Prediction of Acute Kidney Injury After Cardiac Surgery Using Interpretable Machine Learning.

Area of Science:

• Cardiovascular Surgery
• Nephrology
• Artificial Intelligence in Medicine
• Machine Learning

Background:

• Postoperative acute kidney injury (AKI) is a significant complication following surgery, particularly cardiac surgery.
• AKI can lead to extended hospitalization, increased medical costs, and mortality.
• Predicting AKI enables proactive strategies to mitigate patient risk.

Purpose of the Study:

• To develop and evaluate interpretable machine learning models for predicting postoperative AKI.
• To identify key clinical factors influencing AKI risk in surgical patients.

Main Methods:

• Data from 1435 patients across multiple centers were analyzed, encompassing demographics, medical history, lab results, and surgical details.
• Various machine learning algorithms, including Random Forest (RF) and XGBoost, were employed for AKI prediction.
• Local interpretable model-agnostic explanations (LIME) and Shapley methods were used to ensure model interpretability.

Main Results:

• Random Forest (RF) and XGBoost demonstrated superior predictive performance, achieving AUCs of 0.81 and 0.80, respectively.
• Key predictors identified include creatinine (Cr), cardiopulmonary bypass time (CPB time), blood sugar (BS), and albumin (Alb).
• Model interpretability confirmed the clinical relevance of identified risk factors.

Conclusions:

• Machine learning models, specifically RF and XGBoost, can effectively predict postoperative AKI in cardiac surgery patients.
• These predictive capabilities allow clinical teams to implement timely preventive and therapeutic interventions.
• Interpretable predictions enhance trust and facilitate the clinical integration of AI tools for patient care.