Incorporating SULF1 polymorphisms in a pretreatment CT-based radiomic model for predicting platinum resistance in ovarian cancer treatment

[Display omitted] •Early detection of platinum resistance for OC treatment remains challenging.•We built a prediction model incorporating genomic data of SULF1 with CT radiomics.•Our model showed promising prediction efficiency with potential clinical utility.•Radiogenomics is a promising method to noninvasive prediction of chemoresistance. Early detection of platinum resistance for ovarian cancer treatment remains challenging. This study aims to develop a machine learning model incorporating genomic data such as Single-Nucleotide Polymorphisms (SNPs) of Human Sulfatase 1 (SULF1) with a CT radiomic model based on pre-treatment CT images, to predict platinum resistance for ovarian cancer (OC) treatment. A cohort of 102 patients with pathologically confirmed OC was retrospectively enrolled into this study from January 2006 to February 2018. All patients had platinum-based chemotherapy after maximal cyto-reductive surgery. This cohort was separated into two groups according to treatment response, i.e., the group with platinum-resistant disease (PR group) and the group with platinum-sensitive disease (PS group). We genotyped 12 SNPs of SULF1 for all OC patients using Mass Array Method. Radiomic features, SNP data and clinicopathological data of the 102 patients were used to build the differentiation models. The study participants were divided into two cohorts: the training cohort (n = 71) and the validation cohort (n = 31). Feature selection and predictive modeling were performed using least absolute shrinkage and selection operator (LASSO), Random Forest Classifier and Support Vector Machine methods. Model performance for predicting platinum resistance was assessed with respect to its calibration, discrimination, and clinical application. For prediction of platinum resistance, the approach combining the radiomics, clinicopathological data and SNP data demonstrated higher classification efficiency, with an AUC value of 0.993 (95 % CI: 0.83 to 0.98) in the training cohort and 0.967 (95 % CI: 0.83 to 0.98) in validation cohort, than the performance with only the SNPs of SULF1 model (AUC: training, 0.843 [95 %CI: 0.738-0.948]; validation, 0.815 [0.601-1.000]), or with only the radiomic model (AUC: training, 0.874 [95 %CI: 0.789-0.960]; validation, 0.832 [95 %CI: 0.687-0.976]). This integrated approach also showed good calibration and favorable clinical utility. A predictive model combining pretreatment CT radiomics with genomic data such as SNPs of SULF1 could po