Patient Risk Factors for Adverse Drug Events in Hospitalized Patients FREE

ADVERSE DRUG events (ADEs) are an important and costly problem.^1- 4 The Medical Practice Study, which examined patients hospitalized in New York State in 1984, found that almost 1% of patients experienced an injury due to drugs that resulted in disability or prolongation of their hospital stay.^5- 6 Other studies,^7- 17 most of which used the adverse drug reaction as the outcome, have also shown that injuries due to drugs are common among hospitalized patients. The term "adverse drug reaction," as defined by the World Health Organization,¹⁸ excludes events associated with errors, while the term "ADE" includes preventable and nonpreventable events.¹⁹ In earlier reports^20- 21 from the ADE Prevention Study, we found an average of 6.5 ADEs per 100 admissions, of which 28% were preventable.

One potentially attractive strategy for preventing ADEs is to identify prospectively those patients at high risk of an ADE and to target additional resources toward this group. An example of this approach might be that when a patient is determined to be high risk, the patient would be identified so that the pharmacy could pay extra attention to all medications given. Several patient attributes that may make an ADE more likely have been suggested by various investigators.^12,22- 28 However, little empirical data exist on their predictive ability.

Some risk factors for adverse reactions to drugs that have been proposed to date include age,²⁹ number of drugs the patient is receiving,³⁰ and factors that alter drug distribution or metabolism, such as renal or hepatic insufficiency, congestive heart failure, anemia, and alcoholism.³¹ It has also been suggested that a patient who is receiving specific drugs or drugs of a certain class may be prone to having an ADE; in fact, substantial work has been done in evaluating which drugs are most often associated with ADEs or adverse drug reactions.^32- 35 When Karch and Lasagna³² combined several studies, they found that antibiotics caused 42% of adverse drug reactions, but no other group of drugs was responsible for more than 10%. In the Medical Practice Study, antibiotics were also the most common class of drugs associated with drug-related injuries, despite accounting for only 15% of ADEs.³⁵ Several models have been developed for predicting events in patients receiving specific drugs, for example, Landefeld^36- 37 indexes for predicting bleeding in patients beginning anticoagulation therapy. However, to our knowledge, no empirical data exist that allow stratification of hospitalized patients according to likelihood of an ADE across drugs.

Our aims in this study were (1) to develop a practical, efficient method of identifying patients at increased risk of an ADE based solely on information readily available on the hospital's computer database, either at admission or up to the time of the event; and (2) to evaluate a larger spectrum of potential patient risk factors by comparing case patients who experienced an ADE with a matched control group. In the first portion of the study, electronic access to information was available at only 1 of the 2 hospitals we studied, so the other hospital was excluded from this portion of the study. Using this set of limited, yet online information, we were able to compare patients who experienced an ADE against the entire population of patients who were admitted to the surveyed units during the study period. For the second portion of the study, we collected highly detailed data for each patient at 2 hospitals through extensive daily medical record review and verbal inquiry. We were then able to compare this detailed, time-intensive form of data collection against the larger cohort, about whom less information was available, to weigh the relative efficacy of each approach.

Patients surveyed included all adults at 2 large tertiary care hospitals, Brigham and Women's Hospital (726 beds) and Massachusetts General Hospital (846 beds), both in Boston, Mass, admitted to any of 11 units during a 6-month period between February and July 1993, as previously described.²⁰ Study units were a stratified random sample of medical, surgical, and intensive general care units. These units included 5 intensive care units (ICUs) (3 surgical and 2 medical) and 6 general care units (4 medical and 2 surgical). To identify more events, we intentionally oversampled ICUs and excluded obstetric units, because we previously found that ADEs were more common in ICUs than in general care units and that obstetric units had almost no ADEs.¹

The primary outcome of the study was the ADE, which we defined as an injury resulting from medical intervention related to a drug.¹ For example, if a patient with first-degree atrioventricular block were given a β-blocker and then developed complete heart block requiring temporary pacing, this reaction would be an ADE. The secondary outcomes we studied were preventable ADEs and severe ADEs. For example, the ADE would have been recorded as a preventable ADE if the patient in the example who developed atrioventricular block had already been taking a calcium channel blocker, which depressed the atrioventricular node before receiving the β-blocker. An ADE was considered a severe ADE if the consequences were serious, life threatening, or fatal. In prior reports from our group,^20- 21 we also presented data on potential ADEs, which were defined as incidents in which an error was made but no actual harm occurred to the patient; however, we did not attempt to assess risk for potential ADEs in this report.

Two physician reviewers evaluated all incidents independently and classified them according to whether an ADE or potential ADE was present, whether it was preventable, and how severe it was.²⁰ When there were disagreements that affected classifications about the presence of an event, its severity, or its preventability (eg, 1 reviewer scored it as preventable, but the other did not), the reviewers met and reached consensus. If consensus could not be reached, a third reviewer evaluated the incident and made the final decision. Kappa statistics (κ) were calculated and were found to be 0.81 to 0.98 for whether an ADE was present and 0.92 for whether the ADE was preventable; κ statistics were lower, at 0.32 to 0.37, for assessments of severity.²⁰

We collected 2 sets of data: (1) detailed information on the case patients and controls at both hospitals and (2) a limited amount of data available electronically on all patients in the study cohort at 1 of the 2 hospitals. Case patients were defined as all patients with an ADE, and the 2 subcategories of ADEs were preventable ADEs and severe ADEs. Controls were patients on the same unit as the case patient with the most similar preevent length of stay. Thus, controls had the same level of care as the case patients (ICU vs non-ICU), and almost all were on the same service (medicine vs surgery).

Three mechanisms were used for identifying incidents. First, nurses and pharmacists were asked to report incidents to the nurse investigators. Second, a nurse investigator visited each unit at least twice daily on weekdays and solicited information from nurses, pharmacists, and clerical personnel concerning all actual or potential drug-related incidents. Third, a nurse investigator reviewed all medical records at least daily on weekdays.

In the cohort analysis, we compared patients experiencing an ADE in 1 hospital with all patients admitted to any of the surveyed units at that hospital during the study period. Data available online included demographic and administrative information, such as age, sex, race, admission, and subsequent level of care (eg, whether the patient was in an ICU at admission or on the date of the event); admission source (eg, home or nursing home); primary insurer; and diagnosis related group (DRG) weight. In addition, we were able to obtain some medication data, including the number and names of drugs a patient was receiving at admission. Drugs were classified into American Hospital Formulary System (AHFS) categories, including cardiovascular, antiasthmatic, sedative or hypnotic, antisecretory, antitumor, antidepressants, analgesics, diuretics, antibiotics, anticoagulants, antiarrhythmics, muscle relaxants, antipsychotics, electrolyte concentrates such as potassium, and antihypertensives. Laboratory data were also obtained, including serum total bilirubin level as a proxy for hepatic insufficiency, serum urea nitrogen (SUN) and creatinine levels to identify patients with renal failure, serum albumin level, and platelet count. A SUN level of 5.7 mmol/L (16 mg/dL) was used as a cutoff to separate patients with a low SUN level from those with slightly elevated to high SUN levels. Similarly, a platelet count of 50 × 10⁹/L was used as a cutoff to identify patients with a low platelet count. A serum creatinine level of 133 µmol/L (1.5 mg/dL) was chosen as a cutoff point to separate patients with a low serum creatinine level from those with elevated levels, and a serum bilirubin level of 34 µmol/L (2.0 mg/dL) was designated as the cutoff for patients with low vs elevated bilirubin levels. Serum albumin was stratified into 4 levels: less than 20 g/L, 20 to 24 g/L, 25 to 36 g/L, and more than 36 g/L.

In the case-control study, the reviewer was blinded to case status for all risk data obtained from medical record review. We accomplished this by having 1 reviewer identify the case patient and control and then refer the case patients to a second reviewer for data collection but not reveal which individual was a case patient. Data collected included all the information available electronically for the cohort analysis as described earlier and, in addition, clinical data, including severity of illness at the time of the event (measured using the Therapeutic Intervention Scoring System score³⁸), comorbidity (measured using the Charlson comorbidity index³⁹), and extensive drug exposure information, including the number and names of drugs that case patients and controls were receiving at the time of the event. Drugs were classified into AHFS categories, as they were in the cohort study.

Comparisons between categorical variables were made using the χ² test as well as a 2-sided trend test for comparisons in which a variable had multiple ordered categories. Comparisons between normally distributed variables were made using the t test, and nonparametric comparisons were made using the Wilcoxon rank sum test. In the multivariate analyses of the entire cohort, stepwise logistic regression was used, while in the case-control multivariate analyses, comparisons were made using conditional logistic regression to retain the advantages of the paired nature of the data. Analyses were performed using SAS statistical software (SAS Institute Inc, Cary, NC)⁴⁰ except for trend tests, which were performed using StatXact software (CYTEL Software Corp, Cambridge, Mass).⁴¹

In this analysis, there were 139 ADEs and 42 preventable ADEs during the study period, among 2379 total admissions. Among the admissions, 360 lacked initial drug exposure information; while these admissions were retained in the demographic and clinical studies, they were excluded from the drug exposure analyses. We, therefore, had a final number of 113 ADEs and 32 preventable ADEs among 2019 admissions for this portion of the study.

Univariate analyses of demographics and clinical characteristics in the cohort study revealed 4 significant correlates of an ADE (Table 1): hospital service, SUN level of 6.1 mmol/L (17 mg/dL) or greater, platelet count below 50 × 10⁹/L, and albumin category.

For preventable ADEs, a comparison of the demographics and clinical characteristics between case patients and the cohort revealed 5 significant correlates (Table 1). Patients experiencing a preventable ADE were older than the rest of the patient population. In addition, presence of a platelet count below 50 × 10⁹/L, a SUN level of 6.1 mmol/L (17 mg/dL) or greater, and higher albumin levels were all correlated with the presence of an ADE (Table 1). There was also a trend toward finding more preventable ADEs on the medicine service.

In the drug exposure analysis of the cohort group, we found 7 significant correlates of presence of an ADE (Table 2). Individual drug types associated with higher rates of ADEs (Table 2) include diuretics, electrolyte concentrates, antitumor agents, anticoagulants, and ulcer medications. Two summary categories were also correlated with ADE frequency: the average number of different drugs received (7.3 for case patients vs 6.0 for controls; P = .02) and the average number of AHFS classes of drugs received (3.8 for case patients vs 3.3 for controls; P = .02).

Univariate analysis of drug exposure information for preventable ADEs also revealed 7 significant correlates (Table 2). Four of these were also significant in our analysis of all ADEs: a higher frequency of preventable ADEs was associated with use of ulcer medications, use of antidepressants, use of electrolyte concentrates, and the average number of AHFS classes of drugs received (4.3 for case patients vs 3.3 for controls; P = .05). Also associated with preventable events was the use of cardiovascular medications, antihypertensive agents, and diuretics. There was a trend toward an association for the use of antiseizure medications as well.

In multivariate analyses to look for significant correlates of ADEs and preventable ADEs in the cohort study, independent correlates of ADEs were use of electrolytes, use of diuretics, and whether the patient was admitted to a medical ward (Table 3). Multivariate correlates of preventable ADEs were platelet category, use of antidepressants, use of antihypertensive agents, whether the patient was admitted to a medical ward, and use of electrolyte concentrates (Table 3).

In the case-control study, there were 247 ADEs, 70 preventable ADEs, and 106 severe ADEs among 4108 admissions during the study period. If a patient had more than 1 ADE during an admission, only the first episode in the admission was counted. Length-of-stay outliers were identified using Studentized residuals for length of stay.⁴² If an ADE had a residual of −2 or lower or 2 or greater, it was examined. For example, some patients were in the hospital for up to a year often awaiting nursing home placement; in such cases, it was judged that length of stay was not influenced by the occurrence of an ADE. When ADEs in which more than 1 event occurred per admission (n = 40) and length-of-stay outliers (n = 17) were excluded, these figures became 190 ADEs, 60 preventable ADEs, and 84 severe ADEs. Each case patient was paired with a control from the same unit with the most similar preevent length of stay.

When we compared the demographics of the case patients and the controls, there were no significant differences (Table 4). However, when we compared the case patients with all patients admitted to the study units at either of the 2 hospitals, the case patients had a much longer length of stay. In addition, the case patients were much more likely to be admitted to an ICU. On comparing the clinical characteristics of all patients with an ADE with controls in univariate analyses (Table 5), we found no significant correlation between any of the following and presence of an ADE: age, mean comorbidity or severity scores, number of drugs received in the 24 hours before the incident, number of drugs received since admission, number of psychoactive drugs received since admission, presence of altered mental status, or presence of an elevated bilirubin or creatinine level. The same was true for preventable ADEs. For severe ADEs, case patients were older than controls, but no other factors were significantly correlated with presence of an ADE.

A comparison of drug exposure levels between case patients and controls (Table 6) revealed that exposure to 1 or more psychoactive drugs was the only class significantly associated with an ADE. This was the case even though most patients in both groups were exposed to these drugs. Dividing patients into 3 categories based on number of psychoactive drugs received (0, 1, or ≥2) resulted in a more specific high-risk group: 47% of the case patients fell into the group exposed to 2 or more psychoactive drugs vs 38% of the controls (P = .01). For severe ADEs, there was a significant association between cardiovascular drugs. Some examples of drugs included in this group are digoxin, atenolol hydrochloride, diltiazem, nifedipine, and propranolol hydrochloride. For preventable ADEs, there were no significant associations across drugs.

Multivariate analyses of significant correlates of ADEs, severe ADEs, and preventable ADEs in the case-control study revealed few independent predictors. There was 1 independent correlate of ADEs: exposure to 1 or more psychoactive drugs (odds ratio, 2.1; 95% confidence interval, 1.3-3.6). Use of cardiovascular drugs was an independent predictor of severe ADEs (odds ratio, 2.4; 95% confidence interval, 1.3-4.5). There were no independent correlates of preventable ADEs.

Substantial data suggest that ADEs are an important public health problem. In previous studies,^20- 21 we found that these events occurred at a rate of 6.5 per 100 admissions and that 28% were potentially preventable. One attractive approach to preventing events involves risk stratification, in which patients are stratified using prospectively gathered information according to their risk of an event.⁴³ This has been an effective clinical approach for reducing event rates in other areas, for example, for risk of cardiac events in patients undergoing noncardiac surgery.⁴⁴

In this study, we tried to develop a risk stratification model for patients likely to experience an ADE using 2 approaches: a cohort analysis using limited information readily available electronically and a case-control study using more detailed information gathered on case patients and a set of matched controls. However, while the cohort analysis identified a few independent predictors of ADEs and preventable ADEs, they had relatively little power. In fact, almost none of the proposed "risk factors" for ADEs were actually associated with a substantially elevated risk of having an ADE. While such factors as age, multiple drug therapy, and impaired renal function almost certainly increase the risk of an ADE, the magnitude of these risks is probably smaller than has been suggested.

In a number of recent studies, investigators have suggested 3 factors in particular that may predispose patients to ADEs: age,²⁹ polypharmacy,^30,45- 50 and impaired renal function.⁵² Our data suggest that the effect of advancing age may be modest at best. While age emerged as a significant correlate in our cohort analysis, it was not retained in the multivariate studies because of lack of predictive power. One explanation may be that while drug metabolism is clearly altered by age, the drugs and dosages used in older patients are adjusted to account for the patient's age.⁵¹ Regarding polypharmacy as a potential predictor of adverse events, we found only a borderline association between number of drugs received and risk of an ADE. In the case-control analysis, there was no correlation, and in the cohort study, neither average number of drugs received nor average number of AHFS classes of drugs received was an independent predictor in the multivariate analysis. In a previous study,⁵² we found a correlation of ADE rate with drug exposure. Specifically, a significantly higher rate of preventable and potential ADEs was observed in ICU patients vs non-ICU patients, but after adjusting for number of drugs administered, this difference disappeared. It may be the case that being in an ICU is correlated with receiving an increased number of drugs, while polypharmacy is not an independent predictor of having an ADE. Regarding impaired renal function, the cohort study found a significant univariate correlation between SUN levels above 5.7 mmol/L (16 mg/dL) and ADEs and preventable ADEs compared with the entire cohort population at 1 hospital. However, this was not retained as an independent predictor in multivariate analysis. In the case-control analysis, while we had a relatively small number of patients and thus limited power, there was a slight trend toward an effect of renal impairment for preventable ADEs but it was not statistically significant.

The drug exposure data in the case-control analysis suggest that no major drug class was responsible for a disproportionate share of the ADEs, with the possible exception of analgesics. The association between cardiovascular drugs and serious events in the case-control study may be a chance finding given that these drugs were rarely believed to be responsible for individual events in several comparisons, or it may be that use of these drugs was a marker for an underlying condition (eg, cardiovascular instability).

If nonpreventable ADEs occur relatively randomly across the hospitalized population, and a large fraction of preventable ADEs occur as a result of system problems that do not occur substantially more often in one patient group, then the clinical and drug exposure data we obtained in this study are what would be expected. The subgroup of events that might be expected to be predictable are those that occur in the group of patients with impaired drug clearance for some reason. For example, older patients with impaired renal function or impaired liver function may indeed be at higher risk of an ADE. However, our data suggest that this group of patients is relatively small.

The data analyzed in this study suggest that prevention strategies that focus on improving the systems by which drugs are ordered, dispensed, and administered will prevent more events than patient risk stratification strategies. According to this view, the focus for improvement should be on the system involved rather than on individual patients.²¹ These approaches will often involve the use of patient-specific data, for example, in adjusting aminoglycoside dosages based on the presence of renal failure. However, they often will not use patient-specific data, for example, in developing standardized labeling for tubing in situations in which multiple intravenous medications are being administered simultaneously.

This study has several limitations. It was performed at only 2 tertiary care institutions, which may hinder the generalizability of the results to other care settings. Another limitation, specific to the cohort analysis that looked only at information available electronically, is that the specific data elements available online from site to site are likely to vary, and clearly more elements will become available over time. However, the elements we chose are reasonably standard, and even if they are not available within some organizations, they are likely to become available over time. Another issue is that an alternative approach—not our focus in this study—would be to take specific patient groups and look for specific types of events; this way of addressing the problem may prove useful in future studies.

We conclude that risk stratification approaches to identifying patients at high risk of experiencing an ADE in the hospital are unlikely to be productive. Instead, strategies that attempt to reduce the incidence of ADEs by focusing on the systems by which drugs are given, and that incorporate specific suggestions regarding medication choices and dosing, are more likely to be successful in battling this important and costly problem.

Accepted for publication February 18, 1999.

This study was supported by grant R01-HS07107-01 from the Agency for Health Care Policy and Research, Rockville, Md; and the Risk Management Foundation, Cambridge, Mass.

We thank the nurses, pharmacists, physicians, and other personnel on the study units for their support in carrying out the study.

Boston, Mass: Lucian L. Leape, MD; Deborah Servi; Nan Laird, PhD; David W. Bates, MD, MSc; Michael Cotugno, PharmD; Mairead Hickey, RN, PhD; Patricia Hojnowski-Diaz, RN; Sharon Kleefield, PhD; Heather Patterson, PharmD; Stephen Petrycki, RN; Brian Shea, PharmD; Martha Vander Vliet, RN; Jeffrey Cooper, PhD; David J. Cullen, MD, MSc; Harry Demonaco, MS, RPh; Margaret Dempsey Clapp, MS, RPh; Theresa Gallivan, RN; Robert Hallisey, MS, RPh; Jeanette Ives, RN, MSN; Ellen Kinneally, RN; Kathy Porter, RN, MSN; Steven D. Small, MD; Bobbie J. Sweitzer, MD; Taylor Thompson, MD; J. Richard Hackman, PhD; Amy Edmondson. Houston, Tex: Laura A. Petersen, MD, MPH. New York, NY: Glenn Laffel, MD, PhD.

Corresponding author: David W. Bates, MD, MSc, Division of General Medicine and Primary Care, Brigham and Women's Hospital, 75 Francis St, Boston, MA 02115 (e-mail: dwbates@partners.org).