Efficacy of Different ML Models in Predicting Post-Surgical Complications in Patients

Abstract

Post-surgical complications significantly affect patient recovery and healthcare outcomes, making early and accurate prediction critical for timely intervention. This study evaluates the performance of eight supervised machine learning models—Artificial Neural Networks (ANN), Deep Neural Networks (DNN), Random Forest, Support Vector Machines (SVM), Logistic Regression, XGBoost, CatBoost, and ensemble learning techniques—for predicting post-surgical complications. A clinical dataset comprising surgical characteristics and post-procedural outcomes was used, incorporating key features such as DRG family, discharge volume, and complication indicators. Models were trained and tested using a 70–30 split, with one-hot encoding and class-weight adjustments to address data imbalance. Performance was assessed using accuracy, precision, recall, F-beta score, and Cohen’s Kappa. The results indicate that CatBoost achieved the strongest overall performance, with the highest accuracy, recall, and F-beta score. XGBoost and ensemble models also performed well, particularly in identifying high-risk cases, while neural network models demonstrated comparatively lower performance. These findings highlight the potential of tree-based and ensemble learning approaches, especially CatBoost, for developing reliable clinical decision support systems in postoperative care.