Using EEG waveform complexity to characterize the progress of Alzheimer's disease

Abstract

The measurement of electric-potential differences on the human scalp, also known as EEG, has regained a lot of interest in the scientific community throughout the last decade. Improved understanding of the origin of brain-oscillations and rapid progress in the fields of artifact removal and Discrete Time-Series (DTS) analysis have yielded new possibilities to use nonlinear methods in search of diagnostic biomarkers from EEG signals. For people suffering from Alzheimer-s Disease (AD) three common trends concerning the nature of EEG signals have been identified by a wide amount of research groups: a shift of the power spectrum to lower frequencies, a decreased complexity of the signal waveforms and altered synchrony between the EEG records. This thesis investigates the second trend, i. e. the irregularity or complexity of the EEG signals. The choice of methods reflects a selection of the most promising markers according to recent articles and reviews that used complexity based measures. The final implemented algorithms are based on the information theoretic concept of entropy and analyze the signals across multiple temporal scales. The main idea behind the approach is to consider the brain as a black box which - for each sampled timepoint - produces measurable outputs (electric voltages on the scalp) depending on the internal functional- and anatomical structure. Each EEG record is preprocessed to remove artifacts and split into 4s segments called epochs. The methods were applied to and averaged over the resulting DTS of all epochs, and compared between 19 electrodes of all 116 subjects from the PRODEM database classified as probable AD. Several biomarkers were extracted and their predictive strength evaluated by the correlation coefficient (coefficient of determination) of a linear (quadratic) regression model between marker values and the patients- Mini-Mental-State-Examination (MMSE)3 scores. The results of different methods are consistent and confirm the observed trends mentioned above: patients with lower MMSE score and hence more severe AD tend to have less complex waveforms in their EEG records. However this trend is only observable at the original timescale of the signal, which reflects the irregularity between time-steps in the order of the inverse sampling frequency (4ms). With a comparatively big study cohort, this thesis clearly demonstrates the limits of complexity based measures since they show weak trends for the original- and no trends for bigger temporal scales. The main reason for their shortcoming is the high intra-subject variability between epochs of a single patient.