Presentation (CTI)

OBJECTIVES: Lupus affects around 40,000 individuals in Germany. However, identifying patients in prescription datasets without diagnostic codes is challenging, as Lupus treatments—such as anti-inflammatories, antimalarials, steroids, and immunosuppressants—are also used for other conditions.
METHODS: To address this, a dual AI/ML approach was developed combining two models: Split by Indication (SBI) and Eligibility (ELY). The SBI model was trained on longitudinal EMR data (German Disease Analyzer) from 1,000 patients diagnosed with Systemic Lupus Erythematosus (ICD M32). The ELY model was trained on prescription data from 500,000 patients receiving typical Lupus medications, including 2,800 patients treated with Lupus biologics, who served as the target cohort.
RESULTS: The SBI model identified Lupus patients 16 times more effectively than a baseline prevalence model, detecting 14,000 likely cases. Predictive features aligned with clinical expectations, including female gender, hydroxychloroquine use, and absence of methotrexate or anti-TNF therapies. Among patients currently on Lupus biologics, the SBI model predicted Lupus in 80% of cases.
The ELY model achieved 81% recall in identifying target patients and flagged 33,000 additional “opportunity patients”—those not yet on biologics but showing similar treatment patterns. Furthermore, the model predicted 61% of future biologics initiators (within three months from their last prescriptions) as opportunity patients. Combined, the models enabled the prioritization of 6,600 patients most likely to benefit from biologic therapy for Lupus.
CONCLUSIONS: This AI/ML framework enables accurate, confidence-tiered segmentation of patients in low-prevalence diseases without indication-specific medications. It offers valuable applications in both epidemiological research and targeted healthcare strategies.