Real-world evidence analysis of early-stage prostate cancer using natural language processing.

Abstract

e17007 Background: Real-world evidence (RWE) on the epidemiology and management of localized/locally advanced prostate cancer (LPC) is lacking. Here we used EHRead, a natural language processing framework based on SNOMED CT terminology, to extract clinical information from electronic health records (EHRs) to perform a comprehensive description of a real-world cohort of LPC patients. Here we present a follow-up of the project already communicated at ESMO 2020. Methods: This is an observational, multicenter, retrospective study based on the secondary analysis of free-text and structured clinical information found in EHRs from all observed adult patients with LPC between January 2014 and December 2018 in 12 Spanish hospitals. A total of 375 variables were extracted using EHRead to assess patient clinical characteristics, treatment patterns and journey across different prostate cancer stages. Event-free survival (EFS) was analyzed in incident LPC patients with presence of Gleason or PSA and follow-up greater than 0 days. A multivariable Cox-proportional hazards model was applied to obtain adjusted hazard ratios (aHR). Results: We observed 19788 patients with LPC during a median of 3.5 (interquartile range 1.4-7.4) years of follow-up. 52% of patients had low- (LR), 21% intermediate- (IR) and 27% high-risk (HR) disease. 3938 (18%) patients transitioned to other disease stages, leading to castration-resistant prostate cancer (mCRPC) in 4%, 7%, 16% of patients with LR, IR, and HR LPC, respectively. 7884 (42%) patients received radiotherapy (RT), 7021 (38%) prostatectomy (PT), 6061 (31%) androgen deprivation therapy (ADT), and 1034 (6%) active surveillance. Local recurrence was observed in 3% and distant metastasis in 11% (57% bone, 37% visceral). Distant metastases were observed in 7%, 12% and 21% of HR, IR and LR; 8% vs. 9% of PT vs. RT; and 16% vs 6% of ADT vs. no ADT, respectively. EFS was assessed in 7465 (38%) patients, with a 4-year EFS rate of 18%. Median EFS was 22, 13, and 8 months for LR, IR, HR; 23 vs. 12 months for PT vs. RT; and 8 vs. 21 months for ADT vs. no ADT, respectively. A multivariable cox proportional hazards model is presented in the following Table. Conclusions: During the study period, 11% of all LPC patients developed metastatic disease, with only 18% remaining event-free at 4 years. Transition to mCRPC was 4- and 2-fold higher in HR vs. IR and LR disease, respectively. Disease risk, treatment with RT and/or addition of ADT were independently associated with shorter EFS. Our findings suggest that, despite adjuvant treatment, LPC local extension and histopathological characteristics at diagnosis remain critical determinants of patient prognosis. A more detailed analysis in patients with available Gleason and PSA is underway. [Table: see text]