A Machine Learning–Based Prediction Model for Acute Kidney Injury in Patients With Community-Acquired Pneumonia: Multicenter Validation Study

Figures

• 

  Figure 1. Flowchart for patient selection. CAP: community-acquired pneumonia; CAP-AKI: community-acquired pneumonia–associated acute kidney injury.

• 

  Figure 2. Feature selection based on the sliding windows sequential forward feature selection technique process. The optimal number of features was determined according to the OOB value. A lower OOB value and fewer variables included results in the optimal combination of features (11 at the red circle). OOB: out-of-bag.

• 

  Figure 3. Feature importance derived from the machine learning models. (A) logistic regression model, (B) random forest model, (C) support vector machine model, (D) extreme gradient boosting model, and (E) deep forest model. BUN: blood urea nitrogen; CRP: C-reactive protein; DiasBP: diastolic blood pressure; WBC: white blood cells.

• 

  Figure 4. The performance of machine learning models in the internal validation set. (A) area under receiver operating characteristic curve, (B) calibration curve, and (C) decision curve analysis curve. AUC: area under the receiver operating characteristic curve; DF: deep forest; LR: logistic regression; RF: random forest; SVM: support vector machine; XGBT: extreme gradient boosting.

• 

  Figure 5. The performance of machine learning models in the external validation set. (A) area under receiver operating characteristic curve, (B) calibration curve, (C) decision curve analysis curve. AUC: area under the receiver operating characteristic curve; DF: deep forest; LR: logistic regression; RF: random forest; SVM: support vector machine; XGBT: extreme gradient boosting.

• 

  Figure 6. Shapley additive explanations (SHAP) interpretation and visualization of the deep forest model. (A) SHAP analysis shows a vital ranking of 11 features, and (B) SHAP analysis shows point estimation of features to model output. Each point represents a single patient in the data set. BUN: blood urea nitrogen; CRP: C-reactive protein; DiasBP: diastolic blood pressure; WBC: white blood cells.

• 

  Figure 7. Local interpretable model-agnostic explanations (LIME) explaining individual’s prediction results. The left part of the figure shows predicted results using LIME. The middle part presents the impact of the 11 variables on acute kidney injury (AKI) or no AKI. The length of the bar for each feature indicates the importance (weight) of that feature in making the prediction. A longer bar indicates a feature that contributes more to (A) No AKI or (B) AKI. The right panel shows the critical values of these 11 variables when they had the greatest impact on (A) No AKI or (B) AKI. AKI: acute kidney injury; BUN: blood urea nitrogen; CRP: C-reactive protein; DiasBP: diastolic blood pressure; SHAP: SHapley Additive exPlanations; WBC: white blood cells.

• 

  Figure 8. Web-based prediction platform of community-acquired pneumonia–associated acute kidney injury (CAP-AKI). (A) When 11 variables of a patient with community-acquired pneumonia were entered, the risk of CAP-AKI was calculated to be 56.25 %, and the result was 0, that is, the patient was not a high-risk group of CAP-AKI. (B) When 11 variables of a patient with community-acquired pneumonia were entered, the risk of CAP-AKI was calculated to be 82.75 %, and the result was 1, that is, the patient was a high-risk group of CAP-AKI. We set the risk threshold of our web calculator was 74.50% based on the best F1-score. When the predictive probability exceeds the risk threshold, the output is 1, otherwise the output is 0.