Setting
In France, the physician is entirely responsible for prescriptions, including specification of the brand name of the drug (rather than its international denomination), infusion time and solution for reconstitution of intravenous medication. In this context, a pharmacist must alert the prescriber in cases of unavailability or non-conformity with best practice. However, the prescriber cannot modify the prescription directly, with the exception of replacing one drug with another having the same international denomination. Nurses should administer exactly what is written on the prescribing order.
Georges Pompidou European Hospital (HEGP) is a French tertiary care university hospital with 717 beds. A patient information system, integrating an electronic patient record and a CPOE (Dx-Care®, Medasys™) is implemented throughout the hospital since its inception in 2000. Dx-Care® is at the centre of care delivery. It is used by doctors, pharmacists and nurses:
-
to prescribe laboratory examinations and imaging tests for a patient,
-
to visualize the results of laboratory tests,
-
to establish and to consult nursing schedules,
-
to archive a structured observation,
-
to prescribe drugs,
-
to validate prescriptions by pharmacists (pharmacy validation).
The drug prescription facility is available in 17 medical wards, 506 beds, 70% of the hospital's beds. The remaining wards, which do not use the CPOE are the ones for oncology (15% of hospital beds) and for emergency or intensive care (15%). Pharmacy validation is carried out daily, from Monday till Friday, in 7 wards (148 beds) out of the 17 which use the CPOE: immunology, nephrology, vascular medicine, geriatrics, diabetes care and internal medicine (2 wards). It is performed twice a week in 7 other wards (300 beds), and the drug orders of the 3 remaining wards (58 beds) are not reviewed at all by any pharmacist.
Five full-time pharmacists are involved in pharmacy validation. Nights' prescriptions are reviewed on the next day and weekends' prescriptions are reviewed on Monday if unexpired.
We define a prescription as a list of drug orders made per day by physician for one patient. For each drug order the physician has to precise: drug name, dose, unit, reconstitution process, route and optional annotation in a plain text field. In addition, the physician has to choose frequency and duration of the order. Various types of prescription aid are available: information about reconstitution processes for intravenous drugs, typical orders pre-specified by pharmacists for intravenous drugs and an integrated drug-drug interaction system are to be targeted by the prescriber. Only alert concerning maximum dose for oral drugs are actively targeted by the system without request by the prescriber.
The pharmacist analyses each patient's prescription, drug order by drug order (dose, unit, time to take, route, frequency in a day, reconstitution process) and the prescription as a whole, testing for drug-drug interactions. The pharmacist has access to biological data and patient record. This analysis is performed daily so that a drug order prescribed for 3 days is validated 3 times.
In case of prescribing error, the pharmacist posts a message, which can be visualized by prescribers through an 'accepted' or 'refused' symbol inserted next to the order line. The 'accepted' symbol indicates that the pharmacist agrees with the prescription, unless a comment is added relating to good practice, which may or may not suggest a modification of the prescription line. The 'refused' symbol indicates that the pharmacist disagrees with the prescription line, having identified a potential severe prescribing error and suggesting its correction. The physician may click on the symbol to visualize the pharmacist's comment, but is not obliged to take that comment into account. We define as alert, any line with a 'refused' symbol or an 'accepted' symbol associated with a comment from the pharmacist and which corresponds to a potential prescribing error. The prescriber can choose to ignore the alert and maintain his/her order along with the pharmacist symbol: in this case we considered the alert as overridden. Alternatively, the prescriber may take the alert into account. Then, the prescriber can either discontinue the order or modify it by cancelling the order line and recreating another one. It is therefore possible from the pharmacy validation database to analyse whether or not an alerts has been overridden the following day.
Collection of data
The 7 participating wards are those where pharmacy validation is performed daily. Prescribers of these wards were 24 physicians. All data were collected prospectively by pharmacists while validating the prescriptions. According to French regulation, this study that aimed at improving quality of care did not require to be approved by any research ethic committee. The data collection has been approved by French Data Protection Authority. Pharmacists collected prescribing errors detected for each patient admitted in these 7 wards between June 26 to July 13, 2007. We defined as "
new prescribing error" any error appearing for the first time in the patient's prescription whether at admission or in the following days. If this error was maintained the next day, it was not any more considered as "new". Only new errors were recorded and for each of them, their chronology of appearance during the patient's stay (
i
th
day), their type and degree of severity, as defined according to a commonly used classification (see Additional file
1) [
3,
6,
8,
14,
15]. For each patient admitted during the study period we also collected data regarding renal impairment, hypertension, thromboembolic disease, as these are clinical contexts known to be associated with more prescribing errors [
16,
17].
Since pharmacy validation is provided from Monday to Friday, any prescribing error occurring during the week-end was only collected if still present on following Monday. Thanks to the Dx-Care® program we still had access to the exact date of this error, therefore to its chronology in hospital stay. Our first objective consisted in describing the chronology of all new prescribing errors.
Our secondary objective relates to alerts posted by pharmacists in response to new prescribing errors. The alert was considered as overridden if repeated on following days for the same order, consecutively to the absence of any modification of prescription neither evolution of clinical context.
Statistical analysis
Primary outcome
We first described the rate of new prescribing error according to its chronology in the patient's hospital stay. The day of admission was noted as day one.
Then we modelled the time-errors relationship in the first seven days of stay with multivariable models (we ruled-out prescriptions recorded after the seven first days of stay as the number of new errors decreased significantly after this day inducing estimation problems). We tested patient medical data (renal impairment, hypertension and thromboembolic disease), wards, number of order lines, day of discharge as potential confounders of the model. Renal impairment was defined according to the estimated glomerular filtration rate (less than 79 mL/min/1.73 m2). Hypertension was defined according to the arterial blood pressure (systolic blood pressure of 140 mm Hg or greater or a diastolic blood pressure of 90 mm Hg or greater). Thromboembolic disease was defined by the prescription of an anticoagulant drugs.
We modelled the risk of an error on i
th
day of stay with a Generalized Estimating Equation (GEE) regression model taking account for dependency between repeated measurements in a stay. In addition, a mixed Poisson regression model was performed to examine whether the number of new prescribing errors per order lines was related to i
th
day of stay. We presented the results from the Poisson regression model as estimation of the number of new prescribing errors per 10 order lines since the median number of order lines in a prescription was 7. For both models, a backward selection process with all potential confounders was used to reach the final multivariate model, and the i
th
day of stay was introduced as a discrete or a continuous variable with possibly mathematical transformation according to the minimum deviance criteria.
Secondary outcome
To find which characteristics of new alerts were the most predictive of alert's overriding, we performed a classification and regression tree (CART) analysis. This method developed by Breiman et al. [
18] consisted in algorithms with logical "if-then" conditions for predicting or classifying cases. We chose this method since it produced simple decision rules allowing pharmacists to classify prescribing errors with low or high risk of alert's overriding.
We removed new alerts occurring on the day of discharge or after the fifteenth day of stay because it was not possible to allocate them a status (overridden the next day or not). Additionally, weekends' prescriptions were not present in our sample since the pharmacist didn't review prescriptions on these days. The potential variables of interest included in the model were medical history (renal impairment, hypertension, thromboembolic disease), wards, Anatomical Therapeutic Chemical classification drug, number of order lines, type of alerted error, severity of alerted error and the
i
th
day in the stay. The GINI criterion was used to determine the best split at each node. The tree was pruned according to the one standard error rule with a 50-fold cross validation procedure [
18]. Since we aimed to maximize the number of true positive alerts (i.e. alerts predicted as overridden which would actually be) and to minimize the number false positive alerts (i.e. alerts not overridden the next day which are predicted as overridden), we respectively maximised positive predictive value and specificity introducing weighing factors of misclassification.
All statistical analyses were performed using SAS 9.1 (SAS Institute, Cary, North Carolina, United States) and R software (version 2.7.2).