209 Johnson Pavilion: Chonsaeng Kim, PhD, Bergelson Lab
"Coming to our sensors: Why body language is harder to decode than natural language" Ciprian Crainiceanu, Ph.D., Johns Hopkins University
Abstract: Continuously worn accelerometers provide an unprecedented opportunity for studying when and, especially, how people move in their natural environments. I will introduce movelets, a method for predicting the type of movement using one or multiple tri-axial accelerometers. The fundamental idea is that high density accelerometer data can be partitioned into small time series, called movelets, that can then be easily clustered and recognized using simple statistical tools. The approach is inspired by the speech-recognition literature where each movelet is a word in the language of movement. I will point out the pitfalls of such prediction algorithms in the application to observational studies and propose strong design and calibration approaches to mitigate potential problems. I will also dispel the myth that prediction has to be a black box machine learning approach. Instead, I will show that understanding the measurement, having an appreciation for population-level variability, thinking, and building the predictor space are fundamental. I will also briefly discuss several other research topics of our research group: prediction of energy expenditure, association between measures of activity and health, and normalization methods for wearable devices.