Prediction of postoperative complications after oesophagectomy using machine-learning methods

Area of Science:

• Surgical Oncology
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Oesophagectomy carries a high risk of postoperative complications.
• Accurate prediction of these complications is crucial for patient management.
• Existing risk scores may not fully capture patient-specific risks.

Purpose of the Study:

• To apply and compare various machine learning algorithms for predicting postoperative complications (Clavien-Dindo grade IIIa or higher) after oesophagectomy.
• To evaluate the performance of machine learning models against a current risk score.

Main Methods:

• Retrospective analysis of patients undergoing Ivor Lewis oesophagectomy for oesophageal or gastro-oesophageal junction cancer (2016-2021).
• Comparison of logistic regression, random forest, k-nearest neighbour, support vector machine, and neural network algorithms.
• Performance evaluation using accuracy and area under the curve, compared to the Cologne risk score.

Main Results:

• The neural network achieved the highest overall accuracy (0.688) for predicting Clavien-Dindo grade IIIa or higher complications.
• Neural network also showed superior performance in predicting medical (0.692) and surgical (0.667) complications.
• The neural network's area under the curve was 0.672 for severe complications, 0.695 for medical, and 0.653 for surgical complications.

Conclusions:

• Machine learning, particularly neural networks, offers a powerful approach for predicting postoperative complications following oesophagectomy.
• The developed neural network model outperformed other tested algorithms and the Cologne risk score.
• These findings suggest potential for improved risk stratification and personalized patient care.