Socio-demographic and clinical characteristics of re-presentation to an Australian inner-city emergency department: implications for service delivery

We examined 24 months of data for all attendances at one ED. The definition of re-presentation was established as attendance at the same ED within 28 days from discharge date from hospital (either from the ED or inpatient unit). This definition of re-presentation is congruent with that of hospital re-admission, which is also defined as occurring within 28 days of discharge.

We used a cohort design to extract and retrospectively analyze contemporaneously collected information from a database of ED presentations to one inner-city ED over a 24-month time frame. The 24-month time frame allowed for the possibility that seasonal differences might influence ED attendance rates. Since October 1995, all publicly-funded EDs within the jurisdiction where this study was conducted, are required to provide standardized information about ED presentations to the State Government through the Emergency Minimum Dataset (EMD) [48]. At the study site, reporting to the EMD is facilitated by the Patient Administration System (PAS), which includes clinical data collected for the use of the service. Data collection practices have been refined to be practical from both a clinical and an organizational perspective.

The study was approved by the Human Research Ethics Committee (HREC) of the health service in which it was conducted. The protocol was subsequently registered as a project with external approval by the collaborating university HREC.

The hospital in which this study was carried out is located in a large Australian city with a population of 3.5 million people. The hospital receives its income from State and Federal Government funding and is located within 5 km of the central business district. It is a major teaching hospital, with an ED attendance rate of approximately 32,000 per annum.

Participants included all patients who attended the ED during the study period (January 1, 2003 – December 31, 2004). The only exclusion criteria for the study were patients who had a discharge outcome of death. We considered the inclusion of admitted patients a key factor in the study design since admitted patients often have complex healthcare needs that have implications for discharge planning. Those presentations in December 2004 that had subsequent re-presentation in January 2005 were identified and included in the study. Patient data were retrospectively accessed for analysis from the PAS.

Data in the PAS contain a unique patient identifier code (unit record number) and detailed demographic and clinical information for each ED presentation. In this study the unit record number was used to classify patients into two primary groups; re-presentation to the ED within 28 days of discharge from hospital or no re-presentation. Missing data from PAS were identified when frequencies showed either inconsistent results or missing information. For analysis, missing data were entered into a category labeled as "unknown".

A standardized data abstraction tool was devised by identifying the key variables and their defined categories within the EMD and replicating this information into a Microsoft Access program (Version 2003, Redmond, Washington). The variables and defining categories that were used in the EMD are displayed in Tables 1, 2, 3, and 4.

For this study, homelessness comprised various living arrangements such as residing in a public place, a boarding house/hostel, a homeless shelter or having no fixed address. For each ED presentation at the study site, information about housing status was captured in the PAS by the variable, usual accommodation. To ensure the accuracy of this variable the patients identified as homeless had further examination of their medical records to confirm and identify the level of homelessness.

For the purpose of analysis, Chamberlain's [49] classification of housing status was used to create five categories. Primary homelessness included people living on the streets or in squats. Secondary homelessness included people living in crisis accommodation. Tertiary homelessness included people living in boarding houses. Marginalized housing included people residing in public housing requiring rental assistance or those with other unstable rental arrangements. Stable housing included people living in permanent accommodation.

Due to the volume of data generated by an electronic report over the 24-month period, three abstractors were used to collate the data for those patients who were homeless. Training was conducted over two days and review of abstractor accuracy was performed on a daily basis. One abstractor took the position of leader in the data collection process and examined 80% of all data collected. All abstractors were blinded to the etiologic relationships being investigated in the study.

The abstractors conducted an independent review of the first 20 medical records for the purpose of determining inter-rater reliability. Inter-rater reliability was examined in the coding of the homeless variable. The percentage of agreement for the category homeless or not homeless was 96.7% for 3 raters, 20 cases and 2 categories. The percentage of agreement for the level of homelessness was 73.3% for 3 raters, 20 cases and 5 categories. In 12 of the 20 cases all 3 raters agreed with the 5 categories and with the other 8 cases 2 of the raters agreed. The level of homelessness was the only variable that required an assessment from the abstractors. It was not difficult to assess that someone was homeless but there was a degree of difficulty in determining the level of homelessness.

All data were entered into a Microsoft Access database and were subsequently analyzed using SPSS, version 14.0 (Chicago, IL). The ED presentation data (n = 64,147) were examined by constructing a frequency cross-tabulation of re-presentation (yes/no) against each categorical explanatory variable (factor), and by constructing a table of means for each continuous explanatory variable (covariate). This information containing ED presentation data was aggregated into a patient dataset (n = 40,942) that contained the percentage of attendances resulting in re-presentation for each patient. The patient data were examined by constructing a table of the mean percentage for each factor. Because of the unequal numbers of presentations per patient (min = 1: max = 152), the percentage of all visits resulting in re-presentation was larger than the mean percentage calculated across patients, for each category of each factor. For example, if we consider patients who were pensioners (n = 28,660), 25.9% of their visits resulted in re-presentation, but the mean percentage calculated across these patients was 10.9%. Because the probability of re-presentation is correlated between visits within a patient, neither of the above figures is an appropriate estimate of the true percentage, p. The percentage of all visits is an overestimate of p, because it gives equal weight to each visit, and so is inflated by patients with a large number of visits. The mean percentage across patients is an underestimate of p, because it gives equal weight to each patient, and so effectively ignores the additional information provided by the multiple visits of many patients.

We dealt with the potential problem of underestimation or overestimation of re-presentation by developing a generalized linear mixed model (GLMM) with a binomial distribution. This model uses all the data to estimate the correlation between visits within a patient, and gives an estimate of p which gives an appropriate effective weighting to visits and patients. In the example involving pensioners, the estimate is 17.5%, which is between the two percentages mentioned. For this example, the estimate is about half-way between but in other cases it can be much more toward one side, depending on the correlation for the group of interest. The GLMM used for this study was a logistic regression model with the patient variable fitted as a random effect to account for repeated presentations. The fitting of the model was performed using GenStat (8^th edition, VSN International, UK).

For each explanatory variable on its own, a GLMM was fitted with re-presentation (yes/no) as the response variable. The P-value for testing the hypothesis of no effect was determined by comparing the Wald statistic with the appropriate chi-square distribution. Variables for which the P-value was greater than 0.01 were excluded from further model fitting. All the remaining explanatory variables were entered into a backwards stepwise procedure, in which the least significant variable at each stage was progressively removed from the model, until all variables still in the model were significant at the 0.01 level. The 0.01 cut-off was chosen rather than the more common value of 0.05 because the large number of observations meant that a small P-value could arise from an inconsequential effect size.

The estimated parameters and standard errors from the final model were used to calculate adjusted odds ratio (OR) and associated 95% confidence intervals for each explanatory variable. There were twelve of them – seven related to patient characteristics, and five related to ED visit characteristics. For explanatory variables not in the final model, unadjusted OR and 95% confidence intervals were calculated.

Although the housing status comprised five categories, there were no significant differences between the four homeless levels in terms of their re-presentation percentages, so in the final model they were combined into a single category. However, because of the importance of this variable to our study, a separate model was fitted with all five categories, and adjusted OR and confidence intervals calculated.