Article Text

Download PDFPDF

Bringing science to the art of diagnosis
  1. P Moayyedi
  1. Correspondence to:
    Professor P Moayyedi
    Gastroenterology Division, McMaster University-HSC 4W8, 1200 Main St West, Hamilton, Ontario, Canada L8N 3Z5;

Statistics from

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.

Which gastrointestinal symptoms are useful in distinguishing organic from functional disease?

Patients and clinicians are becoming increasingly intolerant of diagnostic uncertainty. This is reflected in the rising demand for endoscopic procedures1 implicitly suggesting that gastrointestinal symptoms are an unreliable indicator of serious pathology. Unfortunately, the growing array of tests that are being demanded for patients are placing further pressures on already stretched health care budgets. The natural reaction to this is to evaluate whether we can improve on the value of the history to diagnose gastrointestinal disease. This process was started over 30 years ago when researchers such as Card and colleagues2,3 and de Dombal and colleagues4,5 evaluated the use of computers to aid the clinician in making diagnoses in patients with upper gastrointestinal symptoms. The enthusiasm for this approach faded when data suggested that computer improved diagnostic accuracy of computers was not sufficient to prevent investigations.6 The paucity of subsequent research in this field is disappointing as it would be useful to know what information computers were using that enhanced diagnostic acumen. The article by Hammer and colleagues7 in this issue of Gut[see page 666] is therefore refreshing as it prospectively evaluates a wide range of gastrointestinal symptoms to establish which are useful in distinguishing organic from functional disease. The strength of this study is that it evaluates a large relatively unselected group of patients, reducing spectrum and selection bias.8 It also divided patients into those with and without organic disease rather than subdividing the data into different diagnoses. This improves the power of the study and gives the clinician overall information on whether the patient is likely to have an organic disease.

Their data suggest that aspects of the history are valuable diagnostic tools for lower gastrointestinal disease. They identify three “alarm features” that were important independent predictors of organic lower gastrointestinal pathology. These were age over 50 years, male sex, and blood on the toilet paper. They also identified five “non-alarm” features (frequent abdominal pain, radiating pain, pain with loose bowel movements, reflux, and absence of diarrhoea) that independently predicted functional bowel disease. The approach of positively diagnosing functional bowel disease with non-alarm symptoms or positively diagnosing organic disease with alarm symptoms performed equally well. The interesting observation is that putting these two pieces of information together improved the accuracy of the model. This is intuitive, as a clinician is likely to weigh up the absence of alarm symptoms and the presence of positive features indicating a functional bowel disorder before deciding that the patient is at low risk of significant organic disease.

These data will provide useful information for the Rome group9 when they next consider guidelines for the assessment of irritable bowel syndrome (IBS). They have highlighted symptoms similar to the classical Manning criteria that are important in making a positive diagnosis of functional gastrointestinal disease but also other characteristics such as radiating and frequent pain that make IBS more likely. Unfortunately, the model is probably not sufficiently accurate to be used to exclude patients from investigation in clinical practice. Using the prevalence of lower gastrointestinal organic disease in the article, the model had a positive predictive value of 79% and a negative predictive value of 86%. The possibility of missing organic pathology in 14% of patients would lead most clinicians to opt for investigating even if the model predicted a low probability of disease. The accuracy of the model may have been improved if non-gastrointestinal symptoms such as backache and frequency of micturition had been included10 but the concern about using models to predict organic disease is emphasised by the conflicting messages from other studies. For example, three papers also found that rectal bleeding11–13 predicted organic disease whereas two studies14,15 found no statistically significant association. Hammer et al found that weight loss and anorexia were not independent predictors of disease whereas others13,15,16 reported these symptoms were associated with organic pathology. Finally, Hammer and colleagues7 found that severe abdominal pain was associated with a reduced risk of having organic pathology in the univariate analysis. This observation is supported by one study12 whereas another suggested abdominal pain is associated with colorectal disease.17

These inconsistent findings probably relate to differences in the design and populations of the different studies. The quality of the data collection on gastrointestinal symptoms, sample size, and prevalence of underlying organic disease varied between studies. In particular, Hammer et al had very few colorectal cancers in their sample and the main organic disease driving the model was inflammatory bowel disease. If ulcerative colitis was the main disease found, then rectal bleeding and absence of pain will predict pathology and weight loss may not be an important feature in the model. It is interesting that the study12 that also had few colorectal cancers found similar results, whereas the report17 that evaluated a larger number of colorectal cancers found that features such as weight loss and abdominal pain were important. The case mix of disease in the underlying population is likely to have a large influence on the symptoms that computers models will predict are important.

Upper gastrointestinal symptoms are a good example of this. Hammer et al found that weight loss, dysphagia, age, and sex were not predictors of organic pathology whereas other researchers have found these to be significantly associated with the risk of upper gastrointestinal cancer.18,19 Indeed, it is almost inconceivable that male sex and increasing age would not be risk factors for upper gastrointestinal malignancy given what is known about the epidemiology of these cancers.1 Hammer et al did not find that these were important because the majority of their patients had gastro-oesophageal reflux and peptic ulcer and these diseases were driving the model. It is not surprising therefore that heartburn symptoms and aspirin use were found to predict pathology whereas alarm symptoms did not.

Although alarm symptoms may be of more value than is suggested by this paper, the overall message is consistent with other researchers. Factors that are found to be predictive of functional or organic disease are not very useful clinically. The model used by Hammer et al found it very difficult to distinguish between groups and as most patients had organic disease it generally defaulted to assigning patients to the disease group. Others have found the same problem: individual factors may be statistically significantly associated with disease but when they are combined in a model they generally perform very poorly.20 This is particularly true when the models are tested prospectively21 or transferred from a hospital to a primary care setting.22

Science has a great deal to offer the clinician in helping determine a diagnosis from the patient history. Statistical techniques can identify symptoms and signs that are important to consider and those that do not contribute to making a diagnosis. The problem is that a single study is not sufficient. After all, we do not rely on one study to definitively tell us that smoking causes lung cancer or Helicobacter pylori causes distal gastric adenocarcinoma. There is a need for a range of diagnostic studies conducted in different settings using different designs. The common feature of all should be the careful prospective collection of a detailed symptom assessment.23 When sufficient data have been accumulated, a picture will emerge on what are the important features that identify a patients as having organic disease. The problem that is addressed also has to be amenable to this approach. It will be difficult to use the history to guide us in patients with upper gastrointestinal disease but the situation with lower gastrointestinal pathology looks more promising. Insights provided by diagnostic studies may allow us to avoid invasive tests in some patients and identify others as having a high risk of serious colorectal disease that warrants urgent investigation. It is not sufficient to show the sensitivity and specificity of this approach as the ultimate goal is to provide more cost effective health care to patients.24 Randomised controlled trials will therefore be required to assess whether a more thorough clinical history will reduce costs or improve patient outcomes.

Which gastrointestinal symptoms are useful in distinguishing organic from functional disease?


Linked Articles