Lyna ABROUGUI

Prediction of emergency department revisits for intracranial hemorrhage

Background: Emergency department (ED) revisits pose a significant burden on healthcare systems by indicating unresolved clinical issues or the need for additional care. Among these, ED revisits for intracranial hemorrhage (IH) are especially concerning given the condition’s high risk of serious complications. Ensuring early identification of IH, timely initiation of treatment, and effective care transitions from the ED are essential to minimizing the risk of ED revisits and reducing morbidity and mortality. While progress has been made in understanding IH and revisits separately, significant gaps remain in identifying patients at risk of returning to the ED with IH as an outcome. These gaps may arise from challenges in detecting IH at an early stage, recognizing cases where IH develops or progresses over time, or addressing factors that predispose individuals with existing IH to further deterioration. Developing innovative tools to address these gaps is essential for improving outcomes and reducing ED revisits associated with IH.

Objective: This study aims to develop a machine learning algorithm to predict ED revisits due to IH. 

Methods: A publicly available deep learning model was trained to classify the presence or absence of IH using NCCT data. By leveraging patient imaging data with and without IH, we trained a deep learning model based on probabilities to identify and classify IH using a combination of image pixels and DICOM metadata from non-contrast head CTs (NCCTs). The trained model was subsequently applied to a dataset of NCCT scans from patients at the Hôtel-Dieu de Lévis to generate IH probability estimates for each case. The study included patients aged ≥ 65 years who underwent NCCT scans between January 2010 and December 2021. A classical machine learning model was then developed and tested to predict short-term ED revisits with IH as the outcome, using patient clinical data with and without the previously obtained IH probabilities.

Expected results and conclusions: We developed a machine learning algorithm that has the potential to assist emergency physicians and radiologists in identifying patients at high-risk of returning to the ED for a newly developed, or worsening of an existing intracranial hemorrhage. This algorithm could support decision-making about ED discharge and early follow-up to prevent ED revisits for IH. Additionally, it could contribute to improving care transitions, reducing complications for patients, and minimizing ED revisits.