Article Text

Download PDFPDF

Spectrum bias: why generalists and specialists do not connect
  1. Michael Jelinek
  1. St Vincent’s Hospital; Melbourne, Victoria, Australia

    Statistics from


    Coronary artery disease (CAD) exists in a spectrum: it ranges from the early lesions seen in young victims of trauma1 through to sudden cardiac death or myocardial infarction with cardiogenic shock. Yet diagnostic studies that compare the clinical features and diagnostic tests with a reference standard usually act as if the disease was a homogeneous entity. This dichotomous approach to diagnostic accuracy is measured as sensitivity and specificity, likelihood ratios, diagnostic odds ratios, and the area under the receiver-operator characteristic curve.2 Ransahoff and Feinstein recognised that “unless an appropriately broad spectrum is chosen for the diseased and non-diseased patients who comprise the study population,” the diagnostic test may result in spurious estimates of diagnostic performance.3 However, this article argues that it is not possible to have a sufficiently broad spectrum of patients to be of value in a single description of diagnosis, prognosis, and therapy of any common condition such as CAD.

    The clinical and diagnostic features of disease such as CAD change with increasing severity of the disease. Generalists, particularly primary care clinicians, are likely to see less severe manifestations of CAD than specialists. Hence, depending on the level of specialisation of the clinician, disease presentation may create spurious clinical associations or miss legitimate clinical associations.

    The solution proposed for this differing disease presentation in different clinical contexts is to recognise that no single class of clinician can define the clinical features or natural history of disease. There will be differing clinical features, risk factors and prognoses at all levels of specialisation.


    Serious misestimates may occur if a disease is defined by invasive techniques. This is sometimes called “work-up bias” and occurs when patients with abnormal test results are disproportionately referred for invasive testing. This bias results in inflated estimates of test sensitivity and unrealistically low test specificity.3-9 Bayes’ theorem10 or regression analysis11 can be used to correct for “work up” bias. George Diamond has shown that it is still possible to define the natural history of CAD if the cardiologist knows what proportion of patients seen by the referring clinician are actually sent to the cardiologist.10 25 In practice this is never known.


    In the 1970s, mitral valve prolapse was regarded as a serious condition, which was associated with high rates of chest pain, dyspnoea, electrocardiographic abnormalities, arrhythmias, strokes, and the need for mitral valve surgery.12 13 Studies were performed on patients who were referred to a cardiologist for these clinical features and were found to have mitral valve prolapse by echocardiography. However, when strict echocardiographic diagnostic criteria were prospectively applied to a community sample of subjects in Framingham Massachusetts, these same clinical features were not found to have any association with mitral valve prolapse.14

    Similarly, hypertrophic cardiomyopathy was first described as a highly symptomatic and lethal disorder from tertiary referral centres,15 but subsequent community studies have shown that most hypertrophic cardiomyopathy is a relatively benign disorder.16 In these cases, spurious associations were created by the careful investigation of sickest (referred) people without similar study of well people.

    CAD shows a similar spectrum problem. The prognosis of patients with severe ST segment depression on stress testing is poor with a high prevalence of serious CAD, need for early revascularisation, and the risk of sudden cardiac death.17 18 However, in a group of patients who were clinically stable and submitted to 6-monthly exercise tests, Podrid et al showed patients with such positive stress tests had only a 1.4% annual mortality rate.19 The reason for the discrepancy between Podrid’s data and the other studies is that most such studies come from referral centres when the usual indication for such stress tests is the occurrence of new or changing patterns of angina or when patients are being considered for major non- cardiac surgery. Clinically stable patients are normally not submitted to regular stress tests. Hence, there is no natural history of severe myocardial ischaemia without consideration of its clinical context.


    The spurious positive clinical associations were a result of examining or testing patients who were sick and not doing similar studies on those who were not similarly affected. Referral bias may also result in loss of clinically important associations. For example, if all patients submitted to coronary angiography have angina pectoris, then characteristics of chest pain will cease to have diagnostic or prognostic significance when defined by coronary angiography.20 The symptom of angina, particularly if limiting effort tolerance or progressively worsening, has major prognostic significance. The finding of no prognostic disadvantage for anginal patients compared with asymptomatic subjects having stress tests at the Cleveland Clinic is a consequence of that clinic’s practice of usually performing coronary angiography on patients with angina at the clinic, reserving stress tests for patients less symptomatic and at lower risk of death.21

    The diagnosis of congestive heart failure (CHF) is a common problem in medical practice because CHF is a serious condition with no specific diagnostic test. It was hoped that the hormone assay B-type natriuretic peptide (BNP) might provide the breakthrough for the diagnosis of CHF. Indeed, the test was found to be very valuable in the diagnosis of CHF in primary care22 and in the emergency room of hospitals.23 However, the prevalence of serious cardiac disease is high in cardiac practice where the occurrence of CHF might result in major changes of management. The use of BNP to make a diagnosis of CHF in this context has been disappointing, perhaps because cardiologists already have knowledge of the structure and functioning of the heart of the patient through the use of echocardiography.24


    Certain conclusions flow from considering the process of referral. Specialists at every level of referral will see similar patients with similar clinical features and prognoses. If referral is selective, they will have more in common with each other than with the clinicians who refer them patients. A specialist’s experience will have limited applicability for the clinician who refers him the patient. Conversely, the referring doctor’s approach may have less validity for the specialist once the patient has been referred to that clinician.

    Each clinician is describing different manifestations of the same disease. All manifestations of the disease are true and all are different. There is no single unifying description and prognosis of a particular disease independent of referral level. It is important for generalists and specialists to realise and respect that they see a different spectrum of disease with different manifestations and outcomes. Failure to understand this sampling discrepancy has, and will continue to, result in a diagnostic and therapeutic disconnection between these groups.


    The author is grateful for the advice given by George Diamond, Paul Glasziou, and Gordon Guyatt, and many colleagues in Melbourne, Australia, in the preparation of this article.


    View Abstract

    Request Permissions

    If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.