This is not a human subject's study. Protocol registered in error. This protocol studied nursing homes as a unit.

Medications are the single most common form of treatment in the long-term care setting, and often represent the most efficacious (and cost-effective) therapeutic modality used in this clinical setting. However, the residents of nursing homes are at increased risk for experiencing adverse drug events. This risk is increased by the physiologic decline and pharmacologic changes that occur with aging, and also by the special clinical and social circumstances that characterize institutional long-term care. In a study funded by the National Institute on Aging (AG 14472), we have previously determined that adverse drug events are common and often preventable in the nursing home setting and that the more serious the adverse drug event, the more likely it is to be preventable. This study will test whether a computer-based clinical decision support system can lower the rate of adverse drug events (ADEs) and potential ADEs in the long-term care setting. The study design is a randomized trial based in the resident care units of two large long-term care facilities. Within each facility, half of the resident care units will be randomized to an intervention arm receiving the computer-based clinical decision support system which will display warnings, messages, and prompts based on resident and drug use characteristics; with over-rides by the prescriber required for some warnings. Rates of ADEs and potential ADEs will be tracked by the study's on-site clinical pharmacists prior to and during the intervention period. Rates will be compared between units receiving and not receiving the computer¬based clinical decision support system and to baseline, pre-intervention rates in the same units. We will track all project costs directly related to the development and installation of the computer-based clinical decision support system. We will also develop and test the sensitivity and specificity of a computerized adverse drug event monitor and assess the validity of a nursing home resident risk model developed in our prior study of adverse drug events in the nursing home setting.

This study was conducted in two large, academic long-term care facilities located in Connecticut and Ontario, Canada. The two facilities have a combined total of 1,229 beds. Patients residing in areas of the facilities related to short-term care (e.g., subacute care, hospital-level care, or rehabilitation) were not included in the study population.

Each of the facilities had an existing computerized provider order entry system without a computer-based clinical decision support system. At the time of the study, approximately 90% of new medication orders were entered using the system. All medication prescribing was performed by contracted staff; in one of the study facilities this included 27 physicians, nurse practitioners, and physicians' assistants. In the other facility, medication prescribing was performed by 10 physicians.

Across the two long-term care facilities, 26 resident care units, each with existing computer provider order entry, were randomized to having a clinical decision support system (intervention units) or not (control units). Bed size of the resident care units ranged from 20 to 60 beds. An effort was made to match the units according to bed size and general characteristics of the residents on the units. We block randomized within categories including: dementia units, units where mental health and behavioral problems were common among the residents, units where the residents had complex medical needs, and units where the residents had profound deficits in physical function.

On intervention units, prescribers ordering drugs were presented with alerts in the form of warning messages; these alerts were not displayed to prescribers when ordering medications for residents of control units. Although efforts were initially made to limit crossover of prescribers between intervention and control units, over the duration of the study, there were providers working simultaneously on both types of units, both on a temporary (coverage) basis and also permanently.

The clinical decision support system was designed by a team of geriatricians, pharmacists, health services researchers, and information system specialists; the process of developing the clinical decision support system and its components has been described previously. The design principles were: 1) messages should be evidence-based; 2) messages should be perceived as useful and informative by practitioners; and 3) the system should have only modest impact on the time required for the practitioner to complete an order. The team reviewed the types of preventable adverse drug events based on previous research, as well as widely accepted published criteria for suboptimal prescribing in the elderly available at the time of this study. We also reviewed all serious drug-drug interactions from standard pharmaceutical drug interaction databases and included alerts for a limited number of more than 600 potentially serious interactions that were reviewed. Reasons for exclusion of alerts for specific drug interactions included that the medications were not on formulary at the facility, or that the medications were generally never used in elderly patients or in the long-term care setting.

The clinical decision support system was designed to provide alerts in response to selected drug orders whenever the order: 1) involved selected high-severity drug interactions; 2) was for a patient with selected abnormal lab test results that suggested a possible danger related to use of the ordered medication; 3) could lead to adverse effects that require special monitoring in order to identify them early; 4) related to a medication that should be ordered within certain dose ranges to reduce the risk of adverse effects in elderly patients; or 5) should be accompanied by prophylactic measures to proactively address situations where there was a high likelihood of adverse drug effects (e.g., constipation with opioid use). Alerts included specific instructions for laboratory monitoring, as well as less explicit recommendations for reconsidering drug orders and monitoring for possible side effects. A summary of the alerts is provided in the appendix.

The computerized provider order entry system in place at the time of the clinical decision support system implementation was capable of linking laboratory test orders and results and current drug orders in real-time. However, the system had several important limitations, as described previously. It was not capable of combining dose and strength information to determine the total daily dose associated with a drug order; therefore, some alerts displayed when they may not have been necessary (e.g., the medication order was already within the recommended dose range). The underlying software was not capable of distinguishing multiple orders for the same drug in different forms or strengths, or orders that had been cancelled and re-ordered within the same prescribing session. These orders were interpreted as multiple orders for drugs in the same category and triggered a number of inappropriate alerts about drug interactions. For example, an order for erythromycin that the prescriber initiated, cancelled, and then re-ordered within the same ordering session would be interpreted as an interaction signaling the need for an alert for increased risk of QT prolongation. Despite the fact that some triggers were likely to produce a substantial number of these unnecessary alerts, we opted to include them in the system if the potential impact of the type of drug interaction in question was considered clinically important.

For residents on the intervention units, the alerts were displayed in a pop-up box to prescribers in real-time when a drug order was entered. The pop-up boxes were informational; they did not require specific actions from the prescriber and did not produce or revise orders automatically. On the control units, the alerts were not displayed to the prescribers.

Our study was limited to adverse drug events occurring in the long-term care setting. Drug-related incidents were identified through review of medical records in monthly segments performed by trained pharmacist investigators for each eligible long-term care facility resident. These investigators, who were not aware of whether the resident was located on an intervention or a control unit, examined the records for possible drug-related incidents, such as new symptoms or events that might represent an adverse drug event, changes in medication regimens (including acute discontinuations or initiations of medications that might be used to treat a drug-induced event), abnormal laboratory values, and all emergency room transfers and hospitalizations. In addition to periodic reviews, medical records were specially targeted for review based on information derived from selected computer-generated signals including abnormal serum drug levels, abnormal laboratory results, and the use of medications considered to be antidotes for adverse drug effects. Administrative incident reports generated within each participating facility were also reviewed for any indication of a drug-related incident.

The primary outcome of the study was an adverse drug event, defined as an injury resulting from the use of a drug. This definition is consistent with definitions used in previous studies. Adverse drug events may have resulted from medication errors (e.g., errors in ordering, dispensing, administration, and monitoring) or from adverse drug reactions in which there was no error.

As described previously, the between-pharmacist investigator reliability for identifying relevant incidents in medical records was assessed through independent review of the same 10 medical records by each of the two pharmacists. Each identified the same incident in the 10 medical records; one pharmacist identified an additional incident in one record that had not been pre-specified as an incident warranting review.

The possible drug-related incidents were presented by a pharmacist investigator to pairs of physician-reviewers (JHG, JJ, PR, LRH, and CB). These physician-reviewers independently classified incidents using structured implicit review according to the following criteria: whether an adverse drug event was present, the severity of the event, and whether the event was preventable. In determining whether an adverse drug event had occurred, the physician-reviewers considered the temporal relation between the drug exposure and the event, as well as whether the event reflected a known effect of the drug. This structured implicit review process has been used in numerous prior studies relating to adverse drug events across various clinical settings. Physician reviewers were not aware of whether a drug-related incident being reviewed had occurred in a resident of an intervention or a control unit.

The severity of adverse events was categorized as less serious, serious, life-threatening, or fatal. Adverse drug events categorized as less serious included a non-urticarial skin rash, a fall without associated fracture, hemorrhage not requiring transfusion or hospitalization, and oversedation. Examples of events categorized as serious included urticaria, falls with associated fracture, hemorrhage requiring transfusion or hospitalization but without hypotension, and delirium. Examples of life-threatening events include hemorrhage with associated hypotension, hypoglycemic encephalopathy, and acute renal failure. Adverse drug events were considered to be preventable if they were judged to be due to an error and were preventable by any means available and not just in relation to the clinical decision support system. For the purpose of the analysis of the effect of the intervention, any event characterized as serious or greater in severity, was categorized as more severe. All other events were considered less severe.

Preventability was categorized as preventable, probably preventable, probably not preventable, or definitely not preventable; results were collapsed into preventable (preventable and probably preventable) and nonpreventable (probably not preventable and definitely not preventable) categories in the analysis.

When the physician-reviewers disagreed on the classification of an incident regarding the presence of an adverse drug event, its severity, or its preventability, they met and reached consensus; consensus was reached in all instances where there was initial disagreement.

Inclusion Criteria:

• prescriber at the study facilities

Exclusion Criteria:

• not a prescriber at the study facilities