Emergency department (ED) ‘frequent users’ are a sub-group of ED users that make repeated, multiple visits during a given amount of time. Though there is no consensus about definition for frequent users, thresholds in the literature range from two to more than 10 ED visits per year,1 2 while the most common one is more than four ED visits per year.1 2 Frequent users represent a small proportion of ED users but account for a large number of visits.3–5 They often display complex characteristics such as low socioeconomic status combined with physical and mental health issues.6 As such, their ED use is considered suboptimal,7 as the best strategy would be to identify those patients at an earlier stage in their health problem trajectory, in order to treat them more efficiently.8 Furthermore, frequent users’ visits may lead to overcrowding in EDs and decreased quality of care.2 Identifying factors that best describe those users and predict their ED use is therefore an essential task to improve ED care as well as frequent users’ health problems. Adequate statistical tools are needed to that end. Although they are numerous, no literature review has been published yet about statistical tools used for analysing ED frequent users. Therefore, the aim of our study was to draw up a list of statistical tools used in identifying variables associated with frequent use or predicting the risk of becoming a frequent user.

In order to list the statistical tools used in describing variables associated with and prediction of frequent ED use, we conducted a scoping review. We followed the 5-stage methodology of Arksey and O’Malley9 adapted by Levac et al.10 The methodological framework of a scoping review allows ‘mapping rapidly the key concepts underpinning a research area and the main sources and types of evidence available’,11 thus allowing us to identify gaps in the literature and future research opportunities.

The studies’ main characteristics are presented in table 1. Out of 114 studies, 65 were conducted in the USA, 17 in Canada and 8 in Australia (figure 2). The various statistical tools were classified into four main categories: regression, hypothesis testing, machine learning and other tools.

The most exploited statistical tools arguably came from regression analysis. This may be because regression is well established in medical statistics or also because it is the most natural tool when trying to find significant variables to explain a dependent variable (in this case, to be a frequent user). Moreover, it allows predicting easily the risk of a new user becoming a frequent user, depending on its covariates. Other tools from hypothesis testing or machine learning also proved to be popular, although to a much lesser extent. Combining these statistical techniques may help in discovering significant and complementary patterns, compared with using tools from one class only. In our scoping review, two studies mixed statistical tools from regression, hypothesis testing and machine learning.31 36 In those studies, the author evaluated various performance criteria. While logistic regression performed well, other techniques such as random forests or LASSO regression were also competitive. Besides the fact that logistic regression can display modest performances,128 random forests and LASSO regression can complete logistic regression. The first technique can be used to assess the importance of each independent variable in the model, while the second technique can be useful for automatic selection of features. Likewise, using a variety of statistical tools can help complete or confirm results obtained with established methodologies. Different tools from one class can also be mixed in order to achieve different stages of the analysis (for instance, different types of regression82).

The analysis of frequent ED users could benefit from using more machine learning techniques. Those were found to be not as common as regression or hypothesis testing, although they are especially appropriate when dealing with classification, prediction or big data. Tools such as support vector machines (which were used by two studies in this scoping review31 36), artificial neural networks or Bayesian networks are common classifiers and predictors in the artificial intelligence community.129 They are popular for instance in cancer diagnostic and prognosis, which strongly rely on classification and prediction.130–132 In particular, support vector machines, decision trees or self-organising maps can deal with binary outcomes, which is usually the case for frequent use outcomes. They usually require large datasets in order to overcome overfitting, but this is becoming less and less of an issue in health sciences.133 Nevertheless, machine learning tools often use a black box approach as there are many intermediary steps leading to the final solution. While each step usually consists of simple arithmetic operations, their multiple interactions can be more difficult to interpret. In spite of this opacity, they still display good performances in classifying and predicting. In some cases, they may be more accurate than the widely used logistic regression.134 Those methods would thus turn out to be less useful in data exploration.135 Machine learning tools are getting popular in other fields in health sciences, such as critical care,136 cardiology137 or emergency medicine.138 The authors state that their fields would benefit from this growing popularity, though results need to be analysed and interpreted in collaboration with clinicians.

Other tools exist that may also be suitable for describing the associated variables or the prediction of frequent ED users but were not reported in the literature. Among those, principal component analysis (PCA) is a dimensional reduction and visualisation technique, sometimes used with cluster or discriminant analysis.139 Based on all the original explanatory variables, PCA constructs new ones by summing and weighing them differently. More weight is given to relevant variables so that those latter become dominant in the new constructions while still including all variables. For instance, Burgel et al 140 built chronic obstructive pulmonary disease clinical phenotypes by constructing new relevant variables with PCA and by grouping similar subjects in this new space with cluster analysis.140 Moreover, PCA has already been used for the construction of questionnaires and diagnosis tools in a medical context,141 142 both of which can prove useful in the identification of frequent users.

As mentioned, regression techniques were common in the selected studies. Yet, quantile regression (QR)143 was not mentioned. QR is a generalisation of mean regression in the sense that its focus is not only the mean of the dependent variable distribution (such as in classical linear regression) but any quantile of it. QR thus represents an alternative to define frequent users by the high quantiles of ED visit distribution (eg, the 90th quantile). Eight studies25 27 46 48 51 54 62 121 defined frequent users with quantiles, but they did not use QR. QR would allow for finer investigations in the different quantiles of ED users in relationship to the explanatory variables. For instance, the association between age and the number of ED visits may be significantly different across the 10th (low users) and 90th (frequent users) quantiles. Such a heterogeneous association would be uncovered by QR, while usually unseen with a classical mean regression. Ding et al 144 used QR to characterise waiting room and treatment times in EDs.144 They explored the lowest, median and highest of those times and highlighted predictors that were significant only in particular quantiles. Usually, QR requires a continuous dependent variable as opposed to a logistic regression, though it is possible to combine these two regressions.145 Furthermore, defining frequent users by quantiles would allow for better comparison between studies as there is no common definition for frequent users.

Frequent ED users represent a complex issue, and their analysis require adequate statistical tools. In this context, this scoping review shows that some tools are well established, such as logistic regression and χ² test, while others such as support vector machines are less so, though they would deserve to get more attention. It also outlines some research opportunities with other tools not yet explored.