Purpose
Prostate cancer is the fifth leading cause of death in the United States. Genomic testing is essential to guide treatment decisions in patients with metastatic castration resistant prostate cancer (mCRPC), the most advanced stage of prostate cancer. However, identifying mCRPC patients from administrative data is challenging and hinders researchers’ ability to assess testing among these patients. This study aims to develop algorithms using structured data and unstructured data with Natural language processing (NLP) methods to identify veterans by disease stage and hormone sensitivity, and to assess patient characteristics as well as receipt of tumor NGS testing.
Methods
We used biopsy, pathology, and diagnosis codes, to identify veterans with newly diagnosed PC within the Veterans Health Administration (VA) from January 1, 2017 to December 31, 2020. We developed and deployed: 1. A structured algorithm that used medication and Prostate-Specific Antigen (PSA) data to assess hormone sensitivity. 2. NLP tools to extract disease stage and hormone sensitivity from clinical notes. We report descriptive statistics on patient demographics, clinical characteristics, disease status, androgen deprivation therapy (ADT), and receipt of tumor NGS testing.
Results
There were 42,485 veterans with newly diagnosed prostate cancer between 2017-2020. This represented ~0.18% of veterans served in the VA and consisted of Whites (57%), Blacks (33%), and others (10%). During the study period, 3,113 (7.3%) patients had documentation of assessment for intraductal carcinoma, 5,160 (12.1%) had ADT treatment, 1,481 (3.5%) had CRPC, and 3,246 (7.6%) had metastatic disease. Among the 42,485 veterans, 422 received tumor NGS testing within VA, and 300 of them had metastatic disease. NLP tool and structured data algorithm collectively showed that 38% of the 422 tumor NGS testing recipients had mCRPC. Among all newly diagnosed PC patients, White patients had highest rates of tumor-based testing (2.3%), then Native Hawaiians (1.7%), Asians and Blacks (1.2% each), compared to Native Americans (0.4%).
Implications
NLP tools alongside structured data algorithms successfully identified variables required to measure access to tumor NGS testing. Efforts to validate and apply this method is ongoing to assess receipt of precision prostate cancer care in VA.