Aging Clinical and Experimental Research (2020) 32:1025–1034 https://doi.org/10.1007/s40520-020-01478-3 ORIGINAL ARTICLE Operationalization of a frailty index among older adults in the InCHIANTI study: predictive ability for all‑cause and cardiovascular disease mortality Emiel O. Hoogendijk1 · Sari Stenholm2,3 · Luigi Ferrucci4 · Stefania Bandinelli5 · Marco Inzitari6,7 · Matteo Cesari8,9 Received: 19 October 2019 / Accepted: 10 January 2020 / Published online: 31 January 2020 © The Author(s) 2020 Abstract Background The frailty index (FI) is a sensitive instrument to measure the degree of frailty in older adults, and is increas- ingly used in cohort studies on aging. Aims To operationalize an FI among older adults in the “Invecchiare in Chianti” (InCHIANTI) study, and to validate its predictive capacity for mortality. Methods Longitudinal data were used from 1129 InCHIANTI participants aged ≥ 65 years. A 42-item FI was operational- ized following a standard procedure using baseline data (1998/2000). Associations of the FI with 3- and 6-year all-cause and cardiovascular disease (CVD) mortality were studied using Cox regression. Predictive accuracy was estimated by the area under the ROC curve (AUC), for a continuous FI score and for different cut-points. Results The median FI was 0.13 (IQR 0.08–0.21). Scores were higher in women, and at advanced age. The FI was associated with 3- and 6-year all-cause and CVD mortality (HR range per 0.01 FI increase = 1.03–1.07, all p < 0.001). The continuous FI score predicted the mortality outcomes with moderate-to-good accuracy (AUC range 0.72–0.83). When applying FI cut- offs between 0.15 and 0.35, the accuracy of this FI for predicting mortality was moderate (AUC range 0.61–0.76). Overall, the predictive accuracy of the FI was higher in women than in men. Conclusions The FI operationalized in the InCHIANTI study is a good instrument to grade the risk of all-cause mortality and CVD mortality. More measurement properties, such as the responsiveness of this FI when used as outcome measure, should be investigated in future research. Keywords Frail elderly · Frailty index · Deficit accumulation · Risk prediction · Geriatrics Electronic supplementary material The online version of this article (https​://doi.org/10.1007/s4052​0-020-01478​-3) contains supplementary material, which is available to authorized users. 5 * Emiel O. Hoogendijk Geriatric Unit, Azienda Sanitaria di Firenze, Florence, Italy e.hoogendijk@amsterdamumc.nl 6 REFiT Research Group, Vall D’Hebrón Institute of Research 1 and Parc Sanitari Pere Virgili, Barcelona, Spain Department of Epidemiology and Biostatistics, Amsterdam 7 Public Health Research Institute, Amsterdam UMC - Department of Medicine, Universitat Autònoma de Location VU University Medical Center, Amsterdam, Barcelona, Bellaterra, Spain The Netherlands 8 Geriatric Unit, Fondazione IRCCS Ca’ Granda Ospedale 2 Department of Public Health, University of Turku and Turku Maggiore Policlinico, Milan, Italy University Hospital, Turku, Finland 9 Department of Clinical Sciences and Community Health, 3 Centre for Population Health Research, University of Turku University of Milan, Milan, Italy and Turku University Hospital, Turku, Finland 4 National Institute on Aging, Baltimore, MD, USA 13 Vol.:(0123456789) 1026 Aging Clinical and Experimental Research (2020) 32:1025–1034 Introduction study of older adults in Tuscany, Italy. The study is well known for its contribution to the frailty research field, In aging societies, there is increasing attention for the especially with regard to insights into determinants and concept of frailty [1, 2]. This is an age-related condition, onset of physical frailty in later life [22–28]. Nevertheless, which has been defined as a decline in multiple physiologi- all previous InCHIANTI studies measured frailty using the cal systems and increased vulnerability to stressors. Frailty frailty phenotype [29]. An FI has not yet been established is strongly related to various adverse health outcomes, in this cohort study. Therefore, our aim was to operation- such as functional decline, delirium, hospital admission, alize an FI in the InCHIANTI study, and to validate it by and mortality [3]. investigating its predictive ability for all-cause mortality In the past decades, many frailty instruments have been and CVD mortality. Both continuous scores and different developed [4], of which the frailty index (FI) is one of the cut-offs were used to study the predictive accuracy of this most commonly used [5]. The FI defines frailty as a gradi- FI. ent from fit to frail, based on the accumulation of health deficits [5]. A critical mass of health deficits (at least 30), including diseases, disabilities, signs, and symptoms, is Methods used to arithmetically generate a continuous score rang- ing from 0 to 1. As a result, the FI seems to be a sensi- Study sample tive instrument to distinguish the degree of frailty in older adults [6]. The InCHIANTI study is a prospective population-based Because of its sensitivity, the FI is increasingly rec- cohort study among older adults aged 65 years and over ognized as particularly useful for longitudinal studies on in Tuscany, Italy. Participants were randomly selected from aging, as predictor of adverse health events or as outcome population registers in this area. The study started in 1998 measure. As an outcome measure, the FI can be used to and is still ongoing. Follow-up measurements are conducted monitor the individual`s changes in frailty over time [7, every 3 years. The main focus of InCHIANTI is mobility 8]. Recently, FIs have been developed in various longi- decline in later life and related factors. The data collection is tudinal studies on aging, such as the English Longitudi- based on a home interview and clinical measurements at the nal Study of Aging (ELSA) [9], the Longitudinal Aging study clinic. Details on the sampling and design of InCHI- Study Amsterdam (LASA) [10, 11], the Health and Retire- ANTI have been described in a previous publication [30]. ment Study (HRS) [12], the Survey of Health, Aging and The InCHIANTI study was approved by the ethical com- Retirement in Europe (SHARE) [13, 14], and the Gothen- mittee of the Italian National Institute of Research and Care burg H70 Birth Cohort Studies (H70 Studies) [15]. This on Aging. All study participants provided written informed is a promising development, as this enables comparisons consent. between different populations, countries, and settings. In the current study, we included people with valid data at The FI was originally developed to be used as a con- baseline (1998–2000) and valid data on the outcome meas- tinuous scale [16]. However, dichotomizing a measure is ure (mortality) during follow-up. Of the 1155 InCHIANTI sometimes necessary to be able to identify the presence of participants at baseline, 24 (2.1%) had no valid frailty score certain conditions and to inform clinicians in the develop- due to missing data and 2 (0.2%) had no data on level of ment of diagnostic/therapeutic plans. Although cut-offs education. Vital status at follow-up was available for all are controversial, they have also been applied to the FI participants, resulting in a sample of 1129 people that were to determine the presence of frailty. Some studies have included in the analyses. reported an FI cut-off of 0.2 to identify a frail state [17], but a cut-point of 0.25 has most often been used (e.g., Frailty index [16, 18–20]). Interestingly, there is a lack of validation studies for these cut-offs with regard to outcome predic- An FI of health deficits was operationalized following a tion. Therefore, optimal FI cut-offs for predicting specific standard procedure, as described by Searle et al. [17]. In the adverse health outcomes are largely unknown. Further- current study, this was done with data from the InCHIANTI more, most FIs have been validated for all-cause mortal- baseline. However, we made sure that the included vari- ity (e.g., [11, 13]. Very few studies have investigated the ables are available at all InCHIANTI follow-up waves, so predictive validity for specific causes of death, such as that future longitudinal studies may use this FI as well. Vari- cardiovascular disease (CVD) mortality [21]. ables were included in the FI if they met various criteria: (a) The “Invecchiare in Chianti”, (i.e., Aging in the Chi- variables must be health-related deficits, such as symptoms, anti area; InCHIANTI) study is a population-based cohort signs, diseases, or disabilities, all associated with adverse outcomes (b) the deficits are biologically meaningful and 13 Aging Clinical and Experimental Research (2020) 32:1025–1034 1027 represent multiple organ systems, (c) deficits increase with retrieved from regional and municipality registers. CVD age, but do not saturate too early (i.e., high prevalence at mortality was determined by ICD-9 codes 390–459. young age), and (d) variables contain less than 5% missing values. Statistical analysis The questionnaires and clinical measurements of InCHI- ANTI were screened for variables that could be included in Descriptive analyses were conducted to show the charac- the FI. This resulted in a list of health deficits from various teristics of the study sample. T tests, Chi-square tests, and functional domains that were used to construct a 42-item Mann–Whitney tests were performed to determine differ- FI. The FI included the following variables: major medical ences in baseline characteristics by sex. Next, descriptive conditions (hypertension, myocardial infarction, congestive statistics were generated to provide insight into the general heart failure, chronic liver disease, cancer, peripheral arte- characteristics of the FI at baseline. The distribution of rial disease, stroke, Parkinson`s disease, diabetes, chronic the FI was displayed in a histogram. Mean frailty scores lung disease, angina pectoris, and knee/hip arthritis) based were plotted by age and sex. To investigate the predictive on disease ascertainment algorithms (including information ability of the FI, various analyses were performed. First, from medical records, medication use, signs and symptoms, associations of the FI (continuous score) with each outcome medical history, and hospital records); having difficulties (3-year and 6-year all-cause mortality and CVD mortality) with various (instrumental) activities of daily living (ADL/ were studied using Cox proportional hazard models, without IADL, including bathing, dressing/undressing, eating, toi- and with adjustment for confounders. Three models were leting, continence, walking across a small room, walking tested: a crude model (Model 1), a model adjusted for age up/down staircase ten steps, getting in/out of bed, food and sex (Model 2), and a model additionally adjusted for preparation, shopping, heavy housework, using telephone, partner status, educational level, and smoking (Model 3). In lifting/carry shopping bag, using public transportation, the Cox regression analyses, survivors were censored at the medication use, and managing finances); self-rated health end of follow-up (3 or 6 years after baseline). People who assessed with the question “How would you evaluate your died within 3 or 6 years were censored at the time of death. current health?”; five items from the Center for Epidemio- In the analyses on CVD mortality, those who died because of logic Studies Depression (CES-D) scale (feel depressed, feel other reasons than CVD were also censored at time of death. everything is an effort, could not get going, feel lonely, and Second, the predictive accuracy of the continuous FI score feel happy) [31]; four sub-domains of the Mini-Mental State was estimated by the area under the ROC curve (AUC). Examination (MMSE, domains: orientation time, orientation An AUC of 1.0 indicates perfect sensitivity and specificity place, attention, and recall) [32]; self-reported weight loss for the outcome of interest, 0.9–0.99 is excellent, 0.8–0.89 in the last 12 months; physical activity level in the last year; is good, 0.7–0.79 is moderate, and everything below 0.70 and two physical performance measures (gait speed, grip is poor [35]. Next, to find optimal FI cut-offs for predict- strength). Details of the items included in the FI and their ing mortality, we calculated the sensitivity, the specificity, cut-offs are shown in Table 1. and AUCs for various cut-points (between 0.15 and 0.35) In line with previous studies, we only calculated an FI around the commonly used cut-point of 0.25 [16, 18]. All for participants with less than 20% missing variables [11, analyses on predictive ability of the FI for all-cause mortal- 33]. Most older adults in the initial InCHIANTI sample had ity were done for the total population and stratified by sex, no missing variables on the FI (65.4%) or 1–3 missing vari- as sex differences in frailty have often been reported, with ables (27.2%) out of the total of 42 variables. Only 2.1% had higher frailty levels among women [36]. For CVD mortality, more than 20% missing variables. An FI ranging from 0 (no the number of events was too low to perform sex-stratified deficits present) to 1 (all deficits present) was calculated analyses. All analyses were done in SPSS 24 (IBM corp, for individual participants by dividing the sum of the items Armonk, NY, USA). present out of the sum of all the possible ones measured in the FI. To illustrate, if a person presents with 10 altered items out of 42, the corresponding FI score is 10/42 = 0.24. Results Outcomes Baseline characteristics are shown in Table 2. The analytical sample consisted of 1129 older adults, of which 642 (56.9%) Mortality is one of the most studied endpoints in the context were female. The participants had a mean age of 75.2 years of frailty [34]. In the current study, outcome measures were (SD = 7.4) and a mean educational level of 5.3 years 3-year and 6-year all-cause mortality and CVD mortality. (SD = 3.3). The distribution of the FI is displayed in Fig. 1. Vital status, date of death, and cause of death (International The FI is skewed to the right and ranges from 0.01 to 0.72. Classification of Disease, 9­ th revision (ICD-9) codes) were The median FI score was 0.13 (IQR = 0.08–0.21) and the 13 1028 Aging Clinical and Experimental Research (2020) 32:1025–1034 Table 1  Frailty index items and coding No. Deficit Cut-off 1 Hypertension No = 0, possible = 0.5, yes = 1 2 Myocardial infarction No = 0, possible = 0.5, yes = 1 3 Congestive heart failure No = 0, possible = 0.5, yes = 1 4 Chronic liver disease No = 0, possible = 0.5, yes = 1 5 Cancer No = 0, yes = 1 6 Peripheral arterial disease No = 0, possible = 0.5, yes = 1 7 Stroke No = 0, possible/TIA = 0.5, yes = 1 8 Parkinson`s disease No = 0, possible = 0.5, yes = 1 9 Diabetes No = 0, possible = 0.5, yes = 1 10 Chronic lung disease No = 0, possible = 0.5, yes = 1 11 Angina pectoris No = 0, possible = 0.5, yes = 1 12 Knee/hip arthritis No = 0, pain or stiffness = 0.5, pain and stiffness = 1 13 Bathing All these ADL/IADL items were coded as: 14 Dressing/undressing no difficulty = 0 with difficulty but without help = 0.33 15 Eating with some help from another person = 0.66 16 Toileting unable to do it = 1 17 Continence 18 Walking across small room 19 Walking up/down staircase ten steps 20 Getting in/out of bed 21 Food preparation 22 Shopping 23 Heavy housework 24 Using telephone 25 Lifting/carry shopping bag 26 Using public transportation 27 Medication use 28 Managing finances 29 Self-rated health Very good = 0, good = 0.25, fair = 0.50, poor = 0.75, very poor = 1 30 Feel depressed (CES-D) All these CES-D items were coded as: 31 Feel everything is an effort (CES-D) rarely or never = 0 sometimes = 0.33 32 Could not get going (CES-D) occasionally = 0.66 33 Feel lonely (CES-D) often or always = 1 34 Feel happy (CES-D) Often or always = 0, occasionally = 0.33, sometimes = 0.66, rarely or never = 1 35 Orientation time (MMSE) Five correct = 0, one wrong = 0.50, two or more wrong = 1 36 Orientation place (MMSE) Five correct = 0, one wrong = 0.50, two or more wrong = 1 37 Attention (MMSE) Five correct = 0, one or two wrong = 0.50, three or more wrong = 1 38 Recall (MMSE) Three correct = 0, two correct = 0.50, one or zero correct = 1 39 Weight loss No = 0, yes, (weight loss > 10 lbs. in past year) = 1 40 Low physical activity No = 0, yes (hardly any physical activity or < 1 h a week) = 1 41 Slow gait speed Normal = 0, lowest quintile, stratified by height and sex = 1 42 Low grip strength Normal = 0, lowest quintile, stratified by BMI and sex = 1 99% upper limit was 0.64. When applying cut-offs, the prev- Table 3 shows the associations of the continuous FI alence of frailty ranged from 10.4% (≥ 0.35 cut-off) to 39.9% score with all-cause mortality for the total sample as well (≥ 0.15 cut-off). Frailty scores were higher in women than in as after stratification by sex. In the total sample (n = 1129), men (median = 0.14 vs. median = 0.10, p < 0.001). Figure 2 112 people (9.9%) died within 3 years of follow-up shows that the FI scores tend to be higher with advancing and 267 people (23.6%) died within 6 years of follow- age in both men and women. up. Mortality was higher in men than women (3-year 13 Aging Clinical and Experimental Research (2020) 32:1025–1034 1029 Table 2  Baseline characteristics Characteristics Total Men Women p ­valuea n = 1129 n = 487 n = 642 Age, mean (SD) 75.2 (7.4) 74.2 (7.0) 75.9 (7.7) < 0.001 Partner status, n (%) with partner 674 (59.7) 384 (78.9) 290 (45.2) < 0.001 Educational level, years, mean (SD) 5.3 (3.3) 6.2 (3.6) 4.7 (2.9) < 0.001 Smoking, n (%) current smoker 158 (14.0) 103 (21.1) 55 (8.6) < 0.001 ADL disabilities, 0–6, mean (SD) 0.2 (1.0) 0.2 (0.9) 0.3 (1.0) 0.21 CES-D score, 0–60, mean (SD) 13.1 (8.8) 9.9 (7.3) 15.6 (9.0) < 0.001 MMSE score, 0–30, mean (SD) 24.3 (4.9) 25.2 (4.4) 23.7 (5.2) < 0.001 FI score, 0–1, median (IQR) 0.13 (0.08–0.21) 0.10 (0.07–0.17) 0.14 (0.10–0.24) < 0.001 FI cut-offs, n (%) ≥ 0.15 451 (39.9) 144 (29.6) 307 (47.8) < 0.001 ≥ 0.20 301 (26.7) 93 (19.1) 208 (32.4) < 0.001 ≥ 0.25 214 (19.0) 62 (12.7) 152 (23.7) < 0.001 ≥ 0.30 154 (13.6) 47 (9.7) 107 (16.7) < 0.01 ≥ 0.35 117 (10.4) 40 (8.2) 77 (12.0) < 0.05 a Based on t test, Chi-square test, or Mann–Whitney test 200 0.45 Men 180 0.4 Women 160 0.35 140 Frailty Index score 0.3 120 Frequency 0.25 100 0.2 80 0.15 60 0.1 40 0.05 20 0 0 64 69 74 79 84 89 0 0.2 0.4 0.6 0.8 Frailty Index score (0-1) Age (in years) Fig. 1  Distribution of the frailty index at baseline Fig. 2  Association between age and baseline frailty index score for men and ­ womena. aThe circles and squares indicate frailty index scores by age rounded to nearest whole number, and may represent mortality = 11.5% vs. 8.7%; 6-year mortality = 27.9% vs. more than one individual 20.4%, respectively). An increase in all-cause mortality by FI score was observed for both men and women (Fig. 3). Table 3 also shows CVD mortality for the total population. The predictive accuracy of the continuous FI score for the During 3 years and 6 years of follow-up, 53 (4.7%) and studied outcomes (Tables 4 and 5) ranged from moderate to 128 people (11.3%) died because of CVD, respectively. good (AUC range 0.72–0.83), and was a bit higher among The FI showed statistically significant associations with women (AUC range 0.80–0.83) compared to men (AUC all outcomes in all models. For example, 0.01 increment in range 0.72–0.77). For the total population, the AUC was the FI score was associated with 3-year all-cause mortality the same for predicting 3-year and 6-year all-cause mortality (HR 1.04, 95% CI 1.03–1.06) and 3-year CVD mortality (AUC 0.76), and nearly the same for predicting 3-year and (HR 1.05, 95% CI 1.03–1.06) in models adjusted for age, 6-year CVD mortality (AUC 0.79 vs. AUC 0.78). Tables 4 sex, partner status, educational level, and smoking. The and 5 also show the predictive accuracy of different FI cut- associations between the FI score and all-cause mortality offs. For prediction of mortality, all cut-offs between 0.15 were stronger among women than among men (p interac- and 0.35 had moderate accuracy (AUC < 0.80). In the Online tion FI*sex was < 0.05, not shown in table) for both 3-year Appendix, ROC curves and optimal cut-offs for all outcomes and 6-year mortality. are graphically displayed (see Online Appendix 1). The 13 1030 Aging Clinical and Experimental Research (2020) 32:1025–1034 Table 3  Cox regression: associations between frailty index score and mortality All-cause mortality CVD mortality 3 years 6 years 3 years 6 years Total Men Women Total Men Women Total Total n events/total n 112/1129 56/487 56/642 267/1129 136/487 131/642 53/1129 128/1129 HR (95% CI)a HR (95% CI)a HR (95% CI)a HR (95% CI)a HR (95% CI)a HR (95% CI)a HR (95% CI)a HR (95% CI)a Model ­1b 1.06 (1.05– 1.05 (1.04– 1.07 (1.06– 1.06 (1.05– 1.05 (1.04– 1.07 (1.06– 1.07 (1.05– 1.07 (1.06– 1.07) 1.06) 1.09) 1.07) 1.06) 1.08) 1.08) 1.08) Model ­2b 1.04 (1.03– 1.03 (1.02– 1.05 (1.04– 1.04 (1.03– 1.03 (1.02– 1.05 (1.04– 1.05 (1.03– 1.04 (1.03– 1.05) 1.05) 1.07) 1.05) 1.04) 1.06) 1.06) 1.06) Model ­3b 1.04 (1.03– 1.03 (1.02– 1.05 (1.04– 1.04 (1.03– 1.03 (1.02– 1.05 (1.04– 1.05 (1.03– 1.04 (1.03– 1.06) 1.05) 1.07) 1.05) 1.05) 1.06) 1.06) 1.06) HR hazard ratio, 95% CI 95% confidence interval, CVD cardiovascular disease a The hazard ratios indicate change in mortality with an increase of 0.01 on the frailty index b Model 1: unadjusted; Model 2: adjusted for age and sex; Model 3: adjusted for age, sex, partner status, educational level, and smoking; Sex adjustment only in the analysis of the total population optimal cut-off for predicting 3-year mortality was 0.198 mortality. Consistent with results from a large number of (AUC 0.71), for 6-year mortality 0.191 (AUC 0.72), for earlier studies (e.g., [11, 34, 38–40]), a higher FI score in the 3-year CVD mortality 0.201 (AUC 0.76), and for 6-year InCHIANTI study was associated with increased all-cause CVD mortality 0.201 (AUC 0.73). Sex differences can be mortality. A novel finding is that the FI was also related to observed with regard to the optimal FI cut-point for out- CVD mortality. come prediction. For example, for men, the optimal cut-off Although sex differences in frailty have been reported in for predicting 3-year mortality was around ≥ 0.17, while for many studies, relatively little attention has been paid to sex women, the optimal cut-off was around ≥ 0.28 (Tables 4 and differences in the association between frailty and adverse 5, Online Appendix 1). outcomes [36]. A common observation from previous research is that men have higher mortality rates in combi- nation with lower average frailty scores [36]. Our results Discussion showed that, even with greater mortality rates among men, a stronger relationship between the FI score and all-cause In the present study, we operationalized a 42-item FI in the mortality was observed among women. Perhaps, this asso- InCHIANTI study, a large population-based cohort study. ciation is influenced by the greater increase in mortality in We validated this FI for predicting mortality, and investi- the highest FI categories among women, as can be observed gated optimal cut-offs for outcome prediction. Our results in Fig. 3. These findings suggest that a sex-specific approach showed that this FI is associated with 3-year and 6-year all- to frailty is warranted when the FI is used to predict adverse cause and CVD mortality. It is, therefore, a good instru- outcomes. More research is needed to investigate whether ment to grade the risk of mortality in older adults. The FI sex differences are also present for associations with other predicted mortality with moderate-to-good accuracy, and outcomes than mortality. showed slightly better predictive accuracy in women than The predictive accuracy of the continuous FI score for in men. Interestingly, the optimal FI cut-off for predicting mortality was similar to findings from other population- mortality differed between men and women. based cohort studies. We observed AUCs of 0.76 for predict- The characteristics of the FI developed in the InCHIANTI ing both 3-year and 6-year mortality. This was comparable to study are in line with previously published FIs in older pop- the previous work in the SHARE study, which found AUCs ulation-based samples. In fact, our FI had a skewed right for predicting 2-year and 5-year mortality of 0.77 and 0.75, distribution, a 99% upper limit below 0.70, and higher val- respectively [33]. And it is also not very different from other ues in women and persons with more advanced age [11, risk indicators that have been used in older populations, such 17, 37]. The median FI score of 0.13 did not differ much as gait speed measurement, which also showed moderate from the FI that was reported in the Longitudinal Aging predictive accuracy for mortality with AUCs around 0.70 Study Amsterdam (median = 0.16) in the same time period [41]. Nevertheless, it is very likely that the predictive accu- (late 1990s) [11]. Furthermore, we validated the FI against racy of the FI differs across settings and subpopulations. For 13 Aging Clinical and Experimental Research (2020) 32:1025–1034 1031 100 different settings. Future studies should explore to what A Men 3-year mortality extent the predictive ability of the FI in clinical settings 90 6-year mortality differs from that in population-based samples, and whether 80 there is consistency in the predictive ability of the FI within specific settings. It is, for example, possible that the FI is less 70 predictive in long-term care facilities, where average frailty 60 levels are much higher than in other settings [42]. One of the main advantages of using an FI is that the con- 50 tent is not fixed. As long as several conditions are met, such % as the type and number of included health deficits (at least 40 30 items), the combination of health deficits does not matter 30 [17]. The key characteristics of the FI are consistent across data sets with different FI operationalizations [37]. Because 20 of this flexibility, FIs can be constructed with almost any comprehensive health database in both research and clinical 10 settings. The FI that was constructed in the current study 0 contains 42 items from various domains of functioning. <0.10 0.10-0.19 0.20-0.29 0.30-0.39 ≥0.40 However, our aim was not to construct an FI that is com- Frailty Index score pletely fixed. In future research with InCHIANTI data, it is 100 still possible to replace items or to add more items, if this is B Women 3-year mortality needed for specific research questions. 90 6-year mortality The FI is not meant to be dichotomized, as Rockwood 80 et al. described in their paper from 2007 [16]. However, when a cut-point is needed, an arbitrary cut-off of ≥ 0.25 70 has usually been proposed for community-dwelling older adults. Until now, this cut-point has been used in many stud- 60 ies [18–20]. In the present study, we explored the use of vari- 50 ous cut-offs around this commonly used cut-point. Whereas % the predictive accuracy of the continuous FI was moderate- 40 to-good, the dichotomized scores had slightly lower levels of accuracy. Remarkably, we observed sex differences with 30 regard to the optimal cut-off for outcome prediction. The 20 optimal cut-off for men for predicting 3-year mortality was around 0.17, while for women, this was around 0.28. This 10 could just be an expression of the male–female health-sur- 0 vival paradox (i.e., a higher life expectancy that is accom- <0.10 0.10-0.19 0.20-0.29 0.30-0.39 ≥0.40 panied by higher rates of poor health in women compared Frailty Index score to men) [36]. It should be noted that for some higher cut-off values, especially among men, the sensitivity was low. It Fig. 3  Percentage of all-cause mortality at follow-up by baseline depends on the purpose of the study, but if a dichotomized frailty index score for A men and B women FI would be used as screening instrument, one would prefer high sensitivity as this would select the highest number of example, in an earlier Italian study, a slightly higher AUC true positive cases (people that experience the adverse health (0.81) for predicting 2-year mortality was observed [38]. The event). Additional research in other cohorts is, therefore, FI was of similar size as ours (40 items), but the main differ- needed, to further investigate optimal FI cut-offs for specific ence is that our study was conducted in a population-based outcomes. sample, while this previous study was done in a specific The current study has several strengths. We used data sample in a clinical setting. Another study conducted in a from a large population-based cohort of older adults in Italy clinical setting found; however, moderate predictive accu- to validate an FI that was constructed based on existing FI racy for 1-year, 3-year, and 5-year mortality (AUC ≤ 0.75) methodology. Our study expands previous research in vari- when operationalizing a frailty index using data from elec- ous ways. First, this is one the few studies that tests vari- tronic medical records [39]. This highlights the need for ous FI cut-offs for outcome prediction. Second, this study more research into the predictive accuracy of the FI across contributes to insights into sex differences in frailty, by 13 1032 Aging Clinical and Experimental Research (2020) 32:1025–1034 Table 4  Predictive accuracy of continuous frailty index scores and various frailty index cut-offs for 3-year mortality 3-Year all-cause mortality 3-Year CVD mortality Total Men Women Total Sens Spec AUC (95% CI) Sens Spec AUC (95% CI) Sens Spec AUC (95% CI) Sens Spec AUC (95% CI) Continuous 0.76 (0.70–0.81) 0.72 (0.64–0.80) 0.83 (0.77–0.89) 0.79 (0.72–0.87) Cut-off ≥ 0.15 0.72 0.64 0.68 (0.63–0.73) 0.61 0.74 0.68 (0.60–0.75) 0.84 0.56 0.70 (0.63–0.76) 0.77 0.62 0.70 (0.63–0.77) ≥ 0.20 0.64 0.77 0.71 (0.66–0.76) 0.52 0.85 0.68 (0.60–0.77) 0.77 0.72 0.74 (0.68–0.81) 0.75 0.76 0.76 (0.69–0.83) ≥ 0.25 0.54 0.85 0.70 (0.64–0.75) 0.41 0.91 0.66 (0.57–0.75) 0.68 0.81 0.74 (0.69–0.82) 0.62 0.83 0.73 (0.65–0.81) ≥ 0.30 0.48 0.90 0.69 (0.63–0.75) 0.32 0.93 0.63 (0.54–0.71) 0.64 0.88 0.76 (0.68–0.84) 0.55 0.88 0.72 (0.63–0.80) ≥ 0.35 0.43 0.93 0.68 (0.62–0.74) 0.29 0.94 0.62 (0.53–0.70) 0.57 0.92 0.75 (0.67–0.83) 0.49 0.92 0.70 (0.62–0.79) Sens. Sensitivity, Spec. specificity, AUC​area under the ROC curve, CVD cardiovascular disease Table 5  Predictive accuracy of continuous frailty index scores and various frailty index cut-offs for 6-year mortality 6-Year all-cause mortality 6-Year CVD mortality Total Men Women Total Sens Spec AUC (95% CI) Sens Spec AUC (95% CI) Sens Spec AUC (95% CI) Sens Spec AUC (95% CI) Continuous 0.76 (0.73–0.80) 0.77 (0.72–0.82) 0.80 (0.76–0.85) 0.78 (0.73–0.82) Cut-off ≥ 0.15 0.70 0.69 0.69 (0.66–0.73) 0.58 0.81 0.70 (0.64–0.75) 0.82 0.61 0.71 (0.67–0.76) 0.75 0.65 0.70 (0.65–0.75) ≥ 0.20 0.58 0.83 0.71 (0.67–0.75) 0.48 0.92 0.70 (0.64–0.76) 0.69 0.77 0.73 (0.68–0.78) 0.67 0.79 0.73 (0.68–0.78) ≥ 0.25 0.48 0.90 0.69 (0.65–0.73) 0.37 0.97 0.67 (0.61–0.73) 0.60 0.86 0.73 (0.67–0.78) 0.56 0.86 0.71 (0.66–0.77) ≥ 0.30 0.40 0.94 0.67 (0.63–0.71) 0.28 0.97 0.63 (0.57–0.69) 0.52 0.92 0.72 (0.67–0.78) 0.48 0.91 0.69 (0.64–0.75) ≥ 0.35 0.33 0.97 0.65 (0.61–0.69) 0.24 0.98 0.61 (0.55–0.67) 0.42 0.96 0.69 (0.63–0.75) 0.38 0.93 0.65 (0.60–0.71) Sens. sensitivity, Spec. specificity, AUC​area under the ROC curve, CVD cardiovascular disease providing detailed information on the predictive accuracy instrument to select older adults in clinical practice that of the FI for men and women separately. And, finally, in may benefit from additional care or specific interventions addition to all-cause mortality, we have also investigated [39, 40]. However, this field is in an early stage of develop- CVD mortality as outcome. Specific causes of death have ment. For example, in the UK, an electronic frailty index seldom been studied in relation to the FI [21]. (eFI) is currently being implemented for older adults in The study also has some limitations. First, we validated primary care [39]. Based on generic cut-points of the eFI, the FI only for mortality, an important outcome measure, the general practitioner has to carry out additional actions but other outcomes should be considered in future research, for those with moderate-to-severe frailty. Although this is such as falls, functional decline, and healthcare utilization a promising development, it is possible that some refine- [3]. Second, we only have studied the main characteristics ment and further specificity of cut-points for use in clinical and the predictive accuracy of the FI. More measurement practice are needed in the future. The results of the current properties need to be investigated. For example, the instru- study would, for example, implicate that cut-points need ment has great potential to be used as outcome measure. A to be sex-specific, as men are at risk of mortality at lower crucial next step is to test the responsiveness of this FI when FI levels than women. used as outcome in longitudinal research. Another important To conclude, we have operationalized an FI in the direction for future research is the performance of compara- InCHIANTI study and validated this FI for predicting tive research, to see to what extent the predictive ability of mortality. This FI is a good instrument to grade the risk this FI differs from that of other risk indicators for mortality of 3-year and 6-year all-cause and CVD mortality in older or other frailty measures. adults. The predictive accuracy of the FI was slightly The application of the FI is not only limited to research better among women compared to men. Future research settings. It has been suggested that the FI is a good 13 Aging Clinical and Experimental Research (2020) 32:1025–1034 1033 should investigate the responsiveness of the FI when using 7. Cesari M, Gambassi G, Abellan van Kan G, Vellas B (2014) The this instrument as outcome measure. frailty phenotype and the frailty index: different instruments for different purposes. Age Ageing 43:10–12 8. Hoogendijk EO, Rockwood K, Theou O et al (2018) Tracking Acknowledgements The content of this research is solely the responsi- changes in frailty throughout later life: results from a 17-year lon- bility of the authors, and does not reflect the views of the InCHIANTI gitudinal study in the Netherlands. Age Ageing 47:727–733 study group. Emiel O. Hoogendijk was supported by an NWO/ZonMw 9. Wade KF, Marshall A, Vanhoutte B et al (2017) Does pain pre- Veni fellowship [Grant Number 91618067]. This study was in part dict frailty in older men and women? Findings from the English funded by the Intramural Research Program at the National Institute Longitudinal Study of Ageing (ELSA). J Gerontol A 72:403–409 on Aging. The InCHIANTI study baseline (1998–2000) was funded 10. Hoogendijk EO, Deeg DJH, de Breij S et al (2019) The Longitu- by the Italian Ministry of Health (ICS110.1/RF97.71) and in part by dinal Aging Study Amsterdam: cohort update 2019 and additional the U.S. National Institute on Aging, Bethesda, Maryland (Contracts data collections. Eur J Epidemiol. https​://doi.org/10.1007/s1065​ 236 MD 916413 and 236 MD 821336). The InCHIANTI follow-up 1 4-019-00541​-2 (2001–2003) was funded by the U.S. National Institute on Aging (Con- 11. Hoogendijk EO, Theou O, Rockwood K et al (2017) Development tracts: N.1-AG-1–1 and N.1-AG-1–2111); the InCHIANTI follow-up 2 and validation of a frailty index in the Longitudinal Aging Study and 3 studies (2004–2010) were financed by the U.S. National Institute Amsterdam. Aging Clin Exp Res 29:927–933 on Aging (Contract: N01-AG-5–0002). 12. Mezuk B, Lohman MC, Rock AK et al (2016) Trajectories of body mass indices and development of frailty: evidence from Compliance with ethical standards the Health and Retirement Study. Obesity 24:1643–1647 13. Romero-Ortuno R, Kenny RA (2012) The frailty index in Conflict of interest The authors declare that they have no conflict of Europeans: association with age and mortality. Age Ageing interest. 41:684–689 14. Stolz E, Mayerl H, Waxenegger A et al (2017) Impact of socio- Statement of human and animal rights The InCHIANTI study was economic position on frailty trajectories in 10 European countries: approved by the ethical committee of the Italian National Institute of evidence from the Survey of Health, Ageing and Retirement in Research and Care on Aging. Europe (2004–2013). J Epidemiol Community Health 71:73–80 15. Backman K, Joas E, Falk H et al (2017) Changes in the lethality Informed consent Signed informed consent was obtained from all of frailty over 30 years: evidence from two cohorts of 70-year-olds study participants. in Gothenburg Sweden. J Gerontol A 72:945–950 16. Rockwood K, Andrew M, Mitnitski A (2007) A comparison of two approaches to measuring frailty in elderly people. J Gerontol Open Access This article is licensed under a Creative Commons Attri- A 62:738–743 bution 4.0 International License, which permits use, sharing, adapta- 17. Searle SD, Mitnitski A, Gahbauer EA et al (2008) A standard tion, distribution and reproduction in any medium or format, as long procedure for creating a frailty index. BMC Geriatr 8:24 as you give appropriate credit to the original author(s) and the source, 18. Song X, Mitnitski A, Rockwood K (2010) Prevalence and 10-year provide a link to the Creative Commons licence, and indicate if changes outcomes of frailty in older adults in relation to deficit accumula- were made. The images or other third party material in this article are tion. J Am Geriatr Soc 58:681–687 included in the article’s Creative Commons licence, unless indicated 19. Wou F, Gladman JR, Bradshaw L et al (2013) The predictive prop- otherwise in a credit line to the material. If material is not included in erties of frailty-rating scales in the acute medical unit. Age Ageing the article’s Creative Commons licence and your intended use is not 42:776–781 permitted by statutory regulation or exceeds the permitted use, you will 20. Eeles EM, White SV, O’Mahony SM et al (2012) The impact of need to obtain permission directly from the copyright holder. To view a frailty and delirium on mortality in older inpatients. Age Ageing copy of this licence, visit http://creat​iveco​mmons​.org/licen​ses/by/4.0/. 41:412–416 21. Jiang M, Foebel AD, Kuja-Halkola R et al (2017) Frailty index as a predictor of all-cause and cause-specific mortality in a Swedish population-based cohort. Aging 9:2629–2646 References 22. Bartali B, Frongillo EA, Bandinelli S et al (2006) Low nutrient intake is an essential component of frailty in older persons. J Ger- 1. Cesari M, Marzetti E, Thiem U et al (2016) The geriatric manage- ontol A Biol Sci Med Sci 61:589–593 ment of frailty as paradigm of "The end of the disease era". Eur J 23. Cesari M, Leeuwenburgh C, Lauretani F et al (2006) Frailty syn- Intern Med 31:11–14 drome and skeletal muscle: results from the Invecchiare in Chianti 2. Hoogendijk EO, Afilalo J, Ensrud KE et al (2019) Frailty: study. Am J Clin Nutr 83:1142–1148 implications for clinical practice and public health. Lancet 24. Shardell M, D’Adamo C, Alley DE et al (2012) Serum 25-hydrox- 394:1365–1375 yvitamin D, transitions between frailty states, and mortality in 3. Clegg A, Young J, Iliffe S et al (2013) Frailty in elderly people. older adults: the Invecchiare in Chianti Study. J Am Geriatr Soc Lancet 381:752–762 60:256–264 4. Dent E, Kowal P, Hoogendijk EO (2016) Frailty measurement in 25. Talegawkar SA, Bandinelli S, Bandeen-Roche K et al (2012) A research and clinical practice: a review. Eur J Intern Med 31:3–10 higher adherence to a Mediterranean-style diet is inversely asso- 5. Rockwood K, Mitnitski A (2007) Frailty in relation to the accu- ciated with the development of frailty in community-dwelling mulation of deficits. J Gerontol A 62:722–727 elderly men and women. J Nutr 142:2161–2166 6. Rockwood K, Mitnitski A (2011) Frailty defined by deficit accu- 26. Rabassa M, Zamora-Ros R, Urpi-Sarda M et al (2015) Association mulation and geriatric medicine defined by frailty. Clin Geriatr of habitual dietary resveratrol exposure with the development of Med 27:17–26 frailty in older age: the Invecchiare in Chianti study. Am J Clin Nutr 102:1534–1542 27. Urpi-Sarda M, Andres-Lacueva C, Rabassa M et al (2015) The relationship between urinary total polyphenols and the frailty 13 1034 Aging Clinical and Experimental Research (2020) 32:1025–1034 phenotype in a community-dwelling older population: the InCHI- 36. Gordon EH, Peel NM, Samanta M et al (2017) Sex differences ANTI Study. J Gerontol A Biol Sci Med Sci 70:1141–1147 in frailty: a systematic review and meta-analysis. Exp Gerontol 28. Stenholm S, Ferrucci L, Vahtera J et al (2019) Natural course 89:30–40 of frailty components in people who develop frailty syndrome: 37. Rockwood K, Mitnitski A (2006) Limits to deficit accumulation evidence from two cohort studies. J Gerontol A Biol Sci Med Sci in elderly people. Mech Ageing Dev 127:494–496 74:667–674 38. Abete P, Basile C, Bulli G et al (2017) The Italian version of 29. Fried LP, Tangen CM, Walston J et al (2001) Frailty in older the "frailty index" based on deficits in health: a validation study. adults: evidence for a phenotype. J Gerontol A Biol Sci Med Sci Aging Clin Exp Res 29:913–926 56:M146–156 39. Clegg A, Bates C, Young J et al (2016) Development and vali- 30. Ferrucci L, Bandinelli S, Benvenuti E et al (2000) Subsystems dation of an electronic frailty index using routine primary care contributing to the decline in ability to walk: bridging the gap electronic health record data. Age Ageing 45:353–360 between epidemiology and geriatric practice in the InCHIANTI 40. Drubbel I, de Wit NJ, Bleijenberg N et al (2013) Prediction of study. J Am Geriatr Soc 48:1618–1625 adverse health outcomes in older people using a frailty index 31. Radloff L (1977) The CES-D scale: a self-report depression based on routine primary care data. J Gerontol A 68:301–308 scale for research in the general population. Appl Psychol Meas 41. Studenski S, Perera S, Patel K et al (2011) Gait speed and survival 1:385–401 in older adults. JAMA 305:50–58 32. Folstein MF, Folstein SE, McHugh PR (1975) “Mini-mental 42. Rockwood K, McMillan M, Mitnitski A et al (2015) A frailty state”. A practical method for grading the cognitive state of index based on common laboratory tests in comparison with a patients for the clinician. J Psychiatr Res 12:189–198 clinical frailty index for older adults in long-term care facilities. 33. Theou O, Brothers TD, Mitnitski A et al (2013) Operationaliza- J Am Med Dir Assoc 16:842–847 tion of frailty using eight commonly used scales and comparison of their ability to predict all-cause mortality. J Am Geriatr Soc Publisher’s Note Springer Nature remains neutral with regard to 61:1537–1551 jurisdictional claims in published maps and institutional affiliations. 34. Kojima G, Iliffe S, Walters K (2018) Frailty index as a predictor of mortality: a systematic review and meta-analysis. Age Ageing 47:193–200 35. Carter JV, Pan J, Rai SN et al (2016) ROC-ing along: evaluation and interpretation of receiver operating characteristic curves. Sur- gery 159:1638–1645 13