Abstract

Neonatal mortality is still a big health problem worldwide, and in intensive care units, prompt identification of high-risk infants is important to enhance clinical outcomes. Most of the prediction methods used today may fail to accurately capture physiological data that is both heterogeneous and measured at irregular intervals over time and/or varies between medical and healthcare institutions, potentially compromising their reliability and scalability. To mitigate these issues, the study suggests a novel framework for neonatal risk prediction using explainable AI in a cloud-based environment that integrates cutting-edge temporal learning with explainable analytics in an intelligent healthcare setting. A combination of a Temporal Fusion Transformer (TFT) and a Neural Controlled Differential Equation (Neural CDE) network is used to model the long-range temporal dependence and continuous-time physiological dynamics available in neonatal monitoring data. The static clinical features and time-varying physiological signals are encoded temporally in two parallel processors, which are fused together by a temporal invariant fusion mechanism, to obtain detailed representations for patients. A feature-attribution module, which is a part of the Explainable Artificial Intelligence (XAI) module, further increases transparency and aids clinician trust in model predictions. The framework is cloud-enabled, which allows for scalable processing, real-time monitoring, and intelligent decision support in healthcare institutions. The very high AUC-ROC of prediction for mortality (98.63%) and for respiratory failure (98.74%) in an experimental evaluation of a multi-center neonatal intensive care dataset of patients showed good predictive performance, and length of stay was correlated with the model's predictions. The results show the effectiveness, interpretability, and generalizability of the framework to assist in proactive health care management of neonates.