Radiomic profiles via machine learning to predict response, overall survival, and KRAS status in metastatic colorectal tumors treated with FOLFOX/bevacizumab/hydroxychloroquine.

Abstract

e15125 Background: Autophagy is a resistance mechanism to chemotherapy that is inhibited by hydroxychloroquine (HCQ). We have reported a phase II trial of FOLFOX/bevacizumab/HCQ in 28 evaluable patients with untreated metastatic colorectal cancer (mCRC). Overall response rate was 68%, with 11% complete response (CR) and 57% partial response (PR), while stable disease (SD) was seen in 30%. We hypothesize that analysis of CT imaging features via machine learning (ML) will enhance subtle yet important radiographic characteristics, and reveal imaging signatures determinant of outcome and mutational status. Methods: Baseline CT images were collected and 1265 quantitative imaging (QI) features extracted across all the metastatic sites – liver, lung, and lymph nodes – including descriptors of size, morphology, texture, and intensity. Cross-validated sequential feature selection coupled with support vector machine (SVM) was used to determine the most discriminative QI features for our integrative predictor of response and mutational status, and support vector regression (SVR) was used to derive imaging predictors of overall survival. The model predictions were compared with actual clinical results, including response, survival, and genomic aberrations. Results: Various QI features, primarily descriptive of texture and tumor volume, were determined as most important by the ML predictor. Using this signature, our predictor classified “Responder (PR+CR) vs Non-responder (SD)” with an accuracy of 85% [sensitivity(se) = 0.84, specificity(sp) = 0.71, AUC = 0.85]. The QI features were also able to detect KRAS and TP53 mutational status with an accuracy of 88% [se = 0.92, sp = 0.85, AUC = 0.87] and 86% [se = 0.80, sp = 0.90, AUC = 0.84], respectively. The SVM model predicted overall survival greater than the median (32 mo) with an accuracy of 86% [se = 0.93, sp = 0.79, AUC = 0.87]. The Pearson correlation coefficient between the SVR score and overall survival was estimated to be 0.73 (p < 0.0001). Conclusions: Radiomic analysis of baseline CT imaging features analyzed by ML yielded an imaging signature predictive of response, survival, and KRAS and TP53 mutational status. If validated in a larger clinical data set, machine learning may offer a predictive biomarker to aid clinical decision making for mCRC patients.