FIR-LSTM: An Explainable Deep Learning Framework for Predicting Iatrogenic Withdrawal Syndrome in Pediatric Intensive Care Units

Iatrogenic withdrawal syndrome (IWS) is a significant yet underrecognized public health concern for pediatric patients in critical care units, most frequently the result of abrupt cessation or rapid tapering of sedative or opioid medications. Early prediction of IWS is important for timely intervention and improved patient outcomes. In this study, we developed an explainable deep learning model utilizing a unidirectional multilayer long short-term memory (LSTM) network to predict the risk of IWS in pediatric ICU patients. Through longitudinal electronic health records (EHRs), our model analyzes the preceding 24 hours of patient data to predict the likelihood of IWS occurring in the next four hours, providing a real-time risk score. To enhance interpretability and identify key risk factors, we applied layer-wise relevance propagation (LRP) to the LSTM model. The feature importance rankings derived from LRP were validated through multiple experiments. Experimental results show that the model was perfectly calibrated and achieved robust predictive performance, suggesting that the LRP enhanced LSTM model holds significant potential for improving pediatric patient care by facilitating early detection and proactive management of IWS in critical care settings. Implementing this model into a system of alerts for clinicians could lead to significant advances in safer sedative and analgesic use, addressing an under addressed public health issue that impacts not only the United States but also the global community.