A Comparison of Machine Learning-Based Models and a Simple Clinical Bedside Tool to Predict Morbidity and Mortality After Gastrointestinal Cancer Surgery in the Elderly

Frailty in the elderly population is associated with increased vulnerability to stressors, including surgical interventions. This study compared machine learning (ML) models with a clinical bedside tool, the Gastrointestinal Surgery Frailty Index (GiS-FI), for predicting mortality and morbidity in elderly patients undergoing gastrointestinal cancer surgery. In a multicenter analysis of 937 patients aged ≥65 years, the performance of various predictive models including Random Forest (RF), Least Absolute Shrinkage and Selection Operator (LASSO), Stepwise Regression, K-Nearest Neighbors, Neural Network, and Support Vector Machine algorithms were evaluated. The overall 30-day mortality and morbidity rates were 6.1% and 35.7%, respectively. For mortality prediction, the RF model demonstrated superior performance with an AUC of 0.822 (95% CI 0.714-0.931), outperforming the GiS-FI score (AUC = 0.772, 95% CI 0.675-0.868). For morbidity prediction, all models showed more modest discrimination, with stepwise regression and LASSO regression achieving the highest performance (AUCs of 0.652 and 0.647, respectively). Our findings suggest that ML approaches, particularly RF algorithm, offer enhanced predictive accuracy compared to traditional clinical scores for mortality risk assessment in elderly cancer patients undergoing gastrointestinal surgery. These advanced analytical tools could provide valuable decision support for surgical risk stratification in this vulnerable population.