Methodological Challenges for Epidemiologic Studies of Deprescribing at the End of Life

Purpose of Review To describe approaches to measuring deprescribing and associated outcomes in studies of patients approaching end of life (EOL). Recent Findings We reviewed studies published through 2020 that evaluated deprescribing in patients with limited life expectancy and approaching EOL. Deprescribing includes reducing the number of medications, decreasing medication dose(s), and eliminating potentially inappropriate medications. Tools such as STOPPFrail, OncPal, and the Unnecessary Drug Use Measure can facilitate deprescribing. Outcome measures vary and selection of measures should align with the operationalized deprescribing definition used by study investigators. Summary EOL deprescribing considerations include medication appropriateness in the context of patient goals for care, expected benefit from medication given life expectancy, and heightened potential for medication-related harm as death nears. Additional data are needed on how EOL deprescribing impacts patient quality of life, caregiver burden, and out-of-pocket medication-related costs to patients and caregivers. Investigators should design deprescribing studies with this information in mind.


Introduction
Deprescribing is a term referring to the process of reducing inappropriate polypharmacy by discontinuing, withdrawing, or tapering medications under the direction of a healthcare provider [1]. The goal of deprescribing is to minimize potential medication risks and to achieve better patient outcomes [2].
The end of life is an important time for deprescribing, particularly for older adults, because polypharmacy is common [3][4][5] in the face of progressively declining organ function that increases older adults' risk of drug-related harm [6][7][8]. Many medications at end of life are of questionable benefit [9][10][11], yet most medications continue until death, while the number of end-of-life (EOL) symptom management drugs increases [5]. This changing combination of medications contributes to stress for patients and caregivers [12,13].
Studying deprescribing at end of life is important but methodologically challenging. One issue is operationalizing the definition of deprescribing. Another issue is the selection of appropriate outcome measures. Designing valid clinical trials and observational studies must balance numerous factors, including validity, utility, psychometric properties of measures, and relevance to stakeholders.
To address these challenges, we conducted this narrative review of published peer-reviewed literature to describe approaches to the measurement of deprescribing and related outcomes for studies of patients at end of life. The intention is to highlight, for a range of domains of interest to deprescribing stakeholders, what strategies have been used, whether they can be used in studies relying upon primary or secondary data collection approaches, and what may be desirable but needs further research to bring into the deprescribing toolbox. This paper will serve as a resource for investigators conducting studies in this growing area of research. was performed to ensure the inclusion of relevant articles and resources. Heterogeneity across studies precluded pooling data, and herein we present results in narrative format. We first describe several methods for defining deprescribing and then organize relevant outcome measures aggregated by patient and family caregivers and into the following domains: medication-related outcomes, patient-reported outcomes, clinical outcomes, safety outcomes, and cost.

Deprescribing
Deprescribing includes the following: (1) drug discontinuation, (2) deintensification (i.e., gradual dose reduction without discontinuation), or (3) switching from inappropriate to appropriate drugs. Operationalizing the deprescribing definition depends on the study design, objectives, and conceptual definition of deprescribing. Observational studies may describe the prevalence of discontinuation and deintensification, while interventional studies may focus on lowering overall polypharmacy, reducing the number of inappropriate drugs, or stopping a particular class of drugs. Each of these aspects of deprescribing raise different measurement challenges.
Determination of whether discontinuation, deinten sification, or both should be included in a study can depend, in part, on the clinical recommendations for deprescribing which are specific to each drug class. Referencing drug-specific deprescribing guidelines can ensure that the more generalized definition applied (e.g., tapering vs abrupt discontinuation) represents the true prevalence of deprescribing in a population. Drug specific recommendations for tapering vs discontinuation are emerging [14, 15·]. Examples of commonly accepted recommendations for select medications appear in the Supplemental Table. For best practice, investigators may include sensitivity analyses to compare results with various deprescribing definitions.
Below we highlight approaches to operationalizing discontinuation and gradual dose reduction in observational studies.

Discontinuation/Gap Periods in Drug Exposure
One challenge to operationalizing drug discontinuation is the dynamic nature of drug prescribing, the unavailability of medical records for many studies, and the lack of explicit notation stating an unequivocal intention to stop medications when medical records are available. One general approach frequently used to define discontinuation relies on identifying gap periods in medication use. Gap periods have included 7 days [16], 14 days [17], and 30 days [18] between drug fills in published end-of-life studies based on pharmacy dispensing and claims. Non-end-of-life studies have also used 60 days [19] or 90 days [20].
Specific study examples appear in Table 1. In addition, gap periods can be determined by drug-specific pharmacokinetic properties (e.g., drug discontinuation defined as a washout period of five times the drugs half-life, representing complete excretion of the drug and biologically corresponding to an expected physiological response) [21].

Deintensification and Gradual Dose Reduction
Deprescribing may involve dose reduction without discontinuation. Defining and operationalizing gradual dose reduction is challenging because administrative claims data often lack detailed data on drug doses. Thus, operationalizing this aspect of the deprescribing approach remains largely underexplored. One important recent review [22··] highlights specific challenges, while describing innovative composite dosage intensity measures and quantified deintensification within three therapeutic classes: central nervous system agents, antihypertensives, and antidiabetic agents. Of note, each drug class required a different composite measure. For example, antihypertensive drug intensity uses defined daily dose (DDD) or maximum recommended daily dosage in the denominator. This approach has been successfully applied using hospice medication administration records in one Veterans Administration-based study [16] (Table 1).
After defining deprescribing, identifying appropriate outcome measures from the patient and caregiver perspectives is critical.

Patient Outcomes
Outcomes of interest to patients can be considered in terms of medication burden, medication-related clinical outcomes, quality of life effects, and economic impact.

Medication-Related Outcomes
Number of Medications-Number of medications and prevalence of polypharmacy are common deprescribing outcome measures. Polypharmacy is defined as the concurrent use of multiple medications surpassing a given threshold, typically 5 ("polypharmacy") and 10 ("excessive polypharmacy"). In practice, the operationa lization of polypharmacy is inconsistent across studies in terms of the type of data available, length of measurement time, and medications counted towards "polypharmacy." When prescription information is available in patients' complete health records, a pharmacotherapy review can provide an accurate "point prevalence" (i.e., prevalence at a particular point of time) of patients' medication use by day [23] [37] have not yet eliminated the need for manual data entry.

Medication Appropriateness
Types of medications used are also important to consider. Determining medication appropriateness is consistent with the goal of improving the quality and safety of a patient's treatment plan. Medication appropriateness can be broadly defined using either implicit or explicit criteria. Implicit criteria are judgement-based indicators that focus primarily on the patient rather than drugs or diseases [38]. Explicit criteria are clearly defined lists of potentially inappropriate medications (PIMs). Implicit approaches tend to require time consuming review by knowledgeable clinicians [39], thus increasing demand on investigator resources when used.

Implicit Measures
For end-of-life studies, the Unnecessary Drug Use Measure defines unnecessary medications based on a simplification of the Medication Appropriateness Index. This definition includes medications that lack an indication, lack effectiveness, or are used for prolonged duration. One uncontrolled, retrospective study in geriatric palliative care simplified regimens by reducing unnecessary drug use by 65% (from 1.7 to 0.6 per patient) [40·] ( Table 2).

Explicit Measures
The Screening Tool of Older Persons' Prescriptions in Frail adults (STOPPFrail) was a major advance in the field. STOPPFrail, developed in 2017, is an explicit list of 27 medications that may be inappropriate in frail older adults aged 65 years or older with limited life expectancy [41]. It is based on literature review and Delphi consensus process. A study comparing the STOPPFrail to a structured geriatricianled deprescribing process (gold standard) among 100 clinical cases reported a positive predictive value of 89.3% [42]. This approach is clinically applicable and efficient, taking 3 min to apply when reviewing regimens of 10 drugs on average, and has high interrater reliability [42]. In a randomized controlled trial, it successfully reduced medication use [43··]. For cancer, OncPal was developed to aid deprescribing for patients transitioning from curative to palliative care [45·]. Based on a literature review and validated using an expert panel approach, there was excellent agreement (kappa = 0.83) between OncPal and an expert panel review ofdrugs. Among 61 patients with a life expectancy of < 6 months, the OncPal tool identified at least one inappropriate medication in 70% of patients.

Additional Dimensions of Medication Appropriateness
Early studies of EOL medication appropriateness broadly divided medications into symptom management and chronic disease medications [5]. While intuitively appealing, a limitation of this approach is that there is no consensus definition for either category. Studies need to tailor and redefine these categories to meet the constraints of their unique data sources [46].
More recently, medication appropriateness screening tools have sought to extend beyond dichotomous determinations of appropriate vs inappropriate. Such binary measures largely fail to consider the relevant time point of medication use (i.e., drug initiation or continuation) and the level of potential appropriateness (i.e., adequate, questionable, inadequate). Morin et al [47] sought to improve medication appropriateness tools for older adults with a life expectancy (< 3 months). Using a Delphi consensus panel, they identified 14 drug classes deemed "often adequate," 28 "questionable," and 10 "often inadequate" medications for continuation. For medication initiation, the panel identified 10 "often adequate," 23 "questionable," and 23 "often inadequate" drugs. In subsequent work, the authors found that in the last 3 months of life, older adults were taking approximately 9 medications, and approximately one-third continued and 14% initiated at least one drug of questionable clinical benefit [48].

Goal-Concordant Prescribing
The holy grail of care quality in end of life is goal-concordant care [49][50][51]. This concept is a nod to shared decision-making between clinicians and patients based on a partnership of equals [52] and focuses on the patients' goals. This is a radical departure from outdated notions of compliance and adherence that reflect whether a drug is taken according to the clinicians' orders [52]. At the end of life, goal-concordant prescribing would align medications prescribed with the patients' goals for care. Thus, reductions in overall numbers of medications or complexity may not necessarily be desirable if the overall regimen is congruent with and able to achieve stated treatment goals. While conceptually desirable to include in deprescribing studies, how to measure this remains an ongoing challenge [53].

Safety
Medication changes may result in positive or negative outcomes. Some studies use a capture of clinical events such as unplanned hospitalizations and emergency department visits to broadly measure adverse events [54, 55··] (Table 2). Inclusion of death in such measures, however, could be problematic because death is an expected in this terminally ill population. One approach used in a classic deprescribing trial was to compare occurrence of death within 60 days between discontinuation and continuation arms as evidence of safety for stopping the medication of interest [55··].
More specifically, of great interest in deprescribing is the occurrence of adverse drug withdrawal events (ADWEs), defined as a "clinically significant set of symptoms or signs caused by the removal of a drug" [56]. Also of concern are adverse drug events (ADE), defined as "any injury resulting from use of a drug" including noxious or unintended drug responses at normal doses [57]. A recent review of adjudication methods to identify ADWEs and ADEs is helpful to consider when thinking about measuring these outcomes [58··]. This adjudication review summarized approaches used by clinical experts, including the following: a probability scale questionnaire [59]; a Likert-scale based measure of clinical plausibility and implied causality [60]; expert clinical judgment [61]; case assessment combining chronological, clinical, and lab findings [62]; and case assessment using a causal relationships from de-challenge and re-challenge results [63], with a version including ADE avoidability [64]. Other approaches use the following: lab tests, causality, and drug administration timing to detect unknown and unexpected adverse events [65]; patient self-report [66]; or computerized ADE monitoring [67]. Notably, distinguishing between ADWEs and ADEs remains difficult. One strategy proposes to focus on adverse event identification, and then: "Once… identified, greater focus…on determining if it represents an ADE or ADWE by adapting an existing framework to meet the needs identified" [58].
While not recent, it is important to note that Hanlon et al. developed an approach that can be applied to medical record review or patient interviews [68]. In a seminal study of ADWEs [56], medical records were screened for potential ADWE signals, including symptoms, signs, abnormal laboratory results, or clinical events known to be associated with drug withdrawal in published literature. Automation of this, and all ADE/ADWE approaches, to electronic medical records or for clinical data such as the nursing home Minimum Data Set remains a significant challenge, with exploration of natural language processing offering a promising development [69].

Patient-Reported Outcomes
Patient-reported outcomes capture aspects of end-of-life care that are important to patients and their families. Patient-reported outcomes, particularly relevant to end-of-life care, include quality of life and symptom control, both physical and psychological. There are several instruments to measure these outcomes. Fortunately, the Palliative Care Research Cooperative (PCRC) group (https://palliativecareresearch.org/) has developed a measurement tool library that contains 205 relevant and high quality instruments for palliative care research as of August 2020. For each instrument, the library contains a brief instrument description, the relevant disease setting, the target respondent, and administration procedures (number of items, requirement for a trained interviewer, recall period, and average time required). The library also provides an assessment of the instrument quality and key references.
The PCRC measurement tool library contains 26 measures that assess quality of life in end of-life settings. Some instruments are specific to a disease or condition (e.g., Alzheimer's disease, cancer), while others pertain to a given care setting (e.g., nursing home). These instruments range in time needed to complete the assessment from 5 min to 1 h. One example of a high quality-rated measure includes the McGill Quality of Life Questionnaire [70], which measures four subscales including physical symptoms, psychological symptoms, outlook on life, and meaningful existence, and a single item question about physical wellbeing.
The library also includes 12 measures, all rated as high quality, assessing a range of physical symptoms, 8 of which are general measures and not disease specific. One example of a physical symptom instrument is the Edmonton Symptom Assessment System [71]. This 10 item tool assesses physical and psychosocial symptom severity and can be applied in general palliative care settings. In addition, this instrument can be completed in approximately 5 min by a patient or their proxy, providing additional flexibility.
Finally, the library includes 9 measures that capture psychological symptoms, with all but one being general tools for use across disease settings. These tools broadly cover measurement of depression, agitation, mood, and communication ability. One example of a psychological symptom measurement tool is the Center for Epidemiologic Studies Depression Scale (CES-D) [72], which measures current depressive symptomatology in the past week. This 20-item instrument is one of the most widely used to measure depressive symptoms in research studies because of its excellent measurement properties, methodological quality, and consistency of results.

Costs
Cost assessment following deprescribing is rarely performed; if it is, the costs assessed are heterogeneous. Statin discontinuation among patients at end of life reduced medication costs by $3.37 [55], using a national average retail price to calculate the avoided statin costs. A different study found significantly lower medication-related costs after deprescribing using the START/STOPP tool. Researchers counted the number of medications taken on the day of patient assessment and then summed medication costs using available price lists from the Israeli health ministry [73]. An intervention delivered to older adults with advanced cancer describes potential savings of $4200 by avoiding ADEs, plus reducing time to take medications histories or perform medication teaching [74]. The authors assumed that each discontinued medication resulted in the avoidance of a major or minor ADE. Prices were calculated using prices established by the health system in which the study took place.

Caregiver Outcomes
Caregiver involvement is a unique aspect of end-of-life care that deserves attention in deprescribing studies, but these are largely underutilized in studies to date ( Table 2). This can include measuring whether caregivers are open to deprescribing medications if a trusted clinician says stopping is possible [75]. Caregiver attitudes towards deprescribing can be assessed using the revised Patient Attitude Towards Deprescribing (rPATD) questionnaire, caregiver version [76].

Medications Administration and Burden
Caregivers play an important role in medication management for patients, from obtaining and administering medications to monitoring for symptom relief and endorsing medication use. Most caregivers report spending significant time ordering or administrating medication for family members at end of life [77]. Caregivers have described polypharmacy burden and lack of clear direction from healthcare providers as significant sources of distress when caring for a dying loved one [78]. For older adults, the degree of caregiver involvement, hospice enrollment, decedent symptom-related distress, and financial burden are associated with higher caregiver burden [79]. The Family Caregiver Medication Administration Hassles Scale (FCMAHS) measures daily irritants associated with medication-related duties for caregivers; monitoring for an increase in irritations can facilitate interventions before significant caregiver distress occurs [80].

Clinical Outcomes
Medication-related care may invoke anxiety among caregivers stemming from uncertainty around medication dosages and use of "as needed" medications and fear of overdosing patients, or committing medication-related errors that may shorten patient survival [81]. This may result in emergency department visits stemming from caregiver concerns about ability to manage symptoms and medications at home [82]. Factors contributing to caregiver distress include misunderstanding of pain medications, poor communication with clinicians, and inadequate knowledge of medications/assessment of pain symptoms [83]. Multiple tools are available to assess caregiver burden. The Caregiver Quality of Life Index [84] is available with multiple versions, including some specific to a particular condition such as cancer. The Ways of Coping Scale has been used to measure caregiver coping with patient symptoms and health [85]. Some caregivers report significant distress after observing patient symptoms at end of life, which can be assessed using the Stressful Caregiving Response to Experiences of Dying scale [86]. The Self-Perceived Pressure from Informal Care Scale and the Positive Experiences Scale have been used to measure caregiver burden and positive experiences stemming from providing care [87]. Other tools include the Zarit Burden Interview and the Center for Epidemiological Studies Depression Scale (CES-D) [88]. Following the loss of a patient, caregiver grief can be evaluated using the Inventory of Complicated Grief [89].

Costs
Time costs are important, as caregivers provide on average 40-70 h of care a week to patients at the end of life [90,91]. Caregiving-related costs can be calculated by surveying individuals, then assigning costs specific to reported activity. Higginson and colleagues surveyed bereaved caregivers using the Client Service Receipt Inventory to calculate hours spent caregiving. Then, the number of hours was multiplied by average hourly wages and cost of nursing care. In the last 3 months of life, average cost to informal caregivers in the USA was $32,468 (SD $28,578) [92]. An additional tool for cost measurement is the Ambulatory and Home Care Record [93]. Beyond hourly time and associated wages, caregiver diaries can measure out-of-pocket expenses and medication-related costs to caregivers and families. For example, caregivers can record all services used (e.g., caregiver respite not paid for by insurance) and expenses such as parking for doctor's visits [94].

Discussion
This narrative review summarizes existing approaches to the measurement of deprescribing and key outcome variables for studies examining deprescribing at the end of life. For each measure, there are several approaches to consider. The selection of appropriate measure follows from the method and study design, and design follows from the research question and the investigators' vision for the study goals. Some studies implemented randomized clinical trials [43, 55], while others were observational pre-post studies [24,40]. Of note is the small sample size of most of the available studies. As with all epidemiological studies, the final choices depend on the details and the perspective of the stakeholders who will utilize the study results.
Since patients are central to all deprescribing studies, and the goal of deprescribing is to reduce medication burden and improve (or maintain) quality of life [1], we have summarized approaches to capturing these measures as they have been operationalized to date. Somewhat unique to EOL research is the perspective of both the patient and the caregiver. We have summarized measures including burden of medication administration and caregiver quality of life, which rely on primary data collection. While costs are important to patients and caregivers, detailed costing methods are beyond the scope of this paper. Issues such as the distinction between out-of-pocket costs vs total costs and costs versus charges are important to consider in this domain.
We suggest addressing the following issues before the application of polypharmacy as an outcome variable for deprescribing interventions in the end-of-life population.
First, identify the goal of the study and what data are available. Patient-level prescription data from medication charts can provide an accurate "point prevalence" of patients' medication use. With individualized medication history to define polypharmacy, providers could compare the drug list with medical history to initiate and follow the deprescribing process. However, such studies are limited to the end-of-life population in well-defined settings leading to limited external validity, making it difficult to validate the outcome for the deprescribing intervention and to generalize to a larger population. Alternatively, studies using registry or administrative databases can be generalized to a broader population but cannot capture the effect on individualized patient-level-deprescribing processes. Such studies can assess the medications that the patients were prescribed or dispensed, but may not reflect what the patient actually took.
Second, identify the definition of polypharmacy. Three questions should be asked. First, what medications should be included? Although the types of medications counted towards polypharmacy are generally "chronic" medications, they may vary with regard to the specific health conditions of the end-of-life population in the study. Second, how long should medications use be measured? This depends on medication type and patients' conditions. If studies included drugs for symptom management or preventive medications in the total polypharmacy count, does the assessment time length for these drugs differ from the chronic medications that were also included? Lastly, given the baseline medication use of the study population, the common threshold of 5 or 10 medications might not be an optimal cut-off. If many medications are deprescribing targets for an end-of-life population of interest, a more conservative threshold of greater number of medications may be necessary.
For both deprescribing and outcome measures, there are differences in definitions based on whether primary data (e.g., survey) or secondary data (e.g., administrative claims or medical records) are used for assessment. Clinical outcomes such as mortality or healthcare utilization can be garnered from claims and medical records. Patient-reported outcomes such as symptom burden and quality of life are mostly reliant on primary data collection, but investigators are starting to explore methods for using secondary data sources such as the Minimum Data Set for nursing home residents. At the edge of existing methods are theoretically important outcome measures like goal-concordant prescribing that need further study for definition and operationalization from data sources. Data captured by a given resource often limits the range of variables available for inclusion. For example, there may be subtle but important differences between drug prescribing orders from medical records, drug dispensing from pharmacy records, and drug claims from administrative records.
Final selection of measures should follow from basic epidemiologic design, which requires that the investigator weigh measures based on balance of validity, reliability, feasibility, and acceptability. Feasibility and acceptability are particularly challenging in end-of-life research as primary data collection from patients who are seriously ill or dying, and their caregivers who are bereaved, is considered ethically challenging. Given these complex trade-offs, the use of a stakeholder panels can help inform study design decisions.

Limitations
Our review is limited to patient and caregiver perspectives. We did not include measures reflective of other stakeholders, including clinicians and payers [95]. Specifically, we did not review healthcare utilization measures. We focus largely on older adults in this review, which represents the available literature. We do not address the issue of how to define end of-life, which has no widely accepted criteria to operationalize. This is an important issue for identifying the population under study and enumerating an appropriate denominator [18]. Such definitions are important to note because they affect the magnitude of the prevalence of deprescribing and medication appropriateness.

Conclusion
Deprescribing studies at EOL should consider medication types, amounts, and appropriateness in the context of patient goals of care. Additional data are needed on deprescribing at EOL in populations other than older adults. Future research should evaluate the association between deprescribing and patient quality of life, caregiver burden, and out-of-pocket medication-related costs to patients and families.

Supplementary Material
Refer to Web version on PubMed Central for supplementary material.

Administrative data
Machine learning using methods such as national language processing could theoretically detect explicit notation of deliberate deintensification/dose reductions in the medical records  Table 2 Outcome measures in end-of-life deprescribing studies