Background: Medication-prescribing discrepancies are used as a quality measure for patients transferred between sites of care. The objective of this study was to quantify the rate of adverse drug events (ADEs) caused by prescribing discrepancies and the discrimination of an index of high-risk transition drug prescribing.
Methods: We examined medical records of patients transferred between seven nursing homes and three hospitals between 1999 and 2005 in New York and Connecticut for transfer-associated prescribing discrepancies. ADEs caused by discrepancies were determined by two clinician raters. We calculated the fraction of medication discrepancies that caused ADEs in each of 22 drug classes by calculating positive predictive values (PPVs). We calculated the discrimination of a count of high-risk drug discrepancies, selected from published lists of high-risk medications and using observed PPVs.
Results: 208 patients were hospitalised 304 times. Overall, 65 of 1350 prescribing discrepancies caused ADEs, for a PPV of 0.048 (95% CI 0.037 to 0.061). PPVs by drug class ranged from 0 to 0.28. Drug classes with the highest PPVs were opioid analgesics, metronidazole, and non-opioid analgesics. Patients with 0, 1–2 and ⩾3 high-risk discrepancies had a 13%, 23% and 47% chance of experiencing a discrepancy-related ADE, respectively.
Conclusions: Discrepancies in certain drug classes more often caused ADEs than other types of discrepancies in hospitalised nursing-home patients. Information about ADEs caused by medication discrepancies can be used to enhance measurement of care quality, identify high-risk patients and inform the development of decision-support tools at the time of patient transfer.
Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.
Medication discrepancies are common during transfer between sites of care.12 Since they are sometimes the result of lapses in documentation, transcription, and provider–provider or patient–provider communication, they have been used to measure the quality of transfer documentation and communication.3 In addition, medication discrepancies may be a result of medication changes made by providers that do not have a clear clinical rationale (eg, the omission of a patient’s longstanding antidepressant when they are admitted to the hospital for pneumonia and do not have any contraindication to antidepressants). The potential for medication discrepancies to occur during patient transfer between sites of care as a result of errors in communication or decision-making is the rationale behind the establishment of medication reconciliation as a national patient safety standard during patient handoffs by The Joint Commission (formerly The Joint Commission on Accreditation of Healthcare Organizations).4
Medication reconciliation involves creating a complete and accurate prior medication use list, identifying discrepancies between current and prior medication use, and ensuring prescriber awareness of current and prior medication use to inform prescribing decisions. Since 2006, healthcare organisations have adopted a variety of approaches for implementing medication reconciliation and have used resolution of medication discrepancies as a measure of successful implementation and effectiveness.5–11 However, variation in the success of medication reconciliation remains, as a result of (1) difficulties in staffing a task that is labor-intensive, (2) risk of clerical errors during reconciliation, (3) lack of prescriber awareness of reconciliation findings and (4) lack of influence of reconciliation findings on prescriber decision-making.
As a measure, medication discrepancies may be the product of communication, data synthesis and decision-making processes, but it is not a health outcome. Good-quality measures should have a strong link to health outcomes and target those at highest risk.12 Like measures of inappropriate prescribing13 and other prescribing “signals,”1415 only a subset of medication discrepancies may cause adverse drug events (ADEs) and affect health.
The objective of this study was to examine the predictive value of medication discrepancies for ADEs in nursing-home patients transferred to and from the hospital. We examined nursing-home patients (1) because for these patients pre- and post-transfer medication regimens can be determined exactly, avoiding the ambiguity in regimens that sometimes exists with outpatients2 and (2) because nursing-home patients commonly experience intersite transfers and transfer-related problems.1617 We calculated the positive predictive value (PPV) of prescribing discrepancies in specific drug classes for ADEs, created indices of transition drug prescribing and compared their performance in discriminating patients who might and might not experience a discrepancy-related ADE.
Setting and participants
Participants were patients in seven nursing homes in New York and Connecticut who were admitted to one of three hospitals that were the primary referral hospitals for the nursing homes. Four of the nursing homes and two of the hospitals were Veterans Affairs (VA) facilities; the remaining facilities were non-governmental, non-profit facilities. When patients were transferred between VA nursing home and hospital, transfer information was conveyed electronically via the VA’s Computerized Patient Record System (CPRS). When patients were transferred between non-VA nursing home and hospital, handwritten or printed transfer documents were used to communicate patient information in each direction of transfer.
Eligible patients were individuals transferred from nursing home to hospital and admitted, and who remained in the hospital at least 24 h. Individuals who were seen in the emergency department alone were excluded. Individuals were included whether or not they survived to hospital discharge and whether or not they returned to the nursing home from which they originated. Institutional review boards of each study institution approved the study. Since data were collected by retrospective medical record review, a waiver of informed consent was obtained from each institutional review board.
Trained research personnel reviewed nursing home and hospital charts to identify differences in medication regimens between sites. Sources of medication data reviewed were: (1) nursing home and hospital orders, (2) nursing home-to-hospital and hospital-to-nursing-home transfer documents, (3) hospital and nursing-home medication administration information, and (4) hospital discharge instructions. Medication-prescribing instructions from chronologically sequential sources were matched and compared in dosage, route and frequency of administration. Codes were assigned for: (1) no change, (2) increase in daily dose, (3) decrease in daily dose, (4) route change, (5) change from routine to as needed (PRN) administration, (6) change from PRN to routine administration, (7) substitution for a medication with the same indication (excluding substitutions between generic and brand-name versions of the same drug) and (8) discontinuation. Any of codes 2–8 was considered a prescribing discrepancy. Medications were divided into pharmacological classes as shown in table 1. A priori high-risk discrepancies were defined as those in the Institute for Healthcare Improvement’s High Alert medication drug classes (anticoagulants, opioid analgesics, insulin and sedatives)18 and those in high-risk drug classes for nursing-home patients (non-steroidal anti-inflammatory drugs, digoxin, insulin, antipsychotics, sedatives/hypnotics and anticoagulants).1920 Topical agents, vitamins, minerals and most as-needed medications were not included, since they were not considered potential causes of ADEs over the study follow-up period. As previously reported, the inter-rater reliability for recording number and types of medication discrepancies was high, with a weighted kappa21 of 0.89.1
Adverse drug events
Patients were followed for the duration of the hospital stay up to 2 months, and for patients transferred back to the nursing home, for 2 months after nursing-home readmission. Two physicians or one physician and one pharmacist reviewed nursing-home and hospital records for medical incidents that were defined in advance and included new or worse symptomatic conditions (including new or worse bleeding, congestive heart failure, delirium, diarrhoea, dyspnoea, fall, decrease in alertness, incontinence, pain, rash, urinary retention, vomiting), blood-pressure abnormalities (new systolic blood pressure >185 or <95, diastolic blood pressure >105), fever (temperature >38°C), and abnormal tests of kidney function (creatinine increase >0.5), liver function (doubling of alanine aminotransferase or aspartate aminotransferase), or overanticoagulation (international normalized ratio >4.0). Other laboratory abnormalities (eg, hypo- or hyperglycaemia, hyperkalaemia) were recorded if symptomatic or if they caused a cardiac arrhythmia.
Raters then matched each recorded medical incident with a prescribing discrepancy at the time of nursing home-to-hospital and hospital-to-nursing home transfer that physiologically could have caused the incident—if one existed—and rated whether the discrepancy could have caused the incident using structured implicit review. Implicit review criteria included: (1) whether there was a note in the medical record that suggested that a medication discrepancy caused the incident (yes or no), (2) the time interval between incident and discrepancy (ie, timing “plausible” or “improbable”), (3) whether the incident could have been caused by something other than a medication discrepancy (ie, competing causes: “many,” “some” or “few/none”), (4) whether the incident was a known possible reaction to this medication discrepancy (yes or no), and (5) whether the patient’s condition improved after correction of the medication discrepancy (ie, dechallenge response: “none/weak,” “suggestive” or “convincing”).
Each rater rated the certainty that the incident was caused by a medication discrepancy using a six-point Likert scale, with 1 indicating “little or no” certainty and 6 indicating “almost total” certainty.22 The two raters discussed each event and provided a final consensus certainty rating. “Possible,” “probable” and “definite” ADEs were those for which the final certainty ratings were 4, 5 and 6, respectively. Raters further categorised ADEs as (1) asymptomatic, (2) causing temporary symptoms, (3) causing temporary disability, (4) causing a prolonged or an additional hospital stay, (5) causing permanent disability or (6) causing death. Finally, if there was no appropriate clinical rationale for the prescribing discrepancy or the discrepancy deviated from prescribing norms, the ADE was considered to be the result of a prescribing error. Prescribing errors were categorised as (1) wrong omissions, (2) wrong dosages or (3) wrong dosing frequencies.
Characteristics of patients and their hospital stays
Information was collected on patient age, gender, race, presence or absence of dementia, and duration of nursing-home stay from the nursing-home medical record. A score for burden of chronic disease, adapted from Charlson et al,23 was calculated from chronic medical problems listed in the nursing-home medical record. Information on hospital diagnoses, hospital length of stay and time of admission (08:00 to 18:00 Monday to Friday vs off-hours) were obtained from the hospital medical record. A modified Acute Physiology and Chronic Health Evaluation (APACHE) score24 was calculated from initial laboratory data and vital signs in the hospital medical record to ascertain initial illness severity.
More than one hospital admission was allowed per participant. The unit of analysis was hospital admission. The number of prescribing discrepancies was calculated as the sum of prescribing discrepancies during nursing-home-to-hospital and hospital-to-nursing-home transfers. The number of ADEs was calculated as the sum of medical incidents caused by prescribing discrepancies with possible, probable or definite certainty. Positive predictive values (PPVs) were calculated as the number of ADEs caused by discrepancies in a drug class divided by the number of prescribing discrepancies in that class. PPVs were also calculated for “enriched” subgroups of episodes in which a patient experienced a medical incident that is commonly captured by automated data systems (pain, vital sign or laboratory data) and also was exposed to a prescribing discrepancy that physiologically could have caused the incident (eg, pain/analgesic discrepancy). The PPV was calculated as the fraction of such episodes in which the prescribing discrepancy was rated as causing the incident, indicating an ADE.
We ascertained discrimination of three indices of transition drug prescribing for ADE: number of drugs prescribed prior to transfer, number of drug discrepancies after transfer and number of high-risk drug discrepancies after transfer. The number of high-risk drug discrepancies was calculated as the sum of those with PPVs at least as high as those in the a priori high-risk category, that is, all drug classes with a PPV ⩾0.04. The sample was stratified by quartile of each of the three indices, and the percentage with ADE in each quartile was calculated. Unadjusted logistic regression models were estimated in which each prescribing index was the key independent variable, and occurrence of ADE (yes or no) was the dependent variable. Adjusted logistic regression models were estimated with each index as key independent variable; relevant demographic (gender, age), clinical (comorbidity score, APACHE score) and circumstantial (off-hours admission, duration of follow-up) variables as covariates; and occurrence of ADE (yes or no) as dependent variable. Models for drug discrepancies and for high-risk drug discrepancies included the number of pretransfer drugs as a covariate. Findings were similar whether or not we accounted for clustering of observations within patients and facilities; only findings without clustering are shown. The 95% CI, p values, and c statistics were calculated using standard formulae. In the absence of a validation sample for high-risk drug discrepancies, a bootstrap validation was conducted with 1000 repetitions. All analyses were performed using SAS software (Cary, NC).
Two hundred and eight patients were hospitalised 304 times. The characteristics of patients and their hospital stays are shown in table 2. Forty-two per cent of hospitalisations were in the VA setting. The most common reasons for hospital admission were pneumonia, urinary-tract infection, dehydration and exacerbations of congestive heart failure and chronic obstructive pulmonary disease. The median hospital length of stay was 7 days (range 1 to 296). The median length of follow-up for ADE ascertainment was 63 days (range 1 to 120).
Patients received a mean of 6.5 (SD 2.9) medications prior to hospital admission and had a mean of 2.8 (2.1) prescribing discrepancies associated with nursing home-to-hospital transfer. Patients received a mean of 6.1 (3.2) medications prior to hospital discharge and had a mean of 1.5 (1.7) prescribing discrepancies associated with hospital-to-nursing-home transfer. The total number of prescribing discrepancies observed in the study sample was 1350, and the total number of discrepancy-associated ADEs observed was 65. Of these, 51%, 39% and 9% were possible, probable and definite ADEs. Forty-six per cent were asymptomatic ADEs, 42% were associated with temporary symptoms, 10% caused temporary disability, and 3% caused a prolonged or an additional hospital stay. No ADE caused permanent disability or death. Finally, 48% of prescribing discrepancies that caused ADEs were considered to be prescribing errors; 46% of these errors were wrong omissions, 46% were errors in dosing frequency, and 8% were errors in dosage.
Overall, 65 of 1350 prescribing discrepancies caused ADEs for a PPV of 0.048 (95% CI 0.037 to 0.061). Positive predictive values of prescribing discrepancies by drug class are shown in table 1; they ranged from 0 to 0.28. The drug classes with the highest PPVs were opioid analgesics, metronidazole and non-opioid analgesics. Among episodes in which medical incidents commonly captured by automated data systems and suspect prescribing discrepancies both occurred, episode PPVs ranged from 0.07 to 0.37, as shown in table 3.
The results of an examination of the discrimination of three indices of transition drug prescribing—number of drugs prescribed prior to transfer, number of drug discrepancies after transfer and number of high-risk drug discrepancies after transfer—are shown in table 4. The number of high-risk drug discrepancies demonstrated the best discrimination, as demonstrated by the highest c statistic and best-risk gradient. Patients with 0, 1–2 and ⩾3 high-risk discrepancies (representing first quartile, second and third quartiles together, and fourth quartile) had 13%, 23% and 47% chance of experiencing a discrepancy-related ADE, respectively. In a multivariable logistic regression model that included relevant demographic (gender, age), clinical (comorbidity score, APACHE score, number of medications at baseline) and circumstantial (off-hours admission, duration of follow-up) variables as predictors, the number of high-risk discrepancies was the only statistically significant predictor of ADE with an odds ratio (OR) of 1.71 (95% CI 1.28 to 2.28; p = 0.0003), indicating an additional 71% risk of ADE with each additional high-risk discrepancy. Bootstrap validation resulted in an OR of 1.71 (95% CI 1.16 to 2.28) and c statistic of 0.713 (95% CI 0.654 to 0.774).
This study examined the link between medication discrepancies at the time of patient transfer and ADEs as a patient health effect in patients transferred between nursing home and hospital. We found that less than 5% of discrepancies caused ADEs, which is consistent with authoritative reviews that suggest that a small fraction of errors result in harm.2526 However, certain classes of drugs had PPVs substantially higher than 10%, including opioid analgesics, metronidazole and non-opioid analgesics. In addition, an index that was a count of discrepancies in 15 high-risk drug classes at the time of transfer discriminated between those with a lower and higher risk of ADE. Patients with 0, 1–2 and ⩾3 high-risk discrepancies had 13%, 23% and 47% chance of experiencing a discrepancy-related ADE, respectively.
These results suggest a link between medication discrepancies and health outcomes, and the capability of an index of high-risk transition prescribing to identify those at highest risk, which are characteristics of a sound quality measure.12 Sound measurement of drug prescribing during patient handoffs is important because The Joint Commission established medication reconciliation as a national patient safety standard during patient handoffs.4 Our results provide partial support to patient safety organisations that have promulgated medication discrepancies as a measure of the effectiveness of medication reconciliation for inpatients and outpatients.5 On the other hand, our results suggest that prescribing information matched with automated clinical event information would more often identify episodes that were true discrepancy-related ADEs (as shown by higher PPVs in table 3, ranging up to 0.37) and identification of such episodes would be a more accurate measure of the effectiveness of medication reconciliation.
The results of this study also may be used to inform development of decision support tools for nursing homes or hospitals. Tools could be designed to identify high-risk medication discrepancies, identify patients at higher risk of ADE and alert providers taking care of patients who are transferred between sites of care. Targeting discrepancies in medications used for chronic symptomatic conditions may be particularly effective. We and others have previously reported a suggested beneficial effect on pain of medication reconciliation, presumably as a result of improving the continuity of analgesia prescribing.2728 On the other hand, decision-support systems (eg, computerised feedback and reminders) have been shown to have a weak impact on drug prescribing.29 In a previous study, we found that nursing-home providers changed orders corresponding to only 10% of alerted discrepancies.28
This study has important limitations. First, it includes only patients admitted to the hospital from a nursing home. Other groups may have different base rates of overall prescribing discrepancies, prescribing discrepancies by class and discrepancy-related ADE, resulting in different PPV calculations. Second, we did not ascertain ADEs caused by non-discrepant medication use (ie, medications that are continued unchanged after a transfer). Therefore, we could not calculate the negative predictive value of medication discrepancies (the chance of not experiencing an ADE in the absence of a discrepancy) or its sensitivity (the fraction of ADEs that it captures). In this regard, medication discrepancies are just one of several measures being examined as “signals” of ADE; others include abnormal laboratory findings such as an elevated digoxin level and prescription of ADE antidotes such as flumazenil.1415 Medication discrepancies could be used in conjunction with these to approximate ADE. This study is also limited by the lack of a true validation cohort for the index of high-risk prescribing, which was derived in part using data from this study (informed by published lists), as well as by small sample size numbers in some of the drug classes that result in wide PPV confidence intervals.
In summary, discrepancies in certain drug classes, in particular opioid and non-opioid analgesics, more often caused ADEs than other types of discrepancies in hospitalised nursing-home patients. The number of high-risk discrepancies discriminated between patients who experienced a discrepancy-related ADE and those that did not. Information about ADEs caused by medication discrepancies can be used to enhance measurement of care quality, identify high-risk patients and inform development of decision-support tools at the time of patient transfer.
The authors would like to acknowledge the assistance of T Mordiglia and B Fridman in data collection.
Funding: Financial support was provided by the VA Health Services Research and Development Service and the New York State Department of Health. SL is the recipient of a Doris Duke Clinical Research Fellowship.
Competing interests: None.
Ethics approval: Ethics approval was obtained from the Institutional Review Boards of Mount Sinai School of Medicine, James J. Peters VA Medical Center and Jewish Home Lifecare (all in New York City).