Bifurcation of normal & AD brains detected by ensemble learning method applied to longitudinal EEG data

Abstract

In the aged society where the number of Alzheimer’s disease (AD) patients are increasing, delaying the onset or slowing the progress of the disease is important to reduce the socioeconomic costs of the disease. According to the hypothetical time course of AD, the behavior level functional decline is preceded by pathophysiological and structural changes in the brain. Since the examination of pathophysiological change in the brain is costly to be done through regular check-ups, there is an increasing demand for early and inexpensive diagnostic methods. EEG signals contain information about the architecture of the neural network in the brain. EEGs are useful in diagnosing brain conditions such as seizure disorders and psychiatric diseases like schizophrenia but there is no established EEG time course or EEG biomarkers for AD yet. In this study we have investigated the longitudinal change of the brain using EEG, if there are any aspects of EEG that are changing with disease progression, and if so, what the aspects are, and how the change looks like, whether it is abrupt or gradual, and whether it has any characteristic time point. With a longitudinal EEG dataset acquired from AD model (5xFAD) mouse brains and normal mouse brains, we present a new approach to differentiate the change of diseased brain function from the normal by using random forest classifier. On a monthly basis, baseline EEGs and event-related EEGs with auditory stimuli was acquired in multiple trials for each mouse from four brain areas, bilateral frontal and parietal cortices. Individual features of EEG such as resting and evoked state band power, correlation between areas, and phase locking to auditory stimuli were acquired throughout the ages. Age-related EEG features were selected with Spearman’s correlation analysis and divided into training and testing datasets where the training set is acquired at the old age like 34 weeks. A random forest model was trained to distinguish AD mice from control and tested in 5-week windows sliding over the younger age dataset with 1-week step. In a test window, experiments were repeated 30 times with a random state seed for the RF model and the evaluation metrics were obtained by averaging over the results of 30 repetition. In the random forest model with the maximum average accuracy over testing windows, class probabilities were obtained and the divergence of probability distributions for AD and control mice were quantified with Kullback？Leibler divergence. Also, feature importance was determined with Gini impurity scores and visualized with a decision tree. The divergence of the predicted class between normal and AD brains was found dramatically increasing at the age of 6？7 months after birth, while the difference between two started to be observed as early as 4-month after birth which is the earliest age detected with EEG, and at which the memory impairment was reported to start. The features of high importance in predicting the class probability were inter-trial coherence of the responses, especially in gamma band, which band is known to participate in cognitive functions such as perception, attention and memory. We have presented a new method to quantify temporal changes of brain function in AD model mice using machine learning method and shown that the ensemble of EEG features could capture the differentiation of diseased and normal brain states, which could not be done on an individual feature basis.