Subscribe to RSS
DOI: 10.1055/s-0037-1598683
Novel Statistical Techniques for the Prediction of Acute Renal Injury after Cardiac Surgery Procedures: The Rise of the Machines?
Publication History
Publication Date:
03 February 2017 (online)
Introduction: Machine learning techniques are being increasingly used in clinical studies for various regression and classification tasks including risk factor identification, predictive forecasting and analytics. Advantages over conventional techniques include efficient handling of high dimensional datasets, avoidance of overfitting, exploration of non-linear relationships without the need for parametric prespecification and detection of complex variable interactions. Aim of the present study was to investigate how these methods can provide insights into risk classification of the development of acute renal injury (AKI) after cardiac surgical procedures.
Methods: 1.176 consecutive, elective cardiac surgical patients were included in this study. Patients with chronic kidney disease stage 5 were excluded from analyses. AKI was defined according to KDIGO - creatinine criteria. Pre-operative plasma GDF-15, NTproBNP, hsTNT, clinical outcomes, and 30-day and 1-year mortality were recorded. Feature selection was investigated using exhaustive multivariate bidirectional stepwise logistic regression, elastic net regularization (lasso & ridge) logistic regression and ensemble learning methods (random forests, RF).
Results: In terms of performance metrics, elastic net regularized logistic regression (lasso) provided superior association metrics with 0.813 accuracy, 0.956 specificity, 0.187 classification error and 0.773 AUC during k-fold cross validation. Ensemble RF techniques revealed significant non-linear effects and variable interactions: the prognostic ability of increased GDF-15 levels on the development of AKI was significantly more prominent in patients with normal preoperative creatinine levels.
Conclusions: Machine learning techniques offer a valuable toolbox for modern analyses of clinical data providing superior results and advantages over traditional techniques at the cost of perhaps a slightly less direct interpretability of the models and their results.