HIMSS26: Multi-AI Agents That Take Action in Clinical Decision Support
At HIMSS26, Dr. Nathan Moore, medical director of the BJC Healthcare ACO, will show how health systems can move beyond chatbots to AI that actually does the work-pulling data, triaging patients, and nudging clinicians at the right moment.
His session centers on multi-AI agents: large language model-enhanced, self-reflecting, and tool-using systems working together. The focus is practical and specific-automating high-risk patient selection and improving care coordination for end-of-life and advance care planning.
From simple chats to agents that move the workflow
Many teams are stuck at the chatbot stage. Dr. Moore will outline how to design agents that take safe, auditable actions inside clinical workflows without adding noise or extra clicks.
The shift: fewer manual reviews, clearer handoffs, and timely prompts that match clinician intent and patient needs.
Inside BJC's learning reviewer agent
BJC Healthcare previously used a deep learning mortality risk model. Scores then went to a manual review queue where retired clinicians read charts and used clinical intuition to decide who was appropriate for advance care planning outreach-accurate, but slow and costly, and it didn't learn over time.
The learning reviewer agent builds on years of labeled data-more than 35,000 patients with risk scores, clinical notes, and prior reviewer decisions. It uses supervised learning plus live reinforcement from clinician feedback to mimic, then gradually enhance, those reviewer decisions.
Result: faster identification of high-risk patients, reduced manual workload, and tighter care coordination-while keeping clinicians in the loop.
A framework you can adapt
Expect a concrete model for "memory-augmenting" agentic CDS so systems learn from deployment, not just from static training sets. The goal is to fit local practice patterns without breaking safety boundaries.
- Define agent roles: data puller, reviewer, triage coordinator, and notification/nudge agent.
- Add institutional memory: store prior decisions, rationales, and outcomes to inform future recommendations.
- Close the loop: collect clinician feedback in near real time and use it to reinforce or correct agent behavior.
- Stage the rollout: historical testing, "silent mode" shadow runs, then monitored production with clear metrics.
Why this matters now
This approach targets real pressure points: value-based care goals, workforce burnout, and aligning care with patient preferences. End-of-life and advance care planning is a strong proving ground because timing and appropriateness are everything.
What to watch for in your own deployment
- Precision of patient selection and the downstream impact on ACP conversations and outcomes.
- Clinician workload signals: fewer manual chart reviews, fewer unnecessary alerts, smoother handoffs.
- Data access and governance: auditable data pulls, clear provenance, and safe tool permissions.
- Evaluation cadence: frequent post-deployment monitoring and transparent change logs.
Session details
"Multi-Artificial Intelligence Agents for Enhancing Clinical Decision Support" with Dr. Nathan Moore is scheduled for Tuesday, March 10, 10:15-10:45 a.m., Palazzo D Level 5 at HIMSS26 in Las Vegas.
Event info: HIMSS Global Health Conference & Exhibition.
Level up your team's AI fluency
If you're building internal capability for agentic workflows and AI in clinical settings, explore role-based learning paths here: Complete AI Training - Courses by Job.
Your membership also unlocks:
