An Interpretable Machine Learning Model for Early Multitemporal Prediction of Onset of Acute Kidney Injury in Intensive Care Unit Patients with Severe Trauma

Abstract

Background:

Acute Kidney Injury (AKI), a leading organ failure cause in critical patients, demands early high-risk identification to enhance outcomes. Yet comparative analyses of diagnostic and prognostic machine learning (ML) models across multiple post-admission timeframes are lacking.

Methods:

Using MIMIC-IV, we carried out using the Boruta algorithm for feature selection, developing and comparing six ML models to predict AKI risk at 0-24, 24-48, 48-72, 0-48, and 0-72 h post-ICU admission. Model performance was evaluated using the Area Under the Curve (AUC) and confusion matrix. Decision Curve and calibration analyses assessed clinical applicability. We compared models with Sequential Organ Failure Assessment (SOFA) and SAPSII scores to evaluate the accuracy of the ML models. Finally, Shapley Additive Explanations (SHAP) values interpreted and visualized key features of the optimal model.

Results:

Our study involved 2092 trauma Intensive Care Unit (ICU) patients. Using the 17 selected out of the 48 features among trauma patients 24 h after ICU admissions, among the six ML models and two scoring systems, all ML models outperformed SOFA and SAPS II, and the extreme gradient boosting (XGBoost) exhibited the best performance, achieving an AUC of 0.948 (95% CI [0.929-0.966]) for AKI prediction within 24 h of admission, with an AUC of 0.941 ([0.892-0.917]) and 0.878 ([0.863-0.892]) at 0-48 and 0-72 h period, respectively. However, their predictive accuracies were very limited at 24-48 h (AUC 0.602 [0.562-0.643]) and 48-72 h (AUC 0.490 [0.429-0.551]), respectively. Urine output per kilogram per hour at 6 and 12 h and age were the most important features identified through SHAP analysis.

Conclusions:

Our study found ML models excel in diagnosing AKI risk in ICU trauma patients but have limited prognostic accuracy at 24-48 and 48-72 h post-admission. Further research is needed to improve this using time-series ML models with optimal windows.