A GPU Pipeline for Hyperparameter Optimization on IoT Platforms: Comparing Pulse-by-Pulse vs. Fixed Time Windows for Cuffless Blood Pressure Estimation

This work explores the potential benefits of using High Performance Computing Architectures for hyperparameter optimization problems in Internet of Things (IoT) platforms for health environments. Particularly, in this work the problem of selecting the best window size for blood pressure estimation is addressed. Traditional approaches typically rely on fixed time windows to segment physiological signals, yet recent research emphasizes real-time analysis of individual pulse waveforms. The effect of varying time slots on blood pressure estimation methodologies is analyzed, comparing the effectiveness of pulse-by-pulse analysis against fixed time windows while integrating optimization techniques, particularly Bayesian methods, to enhance accuracy with a processing pipeline utilizing XGBoost running on a GPU. The analysis shows notable differences: for Systolic Blood Pressure (SBP), higher window lengths correspond to increased Mean Absolute Error (MAE), whereas for Diastolic Blood Pressure (DBP), higher window lengths align with decreased MAE. This finding prompts further investigation into the underlying factors influencing these trends.