Artificial intelligence-guided perfusion metrics to predict delayed graft function in deceased-donor kidney transplantation

Abstract

This study evaluates whether artificial intelligence (AI) applied to machine-perfusion signals can predict delayed graft function (DGF) in deceased-donor kidney transplantation. Background: DGF remains common and costly; hypothermic machine perfusion (HMP) generates high-fidelity physiologic data (flow, vascular resistance, pressure, temperature) that are underused in real-time decision-making. Methods: We analyzed authentic LifePort Kidney Transporter logs from three donor kidneys (~4 hours/run; >1,200 time points each), linked to verified early clinical outcomes. After cleaning, time-normalization, and feature engineering (e.g., resistance slope, flow-to-resistance ratio, temperature stability, pressure–flow correlation), a Random Forest classifier with 10-fold cross-validation modeled DGF (binary) from device-native features. Results: DGF occurred in one of three grafts (33.3%). Across runs, resistance showed an inverse, near-exponential relationship with flow (Pearson r = −0.87, p < 0.001); mean pressure remained ~30 mmHg and temperatures stabilized within 60 minutes. The AI model achieved AUC = 0.91 with accuracy 86.7%, sensitivity 100%, and specificity 80%, with resistance slope and mean flow contributing most to discrimination. Conclusion: Dynamic perfusion trajectories, captured noninvasively during HMP, encode clinically meaningful information about early graft function; AI converts these signals into an interpretable, real-time risk estimate that could standardize organ acceptance and reduce unnecessary discard, warranting multicenter validation for generalizability.