Abstract: PUB051
Explainable Prediction of 24-Hour AKI Recovery in Trauma Patients Using Time-Series Shapley Additive Explanations
Session Information
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Yi, Yongjin, Dankook University Hospital, Cheonan-si, Chungcheongnam-do, Korea (the Republic of)
- Kim, So Mi, Dankook University Hospital, Cheonan-si, Chungcheongnam-do, Korea (the Republic of)
- Jang, Youngjoo, Dankook University Hospital, Cheonan-si, Chungcheongnam-do, Korea (the Republic of)
Background
Acute kidney injury (AKI) significantly impacts morbidity and mortality in trauma patients, and early recovery is associated with favorable outcomes. This study aimed to develop a machine learning model with explainable artificial intelligence (xAI) to predict AKI recovery within 24 hours among hospitalized trauma patients.
Methods
We performed a retrospective cohort study of 1,156 trauma patients who developed AKI between 2015 and 2024 at the Dankook University Trauma Center. AKI stages were classified according to the KDIGO criteria based on serum creatinine (Stage 1, n=663; Stage 2, n=218; Stage 3, n=275). An Extreme Gradient Boosting (XGBoost) model was trained on clinical and laboratory data from 909 patients and validated on a temporally independent cohort of 247 patients. The primary outcome was AKI recovery at 24 hours. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC), with sensitivity and specificity. Feature contributions and model explanation were analyzed using time-series SHapley Additive exPlanations (SHAP).
Results
The XGBoost model predicted 24-hr AKI recovery with an AUROC of 0.843. At a threshold probability of 0.5, the model achieved a sensitivity of 74.6% and a specificity of 71.2%. Time-series SHAP analysis generated interpretable visualizations of patient-specific clinical trajectories, revealing key factors associated with recovery (Figure 1).
Conclusion
This study demonstrates that an explainable XGBoost model can predict 24-hr AKI recovery in trauma patients using routinely collected clinical data. The integration of time-series SHAP visualizations enhances interpretability, enabling clinicians to understand individualized predictions and identify actionable clinical features to support personalized interventions and improve outcomes.
An example of a single patient during hospitalization: time-series SHAP plots of the eight most contributory features to the model (top), the probability of 24-hr AKI recovery (bottom, black solid line) with the actual AKI recovery (green bars).
Funding
- Government Support – Non-U.S.