Unraveling variations and enhancing prediction of successful sphincter-preserving resection for low rectal cancer: a post hoc analysis of the multicentre LASRE randomized clinical trial

Abstract

Background: Accurate prediction of successful sphincter-preserving resection (SSPR) for low rectal cancer enables peer institutions to scrutinize their own performance and potentially avoid unnecessary permanent colostomy. The aim of this study is to evaluate the variation in SSPR and present the first artificial intelligence (AI) models to predict SSPR in low rectal cancer patients. Study design: This was a retrospective post hoc analysis of a multicenter, noninferiority randomized clinical trial (LASRE, NCT XXXXXX) conducted in 22 tertiary hospitals across China. A total of 604 patients who underwent neoadjuvant chemoradiotherapy (CRT) followed by radical resection of low rectal cancer were included as the study cohort, which was then split into a training set (67%) and a testing set (33%). The primary end point of this post hoc analysis was SSPR, which was defined as meeting all the following criteria: (1) sphincter-preserving resection; (2) complete or nearly complete TME, (3) a clear CRM (distance between margin and tumor of 1 mm or more), and (4) a clear DRM (distance between margin and tumor of 1 mm or more). Seven AI algorithms, namely, support vector machine (SVM), logistic regression (LR), extreme gradient boosting (XGB), light gradient boosting (LGB), decision tree classifier (DTC), random forest (RF) classifier, and multilayer perceptron (MLP), were employed to construct predictive models for SSPR. Evaluation of accuracy in the independent testing set included measures of discrimination, calibration, and clinical applicability. Results: The SSPR rate for the entire cohort was 71.9% (434/604 patients). Significant variation in the rate of SSPR, ranging from 37.7% to 94.4%, was observed among the hospitals. The optimal set of selected features included tumor distance from the anal verge before and after CRT, the occurrence of clinical T downstaging, post-CRT weight and clinical N stage measured by magnetic resonance imaging. The 7 different AI algorithms were developed and applied to the independent testing set. The LR, LGB, MLP and XGB models showed excellent discrimination with AUROC values of 0.825, 0.819, 0.819 and 0.805, respectively. The DTC, RF and SVM models had acceptable discrimination with AUROC values of 0.797, 0.766 and 0.744, respectively. LR and LGB showed the best discrimination, and all 7 AI models had superior overall net benefits within the range of 0.3-0.8 threshold probabilities. Finally, we developed an online calculator based on the LGB model to facilitate clinical use. Conclusions: The rate of SSPR exhibits substantial variation, and the application of AI models has demonstrated the ability to predict SSPR for low rectal cancers with commendable accuracy.