Frailty: More than the sum of its parts?

The state of vulnerability to medical stressors has been codified as “frailty.” It is common, prognostically meaningful, and has been graced with a proportionate explosion in scientific articles exploring its implications for adverse outcomes – including a themed issue of Anesthesia & Analgesia (June, 2020). That frailty predicts adverse outcomes is probably no longer in doubt. But – what is frailty? The answer, it turns out, sometimes feels like a solid “It’s complicated…”.

A quick recap: Frailty is defined as a biological vulnerability to stressors – a deficit of homeostatic capability, perhaps. There is no biomarker, there is no single test, and I suspect that a number of us feel that “I know it when I see it.” And we might: clinician assessment of frailty may¹ (or may not²) be reasonably accurate compared to more involved methods. But, does frailty reflect more than just comorbidity burden – does it tell us more about a patient’s risk for adverse outcomes than their diabetes, stroke history, long medication list, cognitive impairment, low albumin, and history of falls? Is frailty more than the sum of its ICD-10 “present-on-admission” diagnosis codes?

It would be nice to end here with “yes.” But... it’s complicated. Statistically, that frailty is an independent risk factor for adverse outcomes has been shown, but within the typical limitations of associational research: there may (there always, always may) be important factors which were not included in the model, or unanticipated and therefore unmodeled interactions among factors already present. In this case, that would mean an unknown and unmodeled factor that “explains” all of the excess morbidity and mortality which was mathematically attributed to frailty in the theoretical incomplete model.

Since frailty has no pathognomonic biomarker, imaging study, or tissue diagnosis, the gold standard is to use one of two main strategies – conceptual models, essentially – for detecting it. One is the “phenotype” model where deficits in muscle strength, nutrition, exercise tolerance, etc., make the diagnosis: for example, low grip strength, unintentional weight loss, a slow walking speed, and self-reported exhaustion.³ The other is the “deficit accumulation” model – and here it gets awkward – where something about the sheer number of, the sheer burden of comorbid medical conditions, functional limitations, medications, etc. yields a “diagnosis” of frailty.⁴ So, by virtue of these two approaches to “diagnosis,” the answer to the question becomes less clear: is frailty weak muscles or having a bunch of things wrong with you? And if, in today’s electronic practice environment, one wishes to use a list of factors collected in an electronic medical record to “diagnose” frailty as a prognostic factor, is that really a diagnosis or just an intermediate step in a predictive model for adverse outcomes, predicated on the general observation that people with a bunch of things wrong tend, over time, to get sicker?

This isn’t merely an intellectual question for interpretation of the work by Alkadri and colleagues, published in this issue of A&A.⁵ They conducted a meta-analysis of the prognostic power of a diagnosis of frailty, determined mostly via different iterations of the deficit accumulation model (“frailty index”). Some level of gut discomfort with the “prognostic factor / predictive model” might never need to be resolved for someone who uses frailty clinically – after all, whether you call it a syndrome itself or just the accumulated effects of being old and sick does not matter directly to your patient. But, the statistical standards for meta-analysis of a prognostic factor are different from those of a predictive model. There is only one right way to do it, and the right way rests on whether frailty is a prognostic factor itself, or the output of a predictive model (Supplemental Figure 1).

The statistical underpinnings of prognostic factor versus predictive model meta-analysis are beyond the scope of this editorial. But they do force a deep and thoughtful examination of frailty: a quality for which we have not identified a single gold standard, and instead have generated a palette of statistical models and diagnostic strategies of varying complexity – and distressingly little concrete overlap – in an attempt to “see” it.

Authors of the meta-analysis clearly are aware of this discomfort, having contributed to the literature in this area already. In a previously published study, they applied the Clinical Frailty Scale (similar to clinician gestalt, though formally anchored), the Fried Phenotype (a “phenotype” diagnostic method), and the Frailty Index (a “deficit accumulation” diagnostic method) to a cohort of almost 650 older patients undergoing surgery, and found frailty determed by all methods helped predict adverse postoperative outcomes compared to adjustment by health factors alone.⁶ The conclusion, that there was one method (the Clinical Frailty Scale) which was more tightly associated with adverse outcomes compared to the others, is clinically useful but – in a paper intending to compare performance of 3 instruments purportedly measuring the same thing – not surprising. What was potentially surprising is that, while all instruments were useful at predicting adverse outcomes on average, frailty was diagnosed in different individuals depending on which methodology was used (Figure 1).

Figure 1.

Overlap of three different frailty instruments applied to the same population of 645 older adults undergoing elective noncardiac surgery. Shading is roughly proportional to the number of patients classified as frail by each instrument or instruments. (Created using BioVenn;¹⁵ data provided by D. McIsaac.)

It is as though three different troponin assays returned different results for “Is this person having a myocardial infarction?” and yet all three were still helpful, on average, in a group of patients with heart disease. If there is something intrinsic and biological about frailty, wouldn’t we prefer that the methods we use to detect it have perfect, or near perfect, overlap? Are the three instruments purporting to detect frailty – by measuring a clinician’s overall assessment, a phenotype, or the number of things wrong with someone – in fact predictive of outcomes not because they are intrinsically labeling patients as “frail” but because they are built on things which themselves are associated with, or cause, adverse outcomes?

In their present work, Alkadri and colleagues meta-analyzed predictive ability of not 3 but 22 different frailty instruments, of which 84% were “index”-style, in over 3 million perioperative patients. The scale and scope of relevant work in this area is staggering, particularly because all studies were published in 2013 or later. Of the 22 instruments, the number of domains assessed (e.g., functional; cognitive; comorbidities) ranged from 1 to 6, and for index-style assessments, the number of individual deficits which could be counted ranged from 5 to 40. And while odds ratios varied from study to study, and instrument to instrument, the odds of mortality and non-home discharge were about 3.5 for those identified as frail – a level at which use as an individual prognostic factor becomes feasible.⁷

But how is frailty more than the sum of its parts? Or, how do we fit frailty into the prognostic factor versus prognostic model rubric, when we call frailty both a prognostic factor itself and derive it from a number of other prognostic factors (e.g., cardiovascular disease; recent falls; low albumin) using a theoretical or conceptual model? The PROGnosis RESearch Strategy (PROGRESS) framework⁸ was developed to help answer challenging questions, like that one, in prognosis research. In particular, PROGRESS distinguishes between prognostic factor research – the study of “specific factors (such as biomarkers), that are associated with prognosis” – and prognostic model research – “the development, validation, and impact of statistical models that predict individual risk of a future outcome.” One might think that differentiating between these would be easy. For frailty? It’s complicated. Many of the ways we assess frailty are, indeed, model-like; this is particularly apparent for “index”-style frailty instruments, where elements in the frailty diagnosis are added up with equal weight, whether a comorbidity, a subjective impression of fatigue, or a medication count. But frailty also works outside of its “index” clothing: for example, when low grip strength and a slow walking speed differentiate older adults at risk of adverse outcomes from those who are more robust. The frailty “phenotype” is therefore conceptualized as a deficit in reserve, rather than the explicit and identifiable consequence of any conventional comorbid disease state. Maybe the most “factor-like” frailty assessment is the Clinical Frailty Scale: interact with your patient in a regular health setting and the clinician can identify, using an anchored scale, whether they are frail or not. How could that, itself, be a model for perioperative death? And yet, it is.⁹

In addition to being dependent on the classification of frailty as a factor, not a model, the manuscript by Alkadri and colleagues also provides some important independent support for frailty’s role as a factor. The PROGRESS framework also explicitly identifies some things we may ask of a factor:¹⁰ Is there repeated confirmation that the factor is prognostic? (Yes, summarized effectively in the present work by Alkadri and colleagues.⁵) Does the factor retain prognostic value even after adjustment for other prognostic factors? (Yes; note that the summary odds ratios were remarkably stable, around 3.5, whether covariate-adjusted or not⁵). Is there evidence of how the factor fits on the causal pathway from disease to outcome, and an understanding of the biological mechanism involved? (Yes – although this is beyond the scope of this editorial, it has been recently reviewed¹¹.)

The work by Alkadri and colleagues represents a meaningful theoretical step forward in the understanding of perioperative frailty as a meaningful prognostic factor: while there are at least 22 reasonable ways to evaluate frailty using electronic data, overall – whichever instrument is used – you are likely to gain more information about a patient’s risk for adverse outcomes than without a frailty assessment. While this remains a major focus in frailty research, there are no broadly accepted methods for treating or reversing frailty, and there is no guidance about canceling or postponing a needed procedure solely on the basis of frailty. Thus, there are few, if any, theoretical concerns about harm coming to a patient at risk for adverse outcomes on the basis of a known frailty status. Frailty stratification, in contrast, places an onus on perioperative physicians of all stripes to look twice, to think twice, and to plan pre-, intra-, and post-operative care around what we might be able to prevent or better counsel patients and their families about: delirium (which was not an element of this meta-analysis, but should be considered¹²), prolonged hospital length of stay, adverse discharge, and death. For any patient identified as frail, identifying specific domains which are most limited – e.g., nutrition, physical performance, etc. – may help further refine both the prognosis¹³ and identify targeted opportunities for preoperative optimization.¹⁴

Simply knowing that there exist tools which, taken as a body of evidence, can glean frailty from electronic data is a meaningful advance toward applying insights about this somewhat mysterious, but important, prognostic factor to clinical care. Yes, it is much more complicated than troponin, and that uncertainty feels uncomfortable; we are accustomed to hard numbers, hard facts, and hard diagnoses. However, that does not obviate our responsibility, as a specialty, to work toward understanding the complexities underlying frailty in order to provide the best possible perioperative care to some of our most vulnerable patients.

Supplementary Material

Supplemental Data File (.doc, .tif, .pdf, etc., Published Online Only)

Supplemental Figure 1. Explanatory schematics elaborating on whether frailty is an intermediate step in a predictive model, or intrinsically a prognostic factor.

Glossary of Terms:

PROGRESS

PROGnosis RESearch Strategy framework