like yesterday's count than those in previous weeks. New control chart methods for monitoring MROs in. Hospitals. Anthony Morton MSc(Appl), MD, Infection ...
Australian Infection Control
New control chart methods for monitoring MROs in Hospitals Anthony Morton MSc(Appl), MD, Infection Management Services, Princess Alexandra Hospital, Brisbane and
School of Mathematical Sciences, Queensland University of Technology
Michelle Gatton BSc(Hons), PhD, Queensland Institute of Medical Research Edward Tong BA, BSc (Hons), Centre for Health Related Infection Surveillance and Prevention, Queensland Health Archie Clements MVM, PhD, Centre for Health Related Infection Surveillance and Prevention, Queensland Health and
School of Population Health, University of Queensland
Abstract Routine surveillance of colonisations with multiple antibiotic resistant organisms (MROs) is now widespread and these data are increasingly summarised in control charts. The purpose of their analysis in this manner is to provide early warning of outbreaks or to judge the response to system changes designed to reduce colonisation rates. Conventional statistical process control (SPC) charts assume independence of observations. In addition, there needs to be a run of stable, non-trended (stationary) data values to obtain accurate control limits. Colonisation with an MRO is not an independent event as it must involve transmission from a carrier and this can lead to excessive variation. In addition, non-linear trends are often present and MRO prevalence data display temporal correlation. The latter occurs when data at particular times are more like data at related, usually contiguous times, than data from more distant times; thus they are not temporally independent. These characteristics make it difficult to implement conventional SPC charts with MRO data. To overcome these problems, we suggest the use of generalised additive models (GAMs) when there is no temporal correlation, as with new colonisations, and generalised additive mixed models (GAMMs) when temporal correlation is present; as occurs commonly with prevalence data. We illustrate their use with multi-resistant methicillin-resistant Staphylococcus aureus (mMRSA) prevalence and new colonisation data. These methods are able to deal with excess variability, trends and temporal correlation. They are easily implemented in the freely available R software package. Our analysis demonstrates an upward non-linear trend in mMRSA prevalence between January 2004 and October 2006. The mMRSA new colonisation data also display an upward trend between September 2005 and May 2006. Monthly new colonisation rates exceeded the upper control limit in April 2005 and equalled it in May 2006. There was a modest downward trend in the new colonisation rate in the latter part of 2006.
Introduction MROs are an increasing problem in many hospitals causing patient injury and death, increased lengths of stay and economic loss 1. Monthly new colonisations with MROs are routinely monitored using a variety of techniques including control charts 2; this assists with the timely detection of outbreaks and in judging responses to system changes designed to reduce colonisation rates. In addition, MRO prevalence, often referred to as MRO ‘burden’ or ‘colonisation pressure’, is being monitored increasingly because of its importance in transmission 3. Conventional monthly control charts require independence of observations and a run of stable, non-trended (stationary) monthly values, preferrably as many as twenty-four, so that accurate centre lines and control limits can be calculated. Stable, non-trended 14
data require an underlying system that is behaving predictably and is therefore in statistical control. This is usually achieved by the successful implementation of evidence-based systems that are increasingly implemented as ‘bundles’4. These have been successful in reducing rates of surgical site infections and device-related bacteraemias 5. However, to date they have been less successful with MRO colonisations, making runs of stable non-trended (stationary) data difficult to achieve. In addition, MRO colonisation is not an independent process as it is due to transmission from another carrier. Excessive variability is a common consequence of this lack of independence. Temporal correlation (autocorrelation) is a feature of MRO prevalence data. For example, today’s MRO prevalence count tends to be more like yesterday’s count than those in previous weeks. Vol 12
Issue 1
March 2007
Australian Infection Control
All of these factors make the use of conventional control charts difficult and consequently their use may give misleading information. Here we show that GAMs 6 for data that do not demonstrate temporal correlation and GAMMs for data that are temporally correlated provide alternative analysis tools to control charts for these types of data.
Methods Daily prevalence of mMRSA, non-multi-resistant methicillinresistant Staph. aureus (nmMRSA), UK epidemic methicillinresistant Staph. aureus (ukEMRSA), multi-resistant Klebsiella pneumoniae, vanconycin-resistant Enterococcus (VRE) and multiresistant Acinetobacter has been recorded at the Princess Alexandra Hospital (PAH), Brisbane, since November 2003. The data from November 2003 to mid-January 2007 are complete for 89% of those 1169 days. Interpolated daily prevalence values, estimated from the data immediately before and after the gaps, have been substituted for the missing data. Monthly counts of new colonisations with these organisms are the subject of the routine surveillance program at PAH. Swabbing of all ICU patients occurs twice-weekly and once-weekly for patients in other high-risk areas as recommended by the Australian Infection Control Association National Advisory Board. The data for January 2004 to December 2006 inclusive are complete and have been employed in the study. We illustrate the analysis of the mMRSA data using GAMs and GAMMs 6. These models are described in the Appendix. The average mMRSA prevalence for each month was determined and rounded to the nearest integer value. Analysis of these data between January 2004 and December 2006 was then performed. To account for the presence of temporal correlation, we have employed the GAMM in the mgcv package 6 in R 7. In addition an approximate upper two standard deviation equivalent control limit was calculated. There is further description in the Appendix. An occupied bed days (OBDs) denominator was included in the analysis. The monthly number of new mMRSA isolates from January 2004 to December 2006 was analysed using a GAM as described in the Appendix. The required R commands for these analyses are available from the corresponding author. Because some infection control practitioners may have difficulty with a statistics package, as opposed to familiar spreadsheets that are incapable of implementing GAM methods, these commands will be automated in easy-to-use functions.
Results Monthly mMRSA prevalence for the study period is displayed in Figure 1. There was significant temporal correlation (the correlogram is not shown). The GAMM model produced a highly significant result 16
(P