Acute kidney injury (AKI) is a frequent and severe complication among elderly patients and is strongly associated with adverse outcomes, including increased short-term mortality, renal replacement therapy (RRT), and progression to chronic kidney disease (CKD). However, current risk assessment tools are limited by linear assumptions, insufficient incorporation of multimodal predictors, and lack of interpretability.
The primary research question of this project is: Can explainable machine learning models, integrating clinical, biochemical, and proteomic data from the UK Biobank, improve prediction of adverse outcomes in elderly patients with AKI?
The objectives are threefold: (1) to develop and validate prognostic models for 28-day mortality and RRT initiation in AKI patients aged !65 years; (2) to apply feature selection and interpretability methods (e.g., SHAP) to identify the most influential predictors across clinical, laboratory, and proteomic domains; and (3) to derive clinically actionable thresholds for risk stratification to facilitate translation into critical care practice.
The scientific rationale is based on the urgent need for early, precise, and interpretable prediction of outcomes in AKI, which remains a major challenge in nephrology and critical care. By leveraging the depth of UK Biobank data, this project will generate novel insights into AKI prognosis, enhance clinical decision-making, and potentially inform targeted interventions aimed at improving survival and renal outcomes in older adults.