Wearable Device Based Activity Recognition and Prediction for Improved Feedforward Control

Abstract

Recognizing and predicting future activity allows current actions based on likely future actions, such as suggesting a restaurant when someone is about to eat, or reminding someone of a potentially missed meeting. In this work, we seek to predict patient activity using wearable devices to improve control of blood glucose levels in people with Type 1 diabetes. We propose an architecture for activity detection and prediction using data from watch and phone based sensors. We further provide results for key components and show that they perform better than simplistic implementations. Using just gyroscope readings gives 63. 7% activity classification accuracy, with a precision of 87.8% for exercise. Prediction error is reduced by ~66% versus naive implementations.