Machine learning for predicting clinical complete response after total neoadjuvant therapy in advanced rectal cancer

Abstract Background Predicting clinical complete response (cCR) to total neoadjuvant therapy (TNT) helps optimize therapeutic strategies and identify the best candidates for a watch-and-wait strategy. This study aimed to develop a machine learning model to predict cCR to TNT in a mixed curative-intent cohort of patients with advanced rectal cancer, including selected oligometastatic cases treated with curative intent. Methods A retrospective analysis of a prospectively maintained database of consecutive rectal cancer patients treated with TNT with curative intent at three South Australian hospitals between 2019 and 2026 was performed . Least absolute shrinkage and selection operator (LASSO) analysis was used for variable selection, and ten machine learning algorithms were employed to develop prediction models for cCR. Model performance was evaluated using the area under the receiver-operating characteristic curve (AUROC). Results A total of 245 patients were identified, of whom 193 were assigned to the training set and 52 to the external validation set. LASSO identified albumin, clinical T stage, tumour length, systemic immune-inflammation index, extramural vascular invasion, platelet count, and body mass index as key predictors of cCR. Among the ten models, the logistic regression model performed best in the training set, achieving an AUROC of 0.817. In external validation, the multilayer perceptron (MLP) model achieved the highest predictive performance with an AUROC of 0.755. Conclusion Machine learning models based on seven pretreatment variables show promise in accurately predicting cCR following TNT in patients with advanced rectal cancer. On external validation, the MLP model demonstrated the strongest predictive performance with moderate discrimination. However, further validation in larger, non-metastatic/selected populations may be required before clinical implementation.