Use of supervised machine learning algorithms in predicting postoperative mortality in gastrointestinal and HPB surgeries.

Aim of the studyThis study aims to evaluate supervised machine learning algorithms in predicting 90 days post-operative mortality in gastrointestinal and HPB surgeries and comparing them with standard logistic regression methods. MethodsWe evaluated various supervised machine learning classification algorithms like gradient boosting, K-nearest neighbours, random forest, and support vector machines with standard logistic regression methods. We used accuracy and the Receiver operating curve to compare the methods. 60% of the data were used for training, 20% for validation and 20% for testing. We used JASP 0.16.04 by the University of Amsterdam to run machine learning algorithms and statistical analysis. ResultsWe used data from 504 patients who have undergone gastrointestinal and hepatopancreatic biliary surgery between April 2016 and March 2023. We analyzed algorithms for predicting 90 days post-operative mortality based on features like Major surgeries, Surgeries for malignancies, age, CDC grade of surgeries, Intraoperative hypotension, Open vs Laparoscopic surgeries, ASA grade, Emergency surgeries, Operative time, Intraoperative blood product used, colorectal surgeries, small intestinal surgeries, HPB surgeries, upper gastrointestinal surgeries and hernia. Test accuracies were 96% for gradient boosting, 90 % for K-nearest neighbours, 96% for the random forest, 94% for support vector and Areas under the ROC curve were 0.802 for gradient boosting, 0.489 for K-nearest neighbours, 0.934 for random forest and 0.5 for support vector algorithms. Accuracy and Area under the ROC curve with standard logistic regression method were 94% and 0.757. Features of importance in decreasing order were ASA, operative times, blood products, small bowel surgeries and Age. ConclusionSupervised machine learning algorithms particularly gradient boosting and random forest predicted 90 days post-operative mortality more accurately than logistic regression and such models can be part of the preoperative evaluation in gastrointestinal and HPB surgeries.