Population-Level Visual Analytics of Smartphone Sensed Health and Wellness Using Community Phenotypes.

The development of mobile health (mHealth) assessment machine learning models requires data gathering studies in which smartphone sensor data is gathered continuously from users’ phones as they live their lives "In-the-wild". Periodically, participants annotate their sensor data with health, wellness and context labels, which serve as ground truth for machine learning models that can predict a user’s health from their smartphone data. However, as the scale of such studies increases, it becomes difficult to analyze such data and build machine learning models that can work across increasingly diverse, heterogeneous participants. Additionally, non-visual analytics approaches have limited interpretability. This paper innovatively takes a visual analytics approach instead. We propose Visualizing COMmunity Phenotypes (VICOMP), an interactive visual analytics framework for exploring complex population-level smartphone-sensed data. Our approach is based on the concept of Community Phenotypes, which effectively visualizes the groups (or phenotypes) that study participant profiles belong to based on how similar they are (communities). Visual representations of community phenotypes within a large population facilitates sensemaking of group patterns. Using VICOMP, analysts can construct multiple community phenotypes using configurable clustering algorithms of sensed and reported information, and explore and reason about them. VICOMP enables analysts to discover homogenous phenotypical sub-groups within a larger heterogeneous population, for whom group-specific machine learning models are more accurate than one-size-fits-all population-level models. VICOMP depicts community phenotypes using accessible visual metaphors such as superimposed bars to visualize community wellness reports, dimension reduction, projections and heatmaps to represent the distribution of smartphone-sensed features. Connected views facilitate contextualization of health and wellness measures across communities that manifest in smartphone data. The utility of VICOMP is demonstrated by visualizing a real, large scale, dataset collected in-the-wild, containing smartphone sensor data and participant-provided health and wellness ground truth labels. In addition, VICOMP is validated using evaluations with smartphone-sensed context recognition and health experts.