Abstract

Wearable healthcare systems have emerged as an essential component of modern medical monitoring by enabling continuous acquisition of physiological signals and real-time health assessment. Field Programmable Gate Arrays (FPGAs) have become attractive computing platforms for wearable devices because they offer high computational capability, reconfigurability, and lower energy consumption than conventional processors. However, achieving an optimal balance among hardware acceleration, software execution, latency, and power consumption remains a challenging task under dynamic healthcare workloads. Existing hardware-software partitioning approaches primarily employ static optimization strategies that cannot effectively adapt to changing workloads, patient-specific data streams, and resource availability. Consequently, these methods suffer from increased energy consumption, reduced resource utilization, and degraded real-time performance. This research proposes a novel Deep Reinforcement Learning based Adaptive Hardware Software Co Optimization Framework (DRL-AHSCOF) for FPGA-enabled wearable healthcare systems. The proposed framework integrates a Deep Q-Network (DQN) based adaptive decision engine with dynamic hardware-software partitioning, runtime FPGA resource allocation, workload-aware scheduling, and energy-aware optimization. The reinforcement learning agent continuously observes workload characteristics, hardware utilization, memory occupancy, communication overhead, and latency to determine optimal execution policies that maximize overall system efficiency while minimizing energy consumption. Experimental evaluation demonstrates that the proposed DRL-AHSCOF framework significantly improves the performance of FPGA-based wearable healthcare systems. The proposed method reduces energy consumption to 92.74 J, achieving a 26.67% reduction compared with GA-HSO. It decreases execution latency to 30.12 ms, improving response time by 40.66% compared with existing optimization approaches. Furthermore, DRL-AHSCOF achieves 95.34% FPGA resource utilization, 1846 tasks/s throughput, and 99.18% scheduling accuracy, outperforming GA-HSO, PSO-RS, and DL-AS methods. These results confirm that the proposed deep reinforcement learning-based optimization strategy effectively enhances energy efficiency, adaptive resource management, and real- time processing capability for FPGA-enabled wearable healthcare systems.

Authors

Mariam Safar Mohammed Alshahrani¹, M.K. Jayanthi Kannan², Shree Nee Thirumalai Ramesh³
Digital Government Authority of KSA, Riyadh Province, Kingdom of Saudi Arabia¹, VIT Bhopal, India², Manipal University Medical College Malaysia, Malaysia³

Keywords

Deep Reinforcement Learning, FPGA Wearable Healthcare, Hardware Software Co-Optimization, Energy Efficiency, Dynamic Resource Allocation