WatchForce: Wrist-worn PPG and IMU sensing for hand force estimation

Recent advancements in extended reality have reshaped human-computer interaction; yet, force estimation remains a critical missing element for achieving realistic interactions. Unlike traditional electromyography-based systems, which are highly accurate but hindered by usability challenges, this paper introduces WatchForce, a novel wrist-worn system that estimates hand force using photoplethysmography and an inertial measurement unit. The objective is to investigate whether smartwatch-compatible photoplethysmography and inertial measurement unit signals can support three-level force classification and MVC-normalized force regression under controlled experimental conditions.

WatchForce employs a Vision Transformer-based model with a multimodal alignment strategy based on bidirectional InfoNCE. This approach separately encodes photoplethysmography and inertial measurement unit signals and aligns their representations for multimodal force estimation.

WatchForce achieves 86.31% force classification accuracy under the session-level universal-model setting and 88.32% in user-dependent models. Under the same session-level universal-model setting, the regression task yields an nRMSE of 0.1386 on MVC-normalized force values and an R² of 78.24%. In a leave-one-subject-out (LOSO) cross-subject evaluation, the PPG+IMU model achieves 78.50% classification accuracy and 0.1509 nRMSE/68.04% R² for regression, providing initial evidence of generalization to held-out participants within the collected cohort. Additional posture, motion, and lighting evaluations further support robustness under the tested conditions, while broader population-level validation remains an important direction for future work.

WatchForce provides proof-of-concept evidence that smartwatch-compatible PPG and IMU signals can support wrist-based hand-force estimation. Future studies with larger and more diverse cohorts, long-term daily-use evaluation, and commercial smartwatch hardware-level validation will be important for extending these findings toward real-world use.