Interpretable Machine Learning for Electro-encephalography

While behavioral, genetic and psychological markers can provide important information about brain health, research in that area over the last decades has much focused on imaging devices such as magnetic resonance tomography (MRI) to provide non-invasive information about cognitive processes. Unfortunately, MRI based approaches, able to capture the slow changes in blood oxygenation levels, cannot capture electrical brain activity which plays out on a time scale up to three orders of magnitude faster. Electroencephalography (EEG), which has been available in clinical settings for over 60 years, is able to measure brain activity based on rapidly changing electrical potentials measured non-invasively on the scalp. Compared to MRI based research into neurodegeneration, EEG based research has, over the last decade, received much less interest from the machine learning community. But generally, EEG in combination with sophisticated machine learning offers great potential such that neglecting this source of information, compared to MRI or genetics, is not warranted. In collaborating with clinical experts, the ability to link any results provided by machine learning to the existing body of research is especially important as it ultimately provides an intuitive or interpretable understanding. Here, interpretable means the possibility for medical experts to translate the insights provided by a statistical model into a working hypothesis relating to brain function. To this end, we propose in our first contribution a method allowing for ultra-sparse regression which is applied on EEG data in order to identify a small subset of important diagnostic markers highlighting the main differences between healthy brains and brains affected by Parkinson's disease. Our second contribution builds on the idea that in Parkinson's disease impaired functioning of the thalamus causes changes in the complexity of the EEG waveforms. The thalamus is a small region in the center of the brain affected early in the course of the disease. Furthermore, it is believed that the thalamus functions as a pacemaker - akin to a conductor of an orchestra - such that changes in complexity are expressed and quantifiable based on EEG. We use these changes in complexity to show their association with future cognitive decline. In our third contribution we propose an extension of archetypal analysis embedded into a deep neural network. This generative version of archetypal analysis allows to learn an appropriate representation where every sample of a data set can be decomposed into a weighted sum of extreme representatives, the so-called archetypes. This opens up an interesting possibility of interpreting a data set relative to its most extreme representatives. In contrast, clustering algorithms describe a data set relative to its most average representatives. For Parkinson's disease, we show based on deep archetypal analysis, that healthy brains produce archetypes which are different from those produced by brains affected by neurodegeneration.