Background
As the world’s population rapidly ages, marked by striking increases in the proportion of older people and in their life expectancy (World Health Organization, 2011), more persons will experience frailty than ever before. A recent systematic review found that 10.7 per cent of community-dwelling persons aged 65 years and older experience frailty, although the reported prevalence varied greatly across studies (4.0−59.1%) (Collard, Boter, Schoevers, & Oude Voshaar, 2012). Although there is no single agreed-upon definition of frailty, most describe a syndrome characterized by decreased physiological reserve and reduced ability to respond to stressors (such as acute illness), and thus subject to an increased risk of adverse health outcomes, including death (Fried et al., Reference Fried, Tangen, Walston, Newman, Hirsch and Gottdiener2001). Although some researchers characterize frailty in terms of physical attributes (Fried et al., Reference Fried, Tangen, Walston, Newman, Hirsch and Gottdiener2001), others incorporate cognitive, psychological, and social elements in frailty definitions and assessments (Gobbens, Luijkx, Wijnen-Sponselee, & Schols, Reference Gobbens, Luijkx, Wijnen-Sponselee and Schols2010; Rockwood et al., Reference Rockwood, Stadnyk, MacKnight, McDowell, Hebert and Hogan1999). Despite being more common in older persons and those with multi-morbidity, frailty can occur independent of advanced age or specific conditions and disabilities (Fried et al., Reference Fried, Tangen, Walston, Newman, Hirsch and Gottdiener2001).
The identification of frailty is a critical step towards improving the care of older persons. At the clinical and patient levels, identification can lead to increased use of interventions (e.g., exercise, reduction of polypharmacy) that reduce the risk of adverse outcomes and optimize quality of life (Sieliwonczyk, Perkisas, & Vandewoude, Reference Sieliwonczyk, Perkisas and Vandewoude2014). There are numerous practice-level tools to screen for frailty in clinical practice (Rockwood et al., Reference Rockwood, Song, MacKnight, Bergman, Hogan, McDowell and Mitnitski2005; Rolfson, Majumdar, Tsuyuki, Tahir, & Rockwood, Reference Rolfson, Majumdar, Tsuyuki, Tahir and Rockwood2006), although the effectiveness of screening is unknown. At the policy and population levels, identification of persons who are frail may help elucidate the implications and consequences of frailty (e.g., health service use, costs, patient outcomes), identify potential “gaps” in health service organization and delivery, and design medical and social programs and policies to maximize health and independence as people age.
Several research groups have attempted to identify frail populations using administrative health datasets; these datasets have included a combination of physicians’ claims data with information from surveys, institutional continuing care data, and/or home care provision information (Bronskill, Carter, & Costa, Reference Bronskill, Carter and Costa2010; Davidoff et al., Reference Davidoff, Zuckerman, Pandya, Hendrick, Ke, Hurria and Edelman2013; Rosen et al., Reference Rosen, Wu, Chang, Berlowitz, Rakovski, Ash and Moskowitz2001). Thus, their identification algorithms usually require specific data elements not contained or incomplete within many administrative databases (The John Hopkins University, 2014) or clinical assessment data not routinely available in population-based databases (Bronskill et al., Reference Bronskill, Carter and Costa2010). An example of this, in Canada and internationally, is databases containing InterRAI assessment data that may be used to assess “health instability” via the Changes in Health, End-stage disease, Signs, and Symptoms (CHESS) scale (Hirdes, Frijters, & Teare, Reference Hirdes, Frijters and Teare2003), often considered a concept analogous to frailty (Armstrong, Stolee, Hirdes, & Poss, Reference Armstrong, Stolee, Hirdes and Poss2010). Although some research groups (Armstrong et al., Reference Armstrong, Stolee, Hirdes and Poss2010; Campitelli et al., Reference Campitelli, Bronskill, Hogan, Diong, Amuah, Gill and Maxwell2016; Hubbard et al., Reference Hubbard, Peel, Samanta, Gray, Fries, Mitnitski and Rockwood2015) have also used InterRAI data from different care settings to measure frailty based on the frailty index (Rockwood & Mitnitski, Reference Rockwood and Mitnitski2007), this information is not available to the same degree across all Canadian provinces and jurisdictions.
To our knowledge, the only algorithm to identify persons who are frail using common population-based administrative health databases is limited to examining causes of death from death certificate data using diagnostic codes (Fassbender, Fainsinger, Carson, & Finegan, Reference Fassbender, Fainsinger, Carson and Finegan2009). Although this algorithm may be valuable, frailty is a complex construct wherein the diagnosis of a specific condition is only one of many factors involved. Additionally, this algorithm includes many relatively benign conditions (e.g., acute infections) not indicative of frailty on their own, although which may indicate frailty if a person died from this condition (Fassbender et al., Reference Fassbender, Fainsinger, Carson and Finegan2009). Consequently, we sought to develop a more refined frailty “identification rule” using population-based health administrative data that could be readily applied across jurisdictions for living and deceased persons. Given that administrative health databases have the unique potential to provide population-based, unbiased, and efficient measures of quality care (Earle et al., Reference Earle, Park, Lai, Weeks, Ayanian and Block2003; Iezzoni, Reference Iezzoni1997), the ability to identify frailty in administrative health databases enables researchers and health system decision-makers to efficiently measure/monitor health care utilization and quality of care for persons with frailty, regardless of where they live or which services they access.
Methods
This study involved two phases: (1) development of rules and (2) identification using population-based datasets available across Canadian provinces. Ethical approval was granted by the Behavioral Research Ethics Board of the University of British Columbia; Conjoint Health Research Ethics Board of the University of Calgary; St. Michael’s Hospital Research Ethics Board; Comité d’éthique de la recherche du Centre de santé et de services sociaux de la Veille-Capitale [Research Ethics Committee of the Health and Social Services Centre – Veille-Capitale]; Comité d’éthique de la recherche du CHU de Québec [Research Ethics Committee of the Centre hospitalier universitaire de Québec]; and the Nova Scotia (NS) Health Authority Research Ethics Board.
Development of Rules
We drafted preliminary identification rules by consulting literature wherein some form of claims-based data was used (Hoover, Rotermann, Sanmartin, & Bernier, Reference Hoover, Rotermann, Sanmartin and Bernier2013; Kim & Schneeweiss, Reference Kim and Schneeweiss2014). We also incorporated definitions of frailty from aging research (Rockwood et al., Reference Rockwood, Song, MacKnight, Bergman, Hogan, McDowell and Mitnitski2005; Rolfson et al., Reference Rolfson, Majumdar, Tsuyuki, Tahir and Rockwood2006) and markers of service utilization by persons who are frail, such as long-term care (LTC) residency and multiple hospitalizations. Pertinent literature was identified via (1) an initial search of PubMed; (2) consultation with experts in the field (see next paragraph); and (3) a scoping review (led by co-author AMCG) to identify current health care services and models, use of health care resources, and outcomes of care relevant to older adults with frailty. Relevant articles from the latter were reviewed among our team to help with rule development. We limited rule development to variables within two population-based datasets available across Canadian provinces: hospital discharge abstracts and physician claims data. We intentionally focused on specificity over sensitivity, knowing that some persons who are frail would not be identified by using administrative data. No specific criteria were employed to balance specificity over sensitivity. Rather, this balance was considered and fine-tuned through iterative team discussions and consultation with experts in geriatric medicine.
We provided the preliminary rules to geriatricians and researchers with expertise in administrative data, gerontology, and/or end of life care (n = 11). These experts included six practicing physicians with expertise in geriatric medicine (with 4/6 also experts in frailty research) and five researchers with expertise in administrative health data (with 2/5 also experts in using administrative data for end-of-life care research). Through an iterative process, experts were asked whether they agreed with the rules, whether anything was missing, and how they could be improved based on their individual clinical and/or technical knowledge. Following this consultation, we made adjustments to improve specificity and sensitivity, including the use of diagnoses codes corresponding to conditions and functions described in the Clinical Frailty Scale (Rockwood et al., Reference Rockwood, Song, MacKnight, Bergman, Hogan, McDowell and Mitnitski2005) and Edmonton Frail Scale (Rolfson et al., Reference Rolfson, Majumdar, Tsuyuki, Tahir and Rockwood2006); adding a rule for persons receiving palliative care; and limiting the criterion related to number of falls to only falls leading to a hospital admission.
Identification in Population-Based Datasets
The final identification rules were used to identify persons potentially with frailty aged 65 years and older in five participating provinces (British Columbia, Alberta, Ontario, Quebec, Nova Scotia) using two administrative health datasets: the Canadian Institutes of Health Information Discharge Abstract Database (DAD) or Med-Echo in Quebec, and each province’s health insurance database (i.e., physician claims). DAD (Med-Echo in Quebec) includes data related to all hospital discharges in each province (from acute, chronic, and rehabilitation facilities). These databases were linked at the patient level using encrypted health care numbers.
With the exception of Nova Scotia, we identified persons with frailty from two populations: (1) those who had died and were 66 years or older at the time of death (decedents) and (2) living persons 65 years or older (living persons). In Alberta, Ontario, and Quebec, the decedents included all persons with a recorded date of death within fiscal year (FY) 2013−2014. In British Columbia, the decedents included all persons with a recorded date of death between FY 2009−2010 and 2013−2014. In Nova Scotia, the decedents included all persons who died from cancer between FY 2004−2005 and 2008−2009 (an existing linked administrative dataset). In British Columbia, Alberta, Ontario, and Quebec, the living persons were alive throughout FY 2013−2014, which represented the most recent year of administrative health data available at the time of study. The differences across provinces were due to data availability in each province. All records (encounters) in both databases were examined for the time periods of interest. We calculated descriptive statistics for persons identified using each rule, by province. Cohort identification and analyses were carried out separately in each province using a consistent approach.
Results
We developed three identification rules to be used with administrative health data. These are shown in Table 1. In brief, persons with claims data meeting at least one of the following three rules were considered frail: (1) person was a LTC resident; (2) person received palliative care; or (3) person was categorized as meeting at least two of seven domains, which were based on frailty scales, geriatrician discussions, and health service utilization indicators.
Identification rules to identify persons with frailty using administrative health data
a Contact the corresponding author for the associated formats and International Classification of Diseases (ICD) 9 and ICD 10 codes.
Tables 2 and 3 depict the number and proportion of persons identified as frail, as well as frailty identification by specific rule(s), for both decedents and living persons, respectively. The proportion of persons identified as frail was much higher among decedents than living persons. This was consistent across all provinces where data were available for both deceased and living persons. The proportion of persons identified as frail was highest in Alberta (78.1% and 14.7% for decedents and living persons respectively). Otherwise, the proportion of persons identified as frail was similar across provinces, ranging from 58.2 to 68.0 per cent (decedents) and 5.1 to 7.4 per cent (living persons).
Frailty identification by specific rule(s) for decedents
a Persons with frailty identified by each individual identification rule are not mutually exclusive.
b Nova Scotia’s dataset included all seniors (aged 66+ years) who had died of cancer between 2004−2009.
Frailty identification by specific rule(s) for living persons
a Persons with frailty identified by each individual identification rule are not mutually exclusive.
b Nova Scotia data among living frail persons were not available.
c Numbers suppressed due to small cell counts.
Discussion
This study developed identification rules using population-based administrative health databases to identify persons who are frail and applied these rules to decedents and living persons across multiple Canadian provinces. Administrative health databases have the unique potential to provide population-based, unbiased, efficient measures of health care utilization and quality care (Earle et al., Reference Earle, Park, Lai, Weeks, Ayanian and Block2003; Iezzoni, Reference Iezzoni1997). Such data represent a powerful tool towards understanding the impacts of frailty for all persons and not simply those enrolled in specific programs such as a specialized geriatric medicine service.
Although algorithms exist to identify frail populations, to date, none have exclusively used population-based administrative health datasets with information before death. We utilized two databases widely used across Canada, with comparable data structures and fields: hospital discharges and physician claims. Given geriatrician feedback, we included diagnostic codes associated with two clinical frailty scales (Rockwood et al., Reference Rockwood, Song, MacKnight, Bergman, Hogan, McDowell and Mitnitski2005; Rolfson et al., Reference Rolfson, Majumdar, Tsuyuki, Tahir and Rockwood2006) and a limited number of suggested events (e.g., falls) rather than the diagnostic codes used by Fassbender et al. (Reference Fassbender, Fainsinger, Carson and Finegan2009). Feedback from geriatricians, who are experts in frailty care, indicates that our rules have face validity. Further, we found that 5.1 to 14.7 per cent of living persons are frail, depending on province. This falls within the range reported in the literature, particularly studies using a physical phenotype frailty definition where the prevalence of frail community-dwelling persons age 65 years and older ranged from 4.0 to 17.0 per cent (Collard et al., Reference Collard, Boter, Schoevers and Oude Voshaar2012). Guided by the cumulative deficits model of frailty (Mitnitski, Mogilner, & Rockwood, Reference Mitnitski, Mogilner and Rockwood2001), Clegg et al. recently developed and validated a frailty index using primary care electronic medical record (EMR) data (Clegg et al., Reference Clegg, Bates, Young, Ryan, Nichols, Ann Teale and Marshall2016). In two separate validation cohorts, they estimated the prevalence of moderate and severe frailty to be 12 to 16 per cent and 3 to 4 per cent respectively. Our population-based findings also correspond to these EMR-based estimates from primary care.
The first decision rule in our set of rules identifies persons with frailty based on whether they are an LTC resident. A recent systematic review that identified the prevalence of frailty in nursing homes ranged widely from 19.0 to 75.6 per cent (Kojima, Reference Kojima2015). Thus, this decision rule appears contradictory with the review’s findings. Nevertheless, the geriatricians we consulted throughout decision rule development (practicing in four Canadian provinces) felt that the vast majority of LTC residents have a frail health state, and the proportion of LTC residents who are fit would be negligible in the context of population-based identification. It is possible the countries included in the systematic review (Brazil, Spain, Taiwan, Lebanon, Egypt, the Netherlands, and Poland) use LTC differently (i.e., LTC plays a different role in health care or for different cultural reasons) and/or have substantively different admission criteria than in Canada – as a result, the characteristics of those LTC populations may be different from what we typically see in Canada. Two additional issues are notable. One, the two studies in the review that used frailty definitions upon which we largely based our rules (Edmonton Frail Scale, Clinical Frailty Scale) had the highest prevalence of frailty at 74.1 and 75.6 per cent. Two, the pooled estimates of the prevalence of frailty and prefrailty across the nine studies were 52.3 per cent and 40.2 per cent respectively. Thus, it may be that this particular rule (LTC resident) is an indicator of either a frail or a prefrail state.
The second identification rule (receipt of palliative care) was based on the Clinical Frailty Scale, a well-validated instrument for frailty measurement. Specifically, on this 1−9 scale, 8 (Very Severely Frail) and 9 (Terminally Ill) refer to persons approaching the end of life or with a life expectancy of less than six months, who are not otherwise frail. The third rule required an indication of at least two of seven domains based on frailty scales, geriatrician discussions, and health service utilization indicators (see Table 1). Most of these domains were based directly on the Edmonton Frail Scale, another valid measure of frailty. The added domains resulted from consultation with the expert group and involved either a geriatrician service claim or a physician home visit. Although these latter two, on their own, might not indicate frailty, our rule was such that two separate domains had to be met for a person to be identified as frail.
This study has several limitations. First, it was beyond the scope of this study to validate the identification rules; this is a logical next step in the development process. Second, we designed the rules to optimize specificity over sensitivity, but because we did not validate the rules, we cannot estimate the proportion of persons with frailty who were missed/not identified. This focus also means that the rules risked increasing the false positives rate. Third, our approach did not reflect the continuum of frailty as a health state ranging from fitness to frailty. Fourth, administrative health data do not contain all of the attributes that define frailty (e.g., social/living circumstances) or pertinent clinical information (e.g., results of functional assessments). Nonetheless, our findings concur with others’ studies, where frailty was defined via frailty indices using clinical or self-reported data (Clegg et al., Reference Clegg, Bates, Young, Ryan, Nichols, Ann Teale and Marshall2016; Collard et al., Reference Collard, Boter, Schoevers and Oude Voshaar2012).
The application of our identification rules using administrative databases permits assessment of health care utilization, quality of care, and outcomes for persons with frailty regardless of where they live, which programs they are enrolled in, and whether they are alive or deceased. The data fields on which these rules are based are widely available in administrative health databases outside of Canada, making the rules broadly applicable across jurisdictions. Despite limitations, these rules represent an important step forward towards identifying frailty in administrative data; health ministers from Organisation for Economic Co-operation and Development (OECD) countries recently emphasized the need to make better use of such data to measure health system performance to address the needs of our aging populations (OECD, 2017).
Here, we provide a set of rules to identify persons with frailty from common administrative health databases. We encourage other research groups to validate these rules in existing cohorts and/or apply these rules in their administrative datasets to improve them for future application.
