Multi-Horizon Chronic Kidney Disease Trajectory Prediction via Interpretable Stacked Ensemble and Deep Sequential Modeling

Résumé fourni par la source

Accurately forecasting the estimated Glomerular Filtration Rate (eGFR) is crucial for the early, proactive management of Chronic Kidney Disease (CKD). However, predicting how eGFR changes over long time horizons remains hard. This difficulty largely stems from the noisy, often high-dimensional patient data found in Electronic Health Records (EHRs). Existing tools are often limited to short-term or single-step prediction, which restricts their ability to generate reliable long-term forecasts. To overcome this limitation, we propose a stacked ensemble framework based on multi-output regression to improve long-range prediction performance. Our hybrid approach strategically blends powerful models: tree-based learners (like XGBoost, Light-GBM, and CatBoost) alongside deep sequence models (LSTM and Transformer). This setup lets the ensemble simultaneously capture a patient’s static risk profile and learn complex temporal patterns, allowing us to make a joint prediction of eGFR at 3, 6, and 12 months.We tested this ensemble on a large, longitudinal CKD dataset sourced from the MIMIC-IV database (N =5,123 patients over about 10 years). Our model really shone, hitting an R2of 0.778 for the crucial 12-month horizon. This wasn’t just a minor win; it was a solid improvement of ∆R2≈ 0.009 (around 1.2%) over the best single model. The average prediction error (MAE) also dropped from 13.90 to 13.49. Finally, using SHAP for feature attribution, we confirmed that baseline creatinine, patient age, and blood pressure were the most important drivers of the prediction. Ultimately, this ensemble offers a statistically strong and clinically useful tool for long-term CKD progression forecasting.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.