Deep Learning Models for NEAT Activity Detection on Smartwatch
Ankita Dewan, Venkata M. V. Gunturi, Vinayak S. Naik · 2023
Our goal of this research focuses on exploring deep learning classification models for recognizing and differentiating Non-Exercise Activity Thermogenesis (NEAT) activities commonly conducted in home settings. NEAT encompasses energy expenditure in activities beyond sleep, eating, and traditional exercise. Our study identifies and classifies NEAT and non-NEAT activities, including cooking, sweeping, mopping, walking, climbing stairs, watching television, and desk work. Our aim is to create a smartwatch-compatible classification model that offers users regular activity updates. By examining various parameters like window length, we assess their impact on battery depletion rate and classification accuracy. Our research diverges from current activity recognition practices that overlook NEAT activities like cooking, sweeping, and mopping while prioritizing high-frequency data. Instead, we utilize a lower frequency data sampling rate of 10 Hz for accurate classification. Findings indicate that artificial neural networks (ANN) achieve the highest accuracy, followed by 1D-CNN, bidirectional long short-term memory (LSTM), and LSTM. Bidirectional LSTM exhibits the lowest battery depletion rate across most window lengths.