Background
Low anterior resection syndrome (LARS) is a common postoperative complication in patients with colorectal cancer, which seriously affects their postoperative quality of life. At present, the aetiology of LARS is still unclear, but some risk factors have been studied. Accurate prediction and early management of medical intervention are keys to improving the quality of life of such high‐risk patients.
Objectives
Based on machine‐learning methods, this study used the follow‐up results of postoperative patients with colorectal cancer to develop prediction models for LARS and conducted a comparative analysis between the different models.
Methods
A total of 382 patients diagnosed with colorectal cancer and undergoing surgery at West China Hospital from April 2017 to December 2020 were retrospectively selected as the development cohort. Logistic regression, support vector machine, decision tree, random forest and artificial neural network algorithms were used to construct the prediction models of the obtained dataset. The models were internally validated using cross‐validation. The area under the curve and Brier score measures were used to evaluate and compare the differentiation and calibration degrees of the models. The sensitivity, specificity, positive predictive value and negative predictive value of the different models were described for clinical use.
Results
A total of 342 patients were included, the incidence of LARS being 47.4% (162/342) during the six‐month follow‐up. After feature selection, the factors influencing the occurrence of LARS were found to be location, distance, diverting stoma, exsufflation and surgical type. The prediction models based on five machine‐learning methods all showed acceptable performance.
Conclusions
The five models developed based on the machine‐learning methods showed good prediction performance. However, considering the simplicity of clinical use of the model results, the logistic regression model is most recommended. The clinical applicability of these models will also need to be evaluated with external cohort data.