Abstract 452P
Background
In our retrospective analysis, Notch signalling pathway was associated with resistance to anti-vascular endothelial growth factor (VEGF) therapy in patients with metastatic colorectal cancer (mCRC). We tested whether radiomics might select treatment-naïve mCRC patients responding to bevacizumab, beyond clinical and genomic (Notch Intracellular Cleaved Domain (NICD)/JAG1 expression) parameters.
Methods
76 consecutive mCRC patients treated with first-line bevacizumab were retrospectively selected. Immunohistochemistry analysis of tissue microarrays assessed NICD, JAG1, CD44, CD3, CD4, CD8, CD20, DLL3 and DLL4 expression. Abdominal CT scans were imported into a dedicated software for tumor segmentation and extraction of 852 radiomic features (RFs), which were included into machine learning-based predictive models. Pre-processing of RFs included redundant features elimination and standardization; L2 penalized logistic regression with Monte-Carlo cross-validation were implemented for wrapper-based feature selection and model training/test. Three models were developed: clinical/genomic (C/G), radiomic (R) and the comprehensive integrated model (I), which were compared based on ROC-AUC and accuracy metrics.
Results
NICD and JAG1 expression was associated with response to bevacizumab (p<0.05). Using likelihood-ratio test as inclusion criteria, the selected variables were 5 for both C/G and R models, then aggregated into the I model. C/G features included NICD expression, number of involved sites, primitive location, resection of metastases and performance status, while selected RFs belonged to both first- and higher-orders classes. ROC-AUC and accuracy were 0.724 (95%CI:0.722-0.727) and 0.669 (95%CI:0.666-0.671), 0.786 (95%CI:0.784-0.788) and 0.710 (95%CI:0.708-0.713), 0.810 (95%CI:0.808-0.812) and 0.743 (95%CI:0.741-0.745) for C/G, R and I model, respectively.
Conclusions
The integration of clinical, genomic and radiomic features showed the highest performance in predicting response to bevacizumab.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
Azienda Ospedaliero-Universitaria di Parma.
Funding
Fondazione SNUPI.
Disclosure
All authors have declared no conflicts of interest.