A Comparative Study of Machine Learning Algorithms in Predicting Severe Complications after Bariatric Surgery.

Area of Science:

• Medical Informatics
• Surgical Outcomes Research
• Machine Learning in Healthcare

Background:

• Severe obesity presents a growing global public health challenge.
• Predictive models for severe postoperative complications are crucial for bariatric surgery candidates.
• Traditional statistical methods have shown limitations in accuracy for this prediction task.

Purpose of the Study:

• To identify a machine learning (ML) algorithm capable of accurately predicting severe complications post-bariatric surgery.
• To compare the performance of various supervised ML algorithms for this specific clinical application.

Main Methods:

• Trained and compared 29 supervised ML algorithms on a large dataset (37,811 patients, 2010-2014).
• Tested algorithms on a separate cohort (6250 patients, 2015).
• Employed synthetic minority oversampling technique to address data imbalance (3% severe complications).

Main Results:

• Most ML algorithms achieved high accuracy (>90%) and specificity (>90%) in both training and testing datasets.
• No algorithm demonstrated acceptable sensitivity in the test data.
• Deep neural networks (NN) showed a minor but perceptible improvement in performance.

Conclusions:

• Ensemble algorithms outperformed base algorithms in predicting severe postoperative complications.
• Deep neural networks show promise for improving prediction accuracy and warrant further investigation.
• Oversampling techniques are valuable for handling imbalanced datasets in clinical outcome prediction.