Prediction of persistent acute kidney injury in postoperative intensive care unit patients using integrated machine learning: a retrospective cohort study

Abstract

Acute kidney injury (AKI) often occurs in patients in the intensive care unit (ICU). AKI duration is closely related to the prognosis of critically ill patients. Identifying the disease course length in AKI is critical for developing effective individualised treatment. To predict persistent AKI at an early stage based on a machine learning algorithm and integrated models. Overall, 955 patients admitted to the ICU after surgery complicated by AKI were retrospectively evaluated. The occurrence of persistent AKI was predicted using three machine learning methods: a support vector machine (SVM), decision tree, and extreme gradient boosting and with an integrated model. External validation was also performed. The incidence of persistent AKI was 39.4-45.1%. In the internal validation, SVM exhibited the highest area under the receiver operating characteristic curve (AUC) value, followed by the integrated model. In the external validation, the AUC values of the SVM and integrated models were 0.69 and 0.68, respectively, and the model calibration chart revealed that all models had good performance. Critically ill patients with AKI after surgery had high incidence of persistent AKI. Our machine learning model could effectively predict the occurrence of persistent AKI at an early stage.

Figure 1

Flow chart of the study. ICU, intensive care unit; AKI, acute kidney injury.

Figure 2

The relative influence of each machine-learning model in the stacked-ensemble model. SVM, support vector machine; XGBoost, extreme gradient boosting.

Figure 3

Evaluation of model performance in the internal validation dataset. (A) The calibration plot shows the consistency between observed and predicted risks for persistent acute kidney injury. (B) Discrimination of the machine-learning models in the internal validation dataset. SVM, support vector machine; XGBoost, extreme gradient boosting; AUC, area under the curve. The number in parentheses indicates the 95% confidence interval.

Figure 4

Variable-importance ranking in the gradient-boosting machine. RRT, renal replacement therapy; SOFA, Sepsis-related Organ Failure Assessment; Uo_24h, urine volume for 24 h on ICU admission.

Figure 5

Heatmap plot showing the contribution of each variable to the classification of the sample patients. The relative contribution of each variable was calculated using the LIME algorithm. Data of patients 2 and 3 are shown as examples. The red colour indicates that the relevant variable contradicts a given label, while the blue colour indicates support. AKI, acute kidney injury; SOFA, Sepsis-related Organ Failure Assessment; Uo_24h, Urine volume for 24 h on intensive care unit admission; LIME, Local Interpretable Model-Agnostic Explanations.

Figure 6

The LIME feature plot shows the contribution of each variable to the classification of the sample patients. The red colour indicates that the relevant variable contradicts a given label, while the blue colour indicates support. AKI, acute kidney injury; SOFA, Sepsis-related Organ Failure Assessment; Uo_24h, Urine volume for 24 h on intensive care unit admission; LIME, Local Interpretable Model-Agnostic Explanations.