Development and validation of machine-learning model based on dynamic tumor markers in predicting pathological complete response after neoadjuvant chemoradiotherapy in patients with locally advanced rectal cancer: a multicenter cohort study.

Abstract

Objective: In this study, we constructed a new pCR predictor based on dynamic tumor marker changes before and after NCRT, the dynamic tumor marker score (DTMS), and combined it with other clinicopathological features to build a machine-learning model.

Methods: In this retrospective study of patients with LARC between September 2010 and October 2017 at The First Affiliated Hospital of Fujian Medical University (FJMUFAH), Fujian Medical University Union Hospital (FJMUUH), and Fujian Provincial Hospital (FJPH), the DTMS predictor was constructed using logistic regression. Factors associated with pCR were screened using single-factor and multifactorial logistic regression, and 10 machine-learning algorithms were used to construct a pCR prediction model. Additionally, various metrics, including the area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPRC), decision curve analysis, and calibration curves, were obtained to validate the model performance and verified using an external validation set. Finally, SHapley Additive exPlanations (SHAP) values were used to interpret the predictive model. Moreover, we developed a website to facilitate the use of prediction modeling.

Results: After analyzing the data of 892 patients with LARC from FJMUFAH, DTMS, tumor size, N stage, and tumor distance from the anal verge were identified as independent predictive factors for pCR using univariate and multivariate regression analyses. The "extreme gradient boosting" (XGB) model displayed the best performance in the training set, with a mean AUC value of 0.86, an AUPRC value of 0.732, and SHAP values utilized in the analysis. In the two external validation sets, the model yielded AUC values of 0.80 and 0.82, along with corresponding AUPRC values of 0.519 and 0.593, respectively, which were the highest among all ten evaluated models, incorporating the use of SHAP values in the analysis. The model maintained superior predictive efficacy in the external validation cohorts (FJMUUH and FJPH).

Conclusions: As a novel marker based on dynamic changes in CEA and CA19-9 levels, DTMS effectively predicted pCR within the XGB model, providing clinicians with a practical tool for treatment decision-making regarding LARC.