At IMO Health, perioperative terminology migration – the process of mapping complex medical terms to standardized, structured formats (IMO Health concepts) and then mapping to industry codes – is critical for accurate, efficient clinical documentation. Historically, this process has been manual and time-consuming, involving multiple teams of clinical experts, mappers, and tools.
To save time and gain greater efficiency, we have been exploring the targeted and strategic use of artificial intelligence (AI). Our goal is not to replace the rigor built into those labor-intensive workflows, but to augment it with a policy-aware AI layer that accelerates decision-making and reduces manual effort.
To achieve this aim, we’ve designed an agentic AI architecture, built around a supervisor agent that coordinates the entire process of term validation, matching, and creation.
The role of the Supervisor Agent
At the center of these efforts is the Supervisor Agent, which orchestrates the migration workflow. This agent:
- Receives migration requests
- Invokes tools or sub-agents
- Coordinates intermediate steps
- Returns a validated, structured response
The supervisor operates in a reasoning loop according to the following steps:
- Understand the request
- Invoke the appropriate tool or agent
- Evaluate results
- Decide next action
- Repeat until the objective is met
This process enables dynamic, policy-aware orchestration rather than static rule execution.
Tools and resources
To empower the agents within IMO Health’s agentic architecture, we have exposed our internal capabilities through the Model Context Protocol (MCP), a secure interface that allows AI agents to safely interact with internal systems.
Tools (What the system does)
The supervisor and sub-agents use these executable functions to perform specific tasks throughout the migration workflow:
- Lexical matching APIs are used for initial candidate selection based on linguistic patterns
- Clinical migration logic applies standardized rules to transform complex medical terms into structured formats
- IMO Health database lookup enables the real-time retrieval of terminology metadata and existing records
Resources (What the system uses to reason)
- IMO Health editorial policies: Documentation that defines the standards for clinical documentation and terminology mapping. At IMO Health, these policies have been continuously built and refined by clinicians and terminologists for more than 30 years
- Client migration guidelines: Bespoke requirements that ensure the migration meets specific client needs
- Historical mapping guidance: A repository of decision patterns that ensures consistency across projects
The functions of specialized sub-agents
To maintain control and governance, responsibilities of sub-agents within the AI architecture are clearly separated.
Term Inspection Agent
Interprets clinical intent, expands abbreviations, and identifies structural qualifiers before matching begins.
Term Suggestion Agent
Identifies alternate term variations and surfaces candidate concepts for consideration, supported by tool outputs where applicable.
Match to IMO Health Concept Agent
This is where governance is enforced. The agent searches the IMO Health resources, applies IMO Health editorial policies and client migration guidelines, evaluates candidates, and proposes compliant standardized IMO Health concepts.
New IMO Health Concept Creation Agent
When no compliant match exists, this agent drafts a structured new IMO Health concept aligned with editorial standards and prepares it for downstream review.
Revalidation Agent
Performs a final compliance and meaning check before returning the response.
Memory and guardrails
The system uses AWS AgentCore Memory, shared across agents to maintain session context and coordinated reasoning. This ensures structured, governed output – not isolated AI decisions.
Guardrails validate:
- Request structure
- Response format
- Policy compliance
How it works
Early results and impact
The system is already achieving 73% accuracy and has reduced the initial migration effort for terms from several business days to just a few hours. For example, in a single day, we processed over 5,000 terms across 10 clients within three hours – work that would typically require five business days of manual effort.
This represents a significant step forward in modernizing how we manage clinical terminology and data, improving operational efficiency, reducing manual workload, and accelerating turnaround time for clients.
