TNT: Tracking Network dynamics with Tensor factorisations. Application to the human Chronnectome in Alzheimer's disease

Alzheimer's disease is a major problem to UK society. Because of the ageing population, the number of people with dementia will increase dramatically in the next years: from about 850,000 today to 1,000,000 by 2025. The current annual cost of dementia to the UK is £26 billion even not everybody with dementia receives a diagnosis. Alzheimer's disease is the most common cause of dementia and it is particularly difficult to diagnose because there are no objective biomarkers for it and the diagnosis relies on the medical history of the patient. We need better ways to detect and monitor the changes that Alzheimer's disease causes in the brain. To achieve this, we will consider the electroencephalogram (EEG), an affordable piece of equipment that can be used outside hospitals to measure brain activity safely at several locations over the scalp (called "channels"). We will create new signal processing tools to analyse EEG brain networks. Doing so will lead to objective ways to monitor Alzheimer's disease.

Namely, this interdisciplinary project will develop a novel set of processing techniques based on tensor factorisations to inspect how the components of brain activity networks change with time. We will then implement methods to compare the temporal profiles of the components estimated for different groups of people (e.g., healthy people versus patients).

Our project is motivated by the facts that: 1) the EEG can measure fast changes in brain activity, 2) Alzheimer's disease damages brain connections, and 3) preliminary results indicate that Alzheimer's disease affects the temporal behaviour of brain activity.

Indeed, there is an increasing interest in understanding brain activity networks and their evolution with time, as this would open up radically new ways to monitor brain diseases. Promising pilot results have reported in, e.g., Parkinson's and multiple sclerosis but, currently, there are no appropriate ways to inspect how the networks change with time systematically. Instead, we will develop a framework based on tensor factorisations (a set of algebraic and computational techniques to analyse tensors: n-mode data arrays with n>=3) to inspect the components of networks directly from the data without the need for manual intervention. We will then apply it to EEG signals. First, for each person, we will assess the coupling between channels of the EEG as a function of time and frequency. These results naturally fit into a multi-modal representation: a "connectivity tensor". Then, we will decompose the "connectivity tensor" into its underlying components. We will implement constraints to bring previous information into the decompositions, including novel ways to measure the natural organisation of the network components. Finally, we will assess the robustness of the extracted network components and we will inspect how Alzheimer's disease changes them.

We will apply our methods to two different sets of EEG signals measured from patients with Alzheimer's disease, people with mild cognitive impairment (a condition that sometimes precedes Alzheimer's disease), and healthy volunteers. One of the EEG datasets measured the activity of the brain at rest using a small number of channels, whereas the other has been recorded during a short-term memory task that has shown promise in the detection of early Alzheimer's disease with a larger number of EEG channels. Hence, we believe that revealing how the EEG network changes with time during this task could lead to a non-invasive, affordable and portable tool to monitor Alzheimer's disease. Nonetheless, this project will have much wider implications because it will benefit the signal processing, tensor factorisation and network analysis communities and the techniques will be readily applicable to other types of data, both inside and outside clinical settings.

Multimodal time-varying data occur in many contexts, from social networks to brain activity. Inspecting the network connections between different elements in such datasets and how they depend on other variables (e.g., time) would lead to new fundamental information about them. This is the problem we are tackling in this project, where we will develop a new approach to analyse the dynamical behaviour of networks with advanced tensor factorisations that incorporate constraints based on graph theory. We will illustrate it by processing time-varying electroencephalogram (EEG) brain connectivity in Alzheimer's disease (AD).

As a result, this proposal has the potential for very significant and diverse impact on:
+ Technical beneficiaries working in:
- Tensor factorisations and network analysis.
- Signal processing in general.
- Other scientific fields, such as data fusion, network design, etc.
+ General population interested in:
- Social networks.
- Dementia, particularly patients and carers.
+ Clinical and neuroscience communities:
- Neuroscientists and clinical experts.
- Medical equipment companies.

The most immediate impact of our project will be on other researchers. This includes not only the processing of brain activity, but also the more general field of the analysis of time-varying connectivity patterns (e.g., in communication or social networks) and the processing of data collected over distributed sensors. The latter area is of special interest in relation to "Big Data". Researchers working in the processing of biomedical signals will also be interested in our work. We expect it to be readily applicable to the analysis of, for example, electroencephalogram activity in epilepsy and magnetoencephalogram recordings.

Research about the brain is likely to be of interest to the general population. This is further the case because our project incorporates ideas from network science, which have been widely embraced by popular culture (e.g., the idea of the six degrees of separation between any two people in the World). Moreover, we will examine our technique for assessing AD, a condition that has received considerable media attention lately due to the UK government considering AD as 'the key health challenge of this generation'.

Indeed, our methods and results will be of interest to neuroscientists and clinicians working in novel ways to monitor AD, which is a major UK societal problem. Nowadays, in the UK, 850,000 people have dementia, of which AD is the most common cause. The number of people with dementia in the UK will increase to over 1,000,000 by 2025. Dementia has a very large cost to UK society even though not everyone receives a diagnosis. (Actually, only about half of the people with dementia are diagnosed.) The current annual estimated cost of dementia to the UK is £26 billion. Despite the fact that the results of this 13-month study cannot be realistically expected to be applied to the clinic directly, they can lead to major breakthroughs in the understanding of brain function in AD. We believe that this could be of interest to companies manufacturing medical equipment, as they could be interested in exploring the exploitation of our methods (as tools to detect disease patterns in EEG recordings), as well as to clinicians working in the early diagnosis of dementia. This is particularly relevant in the case of the EEG dataset acquired during the visual short-term memory binding task, as this memory test has proven to be specific to AD in different populations. Therefore, we expect that the combination of sensitive EEG analysis and specific memory task will, in the future, lead to a low-cost, non-invasive, sensitive and specific diagnostic test for early AD, which would result in better disease management opportunities for patients. Nonetheless, our processing are general and directly applicable to brain recordings in other diseases.