Classification of Physical Activity States Using the Temporal Dynamics of Pulse Arrival Time

Abstract

Human Activity Recognition (HAR) based on physiological signals offers a direct view into the body's metabolic and autonomic response. This work proposes a Long Short-Term Memory (LSTM) recurrent neural network to classify activity states (resting, walking, running) based exclusively on the temporal dynamics of Pulse Arrival Time (PAT). A central aspect of this study is the physiological analysis of the mean PAT variability (PAT Var Mean), a metric reflecting beat-to-beat autonomic modulation. Using a PhysioNet dataset of 22 individuals, we analyzed how gender and activity intensity affect PAT stability. Results demonstrate that females exhibit higher mean variability at rest compared to males, but this variability decreases drastically in both genders during intense exercise due to sympathetic dominance. The LSTM model, trained to capture these dynamic transitions, achieved 92% accuracy. This approach demonstrates that the temporal signature of PAT, quantified through its mean variability, is a robust indicator of physical exertion level that transcends baseline demographic differences, enabling the development of more efficient and personalized HAR systems.