Radiomic signatures to estimate survival in patients with advanced hepatocellular carcinoma treated with sorafenib: Cancer and Leukemia Group B 80802 (Alliance).

BACKGROUND: Current methods to evaluate therapeutic response in patients with hepatocellular carcinoma (HCC) rely on tumor size and density, which do not always correlate well with survival. We used pretreatment clinical and radiomics variables to predict overall survival (OS) in the randomized phase III CALGB 80802 (Alliance) trial, investigating the efficacy of sorafenib + doxorubicin versus sorafenib alone. MATERIALS AND METHODS: Using machine learning, we analyzed baseline and first follow-up computed tomography (CT) images and associated clinical metadata from patients imaged in February 2010-May 2015, up to November 2015 with follow-up. Adult patients with HCC (n = 129) were randomly assigned to training (n = 92) and validation (n = 37) sets. We assessed the performance of a signature combining CT imaging features and clinical variables using hazard ratios to estimate OS after week 10 (first follow-up). RESULTS: Most patients were male (86.6%) and had bilirubin <2 mg/dl (98.4%), albumin >3.5 g/dl (69.0%), and moderately differentiated HCC (34.1%). Median (interquartile range) age: 58 years (63-71 years), alpha-fetoprotein (AFP): 2.5 ng/ml (26-282 ng/ml), international normalized ratio: 1 (1.1-1.2), Child-Pugh score: 5 (5-6). The highest-performing parsimonious training set signature combined clinical and radiomics features at baseline and week 10. In the validation set, the hazard ratio was 2398 (95% confidence interval 121-47 371) (P < 0.001). The signature's variables, ranked by importance, included baseline clinical features [albumin (1), AFP (2), Child-Pugh (4)], baseline radiomics features [component 17 (3), component 1 (5), component 9 (7), tumor volume (8)], and week 10 radiomics features [delta tumor volume (6)]. CONCLUSION: OS can be accurately predicted in patients with HCC receiving sorafenib by combining certain radiomics features with clinical metadata, centered primarily on baseline characteristics.

Duke Scholars

Author Andrew Benjamin Nixon Medicine, Medical Oncology