LAUSR

PURPOSE Data quality and standardization remain a challenge when analyzing real-world clinical data. We built a clinical research database, using machine learning and natural learning processing, and investigated factors influencing testosterone recovery (T-recovery) in patients with localized prostate cancer (LPC) after initial androgen deprivation therapy (ADT). METHODS Medication and treatment-associated dates missing in structured tables were extracted from patient notes using ConceptMapper, an automated data extraction tool, standardized and curated in Sema4 clinical research database. ADT usage duration was evaluated, and T-recovery in patients with LPC was analyzed by the Kaplan-Meier method and multivariable Cox proportional hazards models. We assessed the prognostic value of post-ADT T-recovery with prostate-specific antigen progression-free survival and failure-free survival. RESULTS In total, 4,125 of 30,832 (13.4%) patients with prostate cancer had medication exclusively from notes with high precision and recall, F1 score ≥ 0.95. Association of dates with medication usage had a F1 score of 0.76. ADT duration estimation had higher accuracy combining information from notes to tables from electronic medical record (70% v 45%). Baseline testosterone was the strongest predictor of T-recovery in these patients. Patients with a baseline testosterone ≥ 300 ng/dL recovered in 9.79 versus 38 months for patients with baseline testosterone < 300 ng/dL (P < .0001). Shorter prostate-specific antigen progression-free interval was observed for patients with T-recovery (≥ 300 ng/dL) at 6 months after ADT cessation compared with patients without T-recovery (< 300 ng/dL; 13.7 v 25.1 months; P = .055). CONCLUSION We augmented structured electronic medical record data with data extracted from notes and improved the accuracy of medication information for patients. ADT exposure and T-recovery in patients with LPC produced results consistent with the literature and clinical experience and illustrates the power of applying machine learning methods to enhance the quality of real-world evidence in answering clinically relevant questions. Machine learning and NLP augments data beyond structured EMR fields when building real-world evidence databases