Assessing the Benefits and Harms of Pharmacotherapy in Older Adults with Frailty: Insights from Pharmacoepidemiologic Studies of Routine Health Care Data

Abstract

The objective of this review is to summarize and appraise the research methodology, emerging findings, and future directions in pharmacoepidemiologic studies assessing the benefits and harms of pharmacotherapies in older adults with different levels of frailty. Older adults living with frailty are at elevated risk for poor health outcomes and adverse effects from pharmacotherapy. However, current evidence is limited due to the under-enrollment of frail older adults and the lack of validated frailty assessments in clinical trials. Recent advancements in measuring frailty in administrative claims and electronic health records (database-derived frailty scores) have enabled researchers to identify patients with frailty and to evaluate the heterogeneity of treatment effects by patients’ frailty levels using routine health care data. When selecting a database-derived frailty score, researchers must consider the type of data (e.g., different coding systems), the length of the predictor assessment period, the extent of validation against clinically validated frailty measures, and the possibility of surveillance bias arising from unequal access to care. We reviewed 13 pharmacoepidemiologic studies published on PubMed from 2013 to 2023 that evaluated the benefits and harms of cardiovascular medications, diabetes medications, anti-neoplastic agents, antipsychotic medications, and vaccines by frailty levels. These studies suggest that, while greater frailty is positively associated with adverse treatment outcomes, older adults with frailty can still benefit from pharmacotherapy. Therefore, we recommend routine frailty subgroup analyses in pharmacoepidemiologic studies. Despite data and design limitations, the findings from such studies may be informative to tailor pharmacotherapy for older adults across the frailty spectrum.

1. INTRODUCTION

Optimizing pharmacotherapy in older adults presents a major challenge to clinicians. Age-related changes in organ function (e.g., reduced liver and kidney function), body composition (e.g., decreased total body water and lean body mass), and the number and affinity of receptors affect pharmacokinetics and pharmacodynamics of the drugs.[1] These age-related changes are heterogeneous among individuals and accelerated by disease process.[2, 3] Most older adults have multiple coexisting chronic conditions and adhering to disease-specific practice guidelines increases the likelihood of drug-disease and drug-drug interactions, treatment burden, and financial costs.[4, 5] In particular, older adults living with frailty—a state of decreased physiologic reserve and increased vulnerability to stressors[6]—are at elevated risk for harms from pharmacotherapy.[7] Concerned about harms, clinicians tend to be hesitant to prescribe newly approved drugs to those with frailty.[8–10] Moreover, those with severe frailty nearing the end of life may not live long enough to benefit from preventive medications.[11] Prescribing a drug to patients in whom meaningful benefits are unlikely is referred to as overtreatment. On the other hand, older adults with frailty have a higher baseline risk for developing poor health outcomes from a target disease of pharmacotherapy (e.g., cardiovascular disease) and, therefore, may derive more benefit than those without frailty.[12, 13] Withholding an effective drug therapy in an attempt to avoid treatment-related adverse events leads to undertreatment of a target disease that can be otherwise managed. Although clinicians make treatment decisions after carefully considering patients’ vulnerability to harm, remaining life expectancy, and risk of poor disease outcomes, empirical evidence to guide such decisions is lacking.

Considering that older adults with frailty constitute the majority of medication users, there is an urgent need to develop an evidence base for more tailored pharmacotherapy that addresses their unique physiological changes and vulnerability to adverse events. Frailty likely affects both pharmacokinetics—due to reduced drug clearance from changes in liver and kidney function and a reduced volume of distribution from weight loss and sarcopenia—and pharmacodynamics—characterized by altered dose-response relationships due to reduced homeostatic reserve and increased vulnerability to stressors. However, the clinical impacts of the pharmacokinetic and pharmacodynamic changes related to frailty remain poorly understood. It is uncertain which frailty assessment tool is best suited for assessing these impacts.[14, 15] Moreover, clinical trials evaluating a pharmacotherapy rarely consider frailty in selecting participants, and endpoints do not include patient-centered outcomes (e.g., treatment burden) or geriatric syndromes (e.g., falls, delirium, or functional status) that are relevant to older adults with frailty.[16] To bridge this critical gap, the International Union of Basic and Clinical Pharmacology Geriatric Committee[17] and other experts from Europe[16] have recently advocated for the integration of frailty into drug development and evaluation for older adults. Their position statements called for real-world evidence on the comparative safety and effectiveness studies of pharmacotherapies. To this end, there has been a growing interest in investigating the benefit-harm balance of individual drugs across the fit-to-frail spectrum utilizing routinely collected health data, recognizing that, beyond a certain level of frailty, the harms of some drugs may outweigh their benefits, while other drugs may provide benefits regardless of patients’ frailty level.[18]

The objective of this review is to summarize and appraise the research methodology, emerging findings, and future directions in pharmacoepidemiologic studies designed at generating evidence on how the benefits and harms of pharmacotherapies vary with patients’ frailty levels. The outline of the review is as follows: we first discuss the findings and limitations of post-hoc analyses of clinical trials to assess the effect of pharmacotherapy by frailty subgroups. To overcome limitations of post-hoc analyses of clinical trials, we discuss the effort to measure frailty from routinely collected health care data (database-derived frailty index) and the potential utility of database-derived frailty index for pharmacoepidemiologic studies using routinely collected health care data. We conclude with evidence gaps and future directions to achieve frailty-guided pharmacotherapy for older adults.

2. LITERATURE SEARCH

We searched PubMed to find English language articles published between January 2013 and December 2023 on 1) the review of database-derived frailty index (search terms: frailty, claims, electronic health records [EHR]); 2) post-hoc analyses of randomized controlled trials evaluating a pharmacotherapy on cardiovascular disease by participants’ frailty levels (search terms: frailty, cardiovascular disease, randomized controlled trial); and 3) pharmacoepidemiologic studies evaluating treatment effect by patients’ frailty levels (search terms: frailty, propensity score, claims, EHR). Because there have been several dedicated reviews on frailty measurement in claims data or EHR data that covered numerous database-derived frailty indices,[18–24] we decided to focus on commonly used models whose original paper was cited more than 100 times as of December 31, 2023, according to Google Scholar.

3. CLINICAL TRIALS FOR ASSESSING THE EFFECT OF PHARMACOTHERAPY IN OLDER ADULTS WITH FRAILTY

Exclusion or substantial under-enrollment of patients with severe frailty is typical in clinical trials of pharmacotherapy based on the investigators’ assessment of patients’ remaining life expectancy or ability to adhere to study visits and procedures.[25, 26] A recent analysis of individual patient data from clinical trials found that mild to moderate frailty was common in trial participants and that they were more likely to experience serious adverse events.[7] Yet most trials do not include a formal frailty assessment.

There are several validated assessment tools for frailty[6, 27]; among them, the Fried physical frailty phenotype[28, 29] and the deficit accumulation frailty approach[30, 31] are most commonly used. The Fried physical frailty phenotype defines frailty as a clinical syndrome that is characterized by having 3 or more features of muscle weakness (measured by handgrip dynamometer), slow gait (measured from timed walks), physical inactivity (measured by a physical activity questionnaire), weight loss (self-report), and fatigue (self-report).[28, 29] The deficit accumulation frailty approach defines frailty as a state of poor health that occurs as a result of accumulation of health deficits.[30, 31] A deficit accumulation frailty index can be calculated as the number of abnormal health-related attributes (e.g., symptoms, diagnoses, activities of daily living disability, cognitive impairment, or abnormal laboratory tests) divided by the total number of attributes assessed (range: 0 to 1; frailty index ≥0.25 is considered as having frailty[32, 33]). Any attributes can be chosen as long as they increase in prevalence with age and represent abnormalities in multiple organ systems.[34] Since most clinical trials of pharmacotherapy do not collect handgrip strength and timed walks, the Fried physical frailty phenotype cannot be measured. However, the flexibility of health deficit selection allows calculation of a deficit accumulation frailty index using the health attributes collected at the trial baseline.[7] The treatment effect can be estimated and compared across the subgroups defined by frailty levels.

Table 1 summarizes post-hoc analyses of 7 clinical trials that assessed the effect of cardiovascular drugs by participants’ frailty level. In 6 trials,[35–40] a deficit accumulation frailty index was calculated using 30[40] to 60 variables[35] collected at baseline. One trial[41] modified the Fried physical frailty phenotype by replacing muscle strength and gait speed measurements with self-reported responses. Regardless of how frailty was defined, none of the post-hoc analyses found a statistically significant difference in the effect estimates by participants’ frailty level.

Table 1.

Post-Hoc Analysis of Clinical Trials to Evaluate the Effect of Cardiovascular Drug Therapy by Frailty Levels

Clinical Trial	Target Condition	Interventions	Primary Outcome	Frailty Tool	Treatment Effect HR (95% CI)a
HYVET (2015)[35]	Hypertension	Indapamide ± perindopril vs placebo	Stroke	Deficit accumulation FI (60 deficits)	Fit: 0.75 (0.40–1.38) Frail: 0.41 (0.10–1.65)
TRILOGY-ACS (2016)[41]	ACS	Prasugrel vs clopidogrel	MACE	Physical frailty phenotype (self-report)	Fit: 0.90 (0.77–1.06) Frail: 0.89 (0.54–1.46)
SPRINT (2016)[36]	Hypertension	Intensive vs standard treatment	MACE	Deficit accumulation FI (37 deficits)	Fit: 0.47 (0.13–1.39) Frail: 0.68 (0.45–1.01)
TOPCAT (2018)[37]	HFpEF	Spironolactone vs placebo	CV death or HF admission	Deficit accumulation FI (39 deficits)	No treatment effect heterogeneity by frailtyb
ENGAGE AF-TIMI 48 (2020)[38]	AF	Edoxaban vs placebo	Stroke or systemic embolism	Deficit accumulation FI (40 deficits)	Fit: 1.03 (0.71–1.49) Frail: 0.54 (0.20–1.50)
DAPA-HF (2022)[39]	HFrEF	Dapagliflozin vs placebo	CV death or HF admission	Deficit accumulation FI (32 deficits)	Fit: 0.72 (0.59–0.89) Frail: 0.71 (0.54–0.93)
DELIVER (2022)[40]	HFrEF	Dapagliflozin vs placebo	CV death or HF admission	Deficit accumulation FI (30 deficits)	Fit: 0.85 (0.68–1.06) Frail: 0.74 (0.61–0.91)

Abbreviations: ACS, acute coronary syndrome; AF, atrial fibrillation; CI, confidence interval; CV, cardiovascular; FI, frailty index; HF, heart failure; HFpEF, heart failure with preserved ejection fraction; HFrEF, heart failure with reduced ejection fraction; HR, hazard ratio; MACE, major adverse cardiovascular events.

^aTreatment effects for the fittest group and the frailest group are shown. None of the clinical trials found statistically significant heterogeneity in treatment effect by frailty levels.

^bQuantitative results were not reported by frailty levels.

Despite the intriguing findings, post-hoc frailty scores are not suitable for guiding treatment decisions based on frailty in clinical practice.[42] The variables used to calculate a frailty score in clinical trials are not always measured in routine clinical practice. Different ad hoc or unvalidated compositions of a deficit accumulation frailty index[43] or modifications to the Fried physical frailty phenotype[44] can introduce measurement error and misclassification of frailty levels. The discrepancy between post-hoc frailty scores and frailty assessment tools used in clinical practice makes it challenging for clinicians to identify patients in their clinical practice to whom the frailty-specific treatment effect estimated from clinical trials can be generalized.

Moreover, a subgroup analysis by frailty levels is typically underpowered, resulting in imprecise effect estimates (Table 1). For example, a post-hoc analysis of ENGAGE AF-TIMI 48 trial[38] reported that the hazard ratio (95% confidence interval) of stroke or systemic embolism comparing edoxaban and placebo was 1.03 (0.71–1.49) for the fittest participants (frailty index <0.12) and 0.54 (0.20–1.50) for the frailest participants (frailty index ≥0.36). However, with less than 2% of trial participants in this frailest category, the estimates with wide 95% confidence intervals do not permit determination of whether edoxaban reduces the primary outcome within each frailty subgroup or whether the effect of edoxaban changes based on patients’ frailty level.

For the forementioned reasons, post-hoc frailty subgroup analyses in clinical trials should be considered hypothesis-generating, pending confirmation by an adequately powered clinical trial with a validated baseline assessment of frailty.[42] However, enrolling and retaining older adults with frailty in clinical trials is challenging. This difficulty was clearly observed in a recent clinical trial that compared a routine invasive vs conservative strategy in older adults with frailty and non-ST elevation myocardial infarction.[45] In this study, the investigators enrolled 167 patients across 13 centers over 3.5 years. Despite their dedicated effort, they were only able to enroll 4 patients with severe frailty, highlighting the challenges in recruiting this subgroup for a clinical trial.

4. FRAILTY MEASUREMENT IN ROUTINE HEALTH CARE DATA FOR PHARMACOEPIDEMIOLOGIC STUDIES

Given the limitations of clinical trials, researchers have utilized data collected from routine health care practice, such as administrative claims data and EHR, to evaluate pharmacotherapies.[46] The purpose of geriatric pharmacoepidemiology is to confirm whether the efficacy and safety observed from clinical trials are similar in older people in routine care settings, including subgroups under-represented in clinical trials, and to identify rare safety signals that clinical trials may not detect due to their limited sample size.[18, 47] It also facilitates the comparative effectiveness and safety analysis of two drugs that are used for similar indications but not yet tested in randomized controlled trials. Although routine health care data provide information on drug utilization (e.g., prescription fills) and clinical outcomes in a broad range of populations, including those with frailty, the absence of specific diagnosis codes for frailty has, until recently, been a major barrier to evaluating the outcomes of pharmacotherapy in this population.

To overcome this barrier, researchers have developed frailty assessment methods for administrative claims data and EHR data (comprehensive reviews on this topic are published elsewhere[18–24]). The characteristics of the eight most commonly used database-derived frailty indices are summarized in Table 2. Five models[48–52] were developed for United States Medicare claims data, one for United States Veterans Affair claims data,[53] one for United Kingdom inpatient claims data,[54] and one for United Kingdom primary care EHR data.[55]

Table 2.

Commonly Used Frailty Scores for Routine Health Care Data

First Author (Year)	Database	Predictor Selection	Predictor Assessment Period	Types of Predictors	Interpretation	Validation Outcomes
Frailty Scores for Administrative Claims Data
Davidoff (2013)[48]	Medicare claims data, United States	Regression-based (reference standard: disability)	12 months	Sex Diagnosis Procedure & health service Medicaid enrollment status Health care utilization	0 to < 0.11: non-frail ≥ 0.11: frail	Clinical frailty assessment Disability Mortality
Faurot (2015)[49, 60]	Medicare claims data, United States	Regression-based (reference standard: disability)	8 months	Age, sex, race Diagnosis Procedure & health service	0 to < 0.05: non-frail ≥ 0.05: frail	Clinical frailty assessment Disability Mortality
Segal (2017)[50, 61] (Johns Hopkins Claims-based Frailty Indicator)	Medicare claims data, United States	Regression-based (reference standard: Fried frailty phenotype)	6 months	Age, sex, race Diagnosis Comorbidity index Health care utilization	0 to < 0.12: non-frail ≥ 0.12: frail	Clinical frailty assessment Disability Mortality
Figueroa (2017)[51, 64]	Medicare claims data, United States	Clinical knowledge (no reference standard)	12 months	Diagnosis Procedure & health service	0 to 1: non-frail ≥ 2: frail	Clinical frailty assessment
Kim (2018)[52, 56, 62]	Medicare claims data, United States	Regression-based (reference standard: deficit accumulation frailty index)	12 months	Diagnosis Procedure & health service	0 to < 0.15: robust 0.15 to < 0.25: pre-frail 0.25 to < 0.35: mildly frail 0.35 to < 0.45: moderately frail ≥ 0.45: severely frail	Clinical frailty assessment Disability Mortality
Gilbert (2018)[54, 63] (Hospital Frailty Risk Score)	Inpatient claims data, England	Cluster analysis (no reference standard; prediction of the frailty cluster)	24 months	Diagnosis	< 5: low risk of frailty 5 to 15: intermediate risk of frailty >15: high risk of frailty	Clinical frailty assessment Mortality
Orkaby (2019)[53, 59] (Veterans Affairs-Frailty Index)	Veterans Affairs claims data, United States	The standard procedure for deficit selection by Searle et al.[34] (no reference standard; direct implementation of deficit accumulation)	36 months	Diagnosis Procedure & health service	0 to ≤ 0.1: robust 0.1 to ≤ 0.2: pre-frail 0.2 to ≤ 0.3: mildly frail 0.3 to ≤ 0.4: moderately frail ≥ 0.4: severely frail	Clinical frailty assessment Mortality
Frailty Scores for EHR Data
Clegg (2016)[55, 57, 65] (electronic Frailty Index)	Primary care EHR data, United Kingdom	The standard procedure for deficit selection by Searle et al.[34] (no reference standard; direct implementation of deficit accumulation)	12 months	Diagnosis Procedure & health service Pharmacy Social determinants of health	0 to 0.12: fit 0.12 < to 0.24: mildly frail 0.24 < to 0.36: moderately frail > 0.36: severely frail	Clinical frailty assessment Mortality

Abbreviation: EHR, electronic health records.

The methodologies employed by researchers to develop these models differed based on whether a reference standard clinical frailty measure was available in their dataset. Due to the lack of clinical frailty assessments in routine health care databases, Figueroa et al.,[51] the Veterans Affairs-Frailty Index by Orkaby et al.,[53] and the electronic Frailty Index by Clegg et al.[55] relied on administrative claims data or EHR data. Figueroa et al.[51] defined frailty-related conditions based on the investigators’ clinical knowledge.[18] Orkaby et al.,[53] and Clegg et al.[55] selected diagnosis codes, procedure or health service codes, or pharmacy codes and calculated a frailty score following the standard deficit accumulation procedure outlined by Searle et al.[34]. Gilbert et al. conducted a cluster analysis using inpatient claims data to define a “frailty” cluster and developed the Hospital Frailty Risk Score using a multivariable logistic model which calculates the probability of being in the frailty cluster.[54] When researchers had access to a large cohort or survey of older adults (in which clinical assessment was conducted) and linked claims data, they applied a regression-based variable selection approach to predict the reference standard from clinical assessment based on diagnosis codes, procedure or health service codes, health care utilization, or sociodemographic information. Different reference standard measures were used: disability (Davidoff index[48] and Faurot index[49]), the Fried physical frailty phenotype (Segal index[50], also known as the Johns Hopkins Claims-based Frailty Indicator), and the deficit accumulation frailty index (Kim index[52]). The Davidoff index and the Faurot index estimate the probability of disability, and the Segal index estimates the probability of having the Fried physical frailty phenotype. The Kim index estimates the deficit accumulation frailty index (range: 0 to 1). All of these frailty models have been tested for the convergent validity (i.e., how well a database-derived frailty index correlates with other clinical frailty measures)[54, 56–61] and predictive validity (i.e., how well a database-derived frailty index predicts poor health outcomes in the future).[48–55, 60–65]

When choosing a database-derived frailty index for pharmacoepidemiologic studies, several factors must be considered. Foremost among these factors is the type of data on hand, as well as the length of the predictor assessment period. Different health systems employ specific coding systems, such as the Healthcare Common Procedure Coding System in the United States and the Read Codes in the United Kingdom. The frailty models that utilize these coding systems cannot be used in a health care system using a different coding system. The length of the predictor assessment period varies from 6 months (Segal index)[50] to 36 months (Orkaby index).[53] The baseline covariates based on a shorter assessment timeframe may be more subject to variable misclassification because a database study usually interprets “lack of a code” as “absence of the condition”, thus translating missing data due to limited data availability into misclassification. On the other hand, requiring a longer predictor assessment period restricts study population to the individuals with complete assessment through this period and reduces the amount of data available for outcome follow-up. It also implies that the model may be less sensitive to detect a change in frailty status over a short period. The length of predictor assessment period may affect the calibration and discrimination for identifying frailty, as well as the predictive ability for mortality.[66] Therefore, a sensitivity analysis using variable predictor assessment periods can be informative.

Another factor is the measurement characteristics of different frailty models. Clinical knowledge-based frailty models may have lower sensitivity and a higher false negative rate of identifying frailty compared with regression-based models.[58] This could result in greater misclassification of frailty status and smaller numbers of people classified as frail. When the regression-based frailty models were compared against clinical frailty assessments in a representative cohort of Medicare beneficiaries that was not used to develop these models, the C statistics (after adjusting for age and sex) for the Fried physical frailty phenotype were 0.73 (Davidoff index), 0.74 (Faurot index), 0.73 (Segal index), and 0.78 (Kim index); the Spearman correlation coefficients with a deficit-accumulation frailty index were 0.22, 0.45, 0.40, and 0.59, respectively.[56] The frailty models that include demographic information[48–50] tend to have poorer performance in predicting frailty status and disability than the model without demographic variables once age and sex are adjusted for in the regression models.[56] Moreover, a crosswalk between the Kim index and nine commonly used clinical frailty measures has been developed to facilitate the clinical translation of research findings using the Kim index in routine health care data.[67]

Regardless of the choice of a database-derived frailty index, it is important to note that its accuracy may be affected by access to care, reimbursement, and coding practice. Diagnosis codes, particularly in EHR, tend to be carried over from visit to visit and accumulate over time. Additionally, health information is more likely to be recorded when an individual is seeking medical attention for worsening health, which leads to a form of surveillance bias called informed presence bias.[68] As a result, database-derived frailty models may be better at capturing instances of worsening frailty than improvements. Nonetheless, it has been observed that a decline or rise in the Kim index over one year correlates with a reduction or increase in the risk of death and health care expenditure in the following year, respectively.[69] Missing data can present a unique challenge in measuring frailty from claims or EHR data. Missingness arises in sick people who do not access health care due to lack of insurance, financial burden, limited social support, or non-adherence as well as healthy people who only need an annual visit. To ensure that all individuals in the dataset had opportunities to have their health conditions recorded, it is advised to require a minimum period of continuous enrollment in medical insurance in claims data[70] or a minimum number of outpatient visits or recording of vital signs in EHR data.[71] To minimize an incomplete capture of health conditions due to discontinuity of care in patients who seek care at more than one hospital, a predictive algorithm has been developed to identify patients who are likely to have high completeness of data within an EHR system.[72]

To date, most database-derived frailty indices have mainly been utilized for research purposes as frailty indices derived from administrative claims data are not readily available during clinical encounters. However, a recent development within the National Health Services in England has seen the integration of Clegg’s electronic Frailty Index into primary care EHR.[73] This allows general practitioners to routinely screen older adults for frailty, facilitating further clinical evaluation, comprehensive geriatric assessment, multidisciplinary interventions, and access to supportive services. While it has a potential to enable a tailored pharmacotherapy in older adults based on their frailty levels, more research and educational efforts are needed to change practice and make clinical impacts.

5. APPLICATIONS OF DATABASE-DERIVED FRAILTY INDEX FOR PHARMACOEPIDEMIOLOGIC STUDIES

Application of a database-derived frailty index in pharmacoepidemiologic studies using routine health care data has been increasing lately. Because frailty is associated with death and poor health outcomes, independently of demographic characteristics and comorbidities,[19–21] a database-derived frailty index can be useful in mitigating confounding by frailty. The utility of a database-derived frailty index for confounding adjustment has been examined by Zhang et al.[74] In their investigation using the United States Medicare claims, the researchers evaluated the association between influenza vaccination and mortality, anticipating a hazard ratio of 1.0 before the influenza season. They found an unadjusted pre-season hazard ratio of 0.61. Such a large reduction in mortality is likely due to confounding; physicians often hesitate to initiate preventive treatments to patients with unclear benefits due to poor prognosis (e.g., frailty). It is presumed that more thorough confounding adjustment would bring the hazard ratio closer to 1.0. In the study, progressive adjustments for confounders moved the hazard ratio to 0.59 after adjusting for age, sex, and race, 0.66 with further adjustment for comorbidities and health care utilization, and 0.68 after adding 20 frailty-related variables from the Faurot index to the model.[49, 60] In a subsequent study, they showed that restricting the study population to the initiators of statins, anti-glaucoma drugs, or beta-blockers was able to remove more than 90% of the bias, resulting in a null association with mortality before influenza season.[75] Although it remains uncertain how effectively other database-derived frailty indices can reduce confounding, these results suggest that adding a database-derived frailty index to demographic characteristics, comorbidities, and health care utilization measures can provide a modest improvement in confounding adjustment. A database-derived frailty index should be used in conjunction with other study design and analytic strategies for confounding adjustment, such as restriction,[74–76] high-dimensional propensity score,[77] propensity score calibration,[78, 79] and external adjustment.[80]

A database-derived frailty index can enable the assessment of treatment effectiveness and safety across different levels of patient frailty. In PubMed, we found 13 studies evaluating the effectiveness and safety of medications (including vaccines) by frailty levels. Table 3 shows the study design characteristics and frailty subgroup treatment effects of cardiovascular medications (5 studies[81–85]), diabetes medications (4 studies[86–89]), anti-neoplastic agents (2 studies[90, 91]), antipsychotic medications (1 study[92]) and vaccines (1 study[93]). In the seven studies conducted in beneficiaries of Medicare or commercial insurance in the United States,[81, 82, 86–89, 93] the Kim index was used to measure frailty. In the four studies conducted in United States veterans,[83, 84, 90, 91] the Orkaby’s Veterans Affairs Frailty Index was used. One study using the Clinical Practice Research Datalink in England applied the Clegg’s electronic Frailty Index and another study using the administrative health and clinical databases of Ontario, Canada, employed the Segal’s Johns Hopkins Claims-based Frailty Indicator.

Table 3.

Pharmacoepidemiologic Studies Evaluating Treatment Effect by Different Levels of Patient Frailty^a

First Author (Year)	Data Source and Study Design	Primary Outcome	Event Rate		Effect Estimate
			Intervention	Comparison
Cardiovascular Medications
Kim (2021)[81]	Data: Medicare claims, United States Population: Beneficiaries ≥ 65 years old, AF, no prior use of oral anticoagulants Exposure: dabigatran, rivaroxaban, warfarin, apixaban Outcome: Composite of death, stroke, or major bleeding Frailty measure: Kim index	Composite clinical events  Robust  Pre-frail  Frail Composite clinical events  Robust  Pre-frail  Frail Composite clinical events  Robust  Pre-frail  Frail	Dabigatran (Per 1000 PY) 25.4 62.1 170.0 Rivaroxaban (Per 1000 PY) 31.6 76.5 200.8 Apixaban (Per 1000 PY) 21.9 54.2 157.6	Warfarin (Per 1000 PY) 31.4 64.1 160.2 Warfarin (Per 1000 PY) 37.2 76.7 219.8 Warfarin (Per 1000 PY) 37.6 86.0 226.2	HR (95% CI) 0.91 (0.68, 0.97) 0.98 (0.90, 1.08) 1.09 (0.96, 1.23) HR (95% CI) 0.88 (0.77, 0.99) 1.04 (0.98, 1.10) 0.96 (0.89, 1.04) HR (95% CI) 0.61 (0.52, 0.71) 0.66 (0.61, 0.70) 0.73 (0.67, 0.80)
Lin (2023)[82]	Data: Medicare claims, United States Population: Beneficiaries ≥ 65 years old, AF, no prior use of oral anticoagulants Exposure: rivaroxaban, warfarin, apixaban Outcome: Loss of days at home 14 days or more in 1 year Frailty measure: Kim index	1-Yr home time loss ≥ 14 d  Robust  Pre-frail  Frail 1-Yr home time loss ≥ 14 d  Robust  Pre-frail  Frail	Rivaroxaban (Per 100) 6.9 17.1 46.0 Warfarin (Per 100) 7.7 18.9 46.9	Apixaban (Per 100) 6.0 16.0 43.4 Apixaban (Per 100) 6.0 16.0 43.4	RD (95% CI) 0.8 (0.5, 1.2) 1.1 (0.7, 1.4) 2.7 (1.9, 3.4) RD (95% CI) 1.7 (1.2, 2.2) 2.9 (2.5, 3.2) 3.6 (2.8, 4.3)
Anderson (2023)[83]	Data: Veterans Affairs and Medicare claims, United States Population: Veterans ≥ 65 years old who were hospitalized for non-cardiovascular reasons and had elevated blood pressure readings Exposure: Intensification of antihypertensive regimen during hospitalization Outcome: Inpatient mortality, ICU transfer, stroke, acute kidney injury, BNP elevation, and troponin elevation Frailty measure: Orkaby index (Veterans Affairs Frailty Index)	Adverse inpatient events  Non-frail  Frail	Intensive (Per 100) NR NR	Non-intensive (Per 100) NR NR	OR (95% CI) 1.26 (1.13, 1.41) 1.29 (1.15, 1.45)
Orkaby (2023)[84]	Data: Veterans Affairs and Medicare claims, United States Population: Veterans ≥ 65 years old, no prior CVD, no prior statin use Exposure: statins Outcome: Death, MACE Frailty measure: Orkaby index (Veterans Affairs Frailty Index)	Death  Non-frail  Frail MACE  Non-frail  Frail	Statin (Per 1000 PY) 48.6 90.4 (Per 1000 PY) 51.7 88.2	No statin (Per 1000 PY) 72.6 130.4 (Per 1000 PY) 60.8 102.0	HR (95% CI) 0.61 (0.61, 0.62) 0.63 (0.62, 0.64) HR (95% CI) 0.87 (0.86, 0.88) 0.90 (0.88, 0.92)
Sheppard (2023)[85]	Data: The Clinical Practice Research Datalink, England Population: Patients ≥ 40 years old, hypertension, no prior antihypertensive treatment Exposure: Antihypertensive treatment Outcome: Serious falls (fall-related hospitalization or death) in 10 years Frailty measure: Clegg index (electronic Frailty Index)	Serious falls  Fit  Mild frailty  Moderate frailty  Severe frailty	Anti-hypertensive (per 100) 2.1 6.6 12.2 16.6	No anti-hypertensive (per 100) 1.2 5.3 8.8 11.5	HR (95% CI) 1.22 (1.18, 1.25) 1.10 (1.06, 1.14) 1.16 (1.07, 1.26) 1.31 (1.09, 1.58)
Diabetes Medications
Dave (2019)[86]	Data: Commercial insurance claims, United States Population: Adults ≥ 18 years old, type 2 diabetes, no prior use of the study drugs Exposure: SGLT2i, GLP1RA, DPP4i Outcome: Severe UTI Frailty measure: Kim index	Severe UTI  Robust  Pre-frail  Frail Severe UTI  Robust  Pre-frail  Frail	SGLT2i (Per 1000 PY) NR NR NR SGLT2i (Per 1000 PY) NR NR NR	DPP4i (Per 1000 PY) NR NR NR GLP1RA (Per 1000 PY) NR NR NR	HR (95% CI) 1.00 (0.46, 2.15) 0.60 (0.33, 1.09) 0.84 (0.49, 1.43) HR (95% CI) 1.12 (0.58, 2.16) 0.59 (0.31, 1.10) 0.64 (0.40, 1.02)
Zhuo (2021)[87]	Data: Medicare claims, United States Population: Beneficiaries ≥ 65 years old, type 2 diabetes, no prior use of the study drugs Exposure: SGLT2i, GLP1RA, DPP4i Outcome: Fractures Frailty measure: Kim index	Fractures  Robust  Pre-frail  Frail Fractures  Robust  Pre-frail  Frail	SGLT2i (Per 1000 PY) 2.9 4.5 10.0 SGLT2i (Per 1000 PY) 2.9 4.5 10.0	DPP4i (Per 1000 PY) 2.2 5.9 10.7 GLP1RA (Per 1000 PY) 2.2 4.3 9.2	HR (95% CI) 1.28 (0.73, 2.25) 0.77 (0.59, 1.01) 0.93 (0.62, 1.41) HR (95% CI) 1.33 (0.74, 2.39) 1.02 (0.75, 1.38) 1.09 (0.69, 1.70)
Htoo (2023)[88]	Data: Medicare claims, United States Population: Beneficiaries ≥ 65 years old, type 2 diabetes, no prior use of the study drugs Exposure: SGLT2i, GLP1RA, DPP4i Outcome: Severe hypoglycemia Frailty measure: Kim index	Severe hypoglycemia  Robust  Pre-frail  Frail Severe hypoglycemia  Robust  Pre-frail  Frail	SGLT2i (Per 1000 PY) 4.1 9.8 28.8 SGLT2i (Per 1000 PY) 5.3 10.6 28.6	DPP4i (Per 1000 PY) 5.2 13.5 34.6 GLP1RA (Per 1000 PY) 5.1 12.1 32.8	HR (95% CI) 0.78 (0.59, 1.03) 0.73 (0.64, 0.83) 0.83 (0.69, 1.00) HR (95% CI) 1.03 (0.79, 1.35) 0.88 (0.77, 0.99) 0.87 (0.74, 1.04)
Kutz (2023)[89]	Data: Medicare claims, United States Population: Beneficiaries ≥ 65 years old, type 2 diabetes, no prior use of the study drugs Exposure: SGLT2i, GLP1RA, DPP4i Outcome: Death, MACE, diabetes drug-related safety events (lower-limb amputation, fracture, hypoglycemia, acute kidney injury, diabetic ketoacidosis, severe genital infection, severe UTI, acute pancreatitis, or non-malignant biliary event) Frailty measure: Kim index	Death or MACE  Robust  Pre-frail  Frail Composite safety events  Robust  Pre-frail  Frail Death or MACE  Robust  Pre-frail  Frail Composite safety events  Robust  Pre-frail  Frail	SGLT2i (Per 1000 PY) 18.7 42.5 109.9 (Per 1000 PY) 21.1 45.0 109.0 GLP1RA (Per 1000 PY) 20.5 49.4 130.3 (Per 1000 PY) 26.2 58.5 141.9	DPP4i (Per 1000 PY) 25.5 59.1 137.1 (Per 1000 PY) 26.4 55.3 120.6 DPP4i (Per 1000 PY) 27.5 67.9 156.2 (Per 1000 PY) 29.4 65.7 138.4	HR (95% CI) 0.74 (0.68, 0.81) 0.72 (0.69, 0.76) 0.79 (0.70, 0.89) HR (95% CI) 0.90 (0.74, 0.87) 0.91 (0.77, 0.86) 0.89 (0.78, 1.01) HR (95% CI) 0.75 (0.68, 0.82) 0.73 (0.70, 0.77) 0.83 (0.76, 0.91) HR (95% CI) 0.89 (0.82, 0.97) 0.89 (0.85, 0.93) 1.01 (0.92, 1.10)
Anti-Neoplastic Agents
DuMontier (2023)[90]	Data: Veterans Affairs and Medicare claims, United States Population: Veterans with new multiple myeloma diagnosis who did not receive bone marrow transplant Exposure: triplet therapy (bortezomib-lenalidomide-dexamethasone), doublet therapy (lenalidomide-dexamethasone) Outcome: Overall survival Frailty measure: Orkaby index (Veterans Affairs Frailty Index)	Overall survival  Non-frail  Mildly frail  Moderate-to-severely frail	Triplet (Per 100 PY) NR NR NR	Doublet (Per 100 PY) NR NR NR	HR (95% CI) 0.86 (0.67, 1.10) 0.80 (0.61, 1.05) 0.74 (0.56, 0.97)
Deol (2023)[91]	Data: Veterans Affairs and Medicare claims, United States Population: Veterans with metastatic castration-resistant prostate cancer Exposure: enzalutamide, abiraterone Outcome: Overall survival Frailty measure: Orkaby index (Veterans Affairs Frailty Index)	Overall survival  Non-frail  Frail	Enzalutamide (median) 27.7 months 20.6 months	Abiraterone (median) 26.1 months 17.1 months	HR (95% CI) 0.93 (0.86, 1.01) 0.85 (0.77, 0.93)
Antipsychotic Medications
Maxwell (2018)[92]	Data: Administrative health and clinical database of Ontario, Canada Population: Home care clients ≥ 65 years with dementia or significant cognitive impairment Exposure: antipsychotics Outcome: Death Frailty measure: 72-item frailty index from home care clinical database	Death  Robust  Frail	Anti-psychotics (Per 100 PM) 2.0 3.8	No anti-psychotics (Per 100 PM) 0.8 2.6	HR (95% CI) At 1 month 3.72 (2.45–5.66) 1.74 (1.40–2.17) At 6 months 1.83 (1.18–2.83) 1.02 (0.75–1.38)
Vaccines
Harris (2023)[93]	Data: Medicare claims linked to customer data from national pharmacy companies, United States Population: Beneficiaries ≥ 66 years old in the community, no COVID-19 in the past 4 weeks Exposure: mRNA COVID-19 vaccine mRNA-1273, BNT162b2 Outcome: Vaccine-related adverse clinical events and COVID-19 diagnosis Frailty measure: Kim index	Facial nerve palsy  Robust  Pre-frail  Frail Thrombocytopenic purpura  Robust  Pre-frail  Frail Pulmonary embolism  Robust  Pre-frail  Frail Myocarditis/pericarditis  Robust  Pre-frail  Frail COVID-19 diagnosis  Robust  Pre-frail  Frail	mRNA-1273 (Per 100000) NR NR NR (Per 100000) NR NR NR (Per 100000) NR NR NR (Per 100000) NR NR NR (Per 100000) NR NR NR	BNT162b2 (Per 100000) NR NR NR (Per 100000) NR NR NR (Per 100000) NR NR NR (Per 100000) NR NR NR (Per 100000) NR NR NR	RR (95% CI) 0.86 (0.75, 0.99) 1.04 (0.93, 1.16) 1.14 (0.89, 1.45) RR (95% CI) 0.89 (0.80, 0.99) 1.03 (0.94, 1.14) 0.96 (0.77, 1.20) RR (95% CI) 0.94 (0.88, 1.00) 0.97 (0.93, 1.01) 1.00 (0.92, 1.08) RR (95% CI) 0.76 (0.56, 1.03) 0.94 (0.77, 1.16) 0.96 (0.58, 1.61) RR (95% CI) 0.85 (0.82, 0.88) 0.83 (0.81, 0.86) 0.94 (0.89, 0.99)

Abbreviations: AF, atrial fibrillation; BNP, B-type natriuretic peptide; CI, confidence interval; CVD, cardiovascular disease; d, days; DPP4i, dipeptidyl peptidase-4 inhibitor; GLP1RA, glucagon-like peptide-1 receptor agonist; HR, hazard ratio; ICU, intensive care unit; MACE, major adverse cardiovascular event; NR, not reported; OR, odds ratio; PM, person-months; PY, person-years; RD, risk difference; RR, risk ratio; SGLT2i, sodium-glucose cotransporter-2 inhibitor; UTI, urinary tract infection; Yr, year.

^aSee Table 2 for the cutpoint to define frailty levels.

5.1. Cardiovascular medications:

Kim et al. examined the rate of a composite clinical events including death, stroke, or major bleeding in Medicare beneficiaries with atrial fibrillation who initiated one of the four oral anticoagulants.[81] Frailty levels were associated with greater incidence of the composite clinical events in all treatment groups. Dabigatran and rivaroxaban were associated with lower rates of clinical events compared with warfarin only among non-frail beneficiaries, and the rates were similar among pre-frail and frail beneficiaries. Apixaban was associated with lower rates of clinical events compared with warfarin in all levels of frailty. Lin et al. compared the risk of losing 14 or more days due to hospitalizations, skilled nursing facility stays, or death in one year after initiating apixaban, rivaroxaban, or warfarin for atrial fibrillation.[82] Compared with the initiators of apixaban, the risk was greater for those initiating rivaroxaban and those initiating warfarin, with greater risk differences as the frailty level increased. These two studies suggest a particularly strong clinical benefit of apixaban in comparison to other direct oral anticoagulants or warfarin in patients with atrial fibrillation and frailty. Anderson et al. evaluated the effect of intensification of antihypertensive regimen in older veterans who were hospitalized for non-cardiovascular reasons and had elevated blood pressure readings within 48 hours of the admission.[83] Patients whose antihypertensive regimen was intensified experienced a higher risk of inpatient adverse clinical events that those whose antihypertensive regimen was not intensified. The increased risk was similar between frail and non-frail patients. Orkaby et al. examined the effect of initiating a statin on preventing major adverse cardiovascular events or death among older veterans without prior history of cardiovascular disease.[84] Statin initiation was associated with a similar relative reduction in the rates of death and major adverse cardiovascular events compared with no statin use between frail and non-frail patients. Because of the higher event rate in frail patients, the absolute rate reduction associated with statin was greater in the frail group than in the non-frail group. Lastly, Sheppard et al. found that the rate of fall-related hospitalization and death was elevated among patients with hypertension who initiated an antihypertensive agent compared with those who did not.[85] The excess risk appeared greater in those with severe frailty.

5.2. Diabetes medications:

Four studies examined the rates of severe urinary tract infection,[86] fractures,[87] severe hypoglycemia,[88] major adverse cardiovascular events,[89] and death[89] among beneficiaries of Medicare and commercial insurance with type 2 diabetes who initiated sodium-glucose cotransporter-2 inhibitors, glucagon-like peptide-1 receptor agonists, or dipeptidyl peptidase-4 inhibitors. Because of the low event rates of severe urinary tract infection, fractures, and severe hypoglycemia, the hazard ratio estimates comparing these drugs had wide 95% confidence intervals including the null association (i.e., no difference between the drugs) and there was no clear evidence for the heterogeneity of treatment effect by frailty levels.[86–88] Kutz et al. reported a similar reduction of death or major adverse cardiovascular events for the initiators of sodium-glucose cotransporter-2 inhibitors compared with those initiating dipeptidyl peptidase-4 inhibitors without an increase in the composite safety events across the frailty levels.[89] The results comparing glucagon-like peptide-1 receptor agonists and dipeptidyl peptidase-4 inhibitors were consistent with those from the analysis comparing sodium-glucose cotransporter-2 inhibitors and dipeptidyl peptidase-4 inhibitors. Collectively, these results suggest the effectiveness and safety of sodium-glucose cotransporter-2 inhibitors and glucagon-like peptide-1 receptor agonists compared with dipeptidyl peptidase-4 inhibitors, regardless of frailty levels. Importantly, the rates of both effectiveness and safety outcomes were more than double among frailty patients relative to either robust or pre-frail patients, regardless of treatment, emphasizing the importance of a separate investigation of drug effects in the frail.

5.3. Anti-neoplastic agents:

DuMontier et al. compared a more intensive regimen of three drugs (bortezomib-lenalidomide-dexamethasone) with a less intensive regimen of two drugs (lenalidomide-dexamethasone) in older veterans who were newly diagnosed with multiple myeloma and were not eligible to receive bone marrow transplant.[90] Interestingly, veterans with moderate-to-severe frailty at the time of diagnosis had a slightly larger reduction in mortality from the triple drug regimen compared with those with no or mild frailty. The investigators speculated that certain clinical manifestation of frailty might be driven by multiple myeloma, which could explain the greater benefit of the triple drug regimen in those with moderate-to-severe frailty. Deol et al. compared the two androgen signaling inhibitors, enzalutamide and abiraterone, in older veterans with metastatic castration-resistant prostate cancer.[91] Enzalutamide was associated with slightly lower mortality compared with abiraterone among frail patients relative to the effect among non-frail patients. The researchers attributed this possibly greater survival benefit of enzalutamide to its lower cardiovascular risk profile, fewer drug-drug interactions, and longer half-life relative to abiraterone. These two studies indicate that older adults with frailty might gain greater survival benefit from a certain antineoplastic regimen than those without frailty.

5.4. Antipsychotic medications:

Maxwell et al. analyzed administrative health and clinical databases from Ontario, Canada, to examine the rate of mortality among home care patients with dementia or significant cognitive impairment who started antipsychotic medications and those who did not.[92] The researchers found that the relative increase in mortality associated with antipsychotic use compared with no use was greater in robust patients than in frail patients.

5.5. Vaccines:

Harris et al. analyzed Medicare claims and customer data from two national pharmacy companies in the United States to determine the comparative effectiveness and safety of two mRNA vaccines, mRNA-1273 and BNT162b2, for Coronavirus Disease 2019 (COVID-19).[93] Among potential adverse events linked to COVID-19 mRNA vaccines, compared with BNT162b2 vaccine, mRNA-1273 vaccine was associated with lower rates of facial nerve palsy and thrombocytopenic purpura and possibly lower rates of myocarditis and pericarditis in robust older adults; the rates were similar between the two vaccines in pre-frail or frail older adults, although confidence intervals were wide among the frail and overlapped those of robust individuals. The rate of COVID-19 was lower in the recipients of mRNA-1273 vaccine than those of BNT162b2 vaccine across all frailty levels, although the hazard ratio was attenuated toward the null in those with frailty.

The illustrative examples discussed above should be interpreted while bearing in mind the limitations inherent in pharmacoepidemiologic studies. Importantly, potential heterogeneity of treatment effects by patients’ frailty status might be due to the differential extent of unmeasured confounding within each level of frailty. As demonstrated from Zhang et al.,[74, 75] a database-derived frailty index can be helpful, yet insufficient for removing confounding. As confounding tends to be greater when frailty subgroup analysis is performed without consideration of the matched pairs in the propensity score-matched overall population, it is recommended to perform propensity score matching within each of frailty subgroup strata.[94, 95] In addition to applying other study design and analytic strategies to reduce confounding,[76–80] sensitivity analyses to assess the robustness of the results to various hypothetical scenarios of unmeasured confounding[96, 97] and calculation of E-value[98] can be useful. As discussed in the previous section, database-derived frailty indices were developed using different clinical frailty measures and have varying degrees of misclassification. Understanding how a database-derived frailty index is related to clinical frailty measures is a prerequisite for clinical translation of frailty subgroup-specific treatment effects. Ideally, the findings from pharmacoepidemiologic studies should be confirmed through an adequately powered randomized controlled trial before being applied in clinical settings. In the absence of such confirmatory trials and given under-representation of older adults with frailty in clinical trials, well-conducted pharmacoepidemiologic studies, when considered alongside existing clinical knowledge, can offer a valuable guidance for pharmacotherapy in older adults.

6. EVIDENCE GAPS AND FUTURE DIRECTIONS

The United States Food and Drug Administration (FDA)[99] and the European Medicines Agency (EMA)[100] have both recognized the importance of enrolling older adults who are representative of the target population that is to be treated with the medications in clinical practice. The FDA issued guidance for enhancing the diversity of clinical trial populations,[101] including the inclusion of older adults in cancer clinical trials.[102] The EMA recommended characterizing physical frailty at trial baseline.[100] Moreover, the International Union of Basic and Clinical Pharmacology Geriatric Committee[17] and experts from Europe[16] have recently published their position statements for the integration of frailty in all phases of drug development and evaluation for older adults. They called for increasing the enrollment of older adults with frailty and functional impairment to enhance external validity of findings, incorporating validated frailty assessment tools as both baseline measures and outcomes, collecting pharmacokinetic and pharmacodynamic data from preclinical and clinical studies, and assessing patient-reported and patient-centered outcomes that are important to those living with frailty (e.g., functional status).

To generate high-quality real-world evidence from pharmacoepidemiologic studies using routine health care data, further improvement in frailty measurement may be possible by combining information from administrative claims and EHR data, particularly through natural language processing and deep learning of unstructured clinical notes.[103–106] Lack of a reference standard frailty measure from clinical assessment in EHR data remains a major challenge in developing such machine learning and artificial intelligence models. Given that EHR data are typically available for only a subset of the individuals represented in claims datasets, there is a need for research into the methods of applying the detailed clinical information from EHR to the entire claims dataset. Understanding the relationship among different database-derived frailty scores[107] and clinical frailty measures[67] is needed to facilitate translation of research findings into clinical practice.

There is growing interest in developing methodological approaches that integrate data from routine health care and randomized controlled trials to generate evidence on medical interventions applicable to routine care populations, including older adults with frailty. Strom et al. linked Medicare claims to data from two randomized controlled trials comparing transcatheter and surgical aortic valve replacement.[108] Although the trials did not collect frailty at baseline, the investigators were able to calculate the Segal index retrospectively from the linked Medicare claims prior to the trial baseline. They found no statistically significant difference in mortality between patients randomly assigned to surgery and those assigned to transcatheter procedures across the frailty levels (hazard ratio [95% confidence interval] for surgery vs transcatheter procedure: 0.94 [0.61, 1.46] in frailty tertile 1; 1.23 [0.83, 1.84] in frailty tertile 2; 0.77 [0.56, 1.07] in frailty tertile 3). This approach minimizes confounding by taking advantage of randomization. However, it encounters frequent limitations due to the exclusion of patients with severe or advanced frailty, inadequate statistical power, limited access to individual-level trial data, and the challenges of data linkage. For accurate data linkage, the sharing of unique identifiers between trialists and data vendors is essential, but this sharing is not often covered by the trial’s informed consent. The practicality and usefulness of this approach still needs to be evaluated in drug trials. Other researchers proposed methods to generalize treatment effect estimates from a trial to a target population, which may include a higher proportion of people with frailty.[109–112] While these methods cannot estimate treatment effects for individuals who were excluded from a trial, they can be useful for predicting treatment effects in a trial-eligible routine care population using routine health care data.

Another area of interest is to develop a prospective drug monitoring system for an early detection of the effectiveness and safety signals for older adults living with frailty. The FDA Sentinel Initiative is proactively monitoring the safety of medical products, including medications and vaccines, by facilitating the analysis of insurance claims and EHR from a large network of health care systems and data partners.[113] When a new drug is approved and introduced to the market, little information is available about the effectiveness and safety of the drug in routine care populations. As the drug is used by more individuals and more clinical outcome data are collected in routine health care data, a propensity score-matched cohort study can be conducted to compare the effectiveness and safety outcomes between the new drug and a standard therapy; this analysis is repeated periodically (e.g., every year) using the most up-to-date data until precise estimates of the signals are obtained.[114, 115] Case studies involving a cyclooxygenase-2 selective non-steroidal anti-inflammatory drug,[116] antiplatelet agents,[117] osteoporosis agents,[116] statins,[118] anticoagulants[119] have demonstrated the potential for early detection of safety signals. This framework, by incorporating a database-derived frailty index, allows for the timely assessment of a new drug’s effectiveness and safety across a spectrum of frailty levels in older adults. A recent study illustrates the feasibility of frailty-specific monitoring of direct oral anticoagulants compared with vitamin K antagonist warfarin in Medicare beneficiaries with atrial fibrillation (unpublished). The effectiveness and safety of direct oral anticoagulants were established within two years following their approval, in both frail and non-frail Medicare beneficiaries. Such evidence could encourage the uptake of these effective and safe new medications among older adults.

Lastly, most studies to date have applied database-derived frailty indices to reduce confounding and evaluate the heterogeneity of treatment effects by frailty levels. More research is needed to investigate frailty incidence or progression as an outcome of drug therapy.[16, 17]

7. CONCLUSIONS

The development of database-derived frailty indices has enabled researchers to identify older adults with frailty who are at increased risk for drug-related adverse events and to evaluate the heterogeneity of treatment effects by patients’ frailty levels using routine health care data. While a greater degree of frailty is associated with an elevated risk of many adverse treatment outcomes as well as a greater risk for effectiveness outcomes, this does not imply that older adults with frailty derive no benefit from pharmacotherapy; they may, in fact, benefit as much as or more than those without frailty. Therefore, we recommend that researchers routinely perform frailty subgroup analyses in pharmacoepidemiologic studies and post-hoc analyses of clinical trials testing a drug therapy. These analyses should be interpreted within the inherent limitations of the data sources and study design. In contexts where evidence from randomized controlled trials is scarce, these results may be informative to tailor pharmacotherapy for older adults across different levels of frailty.

KEY POINTS.

• The development of frailty scores for administrative claims and electronic health records data has enabled researchers to identify older adults with frailty who are at increased risk for drug-related adverse events and to evaluate the heterogeneity of treatment effects by patients’ frailty levels.

• While a greater degree of frailty is associated with an elevated risk of adverse treatment outcomes, it does not imply that older adults with frailty derive no benefit from pharmacotherapy; the benefit they derive from specific pharmacotherapies can vary, sometimes being lesser, similar to, or greater than that experienced by non-frail patients.

• Researchers should routinely perform frailty subgroup analyses in pharmacoepidemiologic studies and post-hoc analyses of clinical trials testing a drug therapy and interpret the results within the inherent limitations of the data sources and study design.

Availability of data and material:

Data sharing is not applicable to this article as no datasets were generated or analyzed during the current study.