Prediction of Pathological Complete Response after Chemoradiation for Locally Advanced Rectal Cancer Using Machine
David Mens1, Soogyeong Shin2, Farhan Akram2
1Department of Surgical Oncology and Gastrointestinal Surgery, Erasmus MC Cancer Institute, Erasmus University Medical Centre, Rotterdam, The Netherlands.
Journal of the Anus, Rectum and Colon
|August 1, 2026
Summary
Accurate prediction of pathological complete response (pCR) after chemoradiotherapy (CRT) in locally advanced rectal cancer (LARC) is crucial. Machine learning models using pretreatment clinical variables showed limited accuracy for predicting pCR.
Area of Science:
- Oncology
- Medical Imaging
- Machine Learning
Background:
- Organ-preserving strategies are increasingly adopted for rectal cancer patients achieving pathological complete response (pCR) after chemoradiotherapy (CRT).
- Accurate pretreatment prediction of pCR is vital for treatment decisions in locally advanced rectal cancer (LARC), especially when R0 resection is feasible and CRT's oncological benefit is uncertain.
- The INTERCEPTOR study aims to enhance pCR prediction using machine learning (ML) models with pretreatment clinical variables.
Purpose of the Study:
- To evaluate the efficacy of machine learning (ML) models in predicting pathological complete response (pCR) in locally advanced rectal cancer (LARC) patients.
- To determine if pretreatment clinical variables can improve pCR prediction beyond the baseline probability.
- To assess the performance of extreme gradient boosting (XGBoost) models in predicting pCR after chemoradiotherapy (CRT).
Main Methods:
- Analysis of 238 LARC patients treated with CRT and surgery between 2004 and 2018.
- Training and validation of extreme gradient boosting (XGBoost) models using pretreatment clinical variables with 5-fold cross-validation.
- Assessment of model performance using AUROC, sensitivity, specificity, and feature importance via SHAP and XGBoost methods.
Main Results:
- Among 238 patients, 12.6% achieved pCR.
- The number of radiologically positive lymph nodes was the strongest predictor but had limited discriminative power (AUROC 0.65 ± 0.04).
- Combining positive node count with other clinical variables yielded only a modest improvement in prediction accuracy.
Conclusions:
- Pretreatment clinical variables alone demonstrate poor-to-fair accuracy in predicting pCR after CRT in LARC.
- Further research is needed to develop more accurate predictive models for pCR in LARC.
- Optimizing pCR prediction can guide organ-preserving strategies and personalize treatment for LARC patients.
