A deep learning–based method to predict the length of stay for patients with traumatic fall injuries in support of physicians' clinical decisions and patient management

Abstract

Accurate estimates of the length of stay (LOS) for patients who suffer traumatic fall injuries are crucial to inform physicians' clinical decisions and patient management. They also have important implications for resource utilization efficiency and cost containment efforts by healthcare organizations. Effective predictions should consider essential relationships across different variables pertaining to patient demographics, clinical history, injury severity, and physiology. A proposed deep learning–based method incorporates these relationships and can predict LOS more accurately, as demonstrated by a comparative evaluation involving 3722 patients who suffered traumatic fall injuries between 2011 and 2017. The results show the superior performance of the proposed method, relative to eleven prevalent methods that represent different analytics approaches. Our method demonstrates superior predictive performance, as manifested by the highest F-measure values and area under the curve. It is particularly efficacious for patients likely in need of longer LOS, which is relatively more important to physicians and healthcare organizations. This study underscores the value of incorporating important relationships and interactions among distinct patient variables to estimate LOS, with a particular emphasis on the inter-disease relationships, physiology-severity interactions, and patient information in clinical notes. The proposed method can be implemented as a decision support system to enhance physicians' clinical decisions and patient management, and improve healthcare organizations' resource planning and utilization efficiency, with nontrivial cost containment implications.