Predicting cognitive states from wearable recordings of autonomic function

Abstract

Wearable devices, for gathering bodily measurements from a variety of physiological sources, have now broadly entered the consumer market. Here, we report the results of a preliminary study that aims to extend the usefulness of data collected from a body metrics device that provides 34 continuous physiological measures from sensors embedded in an exercise shirt. Typically, these measures are used to analyze current physical activity by extracting features from the raw data. We examined the possibility that these measures could also be used to predict future cognitive states by allowing users to train a system to categorize historical physiological and movement data, according to their present indication of a salient cognitive transition in mood, motivation, or behavioral context. Indications in the data were self-reported subjective state labels chosen by the user and then configured as a set of buttons within an IBM-designed and implemented mobile app. Our system builds predictors using supervised machine learning applied to the 10-30 min of continuous measures that precede a label. With these labeled measurements, we reliably predicted cognitive events at a degree of accuracy above chance. The proof-of-concept results reported here establish the feasibility of a system that applies our personalized method, which integrates real-Time data from any wearable sensor with cognitive/behavioral labels submitted privately by a user, to anticipate cognitive changes, and which then issues alerts.