Network Latency Estimation for Telesurgery Using Deep Reinforcement Learning

ABSTRACT In this study, we present a method for the direction and value estimation of network latency in telesurgery systems using deep reinforcement learning (DRL). Network latency can impair surgical teleoperation because the delay in the command and feedback signals affect the system to not perform in real time. We utilized a deep Q‐network agent trained on a set of 60 features extracted from raw latency data. The framework was designed as a two‐stage process: first, by predicting the binary directional trend (up/down) of latency and second, by using the direction as an input to solve for the value estimation. The DRL agent achieved a direction prediction accuracy of 85.8% on a held‐out test set, which represents a 35.8% improvement over a random classifier. The subsequent value estimation achieved a low mean absolute percentage error (MAPE) of 7.23% and a 92.55% accuracy. Notably, the framework provides 85.8% directional accuracy, a capability absent in pure regression methods. While simpler regressors achieve marginally lower MAPE, they cannot predict whether latency will increase or decrease, limiting their utility for proactive control in telesurgery. This approach demonstrates that a DRL‐based direction and value estimation paradigm can provide a better and computationally efficient signal for latency mitigation, offering a practical pathway to improve the resilience of tele‐surgical systems operating over unpredictable networks.