IMO Health's LLM
Combining the stability and efficiency of supervised learning with the accuracy of a domain-specific LLM
Traditional AI approaches to information extraction depend heavily on the quality and quantity of the underlying labeled data. As a result, they are less flexible and require substantial effort to annotate data and retrain models for new tasks or domains. String matching algorithms, another approach for concept matching, pose different challenges. They lack the ability to understand context; are typically prone to errors when handling synonyms, misspellings, or complex phrases; and may require additional preprocessing and rule-based enhancements to handle variability effectively.
IMO Health’s LLM powers our solutions with highly accurate, trustworthy, and explainable answers – while balancing cost. By focusing the LLM on the most difficult questions, while leveraging traditional approaches for simpler queries, our solutions deliver the most cost-effective approaches based on specific client needs. Key capabilities include:
- A federated approach that combines multiple techniques to maximize performance and accuracy at the lowest cost
- LLM service with a framework used in conjunction with supervised machine learning (string similarity and embedded logic)
- Deep training resources augmented with proprietary IMO Health clinical content
LLM service
- Traditional NLP approach enhanced with extensive IMO Health clinical content and ontologies
- LLM with enhanced performance through sophisticated prompt engineering, knowledge graphs, and vector databases rooted in IMO Health’s rich clinical content
- Orchestrator (LLM Foundation Service): Fields requests from various applications and routes to the right LLM agent for processing
- LLM agents: Specific to normalization, agent specific to value set creation
- Vector database representation for IMO lexicals and standardized medical terminologies
- Knowledge graphs for individual medical concepts including anatomy, morphology, gene mutations, drugs, cancer stages and the relationships between them
- RAG (Retrieval-augmented generation): Combines foundational models with real-time, proprietary data stores, like IMO Health concepts and editorial policies
