FARE-AKI

The FARE-AKI (Fair and Accurate Renal Event prediction for Acute Kidney Injury) project aimed to develop a predictive model for future acute kidney injury (AKI) occurrence in hospitalized patients.

The FARE-AKI model was developed using the MIMIC-IV dataset for ICU patients and data from Seoul National University Bundang Hospital for general ward patients.

AKI was defined according to the 2012 KDIGO guidelines, using only creatinine-based criteria. Clinical features were carefully selected, modified, and validated based on clinical relevance and significance.

We identified 10 sensitive variables relevant to the clinical setting. Fairness concerns arose with variables such as days from admission and admission department, which were systematically regularized and adjusted at each stage of the training process. A combination of the Factored Generalized Additive Model (FGAM) and LSTM yielded the highest performance, achieving an F1 score of 0.922 and MCC of 0.915 for general ward patients.

We are preparing to present our study results at an upcoming conference.

• The final model received the Director’s Award from the Medical AI Center at Seoul National University Bundang Hospital on October 23, 2024.