Identification of acute kidney injury after non-cardiac surgery by machine learning leveraging a large electronic health record database

Figures

• 

  Fig. 1. Ranks of native Feature Importance (FI) assigned by different PO-AKI prediction models trained on the complete balanced development cohort. The FI ranks across all models are displayed for the most important features (ranks 1-5) in each non-ensemble (a) and ensemble model (b). Predictors were ranked based on the absolute native FI that was estimated by the respective model and scaled to the largest importance per model. Lower rank values indicate higher importance with rank 1 representing highest importance. Continuous features are shown as dots, categorical features are displayed as triangles, respectively. Abbrev.: age, patient age; anestDurat, duration of anesthesia; ASA, American Society of Anesthesiology physical status; clinic_cat_ACH, general surgery; clinic_cat_GCH, vascular surgery; clinic_cat_NCH, neurology surgery; clinic_cat_TCH, thoracic surgery; clinic_cat_URO, urology surgery; doseInduPropoPerfusor, dose of propofol administered via perfusor during induction; doseSurgSteroInf, dose of sterofundin administered via infusion during induction; eGFRPreSurg, estimated glomerular filtration rate before surgery; liverCirrh, liver cirrhosis; ltmDiu, longterm medication with…

• 

  Fig. 2. Ranks of Shapley value based Feature Importance (FI) assigned by different PO-AKI prediction models trained on the complete balanced development cohort. Shapley values were averaged across all models and subsequently scaled to the highest average importance before descending ranks of importance were assigned. Lower rank values indicate higher importance with rank 1 representing highest importance. Continuous features are shown as dots, categorical features are displayed as triangles, respectively. Abbrev.: age, patient age; anestDurat, duration of anesthesia; ASA, American Society of Anesthesiology physical status; clinic_cat_ACH, general surgery; clinic_cat_URO, urology surgery; doseInduPropoPerfusor, dose of propofol administered via perfusor during induction; eGFRPreSurg, estimated glomerular filtration rate before surgery; ltmDiu, longterm medication with diuretics; stdAnes, standard deviation of MAP during anesthesia.

• 

  Fig. 3. Native Feature Importance (FI) ranks of the five most important features across all perioperative time points of the Random Forest classifier models trained on the balanced development cohort. The plot displays the rank trajectories of those variables that were among the top five features at any time point. Ranks were determined within each model based on native FI values. Lower rank values indicate higher importance with rank 1 representing highest importance. To facilitate comparability across time points, the variable names of features in the MAP time series for different time points were transferred to higher-level categories (labeled by “timepoint”). Abbrev.: age, patient age; anestDurat, duration of anesthesia; ASA, American Society of Anesthesiology physical status; baselineMAP, first measured MAP value; doseInduRemifPerfu, dose of remifentanil administered via perfusor during induction; height, patient height; inductDurat, duration of induction of surgery; kurtosis (timepoint), kurtosis of MAP during corresponding timepoint; mean (timepoint), mean MAP during corresponding timepoint; trend (timepoint), trend of MAP during corresponding timepoint; urgency, level of urgency…

• 

  Fig. 4. Displayed are the Area under the Receiver Operating Characteristic curve (ROC-AUC) (a), F1-score (b), and Balanced Accuracy (BACC) (c) of the Random Forest classifier models trained on the balanced development dataset and evaluated on the validation cohort. Each metric is reported as mean±95% confidence interval based on 100 bootstrap replicates. Time points cover the preoperative phase, induction of anesthesia, selected intraoperative intervals, and the immediate postoperative phase.