Ertas for Medical Coding

The Challenge

Medical coding (the process of translating clinical documentation into standardized diagnosis and procedure codes) is the backbone of healthcare revenue cycle management. Accurate coding determines reimbursement rates, compliance with payer requirements, and regulatory reporting. The ICD-10-CM system alone contains over 72,000 diagnosis codes, and selecting the correct code requires understanding clinical context, specificity requirements, coding guidelines, and payer-specific rules. A single digit difference between codes can mean the difference between a clean claim and a denial.

The medical coding workforce faces a chronic shortage, with the American Health Information Management Association reporting tens of thousands of unfilled coding positions. This shortage creates backlogs that delay revenue, increases error rates as overworked coders rush through charts, and drives up costs as organizations rely on expensive contract coders. Generic AI tools cannot address this gap because medical coding requires simultaneously understanding clinical language, applying complex coding guidelines, and navigating the specificity hierarchy of classification systems: a combination of domain knowledge that general models do not possess.

The Solution

Ertas enables healthcare organizations to fine-tune AI models on their own coded clinical documentation, creating a coding assistant that suggests codes based on the patterns, specialties and payer requirements the organization actually sees. In Ertas Studio, revenue cycle teams train models on de-identified pairs of clinical notes and their verified code assignments, teaching the model the link between clinical language and code selection at the level of specificity their coders achieve. Training runs on Ertas's cloud GPUs, so the training set is uploaded to Ertas; de-identify it before upload.

The fine-tuned model works as a coding assistant: it reads clinical documentation and suggests primary and secondary codes with a short rationale citing the findings behind each one. Exported from Hub and run on the organization's own servers with Ollama or vLLM, it processes live charts inside the network as coders work, and coders review and approve every suggestion. This human-in-the-loop workflow keeps credentialed coders in charge of every code while taking the first pass off their plate.

Key Features

Example Workflow

A 500-bed hospital system processes 2,000 inpatient and 8,000 outpatient encounters weekly. The coding team de-identifies 100,000 historically coded encounters (clinical notes paired with verified ICD-10 and CPT code assignments) and uploads them to Ertas Studio as a JSONL dataset. Senior coders review a sample in Data Craft, fix the rationale on edge cases and approve the examples. The team fine-tunes models for its top 15 service lines, which account for 80% of coding volume, training Llama 3.1 8B and Qwen 2.5 7B in parallel and comparing both against a held-out set of encounters coded by the team. They export the stronger model from Hub, serve it with vLLM on the hospital's own servers, and connect it to their coding workflow through a browser extension. As coders open each chart, the assistant suggests primary and secondary codes with the supporting clinical evidence highlighted in the documentation. Coders accept or correct each suggestion, and the team tracks acceptance rates and denial rates against its own baseline to decide where the model is ready and where it needs more training data.