Drug Administration Errors in Hospital Inpatients: A Systematic Review

Context Drug administration in the hospital setting is the last barrier before a possible error reaches the patient. Objectives We aimed to analyze the prevalence and nature of administration error rate detected by the observation method. Data Sources Embase, MEDLINE, Cochrane Library from 1966 to December 2011 and reference lists of included studies. Study Selection Observational studies, cross-sectional studies, before-and-after studies, and randomized controlled trials that measured the rate of administration errors in inpatients were included. Data Extraction Two reviewers (senior pharmacists) independently identified studies for inclusion. One reviewer extracted the data; the second reviewer checked the data. The main outcome was the error rate calculated as being the number of errors without wrong time errors divided by the Total Opportunity for Errors (TOE, sum of the total number of doses ordered plus the unordered doses given), and multiplied by 100. For studies that reported it, clinical impact was reclassified into four categories from fatal to minor or no impact. Due to a large heterogeneity, results were expressed as median values (interquartile range, IQR), according to their study design. Results Among 2088 studies, a total of 52 reported TOE. Most of the studies were cross-sectional studies (N=46). The median error rate without wrong time errors for the cross-sectional studies using TOE was 10.5% [IQR: 7.3%-21.7%]. No fatal error was observed and most errors were classified as minor in the 18 studies in which clinical impact was analyzed. We did not find any evidence of publication bias. Conclusions Administration errors are frequent among inpatients. The median error rate without wrong time errors for the cross-sectional studies using TOE was about 10%. A standardization of administration error rate using the same denominator (TOE), numerator and types of errors is essential for further publications.


Introduction
Medication errors are common in the hospital setting and can lead to adverse drug events [1]. In MEDMARX, the errors reported included prescription errors (21%), medicationdelivery errors (22%) and administration errors (33%) [2]. Administration, the final step of the medication process, has been less well studied although it directly concerns nurses and patients and is the last barrier before a possible consequence for the patient.
Administration error is defined as a deviation from the physician's medication order as written on the patient's chart [3]. The denominator of error rate is generally the Total Opportunity for Errors (TOE) and could be calculated according to 2 definitions but both definitions produce identical results [4]. First, it is defined as the sum of the total number of doses ordered plus the unordered doses given [5]. According to the second definition, it is the sum of the doses given plus the number of omission errors [6,7]. The clinical impact of administration errors can be evaluated using general classification American Society of Health-system Pharmacists (ASHP) [8] or the National Coordinating Council for Medication Error Reporting and Prevention (NCCMERP) [1]. In order to detect drug administration errors, the observation technique gives more efficient, objective, and reliable results than spontaneous reporting or patient chart reviews [3,9]. However, this technique is very time-consuming and cannot be carried out for very long periods of time. Several studies have evaluated administration errors. No systematic reviews using the observation technique have been published so far [10][11][12][13][14][15][16]. The Institute of Medicine reported an extensive review of the literature on medication errors and errors rates [17] but did not focus on the observation technique.
The aim of this present systematic review is to systematically analyse the published evidence concerning the prevalence, nature and severity of administration errors in hospitals detected by the observation technique.

Search strategy
The following electronic databases were used: MEDLINE, EMBASE, and the Cochrane Library. Our search concerned published studies in any language from January 1966 up to December 2011 ( Figure S1). The reference lists of all selected studies were collected for additional studies.

Inclusion and exclusion criteria
Studies that reported the detection of medication errors using the observation method [18], and the rate of administration errors among hospital inpatients were included. Crosssectional, prospective, and intervention studies that evaluated intervention to reduce errors (before-and-after, randomized controlled) were included. Retrospective studies, abstracts, letters to editors, studies in nursing homes, and studies which evaluated errors for only one drug class were excluded. Administration errors could be detected using spontaneous reporting, review of patient charts or observation. Reporting systems suppose that the person reporting is aware that an error was made. Reviewing patient charts is time consuming [19]. Observation method is considered to be the standard for error detection as it yields more objective and reliable results [3,9,20]. Briefly, an observer follows the nurse giving the medications and notes the administration of each dose. The notes are then compared with the prescription. An error is counted when the nurse does not carry out the order accurately. Results are presented using the TOE or the doses observed as the denominator for error rate.

Study selection and data abstraction
Two reviewers (senior pharmacists) screened the title and abstract of each publication to independently determine eligibility. They screened the full text of the articles and extracted the main outcome independently. One reviewer extracted the data; the second reviewer checked the data. Discrepancies were resolved by consensus between the two reviewers after discussion. If no consensus was achieved, a third researcher arbitrated. Standardised data-extraction forms were used to extract relevant data on inclusion criteria (study design, year and country, study period, hospital setting), methods (sampling, profession of observers, number of observers, definitions used), and the rate, type, and severity of administration errors. Error rates were extracted in the control period for crossover-studies, before intervention for before-andafter studies, and in all groups in the control period in randomized controlled before-after studies (we estimated that during each study, practices may change even in the control groups after intervention). For studies that presented results as the doses observed, authors were contacted for further information so that their results could be expressed as TOE. If the authors had clarified the denominator used, we reevaluated data with TOE. Types of errors were reclassified according to the ASHP classification [8]: omission error (the failure to administer an ordered dose to a patient before the next scheduled dose, if any), wrong time error (administration of medication outside a predefined time interval from its scheduled administration time (this interval should be   established by each individual health care facility)), unauthorized drug error (administration to the patient of medication not authorized by a legitimate prescriber for the patient), wrong dose error (administration to the patient of a dose that is greater than or less than the amount ordered by the prescriber or administration of duplicate doses to the patient, i.e., one or more dosage units in addition to those that were ordered), wrong dosage-form error (administration to the patient of a drug product in a different dosage form than ordered by the prescriber), wrong drug-preparation error (drug product incorrectly formulated or manipulated before administration), wrong administration technique error (inappropriate procedure or improper technique in the administration of a drug), deteriorated drug error (use of expired drugs or improperly stored drugs) and other medication error (including any drug administration errors not fitting into the predefined categories). The ASHP classification did not specify the time limits used to determine a wrong time error: generally 30 or 60 minutes before or after the scheduled prescription time [21]. Authors [3] recommended that studies on medication errors should report both the errors rate with and without timing error as clinicians often consider wrong time error to be minor. Studies that reported clinical impact were analyzed and impact was reclassified by the two reviewers into 4 main categories (fatal, life-threatening, significant, minor or no impact) when possible. Discrepancies were resolved by consensus between the two reviewers after discussion.

Risk of bias or quality of reporting
We assessed the quality of reporting of the studies using an evaluation scale adapted from the STROBE Statement checklist [22]. Each item was filled by one reviewer and was validated by the second reviewer.

Main outcomes
The main outcome was the error rate without wrong time errors and measured at the study level. The secondary outcomes were the error rate including wrong time errors at 60 minutes and clinical impact of errors.
Studies were analyzed separately for administration doses observed and TOE, and according to study designs except for a subgroup exploratory analysis. Egger et al. recommend the results from randomized control trials (RCTs) and nonrandomized control trials (non-RCTs) to be stratified to explore potential heterogeneity due to study designs [23].
The global error rate was calculated for each study as the number of errors (or the number of administration with at least one error) divided by the TOE or the doses observed (multiplied by 100). We chose to present the results with the number of errors as numerator, as the studies presenting the number of administration also presented the number of errors. When the number of errors exceed the number of observations (Zribi Triki [24], Gokhman [25]), we modified the numerator to be equal to the denominator. For each study (when available), we also calculated the error rate by type of error using the number of errors of the considered type in the numerator (and the TOE or doses observed in denominator). In a subgroup exploratory analysis, we compared the error rates in studies evaluating injectable drugs only and oral drugs only; and error rates in intensive care units only versus other units.

Statistical analyses
Given the heterogeneity within the review, we did not undertake a formal meta-analysis. The results were reported using the median error rate across the studies with interquartile range (IQR). The median for the global error rate corresponds to the error rate above which (or equivalently below which) 50% of the studies lie for their observed global error rate. The median for the error rate for a particular type corresponds to the error rate above which (or equivalently below which) 50% of the studies lie for their observed error rate for this particular type (considering only studies reporting this type of error). Heterogeneity was presented using the Cochran Q Chisquared test and Higgins' and Thompsons' I 2 statistic. A value of 0% indicates no observed heterogeneity, and larger values show increasing heterogeneity. I 2 values greater than 75% with low P values from the Cochran Q test were considered as indicating substantial heterogeneity. To study the publication bias, we plotted the study precision (within-study standard error) against the log-odds of error rate without wrong time errors (funnel plot). If studies with large proportions of errors were less likely to be published, the funnel plot would appear asymmetric about the vertical. Funnel plot asymmetry was assessed using Egger's test [26]. Under the null hypothesis of no small study effects, the regression line of the effect against its standard error would be vertical.
The software R was used for forest plots (version 2.12.0); the software Stata was used for funnel plots (StataCorp. 2009. Stata Statistical Software: Release 11. College Station, TX: StataCorp LP; commands "metareg" and "metabias"); and the software SAS was used for descriptive statistics (version 9.2; SAS Institute Inc., Cary, NC, USA). The analysis conformed to the PRISMA checklist (Table S1).

Results
The electronic search initially identified 2088 publications. After screening titles and abstracts, 1967 publications did not meet the inclusion criteria. The remaining 121 publications were obtained in full text and assessed for suitability. Searching of the reference lists of the included publications identified one other study, published before 1966, that was eligible as well as six other publications. Overall, 66 publications were included [6,7,[10][11][12][13][14][15][16]18,24,25, (Figure  1).

Types of errors
For 23 publications, the doses observed was used as denominator to calculate the error. After contacting the authors, the denominator definition was confirmed for 5 studies, was declared to be TOE for 9 studies, and remained unanswered for 9 studies. Overall, we analyzed 14 publications using the doses observed and 52 using TOE. Types of errors could not be extracted for 4 studies: Greengold [45] (not including wrong time errors), Dean [37] (not including wrong time errors), Hynniman [6] and Schnell [7] for which only the number of wrong time errors could be extracted. Wrong time errors and wrong drug-preparation errors were the most frequent types of errors (Table 2).
For the publications using doses observed (Figure 3), the median error rates without wrong time errors were 19.7% [IQR: 9.2%-55.6%] for the 12 cross-sectional studies, 9.9% for the before-and-after study, and 4.8% for the controlled before-andafter study. There was a considerable heterogeneity among the 12 cross-sectional studies (Q=2386.4, p<0.001, I 2 = 99.5%). There was no obvious asymmetry in the funnel plot suggesting no evidence of publication bias (Egger's test: p=0.84) ( Figure  S2/B [57], Teixeira [72], Zribi Triki [24]) did not specify the wrong time delay. These 15 studies were not included in the further analyses.
Error rate including wrong time errors at 60 minutes. For the publications using TOE (Figure 4), the median error rates including wrong time errors were respectively 25.2% [IQR: 12.1%-38.4%] (N=15) for the cross-sectional studies, and 22.5% [IQR: 7.2%-23.8%] (N=7) for the before-and-after studies. For the publications using doses observed ( Figure 5), the median error rates including wrong time errors were respectively 11.8%, 12.9% and 28.0 for the 3 cross-sectional studies, and 19.8% for the before-and-after study. Analyses with wrong time delay at 30 minutes are on demand.
Exploratory analysis for studies using TOE. For this exploratory analysis, we excluded studies evaluating both injectable and oral drugs, and studies where the type of unit was unknown or where the units were mixed (medical, surgical and intensive care

Discussion
We found 66 studies which reported the rate of administration errors detected by the observation method. These studies presented different study designs and did not report the error rate using the same unit for the denominator (TOE or doses observed). In addition these studies did not consider all 9 types of errors from the ASHP, such as wrong time errors. We therefore did not perform a formal metaanalysis. Most of the studies were cross-sectional studies (N=46, 70%). The median error rates without wrong time errors reached 10.5% for the 34 cross-sectional studies using TOE and 19.7% for the 12 cross-sectional studies using doses observed. Median error rates tended to be lower for interventional studies (RCT studies, before-and-after studies, cross-over study).
Clinical impact was evaluated for 32 studies (48%) and classification was heterogeneous. No fatal error was observed and most of the errors observed were minor.
To our knowledge, this is the first systematic review of administration errors detected by the observation method. In a review, Ghaleb [81] identified 32 studies reporting the incidence of medication error, of which 8 concerned drugadministration errors detected by observation. Error rates varied from 0.6% to 27% of administrations but the study selection was not exhaustive. Miller [82] conducted a systematic literature review on medication errors for children between 2000 and 2005; 14 studies evaluated administration errors and none of them used the observation method to determine error. The only study [83] using TOE for error rate showed an administration error rate of 23.5%.
The extreme heterogeneity of studies definitions (definition of error rate with number of errors or number of administrations with at least one error, TOE or doses observed, error classification) is well known [17]. That is why we focused on one error identification methodology (direct observation). But heterogeneity was also found within this methodology. A major source of the high heterogeneity between studies could be the calculation of error rate. The types of errors considered in many studies did not use the ASHP classification and these studies therefore did not consider all possible types of errors. As a consequence, the definition of error rates (numerator and denominator) varied according to the study, in particular the definition of wrong time error. However, the definition of the error rate without wrong time errors remained heterogeneous among the studies because they did not consider all other types of errors. These other types of errors seemed to be less reported because they were less frequent. For example, deteriorated drug errors (use of expired drugs or improperly stored drugs) were reported in only 28% of studies using TOE but the median error rate in these studies was only 0.1%.
Our review has some limitations. We selected studies in which the observation method, either disguised or undisguised, was used to detect drug errors. The disguised observation may be the gold standard to evaluate error rate. However, the information whether the observation was disguised or not was available in only 49 studies (74% of the studies included). There was no evidence of difference between the two types of observation method.
Important characteristics such as the type of units and the route of administration of drugs (injectable and oral) could only be evaluated in an exploratory analysis as the number of studies was too small. However, there did not seem to have difference in the error rates in different subgroups.
We used an evaluation scale adapted from the STROBE Statement checklist [22] to evaluate whether information was reported. However, no conclusion was made as to whether it was done appropriately in the study. We investigated the presence of publication bias using funnel plots. These analyses indicated there was no evidence of publication bias, despite that publication bias in a review may not always result in an asymmetrical funnel plot [84]. The calculation of error rate was not homogenous: some authors used the number of errors and others used the number of administrations with at least one error. We only presented the results with the first definition (number of errors) as most of the studies presented this type of rate. Authors described the number of errors by type and calculated the error rate without wrong time error by subtracting the number of wrong time errors from the total number of errors. When the number of errors exceeded the number of observations, we evaluated that the error rate reached 100%. The error rate appeared high even when it did not include wrong time errors. Information on how serious these errors were and the rigor of error definition was not available in all of the publications that we selected.
Standardization of administration error rate using the same denominator (TOE) and types of errors remains essential for further publications. To calculate an error rate, we recommend excluding some types of errors, and reporting the number of administrations (TOE) with at least one error and the number of errors and their types according to ASHP classification (9 types). We also advise reporting of the time limits used to determine wrong time errors. If the study implicates an evaluation of medication process, we recommend distinguishing the error rate per process (order, dispensing, and preparation/administration). Finally, the characteristics of hospital need to be presented (country, types of units, delivery system, characteristics of nurses observed).
We did not evaluate interventions to improve administration errors which are important pieces of information for the clinician. The Canadian Agency for Drugs and Technologies in Health [85] and the US Agency for Healthcare Research and Quality [86] recommend the use of technologies to reduce administration errors. This may include improving the dispensing process with an automated bedside dispensing machine or unit-dose dispensing, or improving the administration process by using dedicated medication nurses with specific training, education on drug safety, medication charts, or a bar-code medication administration system. The impact of these technologies needs to be more evaluated.
In summary, administration errors are frequent among inpatients. The median error rate without wrong time errors for the cross-sectional studies using TOE was about 10%. No fatal error was observed and most errors were classified as minor in the 18 studies in which clinical impact was analyzed. We did not find any evidence of publication bias. A standardization of administration error rate using the same denominator (TOE), numerator and types of errors is essential for further publications.