Pay Your Dues Members Only Area
Central Surgical Association

49th Annual Meeting

Back to 2021 Abstracts

Parsimonious Machine-Learning Models to Predict Cardiac Surgery Resource Utilization Across a Statewide Collaborative
Arjun Verma, Yas Sanaiha, Zachary Tran, Joseph Hadaya, Richard Shemin, Peyman Benharash
University of California Los Angeles, Los Angeles, California, United States

Rapid and accurate prediction of hospital capacity are paramount to forecasting the feasibility of operations during periods of surge such as the COVID-19 pandemic. While the Society of Thoracic Surgeons risk models provide canonical estimates for parameters such as mortality, they require many data fields and do not account for institutional variations in practice. Therefore, we employed several machine learning algorithms to develop predictive models for mortality and resource use in a cohort of cardiac surgical patients across an academic, state-wide collaborative. We hypothesized machine-learning models with inclusion of a sparse variable set would yield superior discrimination and calibration compared to traditional linear models.

The 2013-19 University of California Cardiac Surgery Consortium repository of records from 2013-2019 was used to identify all adults undergoing isolated coronary artery bypass (CABG), or isolated valve, CABG/valve and multi-valve operations. Patients requiring mechanical circulatory support/ECMO and those with ICU and overall length of stay (LOS) >7 days and 21 days, respectively, were excluded. The primary outcome was model performance as measured by receiver operating characteristics (AUC) and coefficient of determination (R2). Gradient boosting (GBM), random forest (RF) and extreme gradient boosting (XGBoost) were compared to traditional logistic and linear regressions. Seven clinical covariates with the highest feature importance were selected for inclusion in all subsequent models, in addition to operative type and center. We used a 75:25 train:test split with 10-fold cross validation for model development.

Across five participating centers, 8,747 patients met study criteria: 43.2% isolated CABG, 36.2% isolated valve, 13.2% CABG/valve, and 7.4% multi-valve operations. We noted a 30-day mortality rate of 2.5% and a re-operation rate of 9.1%. Blood products were used in 39.3% of patients while the median LOS was 8 (IQR: 6-12) days, while median ICU LOS was 72 (IQR: 46.5-102) hours. A RF model was developed to select important features based on 30-day mortality with 35 candidate covariates (Figure 1). Machine learning algorithms exhibited modest improvement in AUC and R2, compared to traditional regression techniques (Table 1). We found machine learning models to have superior sensitivity in classification of 30-day mortality, although similar calibration was noted for length of stay prediction (Figure 2).

We successfully employed machine learning algorithms to create simple, yet customizable predictive models for mortality and several measures of resource use. Given model performance, such methodology can be readily deployed to create institution-specific models.

Figure 1: Feature importance as indicated by Random Forest Classifier for prediction of 30-day Mortality (C-Statistic: 0.78). Explanatory variables in red were included for all subsequent model development. INR: International Normalized Ratio; CABG: Coronary Artery Bypass Grafting; MI: Myocardial Infarction; PVD: Peripheral Vascular Disorder; CVD: Cardiovascular Disease

Table 1: Comparison of model performance using the C-Statistic (AUC) and coefficient of determination (R2) for discrete and continuous outcomes, respectively. GBM: Gradient Boosting Machine; RF: Random Forest; XGBoost: Extreme Gradient Boosting. *Indicates best performance.

Figure 2: Comparison of Gradient Boosting Machine and Logistic Regression Discrimination for 30-Day Mortality. GBM: Gradient Boosting Machine
Back to 2021 Abstracts