Study design and setting
This observational monocentric study was conducted over a one-year period, from February 24, 2013, to February 23, 2014, in a 18-bed medical ICU of a university-affiliated 760-bed hospital (Saint-Antoine Hospital, AP-HP, Paris, France). Saint-Antoine hospital is a multidisciplinary hospital with specialization in hepato-gastroenterological and onco-hematological diseases and a solid tumor and hematopoietic stem cell transplantation expertise. The hospital is also equipped with a digestive surgical ICU (the present study was not conducted in the latter unit). Patients were followed up from their admission in the ICU through to their discharge from the hospital.
The inclusion criterion was the following: all ICU admissions occurring during the study period. Exclusion criteria were age under 18, patient refusal or impossibility to investigate the treatment or medical history and patients admitted and discharged during the same weekend. Patients admitted for external care (bronchoscopy, renal replacement therapy for chronic renal failure or central venous catheter insertion) were also excluded from the study. Readmissions were considered as new ICU admissions and analyzed accordingly.
The need for a written consent was waived since it was a study on usual care, without any specific intervention. All patients and relatives were informed that anonymous data could be used for academic research. The study was approved by an institutional review board (Commission d’Ethique de la Société de Réanimation de Langue Française, Paris, France).
Two investigators, a pharmacist (PAJ) and an ICU physician (CP), screened independently all ICU admissions during the morning staff meeting from Monday to Friday. After reviewing each medical chart, they independently sorted all included admissions into two groups: the ADE group (ICU admissions due to ADE) and the control group (ICU admissions for a matter other than ADE). In case of disagreement, BG (Professor of Intensive Care) or CF (Professor of Clinical Pharmacy) classified the admission as related or not to an ADE.
For both assessors, the causal relationship between drugs and clinical features was assessed according to chronological, semiological and bibliographical data. Chronological data included the chronology of the events (drug administration or interruption and clinical signs), assessment of drug exposure at the beginning of the first clinical signs, taking into account drug pharmacokinetics. Additionally, clinical consequences (recovering or not) after drug dechallenge/rechallenge were assessed when possible. Semiological data were based on determining the possible etiologies of the observed clinical signs and on specific laboratory test results. Bibliographical data search was mainly based on the Summary Product Characteristics (SmPC), and in the case of lack of information in the SmPC, additional sources of data were used (Micromedex® and/or search in Embase database or MEDLINE database via PubMed).
ADEs were classified as preventable or unpreventable according to the Schumock and Thornton modified criteria (see the subsection Data sources/measurement—preventability in the “Methods” section) . Preventable ADEs were subcategorized into 3 classes according to the cause of the event: drug overuse, underuse and misuse. Drug overuse stood for situations in which potential for harm exceeded the possible benefit. Drug underuse corresponded to failures to detect diseases or to use proven effective treatments. Drug misuse corresponded to an appropriate treatment with occurrence of a preventable complication . Non-compliance and self-medication were considered as misuses.
ICU admissions related to non-compliance and self-medication were included in the ADE group, whereas admissions related to self-poisoning were included in the control group.
For all included ICU admissions
The following data, available in medical records, were collected for every included ICU admission: age, gender, comorbidities, patients’ origin (home or hospital), main reason for admission (according to the 10th International Classification of Diseases)  and LOS or vital status at ICU and hospital discharge. The Simplified Acute Physiology Score II  and the Sequential Organ Failure Assessment  were calculated within a period of 24 h after ICU admission. The McCabe score was used as an integrative index of the severity of underlying medical condition . The need for and the number of organ supports were recorded (invasive or noninvasive mechanical ventilation, vasopressors, massive transfusion or renal replacement therapy).
Patient origin was considered as home if he/she had been admitted into the ICU through the French Emergency Medical Aid Unit or after visiting Emergency Department. In any other cases (i.e., transfer to the ICU from another unit or from another hospital), patient origin was considered as hospital.
In order to identify all drugs prescribed before admission and self-medication and to detect non-compliance, the patients and/or their relatives were questioned and medical records were analyzed. Whenever drug prescriptions were unavailable, general practitioners, specialists or pharmacists in charge of the patients were contacted. Whenever patients were hospitalized before ICU admission, the list of all drugs administered during the hospital stay was retrieved. A dichotomous categorization of drugs was considered according to the delay between ICU admission and drug prescription: either at most a month or more.
ICU and hospital costs are expressed in euros for the year 2013. These were the direct costs of the hospitalization adopting the perspective of the payer, i.e., the total fees of the admissions invoiced to the payer. Costs were directly issued from the administrative database of the hospital and are based on the French Diagnosis-Related Group system specially adapted to ICU admissions including the need for organ support .
Specifically for the ADE group
Drug classes and clinical features
For each drug suspected of being involved in an ADE, the investigators recorded the following data: route of administration, daily schedule, starting time and end of treatment, and biological parameters. Drug classes were coded according to the Anatomical Therapeutic and Chemical (ATC) classification system .
Main ADE-related organ failures were classified according to the System Organ Class (SOC) codes of Medical Dictionary for Regulatory Activities (MedDRA) version no. 18.0 (Additional file 1: Table S1).
Except for admissions due to non-compliance and lack of treatment, the relationship between a drug and an ADE was evaluated using 3 standardized causality assessment methods: the official French method  and 2 international methods, the Naranjo  and the Karch and Lasagna methods . ICU admissions were allocated to the ADE group if a drug was very likely (or certain), likely (or probable) and possibly involved according to at least 2 out of 3 causality assessment methods. Whenever several drugs were involved in a single ICU admission, the strongest causality link was retained to characterize the admission.
The preventability of ADE was assessed according to Schumock and Thornton  with an additive criterion (ADEs due to the lack of treatment were considered preventable) and a modified criterion (ADEs were considered preventable if the type of drug–drug interaction was contraindicated). Whenever several ADEs were involved in a single ICU admission, the admission was considered as unpreventable if at least one ADE was unpreventable.
Both above-mentioned investigators independently identified the leading causes of preventable ADE according to a pre-established list. In the case of several leading causes for a single admission, the principal leading cause was retained. The cause of admission was then subclassified as related to drug underuse, overuse or misuse.
Definitions used for ADE, ADR, ME, assessment methods for preventability and causality
are presented in Additional file 2: Table S2.
As described above, the sorting of patients into one of the two groups (ADE or control) was made according to 3 causality assessment methods. These methods are based on the literature data and on chronological, semiological and pharmacological analysis. Moreover, this sorting was assessed by two independent investigators who both intended the daily ICU medical staff.
Study size was planned to obtain a reasonable confidence interval of the incidence of ICU admissions caused by ADE. Assuming such an incidence to be at 15 % (a rough estimate issued from a pilot experiment conducted in our ICU) and about 700 ICU admissions that would be included in the study in 1 year, the period of a one-year study was retained. According to Agresti and Coull method , such a study would a mean incidence estimate of [95 % confidence interval] 15 % [12.5; 17.9].
Each drug of a chemotherapeutic combination was considered individually for the determination of the number of drugs taken before ICU admission. However, for the determination of the number of drugs involved in the ADE, the generic term “antineoplastic agents” was used and the chemotherapeutic combination was counted as one drug in the case of drug-induced neutropenia or tumor lysis syndrome. Indeed, these ADEs were considered as a consequence of a combination rather than of a specific antineoplastic agent.
The number of comorbidities, organ supports and drugs involved in the ADEs were categorized into four groups (0, 1, 2 and ≥3) and were analyzed as categorical variables.
Hospital LOS, mortality and costs were censured at discharge from Saint-Antoine hospital.
The incidence of ADE-related admissions was calculated as the ratio of the number ADE-related admissions to the total number of included ICU admissions.
Results on quantitative variables are presented as medians and inter-quartile ranges, and those on qualitative variables are presented as numbers and associated percentages. Group comparisons on quantitative and qualitative variables were performed with the Mann–Whitney and Wilcoxon test and the Fisher exact test, respectively. Since multiple comparisons involving 3 groups were performed, the threshold P value of 0.0166 was considered for statistical significance, according to the applied Bonferroni correction. Inter-rater agreement (ADE vs control group, 3 causality assessment methods and preventability assessment) was assessed with the kappa test . Statistical analysis was performed with R Studio software (version 0.98.490).
As our study was focused on the burden of admissions caused by unintentional medication related problems, all other admissions including self-poisoning were put in the control group. Nevertheless, a sensitivity analysis excluding self-poisoning cases was conducted in order to (1) be consistent with previous studies which had excluded such cases [7, 11, 13, 14, 17] and (2) explore the impact of self-poisoning cases on the study results.