MedARC Team: Collaborative Medical AI

• MedARC Team is a research group that develops collaborative, agentic, and augmented intelligence systems to enhance team-based medical decision-making, clinical training, and operational support.
• The group employs adaptive multi-agent frameworks, participatory AR design, and edge computing to achieve significant improvements in medical accuracy and workflow efficiency.
• Their innovative integration of human teamwork principles with advanced computational tools has demonstrated measurable gains in medical QA benchmarks and clinical outcomes.

The MedARC Team is a research group focused on advancing collaborative, agentic, and augmented intelligence methods for complex team-based medical decision-making, clinical training, and operational support in healthcare environments. Their body of work spans medical multi-agent systems, collaborative AI, edge computing for real-time medical support, and the integration of advanced augmented reality (AR) technologies in both clinical and research settings.

1. Agentic Multi-Agent Frameworks in Medicine

MedARC’s most prominent contributions are in the domain of medical multi-agent teams—computational frameworks that operationalize human organizational psychology models to enhance medical decision-making by LLMs. Notably, "TeamMedAgents: Enhancing Medical Decision-Making of LLMs Through Structured Teamwork" (Mishra et al., 11 Aug 2025) systematically translates Salas et al.’s “Big Five” teamwork theory into agentic modules for LLMs, including team leadership, mutual monitoring, team orientation, shared mental models, closed-loop communication, and dynamic trust adjustments. Each of these is implemented as an independent class, activatable per task requirement.

Key architectural features:

• Adaptive agent recruitment: A “recruiter” module matches the reasoning task to an optimal set of expert agents (typically 2–5) with complementary medical specialties.
• Teamwork modules: Individual modules can be enabled/disabled based on question complexity, domain, and cohort size.
• Leader weighting: The team leader’s vote is linearly weighted (factor 1.5×) in final decision synthesis for improved hierarchical integration.
• Comprehensive ablations: Systematic studies detail how each teamwork component contributes to performance across medical QA, clinical diagnosis, visual reasoning, and evidence synthesis tasks.

Empirical evaluation demonstrates consistent improvements across seven of eight medical benchmarks, with MedQA accuracy rising from a baseline of 75.0% (single-agent) to a TeamMedAgents configuration achieving 92.6%. Detailed results and optimal teamwork patterns are summarized across datasets (e.g., clinical diagnosis, differential diagnosis, and visual QA require distinct teamwork module combinations).

2. Participatory Design and Co-Design in Clinical Teams

MedARC employs participatory and co-design methodologies to align digital tools with real clinical needs and workflows. For example, in "Co-Designing Augmented Reality Tools for High-Stakes Clinical Teamwork" (Taylor et al., 24 Feb 2025), the team collaborates with emergency workers to design AR head-mounted display (AR-HMD) interfaces. Seven role-specific AR-HMD application scenarios for an Emergency Department (ED) are proposed, including team leader dashboards (aggregating patient and procedural status), role-specific task support (such as dosage calculation for pharmacists and CPR guidance for nurses), procedural timing, and real-time remote teleconsultation for rare emergencies.

Key recommendations stress content minimalism, multi-modal interaction (voice, gesture, gaze), and display adaptation to specific clinical roles to minimize cognitive load and workflow interruption.

3. Real-Time Cognitive Assistance and Edge Intelligence

The MedARC Team advances wearable edge computing for real-time decision support in pre-hospital and field medicine. The “Real-Time Multimodal Cognitive Assistant for Emergency Medical Services” (Weerasinghe et al., 11 Mar 2024) details CognitiveEMS—a wearable cognitive partner that fuses:

• Edge-deployed, fine-tuned Whisper-based speech recognition robust to EMS scene noise (achieving WER of 0.290, compared to SOTA of 0.618)
• Protocol prediction via TinyClinicalBERT fused with graph-based domain knowledge (EMS protocol graphs), reporting Acc@3 near 90.4%
• Context-aware, CLIP-based action/intervention recognition, accurate to 0.727

The system maintains a ~4s end-to-end feedback latency on edge devices (Jetson Nano class). Design emphasizes parallelism, quantization (8-bit CLIP), tiny models, and robust multimodal data streaming to meet performance and reliability requirements in the field.

4. AR, Hybrid, and Multi-User Collaboration Platforms

MedARC contributes a suite of mobile, AR-enabled collaboration tools for clinical planning and training:

• "Multi-User Mobile Augmented Reality for Cardiovascular Surgical Planning" (Mehta et al., 6 Aug 2024): ARCollab is an open-source iOS app supporting multi-user, gesture-based manipulation (scaling, rotation, omni-directional slicing) of 3D heart models. It employs ARKit/RealityKit, Bonjour, and Multipeer Connectivity for group synchronization and shared state updates.
• "HybridCollab: Unifying In-Person and Remote Collaboration for Cardiovascular Surgical Planning in Mobile Augmented Reality" (Mehta et al., 14 Apr 2025): Extends ARCollab’s paradigm to hybrid sessions with both local and remote medical teams via Apple’s GameKit (low-latency, up to 16 simultaneous users), supporting consistent annotation, pointing, and world-anchored multi-user interaction. Demonstrated benefits include improved spatial referencing and collaborative communication in surgical planning.

Technical advances include custom Metal shader slicing, minimal transformation message design for low network overhead, and seamless integration of FaceTime/SharePlay for synchronized remote AR planning.

5. Multi-Agent Clinical Care Algorithms

In "The Application of MATEC (Multi-AI Agent Team Care) Framework in Sepsis Care" (Cho et al., 9 Feb 2025), the team translates the multi-agent team model to sepsis diagnosis and management, with:

• Core doctor agents (specialists) contributing independent diagnostic assessments
• Senior physician agent synthesizing and cross-validating outputs, flagging hallucinations and ensuring guideline compliance
• Health professional agents (nursing, pharmacy, safety/SDOH, QI) providing complementary care plans and checking for compliance with real-world healthcare metrics (e.g., SEP-1)
• Automated risk prediction model agent (utilizing NEWS: $\text{NEWS} = \sum s_i$) for real-time deterioration alerts

Pilot evaluations with attending physicians (N=10) score MATEC high for usefulness and accuracy (median=4/5, P=0.01 and <0.01) with favorable reception among LLM-familiar clinicians. The layered architecture is shown to minimize errors by requiring internal consensus and verification across agent roles.

6. Specialized Clinical AI and Rural and Public Health Tools

MedARC’s agentic frameworks are extended to rural and resource-constrained care via the IMAS system (Gangavarapu et al., 13 Oct 2024), which integrates:

• Multilingual translation (Seamless‑M4T), complexity triage, distributed agentic diagnostic reasoning based on local context/vernacular, final aggregation, and response simplification with safety guardrails
• Fine-tuned Llama 3 (70B) and evaluation on MedQA/PubMedQA/DDXPlus (scores: 78.9/74.1/76.8), with variable performance by language (higher in Telugu/Hindi than Arabic/Swahili)
• All code and resources published for reproducibility and local adaptation

The group also addresses real-world data curation and automation in rehabilitation assessment (Ahmed et al., 2 Jan 2025), leveraging participatory design with clinicians for multi-camera capture, segmentation, human-in-the-loop interface design, and ground-truth creation for supervised/automated movement analysis.

7. Impact and Significance

MedARC's research demonstrates that systematically encoding human teamwork principles and tailoring computational, visual, and interactive systems to fit actual clinical team workflows yields quantifiable improvements in accuracy, decision robustness, and usability across benchmarks and real-world clinical settings. The team’s modular agentic frameworks, AR/edge technologies, and participatory co-design approach lay a foundation for next-generation collaborative medical AI and operational support, with empirical evidence for both quantitative benchmark gains (up to +17.6% in medical QA) and practitioner-validated workflow integration.

Their strategy—with emphasis on modularity, adaptive configuration, and user-centered evaluation—advances the field’s methodological toolkit for deploying collaborative and team-based AI systems in medicine and beyond.