WearPPG‐Former: A Wearable‐Optimized Transformer Using Dynamic Sparse Attention for Motion‐Resilient HR Estimation Under Intense Exercise Scenarios

Photoplethysmography (PPG) signals are widely used in heart rate monitoring via wearable devices, but motion artifact interference and the need for lightweight methods remain the major obstacles to their practical implementation. To address the issues that existing deep learning methods have insufficient anti‐artifact capabilities in strenuous exercise scenarios and excessively high model complexity, making them difficult to adapt to wearable devices, a lightweight Transformer method, called WearPPG‐Former, is proposed for heart rate (HR) estimation. The proposed method incorporates a sparse attention mechanism into transformer modules, which selectively concentrates on heart rate‐associated key feature regions. The design simultaneously reduces computational complexity while enhancing suppression capabilities against motion artifacts, including baseline wander and high‐frequency noise. Experimental results demonstrate that WearPPG‐Former outperforms the state‐of‐the‐art methods in mean absolute error (MAE) on the PPG‐DALIA dataset. Specifically, it achieves a reduced average MAE of 3.18 beats per minute (bpm) under intense exercise scenarios. When deployed on a resource‐constrained wearable embedded platform, the model attains an inference speed of 410.4 ms, fulfilling real‐time monitoring requirements for wearable devices. This delivers an efficient solution for PPG‐based heart rate estimation in dynamic scenarios, thereby advancing the practical implementation of wearable health monitoring technology.