We took three approaches. First, we searched the electronic database MEDLINE (1948 to May 2011) through the OvidSP interface for all articles using the keyword “frailty” (using the term “frail” yielded an unmanageably large literature with little relevance to the present aims). This strategy allowed us to identify articles where this keyword appeared at least once in the title, abstract, or subject heading. Second, the reference sections of the retrieved articles were scrutinized for additional relevant papers by manual searches. Third, we searched our own records which included interrogation of our own relational databases (e.g. Reference Manager, Endnote). This overview followed the guidelines for the Meta-analysis of Observational Studies in Epidemiology (MOOSE) 26 .

We included studies with participants aged 50 years and older at baseline examination in which the authors purport to have measured frailty. Further inclusion criteria were: 1) articles written in English, French, or Spanish; and 2) articles describing the reliability and validity of a frailty instrument.

The reliability and validity were assessed using suggested guidelines 27 28 . Reliability, which determines if a scale measures an entity (here frailty) in a reproducible way, was investigated through the following definitions: internal consistency (the average of the correlations among all items in the measure), intra-rater reliability (the agreement between observations made by the same rater on two different occasions), inter-rater reliability (the agreement between different raters), and test-retest reliability (the agreement between observations on the participants on two occasions separated by an interval of time). Validity – whether the scale in question measures what it purports to – was assessed by criterion and construct validity. Criterion validity refers to how well the instrument predicts an outcome. When frailty and the outcome data are collected simultaneously, the criterion validity is referred to as the concurrent validity. When the outcome data are prospectively collected, it is called predictive validity. Finally, in this context, construct validity refers to the extent to which a frailty measure correlates with factors that are, based on the extant literature, known to have an association (e.g. age, comorbidity, disability, physical capabilities or performances) 27 28 .

To evaluate the level of utilization of a given frailty instrument by researchers, we counted, among the selected articles, the number of publications which had been authored by researchers other than the originators in the periods ≤ 2000, 2001-2005, and ≥2006. In addition to this, we used the Scopus citation database 29 of peer-reviewed literature to analyze the number of citations in original research articles, excluding those cited by the creators of a given frailty instrument, for each frailty scales up to October 2011. In order to have an indication about the level of predictive validity of the identified frailty instruments, estimates – hazard ratios (or relative risks) and odds ratios – for the association between a frailty score and an adverse health outcome, in particular mortality, were examined.

All 27 identified frailty measurements were grouped into three categories (Additional file 1: Table S1): subjective (self-reported items only), objective (inclusion of only directly measured components), or subjective and objective combined (mixed). Eleven of the 27 instruments included only subjective components which were either reported by a participant (self-evaluation) in nine out of 11 cases 30 31 32 33 34 36 38 39 40 , or reported by a clinician or a researcher (hetero-evaluation) 35 37 . Of the 27 frailty instruments, five included only objective components 41 42 43 44 45 . Finally, the remaining 11 instruments included both subjective and objective (mixed) components 46 47 48 49 50 51 52 53 54 55 56 .

Of the 27 frailty assessments, 19 were developed in population-based samples 30 31 32 34 35 36 37 40 41 42 43 44 46 47 48 50 51 53 55 , 7 among hospitalized patients 33 39 45 49 52 54 56 , and 1 without specification 38 . Half of the frailty scales (n=14) were created by research groups in the USA 30 31 36 39 41 42 43 44 46 47 48 49 53 56 , five in Canada 32 34 37 52 54 , three in the Netherlands 33 40 51 , two in Italy 38 45 , and one each in Australia 55 , France 50 , and Sweden 35 . Five of the 27 frailty instruments were adapted from those developed initially to assess functional status 57 , vulnerability 58 , or physical performances 59 60 61 . These were used to measure frailty for the first time by Cacciatore and colleagues 36 , Kanauchi and colleagues 39 , Brown and colleagues 41 , Gill and colleagues 42 , and Bandinelli and colleagues 44 , respectively. Furthermore, recently tested tools assessing frailty such as Static/Dynamic Frailty Index 51 , Study of Osteoporotic Fractures Index 53 , FRAIL scale 55 , and Comprehensive Assessment of Frailty 56 were based on the Fried’s frailty scale 47 and/or the Mitnitski’s Frailty Index 34 .

All identified frailty measures were composed of at least two items, except that of Gerdhem and colleagues 35 where a general assessment of health is made within a 15-second observation by the investigator. Of the subjective and mixed frailty measures, most contained disability and/or comorbidity components. Instruments without disability or comorbidity information were: the 1994 Frailty Measure 31 , Subjective Frailty Score 35 , Tilburg Frailty Indicator 40 all objective measures (Modified Physical Performance Test 41 , Physical Frailty Score 42 , Klein’s frailty index 43 , Short Physical Performance Battery 44 , and Opasich’s frailty scale 45 ), Speechley & Tinetti’s frailty scale 46 , Fried’s frailty scale 47 , Score-Risk Correspondence for dependency 50 , Study of Osteoporotic Fractures Index 53 , and Brief Frailty Index 54 . Further descriptions of characteristics of population and type of components included in each instrument are also provided in (Additional file 1: Table S1).

Additional file 2: Table S2 presents reliability and validity data taken from the original articles and other related articles on the frailty measurements. Three approaches were used for reliability assessment: internal consistency, inter-rater, and test-retest reliability. Concurrent and predictive validity were mainly assessed using outcomes such as mortality, institutionalization, activities of daily living (ADL) disability, hospitalization, and quality of life. Only 7 out of 27 instruments (26%) were found to have had both reliability and validity ascertained 33 35 37 40 43 49 52 .

Of all, 19 instruments had either their reliability or validity assessed. Among them, 4 instruments were tested for validity only once in the original sample/cohort of participants 32 36 55 56 , and the Phenotype of Frailty by Fried and colleagues 47 and the Frailty Index by Mitnitski and colleagues 34 had their concurrent or predictive validity assessed in more than 3 samples/cohorts (17 and 13 samples/cohorts, respectively). One instrument out of 27, the Short Physical Performance Battery, previously used to assess physical functioning 61 , had neither reliability nor validity information in measuring frailty 44 .

Information on the predictive validity was available for 16 instruments. In 69% (n=11/16), the predictive validity was quantified by relating the frailty measure to mortality. With average follow-ups varying from 1 month to 12 years, hazard ratios or relative risks (from Cox regression) or odds ratios (from logistic regression) for mortality risk for frail people relative to those with no record of the condition ranged from 1.21 (95% confidence interval (CI): 0.78; 1.87) to 6.03 (95% CI: 3.00; 12.08) for the Phenotype of Frailty 47 and 1.57 (95% CI: 1.41; 1.74) to 10.53 (95% CI: 7.06; 15.70) for the Frailty Index 34 . The Phenotype of Frailty has been rarely used in a continuous fashion. One exception is Kulminski et al who found an increased mortality risk of 2% (RR=1.02; 95% CI: 1.02; 1.03) for a one unit of increase in this scale. For the Frailty Index, the estimates ranged from 1.008 (95% CI: 1.005; 1.011) to 10.53 (95% CI: 7.06; 15.70). The estimates – hazard ratios (or relative risks) and odds ratios – examining the association between a frailty score and mortality do not allow to affirm which score is the best in the prediction of mortality for several reasons: 1) relative risks and odds ratios are calculated differently 62 ; 2) estimates were assessed in different populations, therefore with different baseline risks; 3) follow-ups and adjustment for confounding factors were heterogeneous. In spite of these limits, the estimates in Additional file 2: Table S2 give a qualitative appreciation on the magnitude of the association between a frailty score and mortality.

Additional file 3: Table S3 presents the number of publications in which a frailty measure had been used by investigators other than those who created it. In 69% of publications, a frailty scale developed by Fried and colleagues 47 was utilized; 12% used the Frailty Index developed by Mitnitski and colleagues 34 ; 4% the Edmonton Frail Scale 52 ; and ≤ 2% used the remaining instruments. This analysis also shows that half the frailty instruments (n=14) have not been employed at all by other researchers 30 35 36 38 43 44 45 48 49 50 51 54 55 56 . Figure 2 displays the number of original research articles based on the Scopus citation database, which referenced one of the 27 frailty instruments: the 3 most cited papers were that of Fried and colleagues, 2001 47 (n=676), Speechley and colleagues, 1991 46 (n=167), and Gill and colleagues, 2002 42 (n=150). The citation rank for Mitnitski and colleagues’ paper, 2002 34 was ninth (n=52).

The Standards for Educational and Psychological Testing 67 , a guideline which describes the best practice in the development of complex measures such as frailty, recommends the reporting of the basic principles of test construction such as reliability and validity. However, this information was available only for a few instruments: CSHA Clinical Frailty Scale 32 and Edmonton Frail Scale 52 . They had acceptable reliability (Kappa coefficient ≥ 0.7) and good concurrent and predictive validity. Two instruments were widely tested for their validity but not reliability: the Frailty Index 34 and the Fried’s scale 47 . Reliability and validity are the most important indicators when selecting one measure over another. However, even among 7 frailty measurements with such information 33 35 37 40 43 49 52 , none of them appear to be recognized as a “gold standard”. Comparing the performances of different frailty scales in predicting an objective health outcome such as mortality was complicated by the use of different confounding factors across studies.

In several studies, investigators have examined the inter-relationships between different measures of frailty. Thus, the Fried’s scale has been compared with the Frailty Index 10 68 69 and the Study of Osteoporotic Fracture index 15 53 using different methods: correlation analyses 69 , comparison of strength of cross-sectional 68 and prospective associations 10 15 , and use of the c-index statistic 53 . The Fried’s scale is moderately well correlated with the Frailty Index 69 , and shows a stronger association with age and sex (important criteria of construct validity 28 ) 68 but a weaker association with mortality 10 . The Fried’s scale and the Study of Osteoporotic Fracture index have a similar strength of association with falls, disability, hospitalization 15 and death 53 . As Streiner and Norman 27 highlighted, we found that it was sometimes difficult to disentangle whether an assessment belongs to concurrent validity or construct validity. Therefore, certain classifications in either category might be arguable.

We attempted to assess the use of a frailty instrument by counting the number of publications that had adopted the instrument other than the original creators. The two instruments which have had their external validity most extensively evaluated against adverse health outcomes were those developed by Fried group (Phenotype of Frailty) and Mitnitski group (Frailty Index). These are based on two different conceptual frameworks. The Fried group has suggested that frailty represents a phenotype which reflects underlying age-related changes in multiple systems. By contrast, the Mitniski group advances that frailty is the accumulation of multiple deficits, with the degree of frailty denoted by the number of such deficits. This highlights that although some investigators recognize that frailty, comorbidity, and disability are distinct entities 28 47 70 , for others they are overlapping. Most reviews or editorials on frailty have implicitly presented the Phenotype of Frailty as standard 63 71 72 73 74 75 76 77 78 79 80 81 whereas for others the standard is the Frailty Index 82 83 . Recommendations from other researchers are more nuanced. For Sternberg and colleagues 84 , the choice depends on the definition and outcomes that best suit the investigators or clinicians responsible for the screening. The European, Canadian and American Geriatric Advisory Panel 66 recommend using a hybrid measure, the “FRAIL” scale, comprising components from both the Phenotype of Frailty and the Frailty Index.

The Fried’s scale 47 has been the most extensively tested for its validity and is the most widely used instrument in frailty research 65 78 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 . Randomized controlled trials have also used the scale to screen elderly participants 24 25 135 136 137 138 139 140 , or as an outcome of interventions 22 23 139 . The Fried’s scale is widely used, allowing comparisons to be made between studies.

The main limitation of our assessment of use of these instruments is that it penalizes the more recently published frailty instruments. However, the Fried’s scale is not the oldest measure in the field and several more recent frailty instruments are either derived or similar to that measure, suggesting that qualities other than duration of availability explain the popularity of this instrument. Another limitation lies in the lack of elimination of articles that may have resulted from the original authors’ circle of influence. For example, some of the articles which report on the use of the Fried’s scale may have been produced from former co-workers who had previously utilized the CHS data – the dataset in which the Fried’s scale was first validated.

In spite of its wide use, the Fried’s scale has some drawbacks common to other frailty instruments. Chiefly, different scales utilize different classification of the individual components. For example, in the Cardiovascular Health Study (CHS), participants were considered positive for weight loss if they reported having lost more than 10 pounds unintentionally in the last year or they objectively lost 5% or more in comparison with the previous year’s body weight 47 . In Women’s Health Aging Study-I, however, a cut-off of 10% in comparison with the self-reported weight at age 60 years 4 was utilized. These important variations in the operationalization of frailty measurement render comparisons of findings between studies as problematic.

In addition to the manual counting procedure to estimate the use of the frailty instruments, we also examined the number of citations in original research articles (excluding those cited by the creators of a given frailty instrument) for the 27 papers describing the frailty instruments. Even though the rank of citations was different for some of the frailty instruments than that of the manual counting, the paper on the Fried’s scale was still the most highly cited. Although the number of citations can be easily accessed, this electronic database search cannot replace the manual counting method as the papers citing the original articles do not necessarily use the tool in question.

Among previously published reviews 66 83 84 141 142 143 144 145 on frailty measures, only one 83 assessed them in terms of reliability and validity. Compared with the De Vries and colleagues’ paper 83 , our review presents additional strengths. First, to evaluate reliability and validity of a given instrument, we have extracted data from other studies, reflecting its level of external validation. Second, to our knowledge, no article has been published on the extent to which frailty measures have been used by other researchers. This finding might reflect the level preference of researchers for a given frailty measurement in the absence of a consensually recognized tool. Moreover, we identified 18 other frailty instruments 30 32 35 36 37 38 40 41 42 43 44 45 46 48 52 54 55 56 , 5 of them created in 2010 and after. Another limitation of our review may lie in the use of a unique keyword “frailty” to identify relevant publications on frailty measurements. One may find such a strategy restrictive, leading to miss some screening tools helping to identify frail elderly. In fact, we included similar frailty instruments than those comprised in the recent reviews 83 84 .