Copenhagen-based healthcare AI company Cortideveloped Symphony AI based on peer-reviewed research from the largest medical coding study of its kind, treating coding as a logic task rather than a labeling problem.
See also: Claude Code: Code leak via Anthropic's npm packaging bug
Medical coding, the process of converting clinical notes, diagnoses, and procedures into standardized alphanumeric codes used for billing, reporting, and public health data, is one of the most error-prone and important administrative tasks in healthcare. The U.S. coding system, ICD-10-CM, contains 70,000 diagnosis codes. Errors are common, costly, and often invisible.
Corti has created a product specifically designed to fix this: Symphony for Medical Coding, a system that claims to outperform models from OpenAI, Anthropic, Amazon, Oracle , and Microsoft by up to 25% on clinical accuracy standards.
The performance difference Corti claims is based on a methodological distinction. Most AI systems approach medical coding as a classification problem: given a clinical note, they predict the most likely code from the training distribution. The problem is that coding guidelines are constantly changing, making historically trained models structurally inadequate. Corti’s approach, developed through a peer-reviewed framework called Code Like Humans, and accepted at EMNLP 2025, treats coding as a logic task.
See also: Vertex AI vulnerability exposes Google Cloud data and files
The system uses four agents in sequence: an evidence extractor that isolates conditions in a clinical note, an index navigator that searches the alphabetical ICD index for candidate codes, a table validator that checks candidates against guidelines, and a code reconciler that follows through and validates the final result. Each step mirrors what a trained human coder would do.
The research was based on 1.8 million patient contacts, making it the largest study of its kind to undergo peer review. The consequences of conventional undercoding are not just financial. Corti cites a peer review study of Danish patient data in which her system identified three times more suicide attempts than were officially coded, cases that existed in clinical notes and medication records but were missed by coders working under time pressure.
When these cases are not counted, health systems cannot track trends, allocate resources, or design effective interventions. The level of coding is not an administrative cost. It is how health systems see themselves.
Symphony for Medical Coding is the first system Corti has built to work in both US coding environments, ICD-10-CM for diagnoses, ICD-10-PCS and CPT for procedures, and European coding environments without local retraining. ICD-10 coverage for Europe, maintained by the WHO, is currently in beta as the company expands to the UK, Germany, France and Denmark.
See also: OpenAI fixes data extraction vulnerability in ChatGPT
The system produces verifiable results: each assigned code is linked to the clinical evidence that supports it.
