PLEADES: Population Level Observation of Smartphone Sensed Symptoms for In-the-wild Data using Clustering

Abstract

Smartphones are increasingly being used for health monitoring. Training of machine learning health models require studies in which smartphone sensor data is gathered passively on subjects’ phones. Subjects live their lives ’In-the-wild” and periodically annotate data with ground truth health labels. While computational approaches such as machine learning produce accurate results, they lack explanations about the complex factors behind the manifestation of health-related symptoms. Additionally, population-level insights are desirable for scalability. We propose Population Level Exploration and Analysis of smartphone DEtected Symptoms (PLEADES), a framework to present smartphone sensed data in linked panes using intuitive data visualizations. PLEADES utilizes clustering and dimension reduction techniques for discovery of groupings of similar days based on smartphone sensor values, across users for population level analyses. PLEADES allows analysts to apply different clustering and projection algorithms to a given dataset and then overlays human-provided contextual and symptom information gathered during data collection studies, which empower the analyst in interpreting findings. Such overlays enable analysts to contextualize the symptoms that manifest in smartphone sensor data. We visualize two real world smartphone-sensed datasets using PLEADES and validate it in an evaluation study with data visualization and human context recognition experts.