Non-exercise activity thermogenesis monitoring using wearable technology

Abstract

Our thesis explores the complex task of distinguishing NEAT (Non-Exercise Activity Thermogenesis) activities from non-NEAT activities in a home environment using data from wearable smartwatch sensors. It presents a multifaceted problem where low-frequency sensor data is used to differentiate thirteen distinct household activities. Existing research often prioritizes high-frequency data or multiple sensors, overlooking fundamental home activities. To address these limitations, this research introduces innovative AI models, capable of achieving superior accuracy while working on low frequency. This dissertation unfolds in three key steps, each contributing to the overall understanding and advancement of NEAT activity recognition. In our first work (Chapter Three), we introduced the Hierarchical Model, tailored to differentiate seven home activities using low-frequency (1Hz) sensor data. Our experiments revealed the model’s remarkable accuracy, outperforming traditional flat models like XGB. Even when the sampling frequency was reduced to 1Hz, the Hierarchical Model maintained substantial accuracy. These results provided a foundation for recognizing NEAT activities using wearable technology, highlighting the potential of our approach. In our second work (Chapter Four) we delved deeper into the system’s development. The system, worn on a smartwatch, identifies and classifies 13 distinct activities, including both physical and sedentary actions. The user initiates data collection, which is transmitted to a central server for interpretation. Key parameters, such as battery depletion rate and data sampling rate, were evaluated, with the goal of future on-device deployment. Notably, our system accommodates unclassified activities through the introduction of the ”OTHERS” class, a versatile approach that can be adjusted based on desired strictness or leniency. In our last work (Chapter Five) we tackled the challenge of distinguishing household activities using low-frequency data without relying on external connectivity. Our innovative Hybrid Model achieved superior accuracy compared to other models, making it more user-friendly in environments without readily available Wi-Fi or paired devices. The choice of a 10-second window length was found to balance accuracy and battery efficiency. Additionally, the deployment of neural networks, including 1d-CNN, LSTM, and Bidirectional LSTM, allowed us to propose a Hybrid Model, integrating TensorFlow Lite (TFLite) for improved performance and efficiency on the smartwatch. This dissertation provides a comprehensive understanding of NEAT activity recognition using wearable technology in a home environment. From the introduction of the Hierarchical Model to the development of the system and the implementation of neural networks, our findings pave the way for future advancements in the field. Distinguishing NEAT and non-NEAT activities has the potential to revolutionize how we monitor and understand daily activities, impacting areas like healthcare, fitness, and beyond.