Assessment of Adverse Drug Events Among Patients in a Tertiary Care Medical Center

Philip E. Johnston; Daniel J. France; Daniel W. Byrne; Harvey J. Murff; Byron Lee; Renee A. Stiles; Theodore Speroff


Am J Health Syst Pharm. 2006;63(22):2218-2227. 

In This Article


Study Site

Vanderbilt University Medical Center (VUMC) is a 658-bed tertiary care medical center in Nashville, Tennessee, and is associated with schools of medicine and nursing. The institution includes a regional level I trauma center and a level IV neonatal intensive care center. Other clinical services include multiple solid organ transplantation specialties, a burn center, and bone marrow transplantation. The department of pharmaceutical services offers full i.v. additive and unit-dose distribution systems, 24-hour service, satellite services in critical care, and decentralized clinical pharmacists assigned to many adult medicine, pediatrics, and critical care units throughout the institution. Pharmacists are members of multiple institutional safety and patient care committees, including those evaluating AEs.

Because this study was initiated as an internal quality-improvement review to assess and report the characteristics of AEs in VUMC, we did not initially seek approval from the university's investigational review board (IRB). The study application was approved by the IRB after completion of the review to enable publication.


An AE was defined as any ADR or medication error reported within VUMC during the period of the study. According to the institution's policy, ADRs included any unwanted or unexpected outcome of drug therapy, including failure to provide an expected response. Medication errors included errors in prescribing, compounding, administration, or monitoring of medication therapy.

During the study period, reporting of AEs was an expected professional activity. VUMC has developed methods of reporting AEs, including secured e-mail, telephone reporting, paper reports, and secure reporting through the CPOE system. Education by the risk management and pharmacy departments has focused on penalty-free reporting. Forms for reporting AEs are stored at each nursing station and in all of the institution's 10 pharmacy locations. E-mail and CPOE were available using several hundred order-entry terminals in the institution. Required data elements (name, medical record number, age, sex, diagnoses, medications, pertinent laboratory tests, subjective and objective findings, and person reporting the AE) on the reporting forms for medication errors differed somewhat from data for ADRs on the forms. For the purpose of this study, data used for analysis were the same. The determination to report an AE can be made by any health care professional caring for the patient, most commonly nursing personnel followed by pharmacists and physicians. The person reporting the AE enters the suspected causative agent. All reports are included in a Web-based, secure database, which can be reviewed by risk management, area supervisors, and other designated personnel in the institution. When an ADR occurred, causality was assessed by the reviewing pharmacist, using his or her own experience and references available in the pharmacy. Traditional methods for determining causality were not used due to lack of follow-up information after the AE occurred.

Severity of AEs was determined after a report was submitted by the pharmacist who evaluated each event. Initially, the Institute for Safe Medication Practices severity scale was adopted but later abandoned because of lack of patient-specific information. Four levels of severity were used in this study: (1) no or minimal injury, (2) patient injury or potential injury, (3) injury causing intervention or extended hospital stay, and (4) major morbidity or death. Categorizing AEs further was considered beyond the scope of this study. Focused surveillance (e.g., tracking of selected medication orders, diseases, patient age, number of doses per day) had not previously revealed beneficial results at our institution and were not considered in this study.

Records for patient admissions that occurred during the study period that did not involve an AE were assigned to a control group. This set of admissions was divided into two equal sets: the training data set and the validation data set. These two data sets were analyzed to determine if they could be compared. Statistical analysis was then conducted with the training data set to determine what risk indicators were present. The validation set was used to test our initial impressions and determine if our indicators were validated or not validated.

Study Population

The study population included all patients admitted to the medical center hospital between January 1, 2000, and June 30, 2002. The unit of analysis was the patient admission; for individuals having multiple admissions with an AE during the study period, the admission involving the AE with the highest severity was selected for inclusion. For patients without an AE, one admission was randomly selected for analysis. Outpatient visits (emergency department, outpatient surgery, clinic, infusion center, and home care visits) were excluded. Control records were screened to ensure that only one admission was selected per patient to avoid multiple reports of the same AE. Duplicate admissions were determined by the database analyst and were excluded with the pharmacist coinvestigator's agreement.

Reports of AEs were excluded from the analysis if a patient was unidentifiable, if the report was generated secondary to an outpatient visit, or if the incident had not been described in the report. When multiple AEs were reported for a single individual, the AE having the highest associated severity was retained for analysis. This decision was made based on the severity scale score assigned by the pharmacist.

Data Collection

The initial data for AEs consisted of AE reports maintained by the pharmacy. Product-problem reports and MedWatch reports for drug products and biologics originating at VUMC were screened and included when appropriate (i.e., reports met study inclusion criteria). MedWatch reports were matched against other AE reports to ensure that each AE was counted only once. The data collected from reporting forms and from hospital database logs included medical record number, sex, age, clinical service, date of occurrence, diagnoses, type of error, suspected medication, and severity of the AE. Additional information was added from the hospital database, including the first 10 diagnoses logged in the patient's record. Pharmacy personnel and quality-improvement personnel independently produced a database of patient admissions for the study period and verified that the patients were the same. Each patient's medical record number, name, and birth date were used to verify inclusion and exclusion of admissions. Once verified, the records were deidentified for security and confidentiality. Each medication involved in an AE was identified by generic name and classified using the American Hospital Formulary Service medication categories.


The dependent variable was defined as the presence or absence of an AE as identified by an AE report. Independent variables included patient age, sex, race, admission source, insurance class, International Classification of Diseases, 9th Revision (ICD-9) codes, and all diagnostic groupings, medication names, and therapeutic categories of those medications. Variables with a frequency of <0.1% were eliminated. After data cleaning and consolidation, 93 variables remained in the analysis. Factor analysis was not performed. We used various data-reduction methods, including recursive partitioning, as described in Harrell,[37] to produce a reasonable number of factors for these models.

Statistical Analysis

Stage 1 of the analysis focused on patient characteristics that had been associated with AEs. Patient age was evaluated using the Mann-Whitney U test to determine which ages or age groups were at significant risk. Age groups, sex, race, admission source, and insurance class were evaluated using the chi-square statistic to determine p values for various diagnoses, individual medications, and medication classes. Further analysis was based on a two-stage logistic regression analysis comparing the training set and validation set, using variables identified in the first phase of analysis as indicators. A three-stage logistic regression model with high-risk indicators was entered on stage 1 and the medications on stage 2. In a separate model, the medication class was added on stage 3. These data allowed the development of a model to statistically evaluate key indicators in the training and validation data sets.


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.