Radiotherapy-based regimens are the main treatment for cervical cancer, which has good curative effect and some toxicity. However, radiation-induced toxicities are not easily detected and may be associated with patient survival. Machine-learning with multiomics model was constructed and evaluated to predict key features, treatment response, and survival outcomes of radiotherapy-induced acute rectal toxicities (RI-ARTs) in cervical cancer. A retrospective analysis was conducted on 370 patients with cervical cancer who received radiotherapy from 2016 to 2023, of whom 231 had RI-ARTs, and 129 had no RI-ARTs. Corresponding CT imaging features, rectal radiotherapy dose-related information, and clinical features were also obtained for individual patients. The cohort was first partitioned into a training set (80%) and a completely independent test set (20%). Screening was then performed through univariate (Mann-Whitney U test/Chi-square test) and multivariate (LASSO) approaches, in which redundant features were eliminated based on Pearson correlation analyses, leading to the final selection of the most appropriate predictive model from among six analyzed machine learning algorithms and deep learning models. Area under the curve (AUC) and survival analyses were then utilized to assess model validity and clinical applicability. Twelve salient features were screened. The best performance model for predicting RI-ARTs was the XGBoost model under the combined model (dosiomics, radiomics, and clinical models), with an AUC of 0.80 (95% CI, 0.68–0.90) in the test cohort. Rectal D0·1cc feature was significantly associated with the incidence of RI-ARTs, and the risk of RI-ARTs was significantly reduced when D0·1cc ≤ 93 Gy. Low XGBoost-based multi-omics model scores relative to the median in outcome cohort was associated with OS (Hazard ratio, 0.31; 95% CI, 0.11–0.88; p = 0.028). The integrated multi-omics model demonstrates potential for non-invasive RI-ART prediction. Exploratory analysis further links RI-ART incidence to survival outcomes. Despite the single-center retrospective constraints, these findings suggest potential utility in clinical decision-making, pending further prospective validation.
Zhu et al. (Thu,) studied this question.