Difficult-to-treat rheumatoid arthritis patient identification: AI-powered mining of unstructured EHR clinical data

Difficult-to-treat rheumatoid arthritis is defined by inadequate response to multiple biologic agents — but this definition lives in the clinical narrative, not in the structured fields of most hospital EHR systems. A physician who documents in free-text that a patient has 'failed three biologics and remains inadequately controlled' has recorded the key diagnostic information — but that patient may be coded in the EHR system as simply 'rheumatoid arthritis' with a list of past medications.

Standard database queries cannot identify these patients. The information is there — but it is locked in unstructured text. For a brand with a product specifically indicated for difficult-to-treat RA, the inability to quantify the true patient population was limiting both the commercial case and the access arguments.

The solution required applying language processing technology to clinical text — not as a research exercise, but as a practical tool for patient identification that could be deployed in partnership with hospital rheumatology departments.

In rheumatology, the patients with the most complex disease are often the ones who are hardest to find in a dataset. Their complexity is documented in words, not in codes — and standard analytics cannot read words.