Modelling dementia progression based on machine learning and simulations

Neurological disease and mental illness together account for ~30% of the national disease burden in the UK, costing the UK economy £70-100 billion per year (4.5-5% of the Gross Domestic Product) including £23 billion for dementia (OECD 2014). With a rapidly ageing population, this burden is increasing. For Korea, prevalence rates in the elderly will rise from 10% now to 15% by 2050. Alzheimer's disease (AD) is the most common form of dementia where age is the most influential of the main risk factors for developing AD. AD is characterized by a continuous process of degradation involving a preclinical stage, followed by a phase of mild cognitive impairment (MCI), which transitions into dementia once the cognitive dysfunction begins to impact significant on day to day function. Experimental evidence indicates that pathophysiological alterations take place in the brain more than a decade before clinical decline. Therefore, the search for biomarkers for early diagnosis and development of disease-modifying treatments is an ongoing and challenging endeavor. The presence of neurofibrillary tangles and amyloid plaques are the main pathological hallmarks of AD. One emerging hypothesis about the progression of AD posits that these toxic proteins originate in a particular area and propagate throughout neural fibers in a prion-like manner. Similar aetiological spread hypotheses have been proffered for Lewy body dementia. Network neuroscience has proven useful for understanding the impact of psychiatric and neurological disorders on brain-wide networks. In particular, it has been shown that AD strongly disturbs connections between nodes, as well as those nodes occupying a central role in the network (hub nodes). Therefore, network changes could be crucial to predict disease progression.

The most ideal time to intervene with disease-modifying treatment is early on before significant neurodegenerative change and neuronal loss has occurred. However, another highly relevant consideration is improvements in subtype diagnosis i.e. determination of the type of neurodegenerative process giving rise to dementia. For Dementia with Lewy Bodies (DLB), the third most common neurodegenerative dementia, diagnosis is currently difficult as symptoms are similar to AD during the early stages of the disease. However, differentiation is crucial as there are different management trajectories for each disease; for example, neuroleptic drugs which are given to AD can be fatal in the DLB group. However, if we have a precise enough predictive model, we may diagnose patients at a very early stage and subtype. Promising preliminary data, using simulation of disease progression, suggest that we may be able to make an early diagnosis even when subtle changes cannot be detected with the current machine learning approach. In summary accurate early and subtype diagnosis of the underlying neurodegenerative cause is becoming increasingly important for ensuring that future disease-modifying treatments can be targeted in individuals before substantive neurodegenerative deficits have occurred.

Going beyond machine learning subtype classification, our study aims to develop a simulation-based model of disease progression that can become a standard clinical tool to predict future disease progression of individual patients and to facilitate early treatment of the disease leading to improved outcomes for patients and reduced overall healthcare costs.

Our basic strategy is to 1) develop machine learning and dynamic model with available public datasets (ADNI, DPUK, Newcastle data) and 2) validate them with the against other cohorts (UK Biobank and Korea data). In this project, we will combine neuroimaging dataset including DWI, resting-sate fMRI, and PET data.

DEVELOPING MACHINE LEARNING APPROACHES
Using structural connectivity as measured with diffusion MRI in prodromal dementia, starting with connectivity in healthy subjects, and using a computer simulation to study the progression over time towards healthy ageing or dementia, we found that pathophysiological alterations associated with dementia become significantly apparent before the onset of symptoms, meeting diagnostic criteria for clinical dementia, indicating a potential biomarker for progression towards dementia. This machine learning approach could be improved further by using deep learning. Indeed, Korea University, in addition to providing further training and test datasets, has already developed a deep learning approach for dementia brain connectivity data. We will extend these approaches to develop a model of disease progression, looking at changes in white matter and gray matter organization, and testing the role of different underlying biological mechanisms through computational modelling.


VALIDATING OUR APPROACH IN A CLINICAL SETTING
This study utilizes existing datasets for training our approach. We use severable studies to ensure that predicting disease progression is reliable across study sites (UK vs. Korea) and patient cohorts. Note that studies, in addition to neuroimaging data, include the complete set of cognitive and clinical scores (MMSE, CAMCOG, UPDRS-III, NPI-Hall, CAF, Cornell-DS) as well as medication information. In a second step, we will use other datasets (UK Biobank and data from Korea University) as test datasets to see whether our approach can deal with novel datasets.

HEALTH OF POPULATION

Mental health, with associated disorders costing the UK economy £70-100 bn per year, is a major concern due to the increased rate of dementia in an ageing population. Dementia is not anymore an acceptable side effect of growing old. There are around 800,000 people with dementia in the UK, costing the UK economy £23 billion a year.

Dementia is a terminal condition, but people can live with it for 7-12 years after diagnosis. Positive outcomes from this study will lead to the development of a low-cost biomarker in dementia which will have significant clinical utility for the diagnosis of early/prodromal dementia as well as more accurate differential diagnosis; longer term outcomes and impacts may be improved treatment response prediction as well as better stratification of responders in disease modifying trials.


COMPUTER MODELS FOR DIAGNOSIS AND TREATMENT

Predictive models informing interventions could facilitate the introduction of new technologies into the healthcare sector by stratifying patient cohorts and improving treatment outcomes.

This is also in line with the need to develop in silico tools mentioned in the 2015 MRC and Innovate UK report 'A non-animal technologies roadmap for the UK - Advancing predictive biology'. Computational models are a new opportunity to reach the 3R objectives of reducing, refining, or replacing animal experiments.


HEALTHCARE SECTOR BUSINESS POTENTIAL

Computer models have huge potential within the healthcare technology sector: Newcastle University spin-off e-Therapeutics PLC uses computational approaches for drug discovery resulting in our School of Computing's number 1 REF 2014 ranking for research impact. In a similar way, we expect that our simulation of human connectome changes will lead to commercial opportunities and already collaborate with BioMax to investigate these opportunities. Therefore, licencing technology to health informatics knowledge management companies such as BioMax will be one option.