Presentation (CTI)

OBJECTIVES: As regulatory agencies look to require computable operational definitions (CODefs) as part of submissions from life sciences companies, it is essential to have ways to create and maintain this content quickly and accurately.
METHODS: The process to develop code list/value sets that accurately represent clinical concepts for multiple EHR, claims and life science data sets can be very time consuming. We developed a novel three-step process to assist expert medical informaticists using Artificial Intelligence (AI): 1. Leveraging Claude.ai we processed and created a concept list from the National Comprehensive Cancer Network (NCCN) Gastric Cancer guidelines. 2. Leveraging Claude.ai we linked the concepts to common terminologies, like ICD-10-CM, LOINC, CPT, and RxNorm. 3. Assisted by Perplexity.ai a non-physician informaticist verified the accuracy of the concept and terminology mappings.
RESULTS: For CPT procedures, we had 101 distinct concepts mapped to 86 distinct code terms with 54 concepts missing a mapped term and 24 concepts with incorrect terms for gastric cancer. For ICD-10-CM diagnoses, we had 9 distinct concepts mapped to 11 distinct terms with no concepts missing a mapping and all 9 verified as correct. For RxNorm medications, we had 81 distinct concepts mapped to 33 distinct terms with 9 concepts missing a mapped term and all terms with correct codes/text, except 12 with hallucinated RxNorm codes but correct medication text. For LOINC labs, we had 3 distinct concepts mapped to 33 distinct terms with 0 concepts missing a mapped term and all terms with correct codes/text except 1 with a hallucinated LOINC code but correct lab text.
CONCLUSIONS: AI was able to accelerate an expert human informaticist’s ability to create and maintain value sets (for diagnoses, medications, and labs) for CODefs quickly and accurately and which are compliant with expectations from regulators / stakeholders.