Rethinking normality in cardiopulmonary exercise testing: why percentile-based thresholds deserve a greater role

Extract

Clinical interpretation of any biological test aimed at assessing the presence of disease depends on careful judgement of the test results in light of the pre-test likelihood of abnormality. At first glance, this seems a relatively trivial task; in practice, however, there are tantalising complexities stemming from multiple sources, such as large test variability, technical issues, reference values that might not be appropriate for the specific population being examined, scarce ancillary information to guide the reader, and critically, how to safely separate “normal” from “abnormal”.

Shareable abstract

A study in the current issue of ERJ Open Research reinforces the need for modernised reference standards for CPET underscored on statistically sound distribution-based approaches https://bit.ly/42o9B9A

Clinical interpretation of any biological test aimed at assessing the presence of disease depends on careful judgement of the test results in light of the pre-test likelihood of abnormality. At first glance, this seems a relatively trivial task; in practice, however, there are tantalising complexities stemming from multiple sources, such as large test variability, technical issues, reference values that might not be appropriate for the specific population being examined, scarce ancillary information to guide the reader, and critically, how to safely separate “normal” from “abnormal”. The latter, in particular, has long been a matter of concern in the interpretation of pulmonary function tests (PFTs) and cardiopulmonary exercise testing (CPET).

In this context, most pulmonologists are familiar with the concept of % predicted, or even simpler, absolute thresholds to define the limits of normality (e.g. 80% for forced expiratory volume in 1s (FEV₁), 120% for total lung capacity and 0.7 for FEV₁/forced vital capacity (FVC)). The key problem with these simplistic approaches is that they fail to consider the entire distribution of values recorded in the reference population; % predicted, for example, is merely a metric of how the observed value differs from the “average” value [1]. Although it is true that a value, say, 40% predicted, is very likely to be abnormal (but not always, e.g. forced expiratory flow at 25–75% FVC in an elderly subject), what about 80%? The higher the % predicted, the more difficult it is to discriminate the subtle yet clinically relevant difference between normalcy and disease. A more statistically sound approach is to use percentiles: a given percentile represents the value below which a given percentage of scores fall. Put simply, percentiles split the data into 100 equal-sized groups and tell us where a score stands relative to other scores. For most purposes, the fifth centile is assumed as the lower limit of normal (or the 95th centile is the upper limit of normal); we tacitly assume that it is very unlikely (but still possible) that such a rare score in the normal population will come from a completely healthy subject (figure 1).

FIGURE 1.

Approaches to judge the normalcy of pulmonary function testing parameters based on a distribution-based framework. a) A variable (forced expiratory volume in 1s (FEV₁)) for which only abnormally low values are clinically relevant. b) A variable (total lung capacity (TLC)) for which both abnormally low and abnormally high values may have clinical implications. Similar statistical concepts are not yet widely used in cardiopulmonary exercise test interpretation, increasing the risk of false-positive and -negative results. See the main text for further elaboration. LLN: lower limit of normal; ULN: upper limit of normal; SEE: standard error of the estimate (derived from the respective regression equations). Reproduced and modified from [1] with permission from the publisher. Data from a) [2] and b) [3].

Translating these concepts to the reality of CPET, in this issue of ERJ Open Research, Staes et al. [4] report the results of a large retrospective analysis of 886 cycle ergometer incremental tests performed in Belgium by 294 individuals without known cardiorespiratory disease and 592 patients with known COPD, fibrotic interstitial lung disease, heart failure and pulmonary hypertension. The authors contrasted commonly used conventional thresholds, either % predicted or absolute cut-offs, with percentile-based thresholds (fifth centile) for peak oxygen uptake (V̇_O2peak), minute ventilation (V̇_E)/carbon dioxide output (V̇_CO2) nadir, peak heart rate, V̇_O2/work rate, and oxygen pulse. They took advantage of the fact that the German SHIP study provided those centiles [5], which is not common across CPET reference frames. They clearly demonstrated that percentile-based thresholds aligned closely with the expected 5% abnormal rate in controls (4–8% across variables), validating SHIP standard references as appropriate limits of normal for Western populations or, at least, the Northern European population. Thus, between-method agreement was high in controls (82% of CPETs concordant across all variables); nevertheless, conventional thresholds overestimated abnormality in ∼10% of the healthy group, most notably for V̇_O2peak and peak heart rate. More concerningly, however, agreement between methods fell to 46–58% across disease groups; in fact, many patients exhibited discordance across two or more variables, emphasising that threshold selection can potentially shift overall interpretation. Collectively, conventional thresholds tended to indicate greater abnormality than percentile-based thresholds; a notable exception was fewer abnormalities in the V̇_E/V̇_CO2 nadir with the former approach.

What are the potential clinical implications of these results? False-positive CPET results with the use of conventional thresholds may trigger additional costly, futile investigations, some with potential iatrogenic harm. Indeed, a falsely abnormal peak V̇_O2/work rate (<8.4 mL·min⁻¹·W⁻¹) and oxygen pulse (<80%) may raise concerns about cardiocirculatory disease, a typical CPET finding that is frequently taken more seriously by the requester. A sizeable fraction of dyspnoeic patients referred to CPET are anxious, depicting evidence of hyperventilation/dysfunctional breathing [6]; a falsely abnormal test result may further fuel this downward spiral. Ventilatory “efficiency”, estimated by the V̇_E/V̇_CO2 nadir, is an important CPET variable for risk stratification; a falsely preserved value (≤34) may preclude patients from receiving optimal care, including being exposed to risky procedures, such as thoracic surgery [7].

Despite the commendable effort of Staes et al. [4], some limitations should be acknowledged. As mentioned, they considered V̇_E/V̇_CO2 nadir >34 as abnormal by the conventional threshold approach; in fact, this is not universally considered the upper limit of normal, which is known to vary with age and sex [8], but rather a useful threshold indicating a higher risk of adverse outcomes in cardiopulmonary diseases [9, 10]. Assuming peak heart rate <90% to be “abnormal” is debatable; most patients limited “by the lungs” will present with relatively large chronotropic reserves, somewhat increasing the likelihood of a false-positive for the conventional threshold. V̇_O2peak <85% predicted is somewhat too high a target for the lower limit of normal, increasing the risk of a false-positive result. The SHIP dataset, although rigorously phenotyped, is relatively modest compared with other population-based initiatives and is derived largely from a Northern European white cohort, limiting generalisability. Additionally, SHIP percentile-based values have not been published for all clinically relevant CPET parameters; peak breathing reserve, for example, remains without validated percentile norms. For variables lacking percentile references (including peak oxygen pulse in individuals on β-blockers), conventional thresholds remain necessary.

CPET is increasingly used across respiratory and cardiac practice [11–13]. The European Respiratory Society has spearheaded commendable initiatives to establish international CPET reference standards, including the development of sounder limits of normality for the main CPET variables [14]. In addition to percentile-based thresholds, z-scores (i.e. how many standard deviations a specific data point is above or below the mean divided by the variability of the estimated mean) are undoubtedly more robust than % predicted alternatives. Percentiles are less sensitive to outliers compared to standard deviations and work well in skewed data (which might be more important for CPET than PFTs); however, they are less precise since large differences in raw scores at the extremes may look apparently small, creating some uncertainty on how to interpret values slightly above or below the fifth or the 95th centile. Novel paradigms for the display and interpretation of CPET data have recently been proposed to improve the accuracy of CPET in uncovering the root causes of exertional dyspnoea. Notably, they are largely underpinned by percentile-based categories of sensory and physiological impairment established by an artificial intelligence algorithm that considers all testing data points (figure 2) [15, 20]. Thus, the momentum for some relevant changes in how we view and interpret the test has gained some new traction over the past few years.

FIGURE 2.

Screenshot of an illustrative example of the advantages of continuous assessment of sensory and mechanical–ventilatory responses throughout the whole cardiopulmonary exercise test (CPET) using percentiles as frames of reference; freely available artificial intelligence (AI)-based software analysed this data (Dynamic Assessment of Dyspnea and Ventilation During Exercise). A 73-year-old male with COPD was referred to CPET due to a previous test indicating the lack of “respiratory limitation” based on normal peak breathing reserve despite the severity of airflow obstruction. b and f) show a high symptom burden, whereas d) confirms a preserved peak breathing reserve (>15%). However, there was clear evidence of “very severe” excessive breathing in the same panel (abnormally low dynamic, submaximal ventilatory reserve) and c and e) “very severe” constrained breathing, i.e. unusually low dynamic, submaximal inspiratory reserves. a) Note an upward inflection on dyspnoea–ventilation when critically high inspiratory constraints were reached earlier on exercise, indicating that the patient was no longer able to translate a high neural drive to breathe to ventilation. These results indicate that relying only on peak breathing reserve was highly misleading; this patient's dyspnoea is indeed ascribable to “the lungs”. The relative intensity of the physiological and sensory abnormalities was quantified by an algorithm (AI-Techniques Loss Function-Based Algorithm for Severity Classification) [15] that identifies the interquartile interval that best describes the observed “trajectory” of the variables as exercise intensity increases. c and e) Commonly used cut-offs to indicate critically low inspiratory reserves [16] are shown.

$$\mathrm{V}{\mathrm{R}}_{\mathrm{dyn}}=\left(1-{\displaystyle \frac{\mathrm{ventilation}\mathrm{at}\mathrm{a}\mathrm{given}\mathrm{exercise}\mathrm{intensity}}{\mathrm{estimated}\mathrm{maximal}\mathrm{voluntary}\mathrm{ventilation}}}\right)\times 100$$

where the estimated maximal voluntary ventilation is calculated as the forced expiratory volume in 1 s (FEV₁)×40.

$$\mathrm{I}{\mathrm{R}}_{\mathrm{dyn}1}=\left(1-{\displaystyle \frac{\mathrm{tidal}\mathrm{volume}}{\mathrm{dynamic}\mathrm{inspiratory}\mathrm{capacity}}}\right)\times 100$$

$$\mathrm{I}{\mathrm{R}}_{\mathrm{dyn}2}=\left(1-{\displaystyle \frac{\mathrm{end}-\mathrm{expiratory}\mathrm{lung}\mathrm{volume}}{\mathrm{TLC}}}\right)\times 100$$

where TLC is the total lung capacity [17, 18]. b and d) Sex- and age-corrected ranges of dyspnoea severity as established in [19]. Ranges of severity based on quartiles; upper limit of normal for symptoms >95th centile; lower limit of normal for mechanical–ventilatory reserves >5th centiles. FVC: forced vital capacity; D_LCO: diffusing capacity of the lung for carbon monoxide. Reproduced and modified from [20] with permission from the publisher.

In keeping with these new perspectives, Staes et al. [4] provide compelling evidence that our current approach to defining abnormality in CPET interpretation requires refinement. Percentile-based thresholds, grounded in population variability, offer a more accurate and statistically coherent foundation than widely used conventional cut-offs. By showing the magnitude and clinical relevance of discrepancies between these methods, the study reinforces the need for modernised reference standards and broader adoption of distribution-based approaches. Specifically, the study underscores that conventional thresholds, many of which are derived from expert opinion or historical convention, are far from ideal and should be applied cautiously when more robust alternatives exist. As CPET continues to expand in scope and clinical value, updating our interpretive framework is not merely beneficial; it is, rather, essential.