Logistic Regression and Machine Learning Algorithms for the Risk Prediction of Perioperative Adverse Cardiovascular Events in Elderly Patients

Abstract

Objectives: Elderly patients with perioperative adverse cardiovascular events (PACEs) may have poorer prognosis and higher mortality. Early identification of patients at risk of developing PACEs is an essential step in preventing and controlling PACEs. To develop and validate models to predict the likelihood of PACEs and to clarify the specific classification of PACEs with different risk stratification for the elderly patients during noncardiac surgery by integrating clinical data, biomarkers, and established risk factors. Most importantly, they help to support clinical decision making and improve patient prognosis. Methods: This retrospective study included elderly in-patients undergoing noncardiac surgery at six hospitals in Chongqing City, China, from March 2020 to July 2021. Logistic regression and machine learning algorithms were used to construct models, which were evaluated by receiver operating characteristic curves, decision curve, calibration curve, sensitivity, specificity, and F1-score were used to interpret the model results. The diagnostic criteria for PACEs encompassed delirium, major adverse cardiovascular events, myocardial injury after noncardiac surgery, new perioperative atrial fibrillation, perioperative acute heart failure, pain and infection, and other cardiovascular events that pose a threat to perioperative safety and influence patient prognosis. PACEs were identified by expert anesthesiologists in accordance with ASA classification on the day preceding surgery. Results: Of the 8309 patients included in the analysis, 1805 were suspected of having PACEs. The logistic regression model was chosen with the area under the curve for 0.895 (95% CI: 0.881–0.908). Pro-BNP, cardiac function grading, and creatinine (Cre) were the most related factors for PACEs, and a new composite indicator PCC was developed by combining the initials of these three indicators in the logistic regression model. The decision curve and calibration curve indicated that this indicator had respectable clinical value. In addition, a machine learning model was built to accurately predict PACEs of different risk stratification in the elderly. Precision-recall curves of the prediction showed low-risk precision was 0.86, and medium-risk precision was 0.870, and high-risk precision was 0.970. The F1 values are all greater than 0.850, especially for the high risk, the prediction effect reaches 0.970. The sensitivity and specificity of the prediction model were 0.736 and 0.973, respectively, indicating that it had the best predictive performance of risk stratification for PACEs. Therefore, we can reasonably assume that our two models can effectively predict the risk of PACEs during noncardiac surgery. This study demonstrated the ability to accurately identify high-risk PACEs patients using an interpretable approach. Conclusion: This study established risk prediction models for patients with PACEs, based on the patients' medical records, with good predictive accuracy. This model is expected to provide a scientific basis for quickly formulating or adjusting the diagnostic and treatment plans for elderly patients and provide clinical strategies for PACEs prevention, intervention, and monitoring, which could potentially reduce the mortality risk of patients with PACEs.