Development and Evaluation of Machine Learning Models to Predict the Risk of Major Cardiac Events and Death for People With Kidney Failure Having Non-Cardiac Surgery.

Abstract

RATIONALE & OBJECTIVE People with kidney failure undergoing non-cardiac surgery face an elevated risk of cardiovascular events and mortality. Existing risk prediction tools for perioperative events are either inaccurate in this population or include many variables that may complicate implementation. We developed and evaluated the performance of simplified machine-learning models for major cardiac events and mortality within 30 days after non-cardiac surgery in patients with kidney failure in Alberta and Manitoba, Canada. STUDY DESIGN Data from Manitoba was split into training (70%), validation (15%), and testing (15%) sets. The training set was used for hyperparameter tuning and model training, the validation set for feature selection and evaluating model performance, and the testing set for final model performance. External evaluation was performed in a cohort from Alberta. SETTING & PARTICIPANTS We included Manitoban adults (≥ 18 years) with kidney failure (eGFR < 15 mL/min/1.73m2 or on maintenance dialysis) undergoing non-cardiac surgery (2007-2019), with evaluation data on adults from Alberta (2005-2019). PREDICTORS Variables included sex, age, surgery type and setting, kidney failure type, chronic conditions, and preoperative laboratory values (albumin and hemoglobin). OUTCOME Composite of acute myocardial infarction, cardiac arrest, ventricular arrhythmia, and all-cause mortality within 30 days of surgery. ANALYTICAL APPROACH Model performance was evaluated using Area Under the Receiver Operating Characteristic Curve (AUC-ROC), Area Under the Precision-Recall Curve (AUC-PR), and calibration. The final XGBoost and random forest models were externally validated using Alberta data. RESULTS Among 4,175 participants (12,082 surgeries), 569 outcomes (5%) were observed. The parsimonious XGBoost model (8 features) showed an AUC-ROC of 0.861 and AUC-PR of 0.304, and the parsimonious random forest model (20 features) estimated an AUC-ROC of 0.863 and AUC-PR of 0.332 in the testing cohort. External validation in Alberta showed similar performance with good calibration. LIMITATIONS Lack of external validation outside Canada. CONCLUSION Our machine learning models were accurate and had improved parsimony over existing regression-based tools. Future work should test these models in other populations and compare them with regression-based models, in addition to assessing the value of these tools for informing risk-guided perioperative care.