A reinforcement learning-guided interpretable method for postoperative sepsis prediction with Hilbert-Schmidt...

Sepsis is a major cause of postoperative morbidity and mortality, and early risk stratification from perioperative electronic health records (EHR) is a representative large-scale, high-dimensional data processing problem that requires models to be accurate, efficient, and clinically interpretable. However, many existing sepsis prediction methods operate as black boxes and rely on extensive temporal monitoring streams, which increases feature dimensionality and computation while limiting transparency.

We propose a reinforcement learning-guided, interpretable feature engineering framework for postoperative sepsis prediction that targets scalable learning on heterogeneous perioperative data. Within an Actor-Critic formulation, feature selection is treated as an action: an Actor network produces a stochastic feature mask over preoperative static variables and intraoperative statistical summaries, while a Critic network performs downstream prediction using a self-attention-based classifier. To benchmark and stabilize learning, we introduce an auxiliary baseline model that incorporates intraoperative temporal signals extracted by a temporal convolutional network (TCN) and regularized using the Hilbert-Schmidt Independence Criterion (HSIC) to encourage non-redundant representations between statistical and temporal feature views. The Actor is optimized to achieve comparable predictive performance to the baseline while using a reduced feature set, improving computational efficiency and supporting instance-level interpretability.

Experiments on a real-world surgical cohort from Southwest Hospital (2014-2018) demonstrate that the proposed framework attains performance comparable to or better than competitive machine learning baselines while selecting fewer input features. On this dataset, our method achieved perfect scores of 1.00 for F1-score, Sensitivity, and Specificity.

The proposed method accurately predicts the occurrence of postoperative sepsis and provides effective instance-level post hoc explanations. These findings offer a novel perspective for postoperative sepsis prediction.