A clinically explainable temporal fusion model for predicting postoperative acute kidney injury

Abstract Objective Postoperative acute kidney injury (AKI) is a common and serious complication, driven by complex perioperative physiologic derangement and hemodynamic instability. There remains a need for a mechanism for deep interaction between preoperative characteristics and intraoperative dynamics, and provide clinical interpretability. Methods We propose an extended Temporal Fusion Transformer for postoperative AKI prediction, designed for heterogeneous clinical data and intraoperative sequences, such as arterial pressure, medications, and fluid infusions. The model is built based on Long Short-Term Memory (LSTM) with the Variable Selection Networks, which adaptively select static and temporal features using gated residual networks, conditioned by static context vectors. Static covariates are encoded as multiple context vectors that guide temporal variable selection, LSTM initialization, and the temporal attention mechanism, modelling complex interactions between baseline risk and intraoperative signal fluctuations. The final representation for outcome prediction was aggregated by using the preoperative state as the attention query. Results The model was evaluated on two public surgical datasets, the Informative Surgical Patient dataset for Innovative Research Environment (INSPIRE) and Medical Informatics Operating Room Vitals and Events Repository (MOVER). Across all experiments, the proposed model achieved better performance and showed stable performance in subgroup analysis. The interpretability results showed the feature importance of static variables and time series at each time step, revealing that the model consistently assigned attention to periods of sustained hemodynamic instability. Conclusions These findings supported the potential for explainable perioperative decision support systems and emphasized the importance of intraoperative signals in AKI risk stratification.