Scaling Up Data Integration and Analysis of Sensor Data for Pediatric Asthma

摘要

Childhood asthma is a serious chronic disease affecting 9.3 percent of the American pediatric population. Current epidemiological research tools are limited in the types of information they can collect and aggregate, which restricts their predictive power. The NIH Pediatric Research using Integrated Sensor Monitoring Systems (PRISMS) program is predicated on the idea that better, more accurate sensing of physiology, environmental exposures, and local context can greatly enhance scientific and clinical understanding towards more effective management of chronic disease. At the PRISMS Data and Software Coordination and Integration Center (DSCIC), we are developing a general data integration and analysis architecture, building upon Apache Kafka and Apache Spark, with several advantages. First, our system enables biomedical researchers to (1) collect streaming data from multiple sensors, including at-home and personal pollution monitors, environmental and weather sources, geospatial features and personal trajectories, and Ecological Momentary Assessments (EMA) data collected from subjects' mobile phones; (2) map these heterogeneous datasets to a common schema to facilitate analysis; and (3) train and apply statistical models over both streaming and historical data at scale. Our target applications include prediction algorithms for asthma exacerbations, identifying key personal-level triggers, and eventually providing closed-loop interventions. We will present an overview of the PRISMS-DSCIC data integration and analysis architecture, demonstrate a statistical analysis use case, and describe how our architecture enabled the development of a novel data mining approach using publicly available spatial data (OpenStreetMap) and meteorological data (Dark Sky) to build an air quality model for predicting the concentration of PM2.5 at a fine spatiotemporal resolution.