OBJECTIVES: Lupus nephritis (LN) is a common kidney-related manifestation of systemic lupus erythematosus. Renal function status (e.g. flares, remission) is not available in claims datasets or standardly structured within medical records challenging the usefulness of common real-world data (RWD) sources to capture disease activity. Recognizing this limitation, we established a real-world dataset by abstracting disease status indicators from unstructured narrative text (i.e. physician notes) and lab results.

METHODS: Data was abstracted from records using natural language processing and human-reviewed machine learning from 174 LN patients recruited between 1 July 2020 and 12 March 2022. Flare and remission events were abstracted from unstructured text mentions and flares were secondarily defined as a 25% increase in serum creatinine from the previous measurement; all events were compared to urine protein creatinine ratio (UPCR), a common outcome used to identify disease activity.

RESULTS: The cohort mean age was 44 (15-78) years, 92% female, 31% Black/African-American, 33% Class III/IV, and mean follow-up of 4.3 years after LN diagnosis.12 (6.9%) had flare events abstracted from narrative text. Flare detection increased to 53 (30.5%) patients when including the lab-based serum creatinine definition. Median (IQR) UPCR was higher within 1 month of a flare mention alone: 800 mg/g (569-2,500) or flare mention plus lab-based definition 569.00 mg/g (301-1,972) and lower within 1 month of ‘complete remission’ mention 166.00 mg/g (107-330). UPCR was highest within 1 month of mentions of ‘partial remission’: 1094 mg/g (809-1,907), though the patient number was low (n=4).

CONCLUSIONS: Narrative text abstraction with computed lab-based indicators may improve RWD LN flare detection and more clearly understand patient journey. Future research should examine how additional unstructured data may improve LN renal flare identification compared to other RWD.