PyCaret machine learning library with three preprocessing steps after eLORETA source estimation predicts Alzheimer's disease

Alzheimer's disease (AD) -the most common form of dementia- begins with mild memory loss and gradually progresses, eventually resulting in a generalized loss of brain function. The pathological changes of AD in the brain cortex begin decades before the onset of symptoms. Improvements in lifestyle habits and disease-modifying treatments before the onset of the disease have been shown to help prevent or delay the onset of AD. However, diagnosis of AD is difficult at the early stage -or even at the prodromal stage [i.e., mild cognitive impairment due to AD (MCIAD)]- since normal aging and other types of dementia also involve memory impairment. Therefore, there is an urgent need to identify markers for the detection of AD at the early stage or pre-onset stage. In this study, we applied exact low-resolution brain electromagnetic tomography (eLORETA) as the source estimation method to electroencephalography (EEG) data. We obtained cortical electrical activity in 96 drug-free AD patients and 147 healthy subjects to train the final model, in addition to activity for 21 MCIAD patients and seven healthy subjects for the purpose of its evaluation. We then applied the low-code machine learning library of PyCaret, with three preprocessing steps (subject-wise normalization, age-difference correction, and log-transformation) to the eLORETA data of AD and healthy subjects. Of the many machine learning classification models used, the linear discriminant analysis (LDA) model showed the highest accuracy, identifying 10 AD patients and 15 healthy subjects with an accuracy of 100.0 %. The LDA model of eLORETA has high transparency and we visualized the discriminant function of the LDA final model using Viewer in eLORETA. Cortical electrical activities in the delta, theta and alpha frequency bands increased in the right dorsolateral prefrontal cortex (DLPFC) regions, as the degree of AD increased (Figs. 2-4). Cortical electrical activity in the beta frequency band decreased in the posterior cingulate cortex (PCC) regions, as the degree of AD increased (Fig. 5) Furthermore, the LDA final model correctly identified 21 MCIAD patients and seven healthy subjects with an accuracy of 96.4 %. Our findings indicate that the LDA final model of eLORETA had the capacity to detect physiological features of AD in EEG data, even before the onset of the disease. Overall, PyCaret with three preprocessing steps after eLORETA source estimation can create an accurate EEG classification model, which makes a significant contribution to the early detection of AD among the many individuals in the general population who remain undiagnosed.