Multicenter validation of a scalable, interpretable, multitask prediction model for multiple clinical outcomes

Figures

• 

  Fig. 1. Study flowchart. This figure illustrates the flowchart of this study, including the inclusion and exclusion criteria, as well as the derivation and validation datasets.

• 

  Fig. 2. Comparisons of AUROC curves in the derivation and validation cohorts. This figure presents AUROC curves for the multitask gradient boosting machine (MT-GBM) model, single-task prediction models, and American Society of Anesthesiologists (ASA) physical status classification in the derivation and validation cohorts. The left, middle, and right columns represent the derivation cohort, Validation Cohort A, and Validation Cohort B, respectively. A Predictions for acute kidney injury, B Predictions for postoperative respiratory failure, and C Predictions for in-hospital mortality. In each plot, the blue, yellow, and green lines indicate the MT-GBM model, single-task prediction model (LightGBM), and ASA classification, respectively. ASA American Society of Anesthesiologists physical status classification, AUROC area under the receiver operating characteristic curve.

• 

  Fig. 3. SHapley Additive explanation (SHAP) summary plots. This figure shows SHAP summary plots presenting the contributions of the 16 features used in the final model to the predictions across all outcomes. A Predictions for postoperative acute kidney injury, B Predictions for postoperative respiratory failure, C Predictions for in-hospital mortality. ASA American Society of Anesthesiologists physical status classification, AST aspartate transaminase, BMI body mass index, Hb hemoglobin, INR international normalized ratio, PT prothrombin time, WBC white blood cell.

• 

  Fig. 4. SHapley Additive exPlanation (SHAP) dependence plots. This figure shows SHAP dependence plots for the 11 selected continuous features in the final model, illustrating the relationship between each feature’s value and its SHAP value in the test dataset of the derivation cohort: A Predictions for postoperative acute kidney injury, B Predictions for postoperative respiratory failure, C Predictions for in-hospital mortality. From left to right: (1) age, (2) duration of anesthesia (DoA), (3) body mass index (BMI), (4) preoperative albumin, (5) preoperative creatinine (sCr), (6) preoperative glucose, (7) aspartate transaminase (AST), (8) preoperative hemoglobin (Hb), (9) preoperative platelet (Plt), (10) preoperative prothrombin time (PTINR), and (11) preoperative white blood cell counts (WBC).