A Hybrid CNN-GRU Model for Real-Time Prediction of Sepsis Clinical Trajectories in the ICU

Abstract

Sepsis progression varies widely among intensive care unit (ICU) patients, making early and accurate trajectory prediction crucial. Traditional models focus on static assessments or binary outcomes, limiting their clinical utility. This study considers clinical trajectories of patients with sepsis for up-to 14 days hospital stay since ICU admission. Using data from 19,177 ICU patients, this study presents a hybrid CNN-GRU model that predicts four distinct clinical trajectories—Fast Decline, Fast Recovery, Slow Recovery, and Delayed Decline—by integrating sequential time-series data with static patient features. The model updates predictions every 2 hours, achieving AUCs of 0.90 (Fast Decline), 0.76 (Delayed Decline), 0.70 (Fast Recovery), and 0.55 (Slow Recovery) on validation data at hour 36. To enhance interpretability, we analyzed feature importance across both modalities. For sequential data, gradient-based saliency identified peripheral capillary oxygen saturation, lactate, positive endexpiratory pressure, temperature, Glasgow Coma Scale, respiratory rate, and mean arterial pressure as key drivers— aligned with known sepsis indicators. For static features, permutation-based analysis highlighted age, ICU stay duration before prediction, comorbidity burden, and pulmonary disease history as top contributors. These insights confirm the model's physiological relevance and support its potential for real-time, explainable sepsis trajectory prediction and risk stratification.