Interpretable Deep Learning Framework for Early Neonatal Sepsis Prediction Using Temporal Physiological Signals

Abstract

The problem of early identification of neonatal sepsis is still one of the challenges of neonatal intensive care units since clinical signs are mild and physiological decline occurs rapidly. Conventional methods of diagnosis usually are not able to reveal the initial signs of infection, and thus it may postpone treatment and pose a threat of complications. Recent technological progress in artificial intelligence has made it possible to develop predictive models that have the ability to analyze complex clinical time-series information. In the paper, an interpretable deep learning model to predict early childhood neonatal sepsis is suggested based on temporal physiological measurements. The recommended solution involves a Long Short-Term Memory (LSTM) network to be able to reveal sequential dependencies within the vital sign data such that the risk of sepsis can be identified early enough. A gradient-based explainability method is also incorporated to enhance transparency and clinical interpretability to determine the input of physiological variables to model predictions. The experimental assessment proves that the proposed model can perform the task of effective prediction and at the same time, offer meaningful explanations in terms of clinical interpretation. “Its findings reveal the promise of explainable artificial intelligence in helping to provide credible clinical decision support systems in neonatal care.