Service Ready Caching and Offloading for Urgency Aware MEC IoMT

ECG-driven IoMT requires low-latency and service-ready edge execution for immediate clinical action. In the case of ECG-driven MEC IoMT, processing a task can occur at any edge node if the required ECG service/model/container is available. But nearest neighbor, least loaded neighbor, and popular caching schemes might not be able to process tasks in emergency conditions due to their inability to handle three factors together—service availability, resource constraints, and urgency. This research paper aims to address this issue by presenting Tiny MedDMARL, which is a lightweight decentralized learning framework for urgent service caching and offloading. The proposed framework makes use of centralized learning and decentralized actor-only execution by running only compressed actors on all MEC nodes. Two-timescale separation ensures that the task offloading occurs faster while service caching takes place slowly. Results of simulation through MIMIC IV ECG-driven tasks reveal 127.17 ms average latency, 79.60% cache hit rate, 17 ms cold-start delay, 6.1% SLA violation, 8.7% cloud offloading, and 96.8% critical task success.