A systematic review of adult admissions to ICUs related to adverse drug events

Adverse drug events (ADE) may lead to hospital admission, and in some cases admission to an ICU is mandatory. We conducted a systematic review dealing with the incidence of ADE requiring ICU admission in adult patients, the reference population being all ICU admissions. Medline, Embase and Web of Science databases were screened from January 1982 to July 2014, using appropriate key words. Only original articles in English reporting the incidence of ADE requiring ICU admission in adult patients among total ICU admissions were included. Article eligibility was assessed by two independent reviewers, a third being involved in cases of disagreement. All reported characteristics (type of ICU, characteristics of patients, incidence of ADE, severity and preventability, drugs involved, causality) in the selected articles were collected for the review. The quality of studies was independently assessed by two reviewers with a specific score that we developed. A meta-analysis was conducted. Inclusion criteria were fulfilled by 11 studies out of the 4,311 identified in the initial literature search. The median (interquartile) quality score was 0.61 (0.44; 0.69). The reported incidences of ADE requiring ICU admission in adult patients ranged from 0.37 to 27.4%, with an associated mortality rate ranging from 2 to 28.1% and a mean length of stay ranging from 2.3 to 6.4 days. Preventable events accounted for 17.5 to 85.7% of the events. Costs and mechanisms at the root of ADE were investigated in only two and five studies, respectively. The forest plot examining the incidence of ADE requiring ICU admission in adult patients was associated with high heterogeneity (I2 statistic >98%), and the shape of the corresponding funnel plot was asymmetric. Heterogeneity across studies concerned many features, including studied populations, events considered, causality assessment methods, definitions of preventability and severity. Despite the heterogeneity of the reports, our review indicates that ICU admission due to ADE is a significant issue that should deserve further interest. The review led us to propose a list of items devoted to the reporting of future studies on ADE requiring ICU admissions. Electronic supplementary material The online version of this article (doi:10.1186/s13054-014-0643-5) contains supplementary material, which is available to authorized users.


Introduction
The incidence of hospital admissions required because of adverse drug events (ADE) has been reported to range from 0.16 to 15.7% with an overall median of 5.3% [1]. Some serious ADE may require ICU admission. Such unplanned admissions may have consequences for the healthcare system because they may deleteriously overload the ICU and restrain ICU access to other patients. Since ICU admissions generate high costs, with per-capita expenditures ranging from US$730 to 7,410 [2], while the need for critical care capacity is increasing [3], a just and effective rationing of critical care is an issue that all western industrialised countries have to consider. The abovementioned context suggests documenting the incidence of medication-related ICU admissions and the drug management problem involved. Identifying specific patterns in the population of patients admitted to the ICU for ADEfor example, in terms of mortality during ICU stay or after dischargealso constitutes an attractive issue. To gain more insight into all of these issues, we performed a systematic review of the literature on ICU admissions required because of ADE. In this review, we adopted an ICU perspective: the reference with which ICU admissions required because of ADE were contrasted had to be all ICU admissions (not the population of hospitalised patients, for example). In the following text, the incidence of drug-related ICU admissions (IDRIA) will be referred to as the ratio of the number of ICU admissions required because of ADE to the total number of ICU admissions.
As a preliminary, some issues on definitions related to ADE have to be tackled. The terminology of events related to drug harm is complex because many terms have been used in the literature [4]. A definition of ADE was proposed by the Institute of Medicine and was recommended by Nebeker and colleagues [4]: 'any injury resulting from medical intervention related to a drug'. This broad definition includes both harm caused by the drug itself (that is, adverse drug reaction (ADR), or harm related to the accumulation of drugs in the case of renal or liver failure) and harm resulting from the use of a drug (that is, medication errors). According to the World Health Organization, an ADR is defined as any noxious, unintended or undesired effect of a drug occurring at dosages administered in humans for prophylaxis, diagnosis or treatment [5]. A medication error is 'any preventable event that may cause or lead to inappropriate medication or patient harm while the medication is in control of health care professional, patient or consumer' [4]. An iatrogenic event (also called an adverse event) will be considered as 'any injury related to medical management, in contrast to complications of disease. Medical management includes all aspects of care, including diagnostic and treatment, failure to diagnose or treat, and the systems and equipment used to deliver care' [6].

Eligibility criteria
The eligibility criteria for inclusion in the review included: original articles in English (reviews, conference abstracts and case reports were excluded); a patient population over 14 years old; hospitalisation in any type of adult ICU (for example, medical, surgical, cardiac, and so forth); main outcome focusing on the ADE as a cause of admission to the ICU, whatever the patients' origin (home, emergency department, hospital ward), the reference population being all ICU admissions; articles included a description of the incidence of all ICU admissions due to ADE, the reference denominator being the total number of ICU admissions; and articles focusing on admissions to the ICU only due to selfpoisoning were excluded.

Information sources
On 4 July 2014 we conducted a combined search in the Embase database, in the Medline database via PubMed and in the Web of Science. Using relevant search terms related to ADE responsible for admissions to the ICU, we searched for all publications in English from 1 January 1982 to 4 July 2014. Complementary searches were made to identify potential additional articles: the reference lists of retrieved articles as well as their citing lists (the latter being issued from the Web of Science) were handsearched.

Search strategy
The queries made in the PubMed, Embase and Web of Science databases are detailed in Additional file 1.

Study selection
The eligibility of each retrieved article was independently assessed by two reviewers (PAJ and PH) on the title, abstract and, if necessary, full text. We a priori decided that in the case of disagreement, a third reviewer (GH) would decide whether to include the study or not.

Data collection process
A standardised data collection sheet was elaborated by three reviewers (PAJ, PH and GH). One reviewer (PAJ) fulfilled this data collection sheet and another (PH) checked the extracted data. In case of disagreements, the third reviewer (GH) would decide.

Data items
The list of data items collected in each study is presented in Table 1. Examination of all features reported was used for proposing a list of items that we judged to deserve a mention in articles related to ADE requiring ICU admissions. We redefined the numerous terms related to ADE mentioned in the selected articles using only three terms, namely ADE, ADR and medication errors (defined in Introduction). All costs were normalised in 2014 US dollars and were discounted at an annual rate of 3%.

Risk of bias assessment
To assess the risks of bias in individual studies, a specific list of items adapted to the scope of ADE responsible for ICU admission was designed by three authors (PAJ, PH and CP). The list was based on a combination of Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) [7] and Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) [8] items. Two of us (PAJ and PH) independently assessed each individual study by granting for each item either 1 point when reporting of corresponding data was complete or 0 points when corresponding data were missing or reporting was incomplete (disagreements were solved by consensus discussion). Scores per study (proportion of items completely reported) were considered for assessing the risk of bias in individual studies. Scores per item (proportion of studies with the item completely reported) were considered for assessing the risk of bias across studies (risk of bias decreases with increasing score values), as was a funnel plot representing study size against the incidence of ADE. The symmetry of the plot was visually assessed.

Summary measures
The main measure of our study was the IDRIA. The IDRIA estimate was either directly extracted from the source document or was calculated. The same process was used for obtaining the preventability rate.

Synthesis of results
Funnel and forest plot analyses were conducted, and risk of bias and heterogeneity across studies were investigated. All analyses were performed using the meta package within R statistical software [9].

Additional analysis Quality assessment
Risk of bias and quality of reporting are two different notions. We assessed the quality of reporting in the selected articles using a score based on the STROBE checklist; that is, on the quality of reporting items that should be addressed in reports of observational studies [7]. If information on a given item was completely reported in the studied article, 1 point was granted; 0 points were attributed if this information was incomplete or missing. Whenever an item was not applicable for a study, the item was not considered for calculating the study's score (that is, number of points granted/number of points considered).
In addition, to propose a more specific assessment of the quality in studies reporting the IDRIA, we designed a specific checklist ( Table 2) derived from the STROBE checklist, and the quality of reporting of the studies based on this checklist was assessed. Item scores were weighted according to our feeling about item importance, with a highest number of points (3 points) attributed to risk of bias items, and with a binary decision for each item (that is, no grading: either all points associated with the item score granted, or none). For each article, the sum of all points was calculated and the total score was defined as the ratio of the number of points granted to the maximum possible score (excluding nonapplicable items). Quality appraisal therefore varied from 0 to 1, with a score increase corresponding to a quality increase. Scoring was independently assessed by two reviewers (PAJ and PH), and we a priori decided that in a case of disagreement a third reviewer (GH) would decide whether to grant the point(s) on a given item.

Sensitivity analysis
To investigate heterogeneity in detail, analyses were conducted in which one or several included studies were removed. Such analyses included funnel and forest plots, and assessment of the I 2 statistic.

Study selection
The initial queries yielded a total of 4,311 initial records ( Figure 1). Among the 50 articles assessed for eligibility, fulltext reading resulted in the exclusion of 39 (see Additional file 2 for exclusion reasons). The whole process resulted in the selection of 11 articles [10][11][12][13][14][15][16][17][18][19][20]. A particular article raised concerns for selection: a multicentre study gathering results from eight ICUs, one of which was a paediatric ICU. However, both reviewers selected this article because of the limited impact of the paediatric ICU in the global results of this multicentre study [15]. A disagreement requiring a third reviewer concerned only one article [12]. Neither the examination of the reference lists of the selected articles nor that of their citing lists (n =183 new references screened and assessed for eligibility) resulted in the retrieval of any additional article finally included in the review.

Study characteristics
The 11 selected studies were published between 1986 and 2012 and were performed in different countries, types of hospital and ICUs (Table 3). Various features were   reported, including incidence, cause, causality, risk factors, severity, preventability, length of stay and costs.

Results of individual studies Population characteristics
Inclusion and exclusion criteria were specified in nine studies and were heterogeneous across studies. The mean length of ICU stay for patients admitted to the ICU because of ADE was reported in five articles [13,[16][17][18]20] and varied from 2.3 days [18] to 6.4 days [16]. In the study by Nazer and colleagues performed in Jordan, the average discounted cost (hospital bill for patient charge normalised in 2014 US dollars) for the management of an ICU admission related to ADE would amount to $10,388 with a median cost of $4,785 [20].
According to the payer's perspective adopted in the study by Darchy and colleagues performed in France, the average discounted cost of an ICU admission related to ADE (including iatrogenic events related to surgery or procedures) would amount to $5,606 [13].
Patients specifically admitted to the ICU for ADE were compared with patients admitted to the ICU for reasons   other than ADE in only three studies [13,16,20], and comparisons involved various criteria from one study to another (see Additional file 3).

Characterisation of adverse drug events responsible for admissions to the ICU
The mortality rate ranged from 2% [18] to 28.1% [20]. The preventability rate varied between 17.5% [20] and 85.7% [17] (Table 5 and Additional file 3). Inter-rater agreement was only reported for the judgement of the preventability of iatrogenic medical events, and only in one study [15] with a corresponding moderate interrater agreement (kappa test =0.5). The leading causes of ADE were reported briefly. Trunet and colleagues underlined that ADE were due to 12 drug-drug interactions and 31 medication errors (28 overdosages and three cases of drug prescription in spite of contraindications) [10]. Darchy and colleagues described the causes of the 30 ADE: inadequate follow-up of therapy in 46.7% (14 cases), error in dose in 26.7% (eight cases), inappropriate drug in 20% (six cases) and failure to use prophylactic treatment in 6.6% (two cases) [13]. Lehmann and colleagues categorised ADE into dosage error in 43% (nine cases), idiosyncratic reaction in 33% (seven cases), frequency error in 10% (two cases), unclassified error in 10% (two cases) and wrong drug to Description of the type of studied ICU?   patient in 5% (one case) [15]. Grenouillet-Delacre and colleagues mentioned that 87 of the 132 ADRs (66%) were caused by interaction between drugs [16]. Finally, anaphylaxis or improper drug use were mentioned in the study by the Italian Group on Intensive Care Evaluation [11]. The most frequent drugs involved in ADE in the 11 studies were cardiovascular, anticoagulant and psychotropic drugs. Most ADE were preventable but the processes that had led to the corresponding drug misuse were poorly detailed.

Causality results
As shown in Table 6, the data required for assessing causality differ from one method to another. Various causality assessments were used, and results on causal assessments of ICU admissions related to ADE were mentioned in only four articles [10,13,16,17].

Risk of bias across studies
Considering the 24 items retained for assessing the risk of bias in/across studies, nine (37%) items were not reported or were incompletely reported in at least one-half of the studies (Figure 2), with five items related to methodology (out of 15) and four items related to results (out of nine).

Synthesis of results
The asymmetric shape of the funnel plot shown in Figure 3 indicates that large studies tend to report low estimates of the IDRIA. A forest plot adopting a random effects model resulted in a final estimate of the IDRIA at 7% (6%; 8%), with a total sample size of 18,241 patients ( Figure 4). However, because of the large heterogeneity of the studies (I 2 statistic >98%), the above final estimate should not be considered a summary measure.

Additional analyses Quality assessment
The quality scores of the included studies varied from 0.36 [11] to 0.78 [18,20], with a median (interquartile range) value of 0.61 (0.59; 0.70), when assessment was based on STROBE items ( Figure 5). The scores varied from 0.33 [12] to 0.80 [16] with a median (interquartile) value of 0.61 (0.44; 0.69) when quality assessment was based on the checklist that we devised for considering specific items on ADE resulting in ICU admission ( Table 2). The differences between the two scores varied highly from one study to another ( Figure 5).

Sensitivity analysis
A first analysis was conducted with removal of the study by Lehmann and colleagues [15] (highest sample size and extremely low estimate of the IDRIA). A second analysis was conducted with the removal of the studies by Lehmann and colleagues and by the Italian Group on Intensive Care Evaluation (the two studies with the highest sample sizes and the lowest estimates of the IDRIA) [11,15]. A third analysis was conducted considering the six studies with the lowest risk of bias [10,13,[16][17][18]20] (see Table 4). The corresponding forest plots of these three analyses (data not shown) still indicated a large heterogeneity across the considered studies, and the I 2 statistic always remained above 96%, confirming such heterogeneity.

Discussion
Our review of published reports regarding ICU admissions related to ADE identified 11 original studies. The IDRIA reported ranged from 0.37% [15] to 27.4% [16].
Vlayen and colleagues proposed a review on the incidence and preventability of adverse events that required intensive care (re)admission [31]. The adverse events considered by Vlayen and colleagues not only included ADE but also surgical procedures and medical devices, the population considered was adult and paediatric patients originating from hospital wards only, and the incidence of events was reported according to the total number of stays in the studied hospital ward(s). The perspective of our review was different: the events considered were strictly ADE as a cause of ICU admission, the patient population was adult patients (whatever the origin of the patient), and the incidence was reported according to the total number of admissions in ICU. These important perspective differences probably explain the fact that nine of the 11 articles retained in our study were not included in Vlayen and colleagues' study selection [31].
Our study outlines several methodological aspects that contribute to the observed wide range of the reported IDRIA, such as study setting (types of patients and ICU under study) and the type of drug-related events considered. The definition of ADE that was proposed by the Institute of Medicine and recommended by Nebeker and colleagues [4] should be adopted as it encompasses all aspects of ADE.
The methods used for assessing preventability of ADE also varied across studies, and either implicit or explicit criteria were used. Excepted for one study gathering a large proportion of febrile neutropenia, which is not preventable [20], the preventability rate in the seven other studies was high (lowest estimate 47.7% [16]). However, a recent systematic review on the methods for assessing the preventability of ADE concludes that there is limited evidence for the validity of the identified instruments, and instrument reliability varied significantly [32]. Moreover, the single study that assessed inter-rater agreement for determining preventability resulted in a moderate agreement [15]. All in all, a reliable and reproducible assessment of preventability remains a challenge at the present time.
Another source of heterogeneity was related to the causality assessment method. Kane-Gill and colleagues compared causality assessment methods (Naranjo, Kramer and Jones) in the ICU [33]. Agreement between the methods depended on whether events were judged retrospectively or were issued from a surveillance monitoring system. At the present time, the degree of agreement/disagreement between methods remains unclear. Most of all, existing tools widely used across different countries have not been customised to assess causality in the ICU. For example, the requirement of a procedure of rechallenge and dechallenge is a key element for assessing the certainty of the causality in the Naranjo, the Karch-Lasagna and the World Health Organization assessment methods. While one can understand the value of such an element with regard to causality strictly speaking, rechallenge is not possible most of the time in patients admitted to the ICU because of the seriousness of the drug-related events and poor patient health status. Moreover, the highest score for causality with these methods implies an improvement of the clinical features after dechallenge whereas recovery after dechallenge may be not achieved in ICU patients (irreversible organ failure, long-lasting effects of ADE). The events and patients involved in intensive care are therefore not well adapted to the above causality scales and, all in all, the design of a specific algorithm tailored to ICU cases would be welcome. Nevertheless, at the present time, the French official method by Begaud and colleagues, in which rechallenge or dechallenge is not a mandatory item for high causality scores, appears the best available tool [30].
Unsurprisingly, the most frequent drugs involved in ADE in the 11 studies are the same drugs that cause admission to emergency departments [34][35][36]. However, in order to better assess the dangerousness of these drugs, the frequencies of these ADE should be contrasted with the prescribing frequencies of the corresponding drugs. Exploring these relationships constitutes an important issue that deserves investigation. These investigations would facilitate the identification of specific patterns related to preventable ADE requiring ICU admissions. In that regard, the use of effective identification and reporting systems based on a consistent terminology should contribute to the design of preventive measures for avoiding such events. In our review, quality assessment of the 11 articles not only included global STROBE recommendations but also items specific to the domain of critical care. The quality of reporting assessed according to STROBE items and according to our proposed items (see Table 2) resulted in identical median scores of 0.61 (see Figure 5), indicating that the quality of reporting for future research is substantially improvable. In addition, the within-study difference between the two above quality scores ( Figure 5) indicates that the quality of reporting with regards to the global form (quality according to STROBE guidelines) may substantially differ from the value of the information reported according to a perspective focusing on the report of ADE in intensive care patients (quality of items specific to the ICU).
In this regard, our corresponding specific items might also be considered as a potential future checklist: experts in critical care could use this list and the associated scores as an initial proposal for developing international guidelines aimed at improving the quality of future research on ADE-related ICU admissions. The use of this checklist should also decrease the risk of bias within and across studies.
This review has some limitations. The fact that only 11 studies were published during the last three decades may be related to two major elements. First, investigations into ADE in the intensive care setting require time-consuming resources and complex data. In that regard, the expanding availability of healthcare systems based on electronic medical records should facilitate future research on large cohorts of patients. Second, publication bias should be considered: one might hypothesise that studies reporting a high frequency of ADE requiring ICU admission are less likely to be published than those associated with a low estimate. The retrieved studies show a wide range for the IDRIA, however, and publication bias may therefore be limited. In addition, we only selected articles written in English, because we did not have the resources available to translate potential articles written in various other languages. Moreover, when considering a given study, only the quality of Incidence of adverse drug events (number of events / study size) Study size Trunet, 1986 [10] IGICE, 1987 [11] Nelson, 1996 [12] Darchy, 1999 [13] Hammerman, 2000 [14] Lehmann, 2005 [15] Grenouillet-Delacre, 2007 [16] Rivkin, 2007 [17] Schwake, 2009 [18] Mercier, 2010 [19] Nazer, 2013 [20] Figure 3 Funnel plot of the incidence of admissions to the ICU required due to adverse drug events [10][11][12][13][14][15][16][17][18][19][20]. IGICE, Italian Group on Intensive Care Evaluation.
reporting was assessed. The major and important limitation of the review concerns the heterogeneity observed across studies for the definitions of events (ADE, ADR) and for the methods used for causality, preventability and severity assessment. Forest plot analyses of the IDRIA reflect such heterogeneity, which remained very high in the sensitivity analyses excluding some studies. Similarly, the asymmetric shape of the funnel plot also raises concerns. Heterogeneity and biases therefore finally constitute the main issue of the present review, in contrast to other reviews for which the final estimate issued from the meta-analysis is the main result.
varied within a large range, from 0.37 to 27.4%. Severity