Learning guided associations of phenotypes and genotypes using high-order multi-modal representations of longitudinal medical data

Abstract

Alzheimer's Disease (AD) is a severe progressive neurodegenerative disorder, threatening the health of millions of people.AD usually progresses along a temporal continuum, initially from a pre-clinical stage, subsequently to mild cognitive impairment (MCI) and ultimately deteriorating to AD. It is estimated that 5.7 million individuals are living with AD and this number is projected to grow to 13.8 million by mid-century, fueled in large part by the aging Baby Boom Generation. The number of AD sufferers worldwide is estimated to be 44 million now and 1 in 85 people will be affected by AD by 2050. With availability of medical devices during the past decades, we now have access to electronic medical records containing a varied set of clinical data coming from multiple sources, including brain imaging scans from different modalities, acquired over time in a longitudinal fashion. In addition, to monitor the progression of AD, different cognitive scores are proposed to clinically indicate the progression at multiple time points. Besides the cognitive scores, behavior assessment are also highly recommended for clinical and severity assessment of dementia. To analyze AD progressions via clinical score and behavior assessment prediction, longitudinal phenotypic measurements have been widely studied. However, these methods fails to handle the missing data challenge in medical records. Higher mortality risk and cognitive impairment hinder older adults from staying in studies that require multiple visits and thus result in incomplete data. In this dissertation, we propose a framework of learning a fixed-length enriched representation for prediction of cognitive outcomes. In our learning framework, we learn a projection for each participant from her or his biomarker records at all available follow-up time points, by which we project the baseline record into a fixed-length vector, regardless of the inconsistent number of brain scans of the participants in a dataset. The enriched representation elegantly combines the baseline biomarkers and all the dynamic imaging measures across time with missing inputs at any time points Armed with the fixed-length biomarker representations, we can take advantage of conventional metric learning methods to predict the cognitive declines of AD patients, which is our theoretical contribution of our works. Besides the prediction of the clinical scores and behaviour assessment, another primary goal of our learning framework is to identify a subset of biomarkers.The identified imaging biomarkers in our work are highly suggestive and strongly agree with existing medical research findings with regard to AD. It demonstrates the correctness and effectiveness of correctness of our proposed learning model and facilitates the understanding of the relationship between MRI measures and cognitive scores or behaviour assessment. This is important for both theoretical research and clinical practices for a better understanding of AD mechanism.