OBJECTIVES: Efficiently identifying and characterizing with precision patients suffering from a heterogeneous condition with diverse manifestations is complex with traditional data sources. This study assesses the ability of an interconnected hospital-specific natural language processing (NLP)-powered solution (Realli) to identify and characterize patients with lupus in France.

METHODS: Upon Ethics and Scientific Committee approval, we searched native hospital patient electronic medical records (“lupus” or “lupique” text strings) and hospital claims (ICD-10-CM L93.X or M32.X).

RESULTS: Between January 2018 and March 2023, we identified 191 adult patients with lupus (mean age 48.1 years; 88% females), with 95.8%, 20.4% and 4.2% of patients who presented with lupus erythematous, nephritic, and cutaneous, respectively and not mutually exclusive (ICD codes, text strings). Biomarker values on lupus-specific or not, such as anti-DNA antibodies, anti-RNP antibodies, Smith antibodies, and C-reactive protein were retrieved in 63.9%, 4.2%, 12.6%, and 54.5% of patients, while disease activity index (i.e., SLEDAI or BILAG) was reported in 9.4% patients. Most prevalent comorbidities were cardiovascular (56.5%), metabolic (22.5%) and psychiatric (16.8%) disorders and infections (8.9%). In addition, renal impairment was identified in 29.8% (all renal ICD-10 codes) and 24.1% (ICD-10 N17.X to N19.X only or creatinine clearance levels <60 mL/min/1.73 m²). Among those latter, 52% required dialysis. Lastly, 26.7% and 15.2% of patients presented with antiphospholipid antibody syndrome or Gougerot-Sjögren syndrome, respectively. Patients received (not concomitantly) cyclophosphamide (4.7%), immunosuppressants (6.8%), glucocorticoids (47.7%), hydroxychloroquine (73.8%), belimumab (14.7%), and/or rituximab (8.4%). Over the study period, patients were hospitalized on average 7 times (median 3).

CONCLUSIONS: A NLP-powered solution connected to multidata hospital sources is effective in characterizing complex pathologies, with a breadth and granularity of information not available in traditional RWE data sources. Learning from this single-center study will be deployed to multihospital systems in France to test the robustness of the findings.