Extended Reality (XR) applications for emergency medical response impose stringent latency requirements that commodity fifth-generation (5G) deployments do not satisfy by default. This paper presents EMERGE-XR, a real-time XR system delivering AI-driven medical guidance to first responders, and addresses the research problem of how Medium Access Control (MAC)-layer parameter tuning on a software-defined 5G gNodeB (gNB) can reliably achieve sub-10 ms round-trip time (RTT) for interactive XR workloads. The system integrates Meta Quest 3 augmented reality (AR) headsets, edge-hosted pose estimation, and a clinical decision-support workflow based on the ABCDE trauma assessment protocol. The clinical guidance component has been functionally demonstrated on the laboratory testbed but has not yet been evaluated in a user study; the experimentally validated contributions of this work concern the network-layer optimization and the profile-switching mechanism, not the clinical or AI-driven components. Through systematic MAC-layer optimization on an on-premises Amarisoft 5G gNB, tuning Scheduling Request (SR) periodicity, the Time Division Duplex (TDD) uplink/downlink slot pattern, and fixed MAC timing parameters, the average RTT was reduced from 24 ms to 9 ms, a 62.5% improvement, with zero packet loss. A lightweight RTT-driven profile-switching mechanism was validated across more than 100 experimental runs. As a work-in-progress contribution, an AI-driven intent-based extension is further proposed, but not yet experimentally validated, in which a contextual reinforcement learning (RL) agent autonomously selects MAC-layer profiles based on application-level Quality of Experience (QoE) signals and network Key Performance Indicators (KPIs). Planned validation over the 6G-BRICKS Cell-Free Massive MIMO testbed at KU Leuven is outlined.