Emergency Room Evaluation for Older Users of Emergency Departments: Predicting Adverse Health Outcomes With Deep Learning Algorithms

An Emergency Department (ED) visit for an older adult is a high-risk medical intervention. Known adverse events (AE) include delirium, prolonged ED or hospital stay, hospitalization, recurrent ED visits and hospital death. These happen in a growing proportion in ED visitors over age 65 are over who are represented in ED visits.

Tools predicting AEs in the ED are of paramount importance to help decision-making on patient triage and disposition. They can help identify areas of unmet needs for seniors in order to develop targeted actions. Multiple scoring systems including "Programme de recherche sur l'intégration des services de maintien de l'autonomie" (PRISMA-7), Identification of Seniors at Risk (ISAR), Clinical Frailty Scale (CFS), Brief Geriatric Assessment (BGA) have extensively been studied in the ED and other settings for various outcomes. These tools rely on a simple scoring system that minimally trained staff can reliably and quickly administer. Doing otherwise is unlikely to be applicable to daily clinical practice.

As prediction accuracy has not significantly improved in the past decade, perhaps new analysis strategies are necessary. The current hype surrounding deep learning comes from better and cheaper hardware and the availability of simple and open-source libraries supported by large companies and a broad community of users. Hence, implementing deep learning (DL) algorithms is now open to a wide range of settings, including medical care in a standard clinical practice. DL has been shown to be more accurate than the average board-certified specialist on very specific tasks. Prediction of various clinical outcomes has produced less dramatic results, perhaps as traditional (non-DL) models already outperformed clinicians for many disease states. Published DL approaches applied to outcome prediction in the ED have focused on acutely ill adults in general, specific conditions or administrative issues such as admitting department or ED overcrowding. None have targeted a specific age group like older ED visitors.

An important caveat to many DL approaches is interpretation of results. To develop interventions based on targeted features associated with AEs in a given model, it has to be somewhat transparent. If multiple layers of NNs improve prediction compared to linear regression, they often provide no clinically relevant insight on how and which variables interact to yield that result.