Summary of Automated Pipeline: (A) EEG collection (B,C) transformation from time domain to frequency domain (D) frequency downsampling (E) removal of low-quality channels (F) automated support vector machine-based artifact detection and removal (G) feature extraction and selection (H) input into ML model (logistic regression). Figure created with BioRender.com.