A novel generative multi-task representation learning approach for predicting postoperative complications in cardiac...

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Computational Biology and Bioinformatics

Background:

• Early detection of surgical complications is crucial for timely intervention and risk management.
• Machine learning (ML) offers potential for identifying and predicting patient risks of postoperative complications.

Purpose of the Study:

• To develop and validate a novel surgical Variational Autoencoder (surgVAE) for predicting postoperative complications.
• To uncover intrinsic patterns in patient data using cross-task and cross-cohort presentation learning.

Main Methods:

• Retrospective cohort study using 4 years of electronic health records (2018-2021).
• Assessed six key postoperative complications in cardiac surgery.
• Compared surgVAE performance against established ML models using 5-fold cross-validation.

Main Results:

• Included 89,246 surgeries, with 6,502 in the cardiac surgery cohort.
• surgVAE outperformed existing ML models, achieving higher macro-averaged AUPRC (0.409) and AUROC (0.831).
• Integrated Gradients identified key preoperative risk factors.

Conclusions:

• The surgVAE framework demonstrated superior discriminatory performance in predicting postoperative complications.
• surgVAE effectively addresses challenges like data complexity and low-frequency events.
• The model provides data-driven predictions and enhances interpretability of patient risk profiles.