NYHA Functional Classification

NYHA Functional Classification in Heart Failure Models and Appraisals

The New York Heart Association scheme places a person with heart failure in one of four ordered classes according to how much everyday physical activity brings on symptoms, judged by the treating clinician at a consultation. Trials and NICE technology appraisals such as TA388 use the class to define who is eligible for treatment. Decision models use the class to define health states or to weight time spent alive, and utility studies estimate a separate EQ-5D value for each class. Göhler and colleagues reported regression-adjusted EQ-5D utilities of 0.90 in class 1 down to 0.60 in class 4 in a trial of patients with heart failure after myocardial infarction, so the share of a cohort assigned to each class drives the QALYs a model produces. The weak point is the rating itself, since Raphael and colleagues found that two cardiologists assessing the same 50 patients agreed on the class for only 54 per cent of them.

The four classes and who assigns them

The classification was set out by the Criteria Committee of the New York Heart Association, whose ninth and last edition of the Nomenclature and Criteria for Diagnosis of Diseases of the Heart and Great Vessels appeared in 1994. Clinical guidelines reproduce the classes in short form. Table 4 of the 2013 ACCF/AHA heart failure guideline, led by Yancy, compares the ACCF/AHA stages with the four NYHA classes, quoted in full for classes 1 and 4 and in shortened form for classes 2 and 3.

There is no questionnaire, item score or recall period. The clinician forms a judgement from the history, which makes the scheme a clinician-reported outcome rather than a patient-reported outcome. The 2013 guideline describes the class as "a subjective assessment by a clinician" that "can change frequently over short periods of time", and contrasts the class with the ACCF/AHA stages A to D, which emphasise the development and progression of disease while the NYHA classes focus on exercise capacity and the symptomatic status of the disease.

The four classes are ordered categories with no defined distance between them, which places the classification on an ordinal scale. Class 3 is worse than class 2, but nothing in the scheme says the gap between classes 2 and 3 equals the gap between classes 3 and 4. Any numeric value attached to a class, such as a utility or a cost, has to come from separate data.

How heart failure models put NYHA classes to work

Heart failure models use the classification as a set of health states, as a weight on time alive, or only to define the population. The choice affects which data a model needs and how a treatment effect on symptoms turns into QALYs.

Each class as a Markov health state

The most direct design makes every class a state in a Markov model, with death as the absorbing state. Wu and colleagues built a model of this kind for sacubitril valsartan against enalapril in China, with five states (NYHA classes 1 to 4 and death), a one-month cycle and a 10-year horizon. A treatment acts through the transition probabilities between classes and towards death, and each state carries its own utility and cost. Göhler and colleagues concluded that NYHA class and the number of cardiovascular rehospitalisations are established proxies for disease progression that can be linked to utilities when used as health states, and they recommended NYHA class where feasible.

Classes inside a single alive state

Some appraisals keep the structure to two states and use NYHA class only to weight time alive. In the ivabradine appraisal (NICE TA267), the manufacturer's Markov cohort model had two states, alive and dead, with monthly cycles and a lifetime horizon, and quality of life differed by NYHA class within the alive state. Utilities without a hospital admission lay between 0.82 in class 1 and 0.46 in class 4, and an admission reduced utility by 0.07, 0.03, 0.08 and 0.21 in classes 1 to 4 respectively. The values came from EQ-5D responses in the SHIFT-PRO substudy, scored with UK population tariff values and analysed with a mixed regression model.

Class as an eligibility rule and supporting evidence

The sacubitril valsartan appraisal (NICE TA388) used a two-state model with alive and dead as the only health states, so the class did not define states at all. Utilities came from a linear mixed regression on PARADIGM-HF EQ-5D data, with predicted scores depending on treatment, baseline characteristics, and the risk of hospitalisation and adverse events. The company justified its assumed utility benefit by noting that more people on sacubitril valsartan than on enalapril in PARADIGM-HF had symptomatic benefit, shown by improvement in KCCQ scores and NYHA class, and the evidence review group questioned the evidence for NYHA class 4, where the sample was small (n=60). NICE recommended the drug in TA388 only for people with NYHA class 2 to 4 symptoms, a left ventricular ejection fraction of 35 per cent or less, and a stable dose of an ACE inhibitor or angiotensin II receptor blocker. NICE TA314 on implantable cardioverter defibrillators and cardiac resynchronisation therapy sets out its recommendations as a table by NYHA class and QRS duration, so the device recommended depends on the class recorded.

Attaching utilities to NYHA classes

A class has no preference weight of its own, so a health state utility value has to be estimated for each class, usually from EQ-5D data collected in a trial and then entered into a cost-utility analysis. Göhler and colleagues used EQ-5D data from the EPHESUS trial, which enrolled patients with heart failure after myocardial infarction, and a multivariate regression adjusted to a 60-year-old woman with no further comorbidities.

Both trends were statistically significant (P < 0.001), and the spread across NYHA classes was wider than the spread across rehospitalisation counts. The TA267 values from SHIFT-PRO run lower at the severe end, 0.46 in class 4, and come from another trial and another regression, so the two sets cannot be swapped without changing results. A model's QALY estimates therefore depend on which utility set is attached to the classes as well as on how many people sit in each class.

Utilities fixed by class can also drift away from reality over a long horizon. In the TA267 manufacturer's model the quality of life for each NYHA class was assumed to stay the same after the trial period and was not based on an ageing population. The evidence review group's critique, summarised in section 3 of the final guidance, noted that this made utilities in later cycles higher than they should be and was assumed to have favoured ivabradine.

Illustrative worked example of a shift in NYHA class distribution

All figures in this example are illustrative. The class utilities are borrowed from the Göhler adjusted estimates for convenience only, and the class distributions and costs are invented. The example covers one year with no deaths and no discounting, so the mean utility over the year equals the QALYs per person.

$$\bar{u} = \sum_{k} p_k u_k$$

where $\bar{u}$ is the mean utility of the cohort, $p_k$ is the share of the cohort in NYHA class $k$ and $u_k$ is the utility attached to class $k$.

1. Standard care. Shares are 40 per cent in class 2, 50 per cent in class 3 and 10 per cent in class 4. The products are 0.40 × 0.83 = 0.332, 0.50 × 0.74 = 0.37 and 0.10 × 0.60 = 0.06, so the mean utility is 0.332 + 0.37 + 0.06 = 0.762.

2. New treatment. Shares move to 55 per cent in class 2, 40 per cent in class 3 and 5 per cent in class 4. The products are 0.55 × 0.83 = 0.4565, 0.40 × 0.74 = 0.296 and 0.05 × 0.60 = 0.03, giving 0.4565 + 0.296 + 0.03 = 0.7825.

3. QALY gain and cost per QALY. The gain is 0.7825 − 0.762 = 0.0205 QALYs per person. With an illustrative extra cost of GBP 410 per person, the ICER is 410 / 0.0205 = 20,000, or GBP 20,000 per QALY.

4. One-class misclassification in the treated arm. In this step unblinded assessors rate 5 per cent of treated patients one class better than a blinded assessor would, so these patients are recorded in class 2 instead of class 3. The utility difference between the two classes is 0.83 − 0.74 = 0.09, and the extra apparent gain is 0.05 × 0.09 = 0.0045. The QALY gain becomes 0.0205 + 0.0045 = 0.025 and the ICER falls to 410 / 0.025 = 16,400, or GBP 16,400 per QALY.

The whole benefit in this example comes from moving people between classes, so a small rating bias of one class in a twentieth of the treated arm cuts the cost per QALY by 18 per cent. That sensitivity follows directly from the reliability problem described below, and a sensitivity analysis on the class distribution is the natural way to expose that weakness.

NYHA class beside patient-reported measures and disease stages

Several neighbouring ideas are easy to confuse with the NYHA class. A health state in a model is a modelling construct, and a NYHA class is one possible definition of such a state. Functional status is the broader construct of what a person can do, which the class grades crudely for one condition through the single lens of symptom-limited activity. The ACCF/AHA stages describe whether heart failure is present and how far it has developed, while the NYHA class describes current symptoms and, in the words of the 2013 guideline, can change frequently.

The Kansas City Cardiomyopathy Questionnaire (KCCQ) asks the patient directly, whereas the NYHA class records the clinician's reading of the patient. The 2022 AHA/ACC/HFSA guideline, led by Heidenreich, describes the class as "a subjective assessment by a clinician" that "can change over time". The NYHA class carries no preference weights, so a NYHA-based model needs a separate utility source before the classes can produce QALYs. TA388 shows the KCCQ and the NYHA class used side by side as supporting evidence for a utility gain.

Reliability and other limitations for economic use

Raphael and colleagues tested the classification in a series of 50 patients assessed by two cardiologists. The cardiologists agreed on the class for only 54 per cent of patients and never disagreed by more than one class, and the authors noted that, in distinguishing class 2 from class 3, a 50 per cent concordance would be expected merely on the basis of probability, which suggests poor agreement. A survey of cardiologists in the same study showed no consistent method for assessing NYHA class. Agreement statistics such as the kappa statistic correct for chance agreement, which is why a raw percentage close to the chance level is weak evidence of reliability.

The 2013 ACCF/AHA guideline states that "reproducibility and validity may be problematic" but that the NYHA functional classification "is an independent predictor of mortality" and is widely used to determine eligibility for certain healthcare services. That combination explains the economic use of the classes despite the doubts. Several practical limits follow for modellers.

• Rating error enters the QALYs. Inter-rater disagreement of one class is measurement error in the state occupancy that drives utilities, and the worked example shows how a small one-sided error moves the ICER.
• Unblinded assessment. When the rater knows the treatment, a class shift can reflect expectation as much as symptoms, a risk that does not arise for mortality.
• Coarse categories. Each class pools people with different symptoms and utilities, so a treatment that helps people within a class without moving them across a boundary shows no benefit in a class-based model.
• Small severe subgroups. Class 4 is often sparse in trials, as the n=60 subgroup in TA388 shows, so class 4 utilities, costs and transitions carry wide uncertainty.
• Fixed utilities over time. Holding class utilities constant over a lifetime horizon ignores ageing, which the evidence review group's critique in TA267 considered to have favoured ivabradine.

Sources

• Criteria Committee of the New York Heart Association. Nomenclature and Criteria for Diagnosis of Diseases of the Heart and Great Vessels. 9th edition. Boston: Little, Brown and Co; 1994:253-256.
• Göhler A, Geisler BP, Manne JM, et al. Utility estimates for decision-analytic modeling in chronic heart failure: health states based on New York Heart Association classes and number of rehospitalizations. Value in Health. 2009;12(1):185-187. doi:10.1111/j.1524-4733.2008.00425.x
• Heidenreich PA, Bozkurt B, Aguilar D, et al. 2022 AHA/ACC/HFSA guideline for the management of heart failure: a report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation. 2022;145(18):e895-e1032. doi:10.1161/CIR.0000000000001063
• National Institute for Health and Care Excellence. Ivabradine for treating chronic heart failure (TA267). Published 28 November 2012. https://www.nice.org.uk/guidance/ta267
• National Institute for Health and Care Excellence. Implantable cardioverter defibrillators and cardiac resynchronisation therapy for arrhythmias and heart failure (TA314). Published 25 June 2014. https://www.nice.org.uk/guidance/ta314
• National Institute for Health and Care Excellence. Sacubitril valsartan for treating symptomatic chronic heart failure with reduced ejection fraction (TA388). Published 27 April 2016. https://www.nice.org.uk/guidance/ta388
• Raphael C, Briscoe C, Davies J, et al. Limitations of the New York Heart Association functional classification system and self-reported walking distances in chronic heart failure. Heart. 2007;93(4):476-482. doi:10.1136/hrt.2006.089656
• Wu Y, Tian S, Rong P, Zhang F, Chen Y, Guo X, Zhou B. Sacubitril-valsartan compared with enalapril for the treatment of heart failure: a decision-analytic Markov model simulation in China. Frontiers in Pharmacology. 2020;11:1101. doi:10.3389/fphar.2020.01101
• Yancy CW, Jessup M, Bozkurt B, et al. 2013 ACCF/AHA guideline for the management of heart failure: a report of the American College of Cardiology Foundation/American Heart Association Task Force on Practice Guidelines. Circulation. 2013;128(16):e240-e327. doi:10.1161/CIR.0b013e31829e8776