Development of a Machine Learning-Based Nutrition-Related Surgical Risk Assessment Model for Older Patients with Gastrointestinal Malignancies.

Abstract

Older patients with gastrointestinal cancer are at a high risk of postoperative complications; however, no accurate preoperative assessment is available. This study developed a prognostic model that leveraged machine learning and multidimensional clinical data to predict postoperative complications in older patients. This study assessed 365 older patients with gastrointestinal cancer who underwent radical surgery at Beijing Hospital. Patients were randomly allocated to training and test sets (7:3 ratio). Multiplex machine learning was used for feature selection and model development. The efficacies of the models were assessed using receiver operating characteristic curves. An imbalance rfsrc + ranger model (IRM) was created using the "shiny" R package. All statistical analyses were performed using R software. The overall rate of postoperative complications was 19.2%. IRM was the most accurate among the 361 models developed using 19 machine learning algorithms and 19 sets of clinical features. Body mass index was the most important variable for predicting postoperative complications in these patients, followed by hemoglobin level, albumin level, and surgical approach. This study developed a nutrition-related surgical risk assessment model that includes malnutrition, comorbidities, and surgical approaches to improve the outcome of older patients with gastrointestinal malignancies, aiding in managing preoperative risk factors and improving surgical safety.