Which Method Best Predicts Postoperative Complications: Deep Learning, Machine Learning, or Conventional Logistic Regression?

Accurate prediction of postoperative complications is critical in surgical care. Recently, deep learning has gained attention and has been applied to various predictive models and image recognition tasks, and researchers are attempting to apply this technology in medicine. This review compares logistic regression, machine learning, and deep learning models used in gastroenterological surgery. There are some studies reporting predictive models with large databases. Among these studies, some studies showed that deep learning outperformed other models, but others reported random forests or gradient boosting methods, a type of machine learning, performed better than the other methods including deep learning. On the other hand, applying image or time-series data are reported to increase the prediction accuracy of postoperative morbidity/mortality, despite relatively small sample sizes. While deep learning shows potential, especially with image and time-series data, it often underperforms on tabular clinical datasets such as current National Clinical Database (NCD). We discuss the limitations of deep learning in term of its "black-box" nature and highlight the need for integrating complex data types to improve model accuracy and interpretability. Incorporating multimodal inputs may enable deep learning to outperform conventional methods and better support clinical decision-making.