AKI Prediction in Septic Shock

Overview

Group project at Vrije Universiteit Amsterdam (M.Sc. AI) developing a predictive model for Acute Kidney Injury (AKI) in patients with septic shock. Approximately two-thirds of septic shock patients develop AKI, often before arriving at the emergency department. Early prediction enables timely interventions like adjusting treatments or initiating renal support.

Research question: Which septic shock patients are at risk of developing a specific stage of AKI (stages 0-3) within the next 12 hours of ICU admission?

Data

Used the AmsterdamUMC database (UMCdb) — 23,106 ICU admissions of 20,109 adult patients recorded between 2003-2016, containing nearly 1.0 billion clinical observations and 5.0 million medication records. Features included vital signs, lab values (creatinine, urine output, eGFR), and demographics.

Approach

• Expanding window training: The training set grows incrementally from 48 hours of ICU data, with the test set being the subsequent 12-hour window. This aligns with the temporal nature of clinical decision-making.
• Decision Tree Classifier: Transparent, interpretable model for clinical trust
• XGBoost: Gradient boosting model as a high-performance approach
• SHAP explainability: Beeswarm plots and feature importance analysis to interpret model predictions for clinical staff
• Uncertainty quantification: Confidence estimates on predictions

Key Findings

• Urine output (CAD), baseline creatinine, and creatinine were consistently the most predictive features across all time windows (60-96 hours)
• Feature importance remained stable over time, reinforcing clinical domain knowledge
• XGBoost outperformed Decision Trees on overall metrics

Technologies

Python, XGBoost, Decision Trees, SHAP, Pandas, scikit-learn, AmsterdamUMC database