Radiomic signatures to predict survival in patients with advanced hepatocellular carcinoma (HCC) treated with sorafenib +/- doxorubicin: Correlative science from CALGB 80802 (Alliance).

343 Background: Alliance/CALGB 80802, a randomized phase III trial, evaluated sorafenib plus doxorubicin vs. doxorubicin in pts with HCC and showed no improvement in median overall survival (OS) (HR[95CI] 1.05[0.83-1.31]) or PFS (HR[95CI] 0.93[0.75-1.16]). In HCC surrogacy of tumor response with OS remains controversial, in part due to varying criteria used for response evaluation (e.g., RECIST1.1 and mRECIST). We evaluated the performance of several models to predict OS using pretreatment clinical and radiomic variables. Methods: In CALBG 80802, we segmented all measurable tumor lesions on sequential CT scans. A lesion’s imaging phenotype was deciphered with 23 uncorrelated quantitative imaging features measured at baseline and week (wk) 10 (first follow-up). An OS landmark survival analysis was conducted at wk 10. Patients were randomly assigned (3:1) to training (n = 92) and validation (n = 37) sets. In a training set, 6 random forest predictive models (6 signatures) used features that best predicted OS using 3 sets of variables: radiomics only (n = 23), clinical only (n = 9), radiomics and clinical (n = 32). Two time points (baseline only or baseline + wk 10) were assessed. Each signature's output was an individualized prediction and a continuous value ranging from 0 to 1 (from most to least favorable predicted OS). The primary endpoint was to compare these models' performance to predict OS using error rate (Harrell's concordance-index) in the validation set. Results: Of the 6 training signatures evaluated, the one achieving the highest performance in the validation set was an 8-feature signature combining radiomics and clinical variables measured at two time points (baseline + wk 10) with an error rate of 35.6%. The variables [rank of importance] (table) selected by the signature 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 wk 10 radiomics features (delta tumor volume[6]). Variable delta tumor volume [6] used a more enhanced estimation of tumor burden at baseline and a delta tumor volumetric measurement; compared to RECIST1.1 measurement of percentage change in unidimensional measurement of a subset of target lesions. The four quartiles of the signature were significantly associated with OS (Log-Rank, P < 0.0001). Conclusions: The selected combined radiomic and clinical composite signature provided the best prediction for OS in the 80802 study patients’ population. It is a suggested way forward to go beyond single anatomic measurement techniques such as RECIST or mRECIST. [Table: see text]

Duke Scholars

Author Andrew Benjamin Nixon Medicine, Medical Oncology