EPID-18. Accuracy of claims-based algorithms in the identification of leptomeningeal disease: a validation study

Abstract Though administrative claims have been shown to reliably identify patients with brain metastases (BM), the validity of claims-based methods for identifying leptomeningeal disease (LMD) remains unestablished. We evaluated the accuracy of claims data in identifying LMD cases among patients diagnosed with LMD at Stanford Hospital (n=145) from 2020 to 2024, using a clinical database derived from neuro-oncology and neurosurgical oncology clinic records that included data on primary cancer type, symptoms, and procedures. A subset of these patients (n=20) was employed as a training set for this study. Using this cohort of confirmed LMD, we performed entity resolution with a large AI-linked claims database (Healthcare Map®) and analyzed the dates of LMD diagnosis, metastasis-related procedures, symptoms, and screening coded in claims. Sensitivity of claims-based LMD identification was assessed against chart-confirmed LMD. Claims-based LMD diagnosis coding alone demonstrated 45% sensitivity within 30 days of chart-confirmed diagnosis. However, incorporating procedure codes (WBRT/SRS/resection) increased sensitivity to 83% (30% of patients). Incorporating coding for specific symptoms also improved accuracy, but to a lesser extent: adding hydrocephalus coding achieved 80% sensitivity, while less-specific symptoms (headache) achieved 33% sensitivity. Combining two code types (procedure + symptom) achieved 100% sensitivity (20% of patients). Notably, 89% of patients with timely LMD coding had concurrent screening or symptom coding, and 78% had metastasis-related procedures within 30 days of diagnosis. These findings demonstrate that claims data can reliably identify LMD when code types (diagnosis + procedures) are combined, mirroring validated algorithms for claims-based identification of BM. However, symptom-based coding lacks precision, highlighting its limitations in capturing LMD without definitive diagnostic coding. This study establishes a framework for leveraging claims in LMD research, enabling large-scale epidemiological studies previously constrained to small cohorts. Future multi-institutional validation will strengthen generalizability, replicating advancements in BM research to address gaps in the understanding of LMD epidemiology.