Development and Validation of an Interpretable Machine Learning Model to Predict Major Adverse Cardiovascular Events After Noncardiac Surgery in Geriatric Patients: a Prospective Study

Jiayu Yu; Xiran Peng; Ruihao Zhou; Tao Zhu; Xuechao Hao

doi:10.1097/JS9.0000000000002203

Development and Validation of an Interpretable Machine Learning Model to Predict Major Adverse Cardiovascular Events After Noncardiac Surgery in Geriatric Patients: a Prospective Study

Int J Surg. 2024 Dec 26. doi: 10.1097/JS9.0000000000002203. Online ahead of print.

Authors

Jiayu Yu¹, Xiran Peng^{1

2}, Ruihao Zhou^{1

2}, Tao Zhu^{1

2}, Xuechao Hao^{1

2}

Affiliations

¹ Department of Anesthesiology, National Clinical Research Center for Geriatrics, West China Hospital, Sichuan University, Chengdu, China.
² Research Unit for Perioperative Stress Assessment and Clinical Decision, Chinese Academy of Medical Sciences West China Hospital, Sichuan University, Chengdu, China.

PMID: 39728602
DOI: 10.1097/JS9.0000000000002203

Abstract

Background: Major adverse cardiovascular events (MACEs) within 30 days following noncardiac surgery are prognostically relevant. Accurate prediction of risk and modifiable risk factors for postoperative MACEs is critical for surgical planning and patient outcomes. We aimed to develop and validate an accurate and easy-to-use machine learning model for predicting postoperative MACEs in geriatric patients undergoing noncardiac surgery.

Materials and methods: The cohort study was conducted at an academic medical center between June 2019 and February 2023. The outcome was postoperative MACEs within 30 days after surgery. Significant predictors were selected using permutation-shuffling. Ten machine learning models were established and compared with the Revised Cardiac Risk Index (RCRI). The SHapley Additive exPlanations algorithm was used to interpret the models.

Results: Of the 18 395 patients included, 354 (1.92%) experienced postoperative MACEs. Eighteen predictors were included in model development. The AutoGluon model outperformed other models and the RCRI with an AUROC of 0.884 (95% CI: 0.878-0.890), accuracy of 0.976 (95% CI: 0.973-0.978), and Brier Score of 0.023 (95% CI: 0.020-0.026). In interpretability analyses, the hemoglobin level was the most important predictor. We identified the relationships between predictors and postoperative MACEs and interaction effects between some predictors. The AutoGluon model has been deployed as a web-based tool for further external validation (https://huggingface.co/spaces/MDC2J/Predicting_postoperative_MACEs).

Conclusion: In this prospective study, the AutoGluon model could accurately predict MACEs after noncardiac surgery in geriatric patients, outperforming existing models and the RCRI. Subsequent interpretability analysis can provide insight into how our model works and help personalize surgical strategies.