Generating natural language summaries from temporal personal health data

Abstract

Within the personal health domain, there is a vast amount of temporal knowledge that can be collected about an individual (e.g., their time-stamped heart rate and step count data) due to the recent surge in the production of health and fitness tracking devices (i.e., smart, wearable technologies such as smart watches). Although everyday individuals have access to their own data via their personal mobile devices, they typically lack the knowledge or tools required to access the underlying patterns hidden within. Without these patterns, there is a sea of information unavailable to them that could potentially aid them in better comprehending their data and utilizing this knowledge to improve their daily routines. The work within this dissertation focuses on closing this gap between the non-expert individual and the meaningful patterns within their temporal personal health data. The basis of this work revolves around the incorporation of automated explainability in the form of natural language summary generation to highlight the behaviors exhibited by an individual and to evaluate them against their defined health goals. In particular, this work showcases the contributions of our time-series summarization framework within the personal health domain, automating this approach via deep learning, and future plans to include unsupervised rule generation.