Development and internal validation population

We developed our FI using data from the Health Search Database (HSD), an Italian primary care database^24,25. We will refer to our FI as the HS-Primary Care FI (hereinafter PC-FI) from now on, to avoid confusion with other FIs. Since 1998, the HSD has collected data from a network of about 800 Italian GPs, who collect and register patients’ routine clinical information. The network of participating GPs is evenly distributed across Italy, covering a population of more than one million primary care patients. The HSD contains information on demographics, clinical diagnoses, drug prescriptions and diagnostic tests, specialist referrals, hospital admissions, and death. Diseases are classified in accordance with the International Classification of Diseases 9^th Revision Clinical Modification (ICD-9CM), drug prescriptions according to the Anatomical Therapeutic Chemical (ATC) system, economic exemptions and diagnostics tests and referrals according to codes issued by the Italian Ministry of Health. For the purposes of our study, we analyzed a random cohort of 308,280 individuals 60 + years old, followed up by the same GP for at least 5 years (median follow-up time 6.9 years) if death did not occur earlier, across a timeframe from 1 January 2013 to 31 December 2019. For this study, the entry date was 1 January 2013 for those that turned 60 before that date or their 60th birthday for the others. Data were randomly split into a training dataset (N = 184,968; 60%) employed for the construction of the FI, and a testing dataset (N = 123,312; 40%) used for the internal validation. The characteristics of the two subsamples were identical as shown in Table S1. In accordance with the Italian legislation, the permission to use anonymized electronic healthcare records is granted for observational epidemiological research as the present one. All methods were performed in accordance with the relevant guidelines and regulations of international Code of Conduct. The study was reported in keeping with the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) recommendations.

External validation population

For the external validation of the PC-FI, we used data from the Swedish National Study on Aging and Care in Kungsholmen (SNAC-K), an ongoing population-based study started in 2001 and involving individuals 60 + years old living in the Kungsholmen district in central Stockholm, Sweden. Thirteen years of follow-up were used for the present study. At baseline (i.e., 2001–2003), 3363 persons (73.3% response rate) underwent a comprehensive assessment using standard questionnaires administered by trained nurses, medical examinations, and instrumental and blood tests to retrieve information about their demographics, and clinical and functional status. Individuals’ information was also linked with the Swedish National Patient Register, Stockholm Regional Outpatient Register and the Swedish Death Register. An in-depth description of the data collection protocols is available elsewhere²⁶. Informed consent was obtained from each participant, or from a proxy in case of cognitive impairment. The study was approved by the Regional Ethical Review Board in Stockholm.

Deficit identification

Individual-level characteristics describing their health status across different organ and systems were considered as potential deficits. Overall, 101 potential deficits were identified in the HSD, based on a clinical appraisal of the information available in the dataset and looking backward to the whole patients’ available health record (Table S2): (a) 64 groups of chronic diseases; (b) 17 signs and symptoms reported in the six months prior to study entry date, (c) 11 related to healthcare utilization in the six months prior to study entry date (e.g., number of GP visits, emergency department admissions, or hospitalizations), (d) six related to acute conditions or drug prescriptions in the six months prior to study entry date (e.g., infectious diseases, oxygen prescription), and (e) three addressing functional or financial difficulties. Chronic diseases were defined based on a previously published list of 60 categories of conditions proposed by Calderón-Larrañaga et al. and identified by means of their ICD-9CM codes²⁷. Signs, symptoms, and acute conditions were also identified through ICD-9CM codes. In the HSD, recently prescribed drugs, diagnostic tests, and specialist referrals are identified by means of the ATC and Italian Ministry of Health codes, respectively. Several exonerations, as registered by GPs through specific codes, were used to identify disability and economical vulnerabilities (i.e., financial exonerations for drugs, or financial exonerations for patients with disability). All deficits were coded to range between 0 and 1. Five potential deficits had 3 levels (i.e., 0, 0.5 and 1) and the remaining included two values (i.e., 0 and 1; Table S2).

In SNAC-K, only those deficits selected for the PC-FI were computed for the external validation. Information on these deficits was obtained through comprehensive interviews, physical examinations, and register data, as reported in Table S3. Diseases were coded using ICD-10 codes, which were mapped to ICD-9 codes via official mapping sheets.

Frailty index construction: the genetic algorithm

To select the deficits to be included in the PC-FI, we employed a genetic algorithm. A genetic algorithm is an optimization algorithm employed to find near-optimal solutions for problems where it is not computationally feasible to evaluate all possible combinations of elements (e.g., the deficits to include in a frailty index). An in-depth explanation of this methodology applied to frailty indices is available elsewhere and in the Supplementary methods²³. Shortly, a genetic algorithm iteratively tests the discriminative ability of a group of frailty indices in the prediction of mortality, selects those exhibiting better performances, and randomly combine them, creating new frailty indices. These newly created frailty indices replace those showing worse predictive performances. At the beginning, the frailty indices evaluated by the genetic algorithm are randomly created and their predictive performance is low, but iteration after iteration, frailty indices exhibiting better and better predictive performances are selected by the algorithm. The algorithm stops when a certain number of iterations is run or when it fails to find a better frailty index. In this study, the number of frailty indices evaluated in each iteration was 1500. The predictive performance for mortality—used as the only outcome to assess models’ fitness—was calculated as the average c-statistic obtained through unadjusted Cox regression models in the whole sample and in specific sex- (male and female), age- (younger and older than 71 years—the median age), and geographic area- (Northern, Central, and Southern Italy) subsamples. All-cause mortality over the whole follow up period was used as outcome.

We ran the genetic algorithm 50 times: in each instance, the training sample was based on different groups of 10,000 randomly chosen participants among those included in the training dataset. To explore the relative importance of each deficit, we counted the number of times each deficit was included in the FI that exhibited the highest average c-statistic in its iteration. The final and best PC-FI was the one conformed by all of the most important deficits and showed the highest average c-statistic in the whole training dataset.

Four cut-offs were identified to stratify the study population in the following categories: “fit”, “mildly frail”, “moderately frail”, “severely frail”. We identified such categories by creating four equally spaced PC-FI intervals using the 99^th percentile of PC-FI as the upper limit¹⁸. The 99th percentile of the PC-FI was 0.28 and the intervals used to categorize frailty were: < 0.07, 0.07 to < 0.14, 0.14 to < 0.21, and ≥ 0.21.

For a secondary analysis, the eFI as proposed by Clegg et al. was calculated in the HSD¹⁸.

Outcomes

Mortality rate over 5 years was the outcome used to train our algorithm during the construction of the FI in HSD. Hospitalization was also tested as an alternative outcome. In the HSD, information on 5-year mortality and first hospitalization was retrieved from the GPs health records, where specific codes and the date of occurrence are used to register such events in their software.

For the external validation of the FI in the SNAC-K cohort, hospitalization and mortality information was retrieved from the Swedish National Patient Register and the Swedish Death Register, respectively. Furthermore, to test the convergent validity of our FI, several clinical and functional outcomes—both cross-sectional and longitudinal—were collected from the SNAC-K population^28,29,30. Disability was defined as the presence of at least one impairment in the activities of daily living (ADL; grooming/personal hygiene, dressing, toileting/continence, transferring/ambulating, and eating) and instrumental ADL (IADL; managing finances, managing transportation, shopping, meal preparation, housecleaning, laundry, use of telephone, managing medications). Physical frailty, is another commonly used conceptual model to assess frailty, and was evaluated according to the frailty phenotype proposed by Fried L. et al., using a cut-off of at least three out of five frailty criteria (i.e., unintentional weight loss, low energy expenditure, self-reported exhaustion, slow gait speed and weak grip strength)¹⁹. Gait speed was timed as participants walked 6 m, or 2.4 m for those who considered themselves slow walkers. Slow gait speed was defined as walking < 0.8 m/s. Unintentional weight loss was defined as the loss of at least one kg within the last three months. Those exercising three times per month or less were said to have low energy expenditure. Self-reported exhaustion was defined as reported fatigue within the last three months. Grip strength (Newtons) was measured in both hands with an electronic dynamometer (Grippit®), using the strongest value of the two. Weak grip strength was classified as the lowest 20% of participants, adjusted by sex and body mass index. The chair stand test was performed by asking participants to fold their arms across their chest and stand up from a seated position five times consecutively as quickly as possible, and the results were expressed in seconds. A test lasting > 17 s indicated poor performance. Dementia diagnosis was based on the DSM-IV criteria. Injurious falls were defined as any fall requiring medical attention. Data on injurious falls during the study period were obtained from diagnoses made at the patients’ hospital discharge and identified through the ICD-10 codes W00 to W19.

Statistical analysis

The characteristics of the study populations were described using counts and proportions, or medians and interquartile ranges, as appropriate. Group differences were investigated using the chi-squared test or the Mann–Whitney test, as appropriate. The associations between the PC-FI and the outcomes (i.e., all-cause mortality and hospitalization) were investigated using unadjusted and adjusted Cox regression models. Death and/or hospitalization (as appropriate), and end of the 5 years follow up were considered censoring events. Time-to-last follow up was further truncated at 1, 3 and 5 years. The proportional hazard assumption (tested according the Schoenfeld residuals test) was ascertained in all analyses. To investigate the discriminative ability of the PC-FI and single deficits in the prediction of mortality in HSD and SNAC-K, we calculated the c-statistic of unadjusted Cox regression models. The association between the four PC-FI categories and mortality was investigated using Cox regression models. Psychometric properties of the PC-FI (i.e., accuracy, sensitivity, specificity, positive predictive value, negative predictive value, positive likelihood ratio and negative likelihood ratio) have been obtained. In the external validation cohort (i.e., SNAC-K), the association of the PC-FI with physical frailty and other outcomes was examined using unadjusted and adjusted logistic regression models. The area under the curve from non-parametric ROC analyses was used to assess the discriminative ability of the PC-FI, based on the unadjusted logistic regression models.

Finally, the percentiles of PC-FI were calculated and graphically represented by age in the whole HSD, stratified by sex. In the same graphs, the estimated risk of death at 5 years for different combinations of PC-FI and chronological age was presented through heatmaps in nomograms. The probability of 3-year mortality was estimated using age and PC-FI as predictors in two second-order polynomial logistic regression models (one for men and one for women). A 5-year mortality predicted probability of 50% was used as a midpoint for the color scale.

Role of the funding source

The funders of this work did not play any role in all the phases of the research.