A machine learning-based model for predicting survival in patients with Rectosigmoid Cancer

TL;DR

This study creates a machine learning model to predict survival in rectosigmoid cancer patients, using clinical factors to help improve clinical decisions.

Contribution

A novel XGBoost-based survival prediction model for rectosigmoid junction cancer patients is developed and validated.

Findings

Seven independent risk factors for survival were identified, including age, gender, diabetes, and tumor characteristics.

The XGBoost model outperformed other machine learning models with AUCs of 0.7856, 0.8484, and 0.796 at 1, 3, and 5 years.

The model showed the lowest Brier scores and best clinical decision benefits via decision curve analysis.

Abstract

The unique anatomical characteristics and blood supply of the rectosigmoid junction confer particular significance to its physiological functions and clinical surgeries. However, research on the prognosis of rectosigmoid junction cancer (RSC) is scarce, and reliable clinical prediction models are lacking.

This retrospective study included 524 patients diagnosed with RSC who were admitted to the Department of Gastrointestinal and Colorectal Surgery at the First Hospital of Jilin University between January 1, 2017, and June 1, 2019. Univariate and multivariate Cox regression analyses were conducted in this study to identify independent risk factors impacting the survival of RSC patients. Subsequently, models were constructed using six different machine learning algorithms. Finally, the discrimination, calibration, and clinical applicability of each model were evaluated to determine the optimal model.

Through univariate and multivariate Cox regression analyses, we identified seven independent risk factors associated with the survival of RSC patients: age (HR = 1.9, 95% CI: 1.3-2.8, P = 0.001), gender (HR = 0.6, 95% CI: 0.4-0.9, P = 0.013), diabetes (HR = 2.0, 95% CI: 1.3-3.1, P = 0.002), tumor differentiation (HR = 2.1, 95% CI: 1.4-3.1, P < 0.001), tumor N stage (HR = 2.02, 95% CI: 1.2-3.4, P = 0.009), distant metastasis (HR = 4.2, 95% CI: 2.7-6.7, P < 0.001), and anastomotic leakage (HR = 2.4, 95% CI: 1.1-5.3, P = 0.034). After evaluating each model, the prediction model based on XGBoost was determined to be the optimal model, with AUC of 0.7856, 0.8484, and 0.796 at 1, 3, and 5 years. It also had the lowest Brier scores at all time points, and decision curve analysis (DCA) demonstrated the best clinical decision benefits compared to other models.

We developed a prediction model based on the optimal machine learning, XGBoost, which can assist clinical decision-making and potentially extend the survival of patients with rectosigmoid junction cancer.

Figures