4722. Natural Language Processing of Radiology Reports to Assess Survival in Patients With Melanoma and Hepatic Metastases on Immunotherapy
Authors * Denotes Presenting Author
  1. Jeeban Das *; Memorial Sloan Kettering Cancer Center
  2. Jordan Eichholz; Memorial Sloan Kettering Cancer Center
  3. Srinivasa Varadan Sevilimedu Veeravalli; Memorial Sloan Kettering Cancer Center
  4. Natalie Gangai; Memorial Sloan Kettering Cancer Center
  5. Danny Khalil; Memorial Sloan Kettering Cancer Center
  6. Michael Postow; Memorial Sloan Kettering Cancer Center
  7. Richard Do; Memorial Sloan Kettering Cancer Center
Through the hepatic microenvironment and promotion of systemic immune tolerance, hepatic metastases may affect systemic anti-tumoral immunity and limit efficacy of immune checkpoint blockade (ICB). Natural language processing (NLP) is a promising tool which can be used to extract specific pertinent text from large imaging data archives to evaluate impact of metastatic spread on overall survival (OS). This study aims to measure OS of advanced melanoma patients treated with immunotherapy, comparing patients with different metastatic patterns, including hepatic involvement, identified by NLP of CT radiology reports.

Materials and Methods:
A retrospective analysis included all patients with advanced melanoma who were treated with ICB at a cancer center and imaged with CT chest, abdomen, and pelvis (CT CAP) between July 2014 and March 2019. Using an NLP model designed to identify metastases from CT reports, all patients were categorized into those without distant metastases (M0) and those with distant metastases with or without hepatic metastases (ML+/ML-), and also according to the American Joint Committee on Cancer (AJCC) criteria: M1a for distant metastasis to skin/soft tissues and/or nonregional lymph nodes, M1b for distant metastasis to lung with or without M1a disease, and M1c for distant metastasis to non-CNS visceral sites with or without M1a or M1b disease.

Of the 778 advanced melanoma patients who received ICB, 430 (55.3%) were in the M0 cohort with median OS of 4 years (95% CI 3.5 - 6.3). In comparison, 348 (44.7%) were metastatic with median OS of 2.3 years (95% CI 1.5 - 2.9), including 203 (26.1%) in the ML-cohort with median OS of 3 years (95% CI 0.9 - 1.5, <em>p</em> = 0.01), and 145 (18.6 %) ML+: median OS of 1.2 years (95% CI 1.7 - 27, <em>p</em> < 0.001). According to AJCC criteria, 12 (1.5%) were stage M1a with median OS of 3.2 years (95% CI 0.6 - 2.9, <em>p</em> = 0.4), 114 (14.6%) were M1b with median OS of 2.2 years (95% CI 0.9 - 1.6, <em>p</em> = 02), and 222 (28.5%) stage M1c with median OS of 1.6 years (95% CI 1.4 - 2.2, <em>p</em> < 0.001).

Patients with advanced melanoma and visceral metastases, and those with hepatic involvement in particular, demonstrated inferior survival when treated with ICB, compared to other patients with metastatic disease. NLP is a useful tool to rapidly investigate prognosis of melanoma patients on immunotherapy with different metastatic patterns, and additional study in this area is warranted.