Introduction
Emergency department (ED) revisit is a well-known quality index for ED medical care and patient safety, in which revisit rates >5% may reflect poor quality of care, and those <1% indicate undue risk aversion.1 Previous studies indicated that ED revisit may increase medical costs, ED crowding and poor prognosis, particularly in patients who require hospital admission, often due to rapid deterioration after ED discharge.2–4 Over the past decade, this concept has become challenging because the factors that influence ED revisit are multifactorial, such as issues related to diagnosis, management, procedural complications and medical adverse effects.3 5 6 Most issues are preventable and do not result in severe outcomes.7 Recent studies of intrinsic factors for high-risk ED revisit focused on patients who received hospital admission or intensive care unit (ICU) care.8–10
To identify potential risk factors for either high-risk ED revisit or unscheduled ED revisit within 72 hours, logistic linear regression models are widely used. Well-known factors for high-risk ED revisit include age, male sex, ambulance transport for return visit, longer ED length of stay, symptoms of dyspnoea or chest pain on ED presentation, triage level 1 or 2, acute change in levels of consciousness and unstable vital signs (tachycardia and/or fever), among others.8 11 12 However, a study of ED revisit has been limited by the use of linear algorithms, such as logistic regression routine, use of administrative data and small sample sizes, in part because the assessment of risk factors is more complicated than that possible with linear association.13
With the development of artificial intelligence, the machine-learning (ML)-based prediction model is used now as a clinical classifier. Lee et al developed an ML framework combining a particle swarm optimisation feature selection algorithm and an optimisation-based discriminant analysis model, to predict ED revisit. Hong et al indicated that gradient-boosting models that leveraged clinical data were superior to traditional logistic regression models built on administrative data to predict ED revisit.14 Hsu et al developed an ML model, the voting classifier model, to predict ED revisit in patients with abdominal pain.15 These works shed light on the use of a prediction model for ED revisit based on an ML algorithm.
In previous ML-based studies, the work by Lee and Hong focused on building a prediction model for general ED revisit, whereas that of Hsu focused on ED revisit and abdominal pain symptoms. All prediction models showed superior prediction performance than that with a traditional logistic regression model. Expanding on these previous works, in the current study, we specifically predict high-risk ED revisit in 72 hours using a large dataset of adult ED revisits, with more than 150 variables extracted per visit from each medical record. Our study used a powerful classification algorithm—the stacked ensemble model. Also, a comprehensive comparison between models and previous reports is presented.