ESTRO 2021 Abstract Book

S73

ESTRO 2021

Results Random Survival Forest Model was prone to overfitting comparing the training and validation fold performances (Table 1). The three remaining ML models achieved similar performance whatever the task on the cross-validation folds. Fast Survival SVM (FSSVM) and Cox Proportional Hazards (CPH) model performances correctly generalized on the test set with values respectively equal to 0.77, 0.72, 0.67, 0.69 for OS, PFS, PRFS and EPRFS for FSSVM and 0.73, 0.75, 0.74, 0.76 for CPH (Table 1). Radiomics features-based models had superior performance compared to the clinical ones (Wilcoxon signed-rank test p-value<0.03).

Conclusion Results suggest that radiomics features extracted from pre-treatment [18F]-FDG PET images could predict survival outcomes using a common HPV signature in CC and ASCC. Validation of initial findings on two independent test cohorts with ASCC and head and neck HPV-induced cancers is in progress. PH-0105 Prediction of clinical complete response in rectal cancer using clinical and radiomics features P. Mbanu 1 , E. Vasquez Osorio 2 , H. Mistry 2 , J. Mercer 3 , L. Malcomson 2 , R. Kochhar 3 , A. Renehan 4 , M. van Herk 4 , M. Saunders 1 1 Christie NHS Foundation Trust, Department of Clinical Oncology, Manchester, United Kingdom; 2 University of Manchester, Division of Cancer Sciences, School of Medical Sciences, Faculty of Biology, Medicine and Health, Manchester Academic Health Science Centre, Manchester, United Kingdom; 3 Christie NHS Foundation Trust, Department of Radiology, Manchester, United Kingdom; 4 University of Manchester, Division of Cancer Sciences, School of Medical Sciences, Faculty of Biology, Medicine and Health, Manchester Academic Health Science Centre, Manchester, United Kingdom Purpose or Objective In patients with rectal cancer, watch and wait for a clinical complete response (cCR) after neoadjuvant chemo-radiotherapy has the potential to avoid major surgery and stoma formation. But, other than tumour size, there are currently no predictors for cCR. To better optimise neoadjuvant strategies, we evaluated the predictive characteristics of clinical and radiomics variables. Materials and Methods From the OnCoRe (The Rectal Cancer Oncological Complete Response Database) database, we performed a matched case-control study in 304 patients (152 cCR; 152 non-cCR) deriving training (N=200), and validation (N=104) sets based on the patient’s date of diagnosis to mimic a prospective study. We collected pre- treatment demographic and routine laboratory parameters. We segmented the gross tumour volume on T2W pre-treatment MR Images, which were normalised using histogram-based normalisation. We extracted 1781 radiomics features per patient (1430 features accepted as stable features for analysis based on ICC >0.9). We used principal component analysis to cluster radiomics features. The ROC AUC was used to evaluate predictive A multivariable clinical model that included full blood count parameters, alkaline phosphatase, albumin, and tumour diameter had modest predictive characteristics. In the radiomics analysis, four clusters were identified using principal components analysis – predictive characteristics of these alone were modest. The addition of radiomics variables to clinical variables marginally improved prediction, and this improvement remained after validation. A significant drop in ROC AUC on the validation cohort with the models containing clinical variables is due to calibration drift, a known phenomenon with clinical variable over time. Patients with cCR were treated from 2008-2013 in the training cohort and 2013-2019 in the validation cohort. power. Results