Machine Learning Approach Using Routine Immediate Postoperative Laboratory Values for Predicting Postoperative Mortality

Abstract

The development of new surgical instruments and techniques for surgical treatment has facilitated an increase in the number of patients undergoing surgery. As the proportion of elderly patients and patients with delayed recovery also increases, reducing the additional burden of clinical outcomes after surgery is a new challenge for perioperative management. While numerous models have been created to predict mortality and to stratify risks in the preoperative period, few studies evaluated postoperative risk factors that significantly affects survival after surgery. This study aimed to develop prediction models using routine immediate postoperative laboratory values for predicting postoperative mortality. The databases of two tertiary hospitals were used in the current study: one to develop model and the other for external validation of the final models. LASSO logistic regression, random forest, deep neural network, and XGBoost algorhythms were utilized to create model. Following creation of the models based on the laboratory values of postoperative blood tests, their efficiency were demonstrated through comparison using SASA scoring system. There was a total of 3,817 patients with postoperative blood test values. All the developed models outperformed the SASA model. Furthermore, the best AUROC of 0.82 and AUPRC of 0.13 was obtained by the random forest model, and the phosphorus level was the most contributing factor to the random forest model. The machine learning models that were trained on postoperative laboratory values outperformed the previous approaches in regards to predicting postoperative 30-day mortality, which indicates that the current model may be a useful tool to identify patients who are at increased risk of postoperative mortality.