Applications of blind source separation to the magnetoencephalogram background activity in alzheimer's disease

Resumo

In this Doctoral Thesis, Magnetoencephalogram (MEG) background activity from 36 patients with a diagnosis of probable Alzheimer's Disease (AD) and 26 healthy elderly control subjects has been analysed with Blind Source Separation (BSS) methods. Our aim was to apply BSS techniques to help in the analysis and interpretation of this kind of brain activity, paying special attention to AD. The MEG is the non-invasive recording of the tiny magnetic fields generated by the neurons. This neurophysiological technique measures the brain cortex activity directly, without interpreting the information on the basis of vascular or metabolic changes. Its temporal resolution is high and the magnetic recordings neither depend on any reference point nor are affected by the resistive properties of extra-cerebral tissues. On the other hand, the MEG apparatus needs superconductive materials and magnetically shielded rooms to properly acquire the brain signals. This has prevented any widespread use of this technique to record the brain activity. Despite the fact that difficulties are faced in the analysis of these signals, the MEG could provide relevant information about diverse brain states and diseases, such as AD. AD is a progressive neurodegenerative disorder. It causes memory loss and other cognitive and behavioural symptoms that impair the activities of daily living. AD is the most common dementia in the Western World as it accounts for 50% to 60% of all cases. It shows an almost exponential increase with age. As a result, its prevalence in people over 85 years is between 24% and 33%. In clinical practice, AD must be differentiated from other dementias, though a definite diagnosis can only be made by necropsy. The criteria for AD diagnosis largely depend on the exclusion of other disorders. Of note is that the accuracy in the clinical diagnosis is limited, with sensitivity of around 80% and specificity of 70%. Thus, it is important to develop new approaches that might help in AD detection. The term BSS denotes a set of techniques useful to decompose multichannel recordings into their constituent underlying components. The BSS defines a generative model for the measurements and it tries to estimate the inner components (or sources) by making a few general assumptions about the data. The most important hypothesis is that the BSS components are mutually independent or, alternatively, that they are mutually decorrelated over time. BSS extracts the sources by exploiting this assumption. Thanks to this ability, the application of BSS techniques to MEGs can help to inspect and analyse these biomedical signals from new perspectives. Thus, this could provide us with both novel methodologies to deal with problems encountered in the processing of these recordings and relevant information about brain activity. For these reasons, in this Doctoral Thesis, MEG data were processed with BSS techniques in the context of four different applications: * The decompositions of real MEGs computed with five common BSS algorithms were compared to assess their degree of similarity. The results showed that the most consistent (i.e., similar) pair of algorithms was AMUSE-SOBI, followed by JADE-FastICA. Additionally, the overall level of similarity increased as longer signals were decomposed. * The ability of several combinations of BSS algorithm, epoch length and artefact detection metric to automatically reduce the cardiac, ocular and power line artefacts in the MEGs was assessed. The results indicated that a Constrained Blind Source Separation (cBSS) approach was suitable to remove the cardiac activity. Additionally, a combination of artefact detection metrics based on entropy or power criteria with AMUSE or SOBI could help to reduce the ocular contamination. Finally, the electrical noise could be reduced by means of a spectral metric and AMUSE. * The ability of a BSS preprocessing to improve the separation between AD patients and controls' spectral and non-linear features from MEGs was measured. Ordering criteria were defined to straightforwardly compare the BSS components of different subjects. The comparisons between the classifications derived from the unpreprocessed and the BSS preprocessed MEG signals revealed rises in the areas under the ROC curves associated with the features between 0.023 and 0.227 and accuracy increases of up to 22.6% in the best cases. These corresponded to preprocessings developed using AMUSE and SOBI with a spectral ordering of the components. * An adaptive framework to extract rhythmical brain activity from diverse scalp regions was introduced. This employed the local adaptiveness of an Empirical Mode Decomposition to estimate oscillating brain activity in several bands and a cBSS to extract the actual brain rhythms from each region. The statistical analysis suggested that AD might affect the Spectral Coherence (Coh(f)) between regions, although the results were not significant. However, a leave-one-out classification analysis based on the Coh(f) computed from spectral bands classified the AD patients versus the elderly control subjects with an accuracy equal to 96.8%. In summary, the findings of this Doctoral Thesis suggest the utility of BSS to help in the processing of MEG background activity and in the identification and characterisation of AD. Therefore, BSS may be an important tool to analyse this kind of biomedical recordings. Nevertheless, further investigations are needed to confirm our results.