Visual Analytics for Smartphone Health Phenotyping

Mansoor, Hamid

Etd

Visual Analytics for Smartphone Health Phenotyping

Public Deposited

The healthcare system in the USA is schedule driven, with infrequent assessments that often result in late diagnoses and poor health outcomes. Consequently, researchers are investigating ways to monitor health passively and continuously for preventive interventions. Smartphones are sensor-rich, ubiquitously owned and users’ interactions with them can predict their physical and mental health status. This novel paradigm is called smartphone health phenotyping. However, several issues arise. Ground truth labels provided by the user can be wrong or missing and smartphone sensor data is complex and challenging to interpret. Interactive Visual Analytics (IVA) can assist in correcting and assigning labels to improve data quality during model development and in making sense of user behaviors after deployment. In this dissertation, a suite of IVA frameworks that advance the state-of-the-art in smartphone-sensed health data analyses is proposed. The frameworks aid analysts at various stages of analysis including improving label quality, guiding model steering for health predictors, and health-based insights. Prior IVA work for healthcare has typically analyzed structured datasets such as electronic health reports, which have clear ground truth labels. Moreover, smartphones are not dedicated devices for health monitoring. Machine learning is often used for ailment detection from smartphone-sensed data, which while accurate, is not always explainable. Additional challenges include weak labels provided by users as they live their lives. The COMEX and DELFI IVA frameworks present smartphone-sensed data in linked panes for label correction and assignment, using intuitive encodings of anomaly scores and label predictions along with other contextual information. PLEADES is an IVA framework for exploratory data analysis of smartphone-sensed data, enabling analysts to apply their expertise to guide machine learning model development. PLEADES utilizes clustering and projection and overlays symptom labels to highlight semantic associations between sensor data and human-provided labels. ARGUS and INTOSIS are IVA frameworks that use intuitive glyphs and visual metaphors that allow analysts to gain health-based intuition from contextual factors such as movement patterns and apply semantic/human understandable labels to such data for reproducible analysis. INPHOVIS is an IVA framework to present a complete picture of a person's smartphone-sensed phenotype and predictive health and wellness behaviors. INPHOVIS summarizes sensed behavioral patterns and enables further exploration in linked views that show links between sensed patterns and reported health symptoms, to facilitate understanding of a participant’s phenotype. Finally, VICOMP is an IVA framework for performing population-level analysis of smartphone-sensed data. VICOMP utilizes “Community Phenotypes'', an IVA metaphor to highlight health and wellness measures across different communities and to find optimal groupings for machine learning classifier building of those measures. All proposed frameworks involved detailed goal and task analyses, design of visual metaphors and rigorous evaluation involving use cases with feedback from experts in a diverse field of studies including data science, computer science and health psychology.

Creator