Natural Language Processing of Radiology Reports to Assess Survival in Patients with Advanced Melanoma

TL;DR

This study uses natural language processing to analyze radiology reports and shows that liver metastases in advanced melanoma patients significantly worsen survival, even with immunotherapy.

Contribution

The study introduces a novel NLP-based method to extract metastatic patterns from radiology reports and assess their impact on survival in melanoma patients.

Findings

Patients with liver metastases (M1L+) had significantly worse overall survival compared to those without (M1L−).

NLP effectively classified metastatic patterns and confirmed the poor prognosis of hepatic metastases in melanoma patients.

Immunotherapy-treated patients with liver metastases still had lower survival rates than those without.

Abstract

Advanced melanoma therapeutic outcomes have improved markedly with immunotherapy, but with variable response across patients with different metastatic patterns. Understanding the impact of specific sites of metastatic disease on survival in advanced melanoma, in particular understanding whether liver metastases have a deleterious effect, has great clinical significance. Natural language processing (NLP) allows for text extraction from a large imaging dataset to evaluate the impact of the pattern of metastatic spread. We identified 2239 patients with advanced melanoma and CT imaging using NLP and classified them according to AJCC staging criteria as well as alternative criteria indicating whether liver metastases were present (M1L+) or not (M1L−). Whether using AJCC or alternative criteria, overall survival (OS) was poorest for the M1L+ group (median OS 0.69 years and 1.4 years for the entire cohort and immunotherapy-treated subset, respectively) versus 1.8 years and 2.9 years for the M1L− group. NLP can rapidly evaluate the prognosis of melanoma patients with different metastatic patterns, confirming inferior OS seen in patients with hepatic metastases.

Background/Objectives: To use natural language processing (NLP) to extract large-scale data from the CT radiology reports of patients with advanced melanoma treated with immunotherapy and to determine whether liver metastases affect survival. Methods: Patient criteria (M1 disease subclassified into M1a, M1b, or M1c) as well as alternative criteria (M1 with advanced melanoma, imaged with CT chest, abdomen, and pelvis from July 2014–March 2019) were included retrospectively. NLP was used to identify metastases from CT reports, and then patients were classified according to American Joint Committee on Cancer (AJCC) staging disease subclassified into M1L+ or M1L−, indicating whether liver metastases were present or not). Statistical analysis included constructing Kaplan–Meier survival curves and calculating hazard ratios (HRs). Results: 2239 patients were included (mean age, 63 years). Whether using AJCC or alternative criteria, overall survival (OS) was poorest for M1L+ (entire cohort median OS, 0.69 years [95% CI: 0.60–0.82]; immunotherapy cohort median OS, 1.4 years [95% CI: 0.92–2.0]) compared to M1L− (entire cohort median OS, 1.8 years [95% CI: 1.4–2.2]; immunotherapy cohort median OS; M1L−, 2.9 years [95% CI: 2.3–3.9]). The median HR for M1L+ (median HR, 5.35 [95% CI: 4.59–6.24]) was higher than that for M0 (p < 0.001). The median HR for M1L+ (median HR, 2.13 [95% CI: 1.65–2.64]) was higher than that for M0 (p < 0.01). Conclusions: Patients with advanced melanoma, particularly those with liver metastases, demonstrated inferior survival, even when treated with immunotherapy.

Figures