Biostatistics Seminar Series: Sandy Eckel, PhD

Title: "Statistical and machine learning methods for using sensor monitoring systems in environmental epidemiology" Abstract: Sensor technologies are increasingly being incorporated in environmental epidemiology studies, but traditional statistical methods often cannot handle or fully exploit the unique features of these data. The Pediatric Research using Integrated Sensor Monitoring Systems (PRISMS) program—sponsored by the US National Institute of Biomedical Imaging and Bio-engineering—has aimed to develop sensor-based, integrated health monitoring systems for measuring environmental (e.g., personal air pollution exposure), physiological, and behavioral factors in epidemiological studies of pediatric asthma. As part of the PRISMS Data and Software Coordination and Integration Center, our goal has been to develop new data analysis frameworks for predicting asthma exacerbations from sensor data. We are developing methods for data from two complementary study designs: a pilot study of 6 adults wearing several sensors for up to a week and a long-term study (up to 18 months) of households including asthmatic adults and children with stationary monitors located inside and outside the homes. Our basic data analysis pipeline consists of: noise filtering, temporal alignment and data integration, feature engineering (transformation, windowing, and summarization) and health models. To better characterize exposure-response associations, we are investigating a shapelet approach to identify key patterns of exposure within a day, and a novel variant of self-organizing maps to explore diurnal trends in exposure. Data arising from sensor-based monitoring systems offer an exciting new paradigm for clinical and epidemiological studies but require modern statistical approaches and many open questions remain.