Abstract
Dementia patients experience impaired gait, cardiovascular instability, and elevated fall risk,creating a need for continuous physiological monitoring through noninvasive wearable devices. Although recent wearables incorporate multi-sensor fusion into track movement and cardiovascular function, their performance is constrained by limited battery resources and thermal restrictions. Existing Dynamic Voltage and Frequency Scaling (DVFS) policies adapt to CPU utilization reactively and ignore physiological workload indicators, resulting in unnecessary energy use during inactivity and inadequate responsiveness during critical activity. This work introduces Pulse & Motion, a machine-learning-guided DVFS framework that leverages physiological signals to anticipate power demand in wearable processors. Using heart rate, skin temperature, and three-axis inertial measurements from the PAMAP2 dataset, sensor traces were clustered into four movement-intensity states (rest, light, moderate, intense). Each cluster was associated with realistic multicore power profiles generated using the SniperSim architecture simulator and McPAT modeling framework. A Random Forest regressor trained on these labeled features achieved R² = 0.9327, outperforming Support Vector Regression (SVR) and Multi-Layer Perceptron (MLP) models. A four-tier DVFS policy was derived from the learned power estimates, enabling proactive frequency scaling based on predicted physiological state. Simulation results demonstrate an average energy reduction of approximately 15%, achieved without compromising latency or sensor responsiveness. These findings indicate that DVFS driven by physiological workload prediction can extend battery life in dementia-care wearables while maintaining continuous monitoring reliability.