Continuous-Time Transformer with Adaptive Mutation-Locking for Early Prediction of Acute Kidney Injury.

Abstract

Accurate early prediction of acute kidney injury is critical, yet existing deep learning models operate as opaque black boxes that rely on uninterpretable correlations. Therefore, this study develops a transparent temporal attribution architecture to uncover precise pathophysiological triggers. We propose the CT-Former framework, which integrates continuous time neural networks with a transformer to preserve irregular patient trajectories without biased data imputation. Furthermore, a two-stage temporal attribution decoupling module purifies the global attention field. This module generates an explicit temporal attribution graph to identify sudden critical time steps. The network locks onto these acute mutational triggers before fusing this temporal signal with the background context via an adaptive gate. Evaluated on a MIMIC-IV cohort of 18419 patients, the CT-Former achieves the highest predictive performance among the evaluated baselines across all prediction time windows. Additionally, the learned temporal attribution matrix demonstrates structural consistency with independent interpretability evaluations. The CT-Former establishes a transparent temporal decision pathway that delivers exceptional predictive robustness while entirely avoiding spurious statistical correlations. This explicitly interpretable framework represents a promising step toward interpretable prognostic tools for critically ill patients.