Explainable machine learning-based 28-day mortality prediction model for elderly patients with acute kidney injury

Area of Science:

• Medical Informatics
• Geriatric Medicine
• Nephrology

Background:

• Elderly patients with acute kidney injury (AKI) have a high mortality risk.
• Early prediction of mortality is crucial for timely intervention and resource allocation.
• Machine learning (ML) offers potential for developing accurate predictive models.

Purpose of the Study:

• To develop and validate ML models for predicting 28-day mortality in elderly AKI patients.
• To identify key predictors of mortality in this population.
• To enhance clinical decision-making through predictive healthcare analytics.

Main Methods:

• Retrospective analysis of 1290 elderly AKI patients from PLAGH (2008-2018).
• Comparison of five ML algorithms (L2-logistic, LASSO, XGBoost, RF, MLP) with oversampling techniques.
• Performance evaluation using Area Under the Receiver Operating Characteristic Curve (AUC); interpretability via SHapley Additive exPlanations (SHAP).

Main Results:

• The XGBoost model with random oversampling achieved an AUC of 0.8659 in the validation cohort.
• External validation in the eICU-CRD dataset yielded an AUC of 0.6317.
• Top predictors for 28-day mortality included Mechanical Ventilation, Peak serum creatinine, AKI Stage, urine protein, and serum albumin.

Conclusions:

• ML models, particularly XGBoost, can effectively predict mortality in elderly AKI patients.
• Predictive models enhance clinical decision-making and can improve patient outcomes.
• Key factors identified by SHAP analysis provide insights for targeted interventions.