Clinical risk-aware reinforcement learning for latency-constrained healthcare IoT scheduling
TL;DR
CRAI-LCS, a clinical risk-aware Reinforcement learning (RL) framework for latency-constrained scheduling in HIoT systems, combines data-driven clinical risk estimation, deadline-violation prediction, and RL-based scheduling to dynamically prioritize high-risk tasks while efficiently managing system resources.
Abstract
The rapid growth of the Healthcare Internet of Things (HIoT) has enabled continuous, real-time patient monitoring through wearable and bedside devices. These systems generate time-sensitive physiological data that are essential for the early detection of critical conditions such as arrhythmias and hypoxia. However, conventional cloud-centric architectures introduce significant end-to-end latency, which can compromise timely clinical response in safety-critical scenarios. Mobile Edge Computing (MEC) mitigates this limitation by bringing computation closer to data sources; yet, existing scheduling approaches remain largely system-centric and do not adequately incorporate patient-specific clinical risk into their decision-making. To address this gap, this paper presents CRAI-LCS, a clinical risk-aware Reinforcement learning (RL) framework for latency-constrained scheduling in HIoT systems. Unlike prior simulation-driven studies, CRAI-LCS integrates real physiological data from the PhysioNet MIT-BIH Arrhythmia database to construct realistic, data-driven workloads. Specifically, electrocardiogram (ECG) signals are segmented into time-windowed tasks with clinically grounded characteristics, including input size, computational demand, and urgency-aware deadlines. The framework combines data-driven clinical risk estimation, deadline-violation prediction, and RL-based scheduling to dynamically prioritize high-risk tasks while efficiently managing system resources. Experimental results demonstrate that CRAI-LCS consistently outperforms baseline approaches in terms of latency, deadline compliance, and resource utilization under realistic workload conditions. Ablation studies further confirm the individual contributions of the clinical risk-awareness and predictive scheduling components. Overall, these findings highlight the importance of incorporating real physiological data into scheduling design, providing a more reliable and clinically relevant foundation for next-generation healthcare edge intelligence systems.
