Alcohol use disorder (AUD) affects millions of patients, but a significant portion never receive a formal diagnosis, and without one, they’re unlikely to get the treatment they need. Our new study, to be presented at ISPOR Europe 2026, examines why diagnosis codes alone aren’t enough, and how combining structured real-world data with LLM-extracted clinical notes can surface patients who would otherwise remain invisible to research, care pathway planning, and medical and commercial field strategies.
Why Diagnosis Codes Provide an Incomplete View of AUD
Unlike many chronic conditions, substance use disorders like AUD carry a stigma that shapes not only how patients talk to their physicians, but also how those interactions get documented and captured downstream. It’s also inherently subjective: Diagnosis depends heavily on what patients disclose and on what is required for billing and reimbursement reasons, and much of the most clinically meaningful information lives in physician notes, not billing records.
Claims data captures what gets reimbursed, but it doesn’t capture the healthcare provider-patient conversation. When a physician asks a patient how many drinks they’re having per week, what their triggers are, or whether there’s a family history of addiction, those details live in the notes. The result is a structural gap: healthcare providers, commercial or medical field teams and researchers working from claims data alone are seeing only a fraction of the true AUD population and are often catching patients late, after comorbidities have already progressed. By mining clinical notes using human-in-the-loop LLM extraction from NorstellaLinQ real-world electronic health record data, the linked real-world data foundation for biopharma, we identified patients who were presenting with clear AUD indicators but had never received a formal billable-coded diagnosis.
85,000+ Patients That Claims Data Missed
Our findings were striking: Diagnosis codes alone underestimated the identifiable AUD population, with clinical note analysis surfacing an additional 85,000 patients, including patients with AUD-related indicators and no recorded diagnosis. These patients had relevant information documented during healthcare encounters, despite the absence of an AUD diagnosis code in the study data.
What made Norstella’s approach particularly powerful was the granularity and breadth it unlocked. Researchers could characterize documented alcohol consumption patterns, assess the presence of screening-related information, and examine the timing of these observations relative to a recorded AUD diagnosis. Examining the timing of documented indicators relative to diagnosis may help researchers better understand gaps in recognition and potential opportunities for earlier clinical assessment.
Why This Matters for Care, Coverage, and Policy
The implications extend well beyond research. A more complete understanding of the AUD population can inform research into diagnosis and treatment patterns, including where patients may encounter gaps in screening, assessment, or care. Payers require it. Without accurate population identification, patients who need treatment simply don’t receive it.
For organizations working on care pathway design, disease burden estimation, or HTA submissions, accurate AUD population sizing is foundational. This study demonstrates that relying on claims data alone systematically undercounts that population, and that the gap is addressable.
For medical affairs and commercial teams, aggregate insights into consumption patterns, screening history, and documented risk factors can inform HCP education and engagement strategies. The goal isn’t just finding more patients; it’s understanding what’s different about the journeys of those who get diagnosed versus those who don’t, and using that to drive earlier, better-informed intervention.
The Bigger Picture
These findings have implications beyond AUD. For conditions where symptoms, risk factors, and clinical assessments are documented in notes but incompletely captured in diagnosis codes, structured data alone may provide a limited view of the patient population and its care journey. By combining linked NorstellaLinQ real-world data with Norstella’s human-in-the-loop LLM extraction, researchers can examine both coded diagnoses and the clinical context surrounding them. This broader view can inform disease burden estimates, reveal differences in diagnosis and treatment pathways, and guide medical and commercial strategies.
The value extends beyond identifying additional patients: it lies in understanding which populations are missed, what their documented histories reveal, and how those insights can inform research, care planning, and field strategy.
Find out how NorstellaLinQ can bring you closer to the patient than ever before. Contact us.
