Development and Validation of Machine Learning Models for Adverse Events after Cardiac Surgery

Area of Science:

• Cardiovascular Surgery
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Early recognition of adverse events post-cardiac surgery is critical for effective treatment.
• The current Society of Thoracic Surgery (STS) risk model has limitations in predicting adverse events and its applicability is restricted to less than 80% of cardiac surgeries.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting postoperative outcomes in cardiac surgery patients.
• To compare the performance of the developed ML model against the established STS risk model.

Main Methods:

• Machine learning models, termed the Roux-MMC model, were developed using a retrospective cohort (n=9,841) from the STS Adult Cardiac Surgery Database (ACSD) (2012-2021).
• The Roux-MMC model was further validated on a prospective cohort (n=2,305) (2022-2024).
• Model performance was evaluated by comparing the area under the receiver-operating curve (AUROC) for predicting eight key postoperative outcomes against the STS model.

Main Results:

• The Roux-MMC model demonstrated superior performance compared to the STS model across all eight predicted postoperative outcomes in the prospective cohort.
• Specific AUROC values for the Roux-MMC model in the prospective cohort ranged from 0.818 for short length of stay (SLOS) to 0.911 for prolonged ventilation.
• The Roux-MMC model demonstrated wider applicability, covering all cardiac surgery patients, unlike the STS model which applied to only 65-77% of patients.

Conclusions:

• The developed Roux-MMC machine learning model effectively predicts eight adverse postoperative outcomes in cardiac surgery patients.
• The Roux-MMC model significantly outperforms the STS model and is applicable to all cardiac surgery patients.
• Given its development on the STS ACSD, the Roux-MMC model has the potential for widespread implementation in hospitals nationwide.