The predictive capacity of the HAT was high across all external validation cohorts as well as in the harmonized dataset and was similar in magnitude and direction to that observed in the original development dataset (i.e., SNAC-K), both for mortality as well as unplanned hospital admissions. Even if the predictive capacity was affected by sample size, the method to construct HAT scores was replicable in other settings. Moreover, the performance of the HAT was high in the harmonized dataset, which integrated populations of different age, educational levels, geographic locations, and urbanization levels, similar to the Swedish older population. Furthermore, the HAT-based geriatric health charts are visualization tools that may enable clinicians to view, understand, and communicate HAT scores and their evolution to their patients. The prognosis of different outcomes can be displayed in the charts to guide clinical decisions and treatment plans of patients.

Few geriatric health assessment instruments have been externally validated, and validation methods vary across studies [23]. Moreover, instruments have usually been developed in inpatient and specialist outpatient clinical environments, using short-term outcomes such as length of hospital stay, 30-day hospital readmissions, or 1-year mortality [24]. Although the HAT was originally aimed at assessing the general health status of older individuals, its components represent key functional domains that overlap with the concept of frailty.

Several tools have been developed to detect frailty, based either on objective or patient-reported indicators, but no consensus exists regarding the tools with best predictive performance [25]. Starting with one of the most well studied instruments, the Frailty Index (FI) by Rockwood and Mitnitski [26], it showed sufficiently good predictive performance in terms of hospitalization and mortality in community-dwelling adults according to the umbrella review by Apóstolo et al. [27], which summarized the existing evidence on the performance of several frailty indices. In a comparative study by Hogan et al. [28], the FI seemed to have modest predictive ability for mortality and hospitalization, with AUC ranging between 0.65 and 0.73 for death and 0.58 and 0.64 for hospitalization in multivariate analyses adjusting for age, sex, and number of chronic conditions. However, according to the findings from the WHO Study on global AGEing and adult health (SAGE) in Shanghai, the AUC of FI for 4-year mortality was over 0.75, but the AUC for 4-year hospitalization was low (AUC 0.53–0.57) [29]. In a study by Vetrano et al. [30], a data-driven Primary Care Frailty Index (PC-FI) including 25 health deficits was developed, which showed a predictive capacity comparable but slightly lower to that of the HAT (c-statistic range 0.74–0.84 for mortality and 0.59–0.69 for hospitalization).

Other instruments have also been proposed and tested. The Edmonton Frailty Scale (EFS) has shown good validity and reliability [31], but its use has been limited to the inpatient environment and for disease-specific populations. In a systematic review assessing the construct validity of frailty instruments [32], the EFS and the Clinical Frailty Scale (CFS) were recommended for routine clinical use because of their short administration time and good construct validity. However, no data regarding their predictive ability were given. In a recent multicenter prospective cohort study [33], the authors compared the predictive capacity of the CFS and the Hospital Frailty Risk Score (HFRS) in critically ill patients, and the AUC values for 1-year mortality were 0.66 and 0.63, respectively, and 0.70 for both tools together.

In summary, the predictive capacity of commonly used frailty assessment instruments, whether in the research or clinical settings, is inferior or at best equal to that of the HAT. Unlike in our study, many of the cited papers adjusted for age, sex, and/or other variables, which falsely increases the predictive capacity of the studied tools.

The assessment of health status and risk of frailty over time using the HAT could enable clinicians to intervene at the right time and for the right person to slow down the further deterioration of functioning, maintaining capacity or even reversing a declining health trajectory. This can be projected in all four levels of prevention. Older healthy individuals with high HAT scores could get primary preventative advice and be monitored for risk factors for frailty, both at the primary care level as well as through higher level public-health policy. Ki et al. [34] developed a framework for preventing frailty that comprised different domains among which are physical activity, resilience, and management of chronic diseases. In terms of secondary prevention, patients with prefrailty or with reversible frailty could be identified. In recent systematic and scoping reviews, it has been shown that complex primary care interventions, including physical and nutritional counseling and comprehensive geriatric assessment, may effectively reverse frailty and postpone the transition to frailty in prefrail individuals [4, 5]. Unfortunately, though, frailty can also be irreversible. This group of patients with irreversible frailty can get help to reduce the incidence of complications such as disability and dementia and to maintain the best possible quality of life [5], along the spectrum of tertiary and quaternary prevention. Consequently, a geriatric health assessment tool, such as the HAT, which can accurately predict short- and long-term risk of negative outcomes may be of great help for primary care units to better target the level and intensity of prevention and care to their older patients.

Many diagnostic, preventive, and/or therapeutic decisions taken in primary care are done based on patients’ chronological age. In lieu, the HAT has potential to support clinical decision-making based on biological age. For example, screening for colorectal cancer is not recommended for individuals over the age of 75 [35], and PSA testing for cancer is currently not motivated for men over the age of 70 [36]. Primary prevention medications such as statin treatment for modification of the risk for coronary artery disease or other vascular diseases have been questioned after the age of 75 in recent studies [37]. However, such decisions should be individualized, rather than age restricted, based on accurate geriatric health assessments as claimed for in recent initiatives against ageism. Authors of a recent systematic review [38] showed that ageism led to significantly worse health outcomes in 95.5% of the studies included in the review. Strategies to reduce ageistic approaches to treating older patients are warranted.

The HAT was externally validated in Swedish aging cohorts representing different socio-demographic and urbanization levels. The results of our study have been additionally interpreted after age- and sex-stratified analyses and considering both short- and long-term outcomes. Beyond the individual cohorts, using a harmonized dataset increased the sample size enabling an optimal performance of the statistical models and recalibration of the HAT. The visualization tool proposed may strengthen the clinical applicability of the HAT.

Nevertheless, some limitations need to be acknowledged. Within the SNAC-B cohort, there were two major discrepancies in the measurement of the individual health indicators. First, gait speed was not measured at baseline, which required imputing the values, and second, it was not possible to differentiate between unplanned and planned hospital admissions. However, to alleviate the impact of such limitations, the IPD-MA was done with and without SNAC-B. The results showed that the statistical measures taken did not have any negative effects on the overall results, and that SNAC-B performed similarly as expected based on sample size and distribution. Inconsistencies due to measurement heterogeneity across cohorts were documented. However, such discrepancies reflect the real-world clinical environment and are expected to have minimal impact following standardization and harmonization.

When stratifying by age, the predictive capacity of the HAT decreased, more so among the younger group. Higher functional resilience among the younger-old group could limit the ability to capture poor outcomes by the different components of the HAT [16]. This is evidenced by differences in the distribution of outcome events among the age groups. Complementarily, although the TIF showed high performance across the whole spectrum of the latent health construct, the information density was lower for levels of poorer health. Given that the manifestation of poor health differs among the older and younger groups [14], this might affect the performance of the HAT and could explain the lower predictive capacity when stratifying by age. Additional research is warranted to understand why these differences are observed and their impact on the tools’ predictive capacity.

Lastly, even if the external validity of the HAT was overall very good across different Swedish aging cohorts, it still needs to be tested further in routine clinical (and ideally primary care) practice, both in terms of its predictive capacity as well as its feasibility and acceptability among healthcare professionals and patients.