NEAT activity detection using smartwatch

Abstract

This paper presents a system for distinguishing non-exercise activity thermogenesis (NEAT) and non-NEAT activities at home. NEAT includes energy expended on activities apart from sleep, eating, or traditional exercise. Our study focuses on specific NEAT activities like cooking, sweeping, mopping, walking, climbing, and descending, as well as non-NEAT activities such as eating, driving, working on a laptop, texting, cycling, and watching TV/idle time. We analyse parameters like classification features, upload rate, data sampling frequency, and window length, and their impact on battery depletion rate and classification accuracy. Previous research has not adequately addressed NEAT activities like cooking, sweeping, and mopping. Our study uses lower frequency data sampling (10 Hz and 1 Hz). Findings suggest using statistical features, sampling at 1 Hz, and maximising upload rate and window length for optimal battery efficiency (33,000 milliamperes per hour, 87% accuracy). For highest accuracy, use ECDF features, sample at 10 Hz, and a window length of six seconds or more (37,000 milliamperes per hour, 97% accuracy).