Early Postoperative Prediction of Complications and Readmission After Colorectal Cancer Surgery Using an Artificial Neural Network

BACKGROUND: Early predictors of postoperative complications can risk-stratify patients undergoing colorectal cancer surgery. However, conventional regression models have limited power to identify complex nonlinear relationships among a large set of variables. We developed artificial neural network models to optimize the prediction of major postoperative complications and risk of readmission in patients undergoing colorectal cancer surgery. OBJECTIVE: This study aimed to develop an artificial neural network model to predict postoperative complications using postoperative laboratory values and compare the accuracy of models to standard regression methods. DESIGN: This retrospective study included patients who underwent elective colorectal cancer resection between January 1, 2016, and July 31, 2021. Clinical data, cancer stage, and laboratory data from postoperative days 1 to 3 were collected. Complications and readmission risk models were created using multivariable logistic regression and single-layer neural networks. SETTING: National Cancer Institute-Designated Comprehensive Cancer Center. PATIENTS: Adult patients with colorectal cancer. MAIN OUTCOME MEASURES: The accuracy of predicting postoperative major complications, readmissions, and anastomotic leaks using the area under the receiver operating characteristic curve. RESULTS: Neural networks had larger areas under the curve for predicting major complications compared to regression models (neural network 0.811; regression model 0.724, p < 0.001). Neural networks also showed an advantage in predicting anastomotic leak (p = 0.036) and readmission using postoperative day 1 to 2 values (p = 0.014). LIMITATIONS: Single-center, retrospective design limited to cancer operations. CONCLUSIONS: In this study, we generated a set of models for the early prediction of complications after colorectal surgery. The neural network models provided greater discrimination than the models based on traditional logistic regression. These models may allow for early detection of postoperative complications as early as postoperative day 2.