Multi-View Learning Using Dependency Models For Medical Decision Making
COM2 Level 4
Executive Classroom, COM2-04-02
Abstract:
Learning from multiple sources of data that describe the same entity/phenomenon is called data fusion. Data collected from varied modalities or features extracted through separate procedures present distinct views of information about an entity. Complementary views contain different yet partly independent information and, provide a fuller picture beneficial for analytic tasks.
Data heterogeneity, high dimensionality and, statistical dependencies among the views are major hurdles to overcome in fusion. Heterogeneity is dealt with specific conversions between distributional representations and high dimensionality through feature selection. Statistical dependencies arise due to the common origin of the views and involve both domain and data-driven definitions. Dependencies can either be directional when a view is conditional on another or mutual when they are correlated. Recent data-driven fusion studies explore mainly the correlation between views. In this thesis, we present fusion techniques that utilize a middle path, to learn data-driven correlations with domain-driven causal constraints. We aim to fuse heterogeneous, multi-view data, possibly from different modalities, using probabilistic models that describe the correlative, causative and complementary semantics of their dependencies.
Dementia is a broad spectrum of age-related neurodegenerative disorders that severely affect a person's cognitive capabilities, rendering them incapable of performing activities of daily living. While multiple procedures such as brain imaging, cerebrospinal fluid tests etc., are useful for diagnosis, various factors from family and medical history to demographic, lifestyle and dietary patterns of individuals are believed to be linked to dementia. Early detection of this incurable disease is of great practical interest to clinicians, to administer timely interventions that possibly delay its progression. Associating dementia with potential risk/preventive factors and symptoms support disease detection, progress tracking and intervention planning.
We apply our methods to make critical decisions in disease management in the absence of sophisticated imaging and invasive markers, with easily available data from self/caregiver-reported questionnaires. This is especially useful in a community healthcare setting with limited access to advanced medical facilities. We learn a disease model for dementia from different views of subject information about a subject, including their past medical profiles and current health indicators. Views in a disease model have two roles; while the subject's background and lifestyle 'contribute' to the disease state, diagnostic markers 'indicate' the presence of the disease. We develop a methodological framework to fuse multi-view medical data that acknowledges the i) heterogeneity in data distributions, ii) roles of the views and, iii) semantics of dependencies, using probabilistic graphical models.
First, we show the inadequacy of a single diagnostic modality for the tasks in dementia management. We present a data-driven dependency mining framework which fuses multiple views at higher levels of abstraction. With this framework, we suggest the integrative analysis of multiple views of dementia markers through an interpretation of their consensus and complementarity. Next, we propose a methodology to incorporate domain knowledge into the definition of dependencies at the feature-level as a step towards 'Explainable Artificial Intelligence (XAI)'. Finally, we propose a unifying approach to model simultaneously the probabilistic associations within and between views and understand the data generation semantics behind their fusion. We hypothesize the dependencies as results of the latent consensus between concepts extracted from multi-view subject data. We plan to extend our work to include longitudinal assessments to model disease progression.