Temporal Intelligence in Wearable Health: Towards Robust and Reliable Gesture-Based AI Frameworks

Abstract

The widespread adoption of smartwatches and wrist-worn wearables presents a unique opportunity to enable continuous, privacy-preserving health monitoring in everyday life. This dissertation develops a series of AI-driven frameworks that leverage inertial and multimodal sensor data from commodity smartwatches to monitor clinically significant health behaviors. Rather than relying on intrusive audio or video recordings, the proposed systems interpret characteristic hand and arm gestures associated with the health events, enabling the device to act intelligently and selectively without compromising user privacy. Across all the frameworks, the work progressively advances from lightweight models to compact deep learning architectures and transformer-based models with modern positional encoding strategies, each tailored to the unique temporal and motion characteristics of the target behavior. Collectively, this work establishes that temporal intelligence embedded in wrist-worn devices can serve as a practical, scalable, and privacy-respecting foundation for longitudinal health monitoring, with meaningful implications for early intervention, medication adherence, and clinical decision support.