Frailty trajectories to identify end of life: a longitudinal population-based study

An increasing number of older people are now dying with a diagnosis of frailty. In high-income countries, the estimated prevalence of frailty is 11% for people aged over 65 years, rising to 25–50% for people over 85 [1]. Frailty is characterised by an accumulation of deficits and reduced strength, endurance and physiological function [2, 3]. It is associated with a range of adverse outcomes, including falls, delirium, disability and mortality [4‐8].

Timely recognition of the end-of-life phase is fundamental to the provision of palliative care, as it allows clinicians to introduce generalist or specialist services, to discuss preferences and make advance plans [9‐11]. This may be particularly challenging for patients with frailty, where trajectories of decline are gradual and slow, and patients may not have a recognised life-limiting diagnosis [12, 13]. Despite calls for greater awareness of the benefits of palliative care for people with frailty [14], recognition remains incomplete [15].

Since 2017, general practitioners (GPs) in England have been required to identify and review patients with moderate and severe frailty using an appropriate tool. One of the instruments recommended by NHS England is the electronic frailty index (eFI) [16, 17]. The eFI is a cumulative deficit measure of frailty that uses over 2000 Read codes to calculate a frailty score based on a variety of symptoms, diagnoses and observations recorded in primary care electronic health records [5]. It is available in GP software systems in the UK, and can be calculated automatically from patients’ records at any time. Previous research has shown that frailty measured by the eFI at a single time point can predict mortality at a population level [17], but it is not effective at identifying individuals at short-term risk of all-cause mortality before this would be clinically apparent [18]. Frailty is associated with increasing age, but there is also wide inter-individual variation in how deficits are accumulated over time, and studies have demonstrated different longitudinal trajectories of frailty [19, 20]. However, to date, few studies have examined the association between longitudinal change in frailty and mortality. Where studies have been carried out, they have produced population level results with limited clinical application. For example, recent analysis of data from a US cohort identified distinct frailty trajectories and their associated mortality risk, using scores calculated at annual intervals over 8 years [21].

In this study, we used eFI measured at monthly intervals to examine longitudinal change in frailty in the year before death. The aim was to determine if changes in frailty measured by the eFI could be useful in primary care to indicate increased risk of dying and the need to consider palliative care.

This study analysed information on 26,298 individuals (13,149 cases and 13,149 controls), 14,620 (55.6%) of whom were female and 11,678 (44.4%) were male. The mean age at death for cases was 85.14 (SD 5.98) years with a pseudo age of 85.66 (SD 5.98) years for controls, as expected by design. Age-adjusted mean eFI at baseline was significantly higher (p < 0.001) for cases (eFI = 0.27, 9.8 deficits) than for controls (eFI = 0.23, 8.3 deficits), as expected from previous work [17, 18], and mean eFI for females (eFI = 0.23, 8.3 deficits) was significantly higher (p < 0.001) at baseline than for males (eFI = 0.22, 8.0 deficits) when adjusted for age and case/control status. Table 1 shows the demographic characteristics of the study sample.

The latent growth curve model that best described the shape of frailty change over the study period included freely estimated intercepts, slopes and a quadratic polynomial term for time (see Additional file 1: Table S1 for parameter descriptions and model fit statistics). The estimated mean trajectory of eFI increased by 0.002 (95% CI 0.001–0.002, 0.1 deficits) per month (at study baseline) from a mean of 0.252 (95% CI 0.251–0.253, 9.1 deficits). The quadratic term was small (eFI = 0.00003, 95% CI 0.00003–0.00004) but improved the model fit and was retained to allow flexibility in class structures during the mixture modelling phase. Large variance in the intercept (eFI = 0.128) and slope (eFI = 0.278) was observed, suggesting individual heterogeneity in trajectories. There was some evidence that higher baseline frailty was related to slower frailty increase over time (intercept slope covariance − 0.109, p < 0.006).

The quadratic model was extended to investigate the impact of age and sex on the trajectories (Table 2 and Additional file 1: Table S2 for model fit and parameter descriptions). In the full model (Additional file 1: Table S2), eFI at study baseline increased by 0.0047 (95% CI 0.0044–0.0049, 0.2 deficits; p < 0.001) for each year of age, and men had a mean eFI of 0.0016 (95% CI 0.00133–0.00187, 0.6 deficits; p < 0.001) lower than women. Because there was little evidence that sex or age influenced the trajectory of frailty over time (for each year of age, eFI increased by 0.00002 per month, 95% CI 0.00001–0.00003, < 0.001 deficits, p = 0.003, all other coefficients p > 0.05), the model regressing the intercept only on age and sex, including slopes and quadratic terms (Table 2), was carried forward to investigate potential latent class trajectories.

A three-class latent growth mixture model model (Table 2) was chosen as the most parsimonious with good fit and clinically relevant classes (Additional file 1: Table S3). The most common ‘stable’ class contained 76.6% of the sample and showed little change in frailty over time, with eFI increasing by 0.001 per month (0.04 deficits, ‘stable’ class) from a baseline eFI of 0.260 (9.4 deficits). A ‘moderate growth’ class containing 21.2% of the sample demonstrated an eFI increase of 0.0045 (0.2 deficits) per month from a slightly higher baseline eFI of 0.262 (9.4 deficits), and a small ‘rapidly rising’ class containing 2.2% of the sample showed a sharp, distinct increase in eFI of 0.01 per month (0.4 deficits) on average over the year from a lower baseline eFI of 0.206 (7.4 deficits). The majority of individuals in this class accrued deficits quickly, with a 0.02 eFI increase per month over the first 3 months (s = 0.023, 95% CI 0.021–0.025), but the quadratic component (q = − 0.001, 95% CI – 0.001 to – 0.001) indicates that they then experienced a period of relative stability; this group had an average increase in eFI over 1 year 14 times higher than the stable class. Age and sex were associated with baseline frailty (eFI = 0.005/0.2 deficits per year of age, eFI = 0.016/0.6 deficits higher for women) across classes, but were only weakly related to class membership (all ORs approximately 1.0).

Compared with the ‘stable’ class both the ‘rapidly rising’ class (OR 2.84, 95% CI 2.34–3.45) and the ‘moderate growth’ class (OR 1.65, 95% CI 1.54–1.76) were found to be associated with mortality. Figure 1 displays the mean and observed individual trajectories for each class and the association between class and mortality risk.

To understand the population characteristics, the posterior probabilities for being in each class were reweighted to reflect the general population. Compared with the ‘stable class’, both the ‘rapidly rising’ class (OR 3.02, 95% CI 2.49–3.65) and the ‘moderate growth’ class (OR 1.58, 95% CI 1.49–1.67) were associated with an increased chance of mortality. Table 3 contains demographic information for the members of each class. Reweighting the controls to the population suggests that 1.1% of the population exhibited rapidly rising frailty; using this group to predict mortality within the year was highly specific (99.1% specificity), with 19.8% positive predictive value and 93.3% negative predictive value, but with low sensitivity (3.2%). Addition of the ‘moderately increasing’ class increased the sensitivity (27.8%) and negative predictive value (93.9%) at the expense of a decreased specificity (82.3%) and positive predictive value (10.4%).

We have identified a distinct frailty trajectory that is associated with an increased risk of mortality amongst older adults. Individuals with rapidly rising frailty (average monthly rise in eFI of 0.01 or accrual of 0.4 deficits) are twice as likely to die within 12 months as individuals with stable frailty. Individuals in this class experience an initial monthly increase in eFI of 0.02, prior to slowing around 8–9 months prior to death. This rapidly rising class represents a small but important proportion of the population who may not be recognised as being at high risk of death, as they started from a lower mean baseline eFI than the rest of the population [17]. The ‘moderately increasing frailty’ class are also at a greater risk of mortality, with an average rise of 0.005 in eFI per month associated with a 58% increase in mortality.

Our study demonstrated that, for a small proportion of the population, significant changes in frailty occur over 1 year. Individuals with lower levels of frailty who experience a rapid rise have a greater risk of death over 1 year than individuals with higher, but stable frailty.

The association between severity of frailty measured at a single time point and risk of mortality has been robustly demonstrated at the population level [23]. However, our previous work found that this association did not translate into a useful measure for predicting individual risk of mortality when frailty is measured at a single time point, even close to death [18]. Building on previous work that has examined the relationship between longitudinal change in frailty and mortality at yearly intervals [21, 24], we have been able to measure frailty changes at monthly intervals over the final year of life. Our findings suggest that some individuals are resilient to higher but stable levels of frailty, and that changes in frailty status may be a more useful indicator of mortality risk. The positive predictive value and sensitivity for the rapidly rising trajectory mean that recognition of this trajectory alone may not be a reliable predictor of imminent death. However, it may be useful to support clinical judgement and as a risk stratification tool, especially as many individuals who follow this trajectory may not be considered as ‘at risk’ because of their low levels of baseline frailty. The Q Mortality risk prediction algorithm, which incorporates a frailty index and was designed specifically to assess mortality risk, has better cross-sectional prediction of mortality than the eFI. However, the frailty index implemented in Q mortality does not currently include a weighting for risk based on changes in frailty scores over time [25].

This longitudinal population-based study demonstrates that it is possible to use a frailty index calculated within electronic healthcare records to identify people who are at a higher risk of dying within 1 year. This has potential application in health services to support clinicians in identifying older adults dying with frailty who may have been overlooked by traditional approaches and to help ensure appropriate care is offered. The rapid initial rise in eFI from a low baseline, followed by a plateau, may be particularly useful in primary care, promoting recognition of people at risk of death, but also allowing time for intervention.