Validation of a treatment-based algorithm to infer biomarker status in real-world data.

Abstract

e23019 Background: Real-world (RW) precision oncology studies require accurate identification of actionable biomarkers, yet structured electronic health record (EHR) biomarker fields are often inconsistently captured. Treatment exposure can serve as a proxy for biomarker positivity when restricted to biomarker-defined therapies, but the validity of this approach in RW data is not well characterized. NTRK gene fusions and BRAF V600E mutations are clinically actionable biomarkers with tissue-agnostic targeted therapy approvals, offering complementary contexts to assess treatment-based inference. This study developed and validated treatment-based inference algorithms for NTRK fusion–positive and BRAF V600E–mutated metastatic solid tumor patients (pts) and assessed the incremental contribution of treatment-based inference (ICTBI) beyond structured biomarker fields in community oncology. Methods: Using structured RW data from the US Oncology Network iKnowMed EHR database, analyses were conducted separately for NTRK fusion and BRAF V600E using algorithm-specific validation cohorts (VCs; pts with interpretable structured biomarker results) and application cohorts (ACs; pts receiving biomarker-restricted targeted therapies). The study period spanned biomarker-specific FDA approval dates through 12/31/2025. Qualifying therapies included larotrectinib or entrectinib for NTRK fusion and dabrafenib + trametinib for BRAF V600E. Within VCs, algorithm-inferred biomarker status was assigned based solely on qualifying therapy exposure, with structured biomarker fields masked. Inferred status was compared with documented biomarker results to estimate positive predictive value (PPV), sensitivity, and specificity. Within ACs, ICTBI was calculated as the proportion of biomarker-positive pts identified exclusively through treatment exposure. Results: In the NTRK fusion VC (n = 34,765), treatment-based inference demonstrated high specificity (99.98%) and PPV (84.31%) with low sensitivity (7.68%), consistent with biomarker rarity and limited treatment exposure. In the NTRK AC (n = 90), 52% of biomarker-positive pts were identified exclusively through treatment-based inference, indicating substantial ICTBI. In the BRAF V600E VC (n = 6,021), treatment-based inference showed very high PPV (98.99%) and specificity (99.92%) with moderate sensitivity (39.42%), reflecting broader targeted therapy uptake. In the BRAF V600E AC (n = 643), 44% of biomarker-positive pts were identified exclusively via treatment-based inference. Conclusions: Treatment-based inference demonstrated high validity and meaningfully improved identification of biomarker-positive pts. Incorporating this approach has the potential to enhance RW cohort completeness, improve the feasibility of biomarker-driven studies, and support more robust evaluation of precision oncology therapies in community practice.