Kidney Week

Abstract: TH-PO003

Development and Validation of Machine-Learning Model to Predict the Risk of Major Cardiovascular Events and Death for Patients with Kidney Failure Having Noncardiac Surgery

Session Information

• Augmented Intelligence for Prediction and Image Analysis
  October 24, 2024 | Location: Exhibit Hall, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Augmented Intelligence, Digital Health, and Data Science

• 300 Augmented Intelligence, Digital Health, and Data Science

Authors

• Pabla, Gurpreet S., University of Manitoba Faculty of Health Sciences, Winnipeg, Manitoba, Canada

Gurpreet S. Pabla, MS, MSc

• Tangri, Navdeep, University of Manitoba Max Rady College of Medicine, Winnipeg, Manitoba, Canada

Navdeep Tangri, MD, PhD

• Harrison, Tyrone, University of Calgary Cumming School of Medicine, Calgary, Alberta, Canada

Tyrone Harrison, MD, PhD

• Ferguson, Thomas W., Seven Oaks General Hospital, Winnipeg, Manitoba, Canada

Thomas W. Ferguson, MS, MSc

• Sevinc, Emir, University of Calgary Cumming School of Medicine, Calgary, Alberta, Canada

Emir Sevinc, MSc

• Whitlock, Reid, Seven Oaks General Hospital, Winnipeg, Manitoba, Canada

Reid Whitlock, MSc

Background

Patients with kidney failure undergoing non-cardiac surgery face significantly higher risk of adverse cardiovascular events and mortality compared to those with normal kidney function. Existing risk prediction tools are limited in estimating these risks for kidney failure patients. We developed and validated a machine-learning model for major cardiovascular events and mortality in kidney failure patients within 30 days of undergoing outpatient or inpatient non-cardiac surgery in Alberta and Manitoba, Canada.

Methods

Derivation data was sourced from Manitoba Health, including adults (≥ 18 years) with kidney failure (eGFR < 15 mL/min/1.73m² or on maintenance dialysis) undergoing non-cardiac surgery between April 1, 2007, and December 31, 2019. We focused on a composite outcome of acute myocardial infarction, cardiac arrest, ventricular arrhythmia, and all-cause mortality. Data was split into 70% for training, 15% for validation, and 15% for testing. The training set was used to tune the hyperparameters and train the models; the validation dataset was used for feature selection and evaluate model performance, while the testing set evaluated the model’s final performance. The model's performance was evaluated using C-statistics, Area Under the Precision-Recall Curve (AUC-PR), calibration plots, and Brier Score. We used XGBoost and Random Forest, selecting a model with reasonable and balanced C-statistics and AUC-PR. The final model was externally tested using Alberta data.

Results

We identified 12,082 surgeries and 569 outcomes. The final model (XGBoost) included surgery type, surgery setting (emergency inpatient, outpatient), history of myocardial infarction, albumin, and hemoglobin levels. It had an estimated C-statistic of 0.86, an AUC-PR of 0.30, and a Brier score of 0.04 in the testing cohort. External testing in Alberta showed similar performance. Calibration plots demonstrated excellent calibration, except for underestimation at the highest predicted risks.

Conclusion

Our XGBoost model for adverse peri-operative outcomes in patients with kidney failure demonstrated good performance, with improved parsimony compared to existing tools. Future work should compare these tools and test the impact of risk-guided approaches to perioperative care.

Funding

• Government Support – Non-U.S.