3020 - Multi-Modality Machine Learning Prediction Model of Hematologic Toxicity in Neoadjuvant Chemoradiotherapy for Locally Advanced Rectal Cancer

B. Li¹, Y. Li¹, X. Zhen², and W. Xiao¹; ¹Department of Radiation Oncology, Sun Yat-sen University Cancer Center, Guangzhou, China, ²School of Biomedical Engineering, Southern Medical University, Guangzhou, China

Purpose/Objective(s): Hematologic toxicity (HT) is a severe adverse effect of Neoadjuvant chemoradiotherapy (CRT) for local advanced rectal cancer (LARC), due to myelosuppression during the treatment. Hence, there is an urgent necessity to accurately predict the occurrence of HT before radiotherapy. The purpose of this study to develop an multi-modality machine learning HT prediction model, using multiple classifiers and feature modalities, to guide the treatment adaption to avoid the occurrence of HT. Materials/

Methods: A total of 279 LARC patients who received CRT at Sun Yat-sen University Cancer Center (SYSUCC) from Feburary 2014 to December 2018, were analyzed restrospectively in this study, including 61 patients with severe effect of grade 3+ HT. Three main modalities, includes clinical demographic, dosimetric and radiomics features, were extracted to identify the potential indicators of HT occurrence. Specifically, the dosimetric and radiomics features were extracted from the pelvic bony structures as the regions of interest (ROI) in planning CT images. A multi-criteria decision-making based classifier fusion algorithm (MCF) was applied to select the most predictive features and construct the multi-modality fusion model. In the model construction, a five-fold cross validation was applied for training process. In each fold, feature selection was performed only on the training dataset, and the selected features were fed for further classifier fusion and validated on the testing dataset.
Results: Model performance metric of area under curve (AUC) of the three single-modality models that constructed with only clinical demographic, dosimetric and radiomics features, were 0.627 (95%CI, 0.515-0.800), 0.643 (95%CI, 0.530-0.833) and 0.606 (95%CI, 0.518-0.775) respectively. In comparison, the proposed multi-modality improved the AUC to 0.681 (95%CI, 0.528-0.889), and the accuracy, sensitivity and specificity were 0.667 (95%CI, 0.500-0.846), 0.761 (95%CI, 0.556-0.999) and 0.447 (95%CI, 0.200-0.750) respectively.
Conclusion: This study proposed a multi-modality machine learning model for HT prediction, demonstrating the advantage of multi-modality fusion model over the single-modality models. This proposed method could effectively predict HT occurrence and thus improve the clinical outcome of CRT for LARC patients.