Predicting the risk of radiation enteritis in cervical cancer patients using inflammatory markers through machine learning

Abstract

Abstract Methods A total of 228 patients diagnosed with cervical squamous cell carcinoma of stage IB-IIIB and undergoing radical RT were included in the study. The LASSO method was utilized to identify crucial features associated with RE. Clinical characteristics and inflammatory markers pre/post-treatment were used to develop five machine learning models, comprising a training set and validation set (80% of participants), which were then assessed in the remaining study sample (20% of participants). The area under the receiver-operating characteristic curve (AUROC) and Brier scores were employed to compare the prediction performances of different models. The Random Forest (RF) Classifier model was employed for predicting RE, with interpretation provided by the SHapley Additive exPlanations (SHAP) package. (3) Results The RF model demonstrated superior performance compared to other classifier models in the training set (area under the curve [AUC]: 1.000, 95% confidence interval [CI]: 1.000–1.000) and the validation set (AUC: 0.757, 95% CI: 0.636–0.878). Additionally, this model achieved the lowest Brier Score (0.163). Nine crucial variables, including LMR, Pre_N, Post_P, FIGO, Post_L, Post_Hb, UBI, Tumor, and DPT_high, were selected. (4) Conclusions This pioneering study's predictive model will enhance understanding of the risk of RE and provide clinicians with a valuable tool based on inflammatory markers (4 factors) and clinical parameters (5 factors) in cervical cancer for guiding patient treatment.