Author: Sijuan Huang, Yongbao Li 👨🔬
Affiliation: Sun Yat-sen University Cancer Center; State Key Laboratory of Oncology in South China; Collaborative Innovation Center for Cancer Medicine; Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangzhou, Guangdong, 510060, China, Sun-Yat sen University Cancer Center 🌍
Purpose: Hematologic toxicity (HT) is one of the most prevalent treatment-related toxicities experienced by locally advanced cervical cancer (LACC) patients receiving radiotherapy (RT). This study aimed to develop a prediction model for severe acute HT, combining radiomic features and traditional features.
Methods: 187 LACC patients were retrospectively recruited in our center. Acute HT was assessed weekly during RT. Traditional features, clinical features (body mass index, chemotherapy, brachytherapy, stage). DVH (Dose Volume Histogram) features of bone marrow (BM), and radiomics of BM were analyzed. Four models were: 1) only clinical feature; 2) only DVH; 3) combination of clinical and DVH features; 4) multi-omics: combination of clinical, DVH features and radiomics of BM, as the input of the machine learning (ML), to prediction acute HT. Four ML methods: (random forest (RF), logistic regression, naive Bayes, extreme gradient boosting (XGBoost)).
Results: A total of 127 patients were included in the analysis. 60 patients (47.24%) suffered grade ≥3 acute HT. 8 radiomic features, 2 clinical features and 2 DVH features were obtained by Pearson correlation Coefficient and random forest variable importance analysis. For all of the 4 ML methods, the multi-omic model combining chemotherapy status, radiomic and DVH features, outperformed the conventional model, with AUC (area under the curve) 1.0 (RF), 0.777 (Logistic), 0.805 (Naive Bayes) and 1.0 (XGBoost), compared to the clinical and DVH features model with 0.976 (RF), 0.736 (Logistic), 0.715 (Naive Bayes) and 0.888 (XGBoost), the clinical model achieving AUC 0.864 (RF), 0.713 (Logistic), 0.788 (Naive Bayes), and 0.88 (XGBoost), compared with DVH model AUC 0.861 (RF), 0.552 (Logistic), 0.531 (Naive Bayes), and 0.794 (XGBoost).
Conclusion: A multi-omic model with clinical, radiomic and DVH features achieved better prediction results for acute HT caused by RT for LACC. Further research may use 3D dosiomic features to develop more accurate probability models.