Now let's think about how to go about conducting a diagnostic study. Let's assume that we've already identified the clinical problem, derived a diagnostic research question, decided on the diagnostic determinants or tests we're interested in, and determined which reference standard we'll use to determine the disease status of our participants. In our study, we'll need to collect this information for a number of patients, possibly hundreds or even more, depending on the research question. There are a number of ways that epidemiologists collect data in their studies. And for diagnostic research, some designs are more appropriate than others. First, we need to consider the time frame of the diagnostic process. In most epidemiological research, we're interested in seeing how certain factors are associated with or influence health outcomes over a period of time. This is commonly the case for etiologic research, and is absolutely necessary for prognostic and therapeutic research. In diagnostic research, however, we want to assess how well a combination of tests works in a patient, at a specific point in time, typically when patients consult their doctor. Consequently, diagnostic research is cross-sectional by nature. To actually recruit and collect information on all diagnostic determinants and the reference standard in a single patient may take days or weeks. But this information will be applied to assess which diagnostic determinants best predict the presence of the diagnosis at that point in time, that is, when the patients present their complaints. In such a cross-sectional approach, there’s no need for patient follow-up. There are exceptions to the rule, however, where you might be interested in following patients over time in a diagnostic setting. You might be interested in knowing what impact the use of a new diagnostic rule will actually have on a patient's prognosis. This kind of research is best described as diagnostic intervention research. It aims to answer a very different type of research question to those that we have been focusing on when discussing diagnostic research. To really assess health impact, it’s no longer sufficient to describe the relationship between your diagnostic test and the presence of the disease. Instead, this kind of research is causal by nature. You want to investigate whether your diagnostic strategy causes the prognosis of your patients to improve, possibly due to a faster or more accurate diagnostic process. The same principles then apply as in other types of intervention research. We won't be focusing on diagnostic intervention research this week, but you'll hear more about intervention research in week four. Another aspect of data collection to consider is whether information should be collected prospectively or retrospectively using registries or other pre-existing databases. While a retrospective approach could be less time-consuming and less expensive than recruiting new patients and collecting that information directly, it may be very difficult to find a group of patients in whom all of the tests of interest, as well as the reference standard, have been performed in a standardized way. Diagnostic research should therefore ideally be conducted prospectively, in a manner that reflects clinical practice. So let's imagine we've begun our diagnostic study to find the most accurate combination of less-invasive tests for cow milk allergy in children. We conduct a cross-sectional study assessing incoming patients at several primary care practices over a period of 12 months. The reason we do this over 12 months is not follow those patients over those 12 months, but instead, to give us enough time to actually include a large enough number of patients in our study. We now have information about a range of diagnostic determinants, including general patient information, medical history, signs and symptoms, and blood markers. We also have the confirmed diagnosis of the patients based on the results of the oral food challenge test. And we have all of this information for a couple of hundred children. What should we do next? The aim of our research was to see if we could find a less invasive set of diagnostic tests that could accurately detect cow milk allergy. And ideally, we would want to present the information we find in a form that practitioners can use. One way we can do this is by using our newly collected data to develop a mathematical model known as a prediction rule. Such a model would use the results of the diagnostic test we are interested in, and could be applied to future patients to estimate the probability of having a cow milk allergy. These kind of models are commonly developed using multi-variable regression techniques such as logistic regression. We need to remember at this point that as diagnostic research is descriptive by nature, our regression model won't be used in the same way as models are used in etiologic research, where confounding is often adjusted for. Instead, all of the components of our model are important to us and provide useful information to help us estimate the probability of the presence of a diagnosis. So how should we actually go about building a new diagnostic model? Which tests should we include in our model? Statistical selection methods can be used to try to determine the best combination of diagnostic determinants in an automated manner. But it's generally a much better idea to try to select combination of determinants that are likely to be most useful, based on evidence from the literature, as well as what makes sense as a realistic diagnostic work up, for example, in terms of feasibility. The resulting prediction model can then be adapted in a number of ways to make it easier to use in clinical practice. For example, by turning it into a sum score, so that patients with a score above a certain threshold would then be diagnosed with the disease. There are several famous examples of diagnostic rules that are currently used in practice, such as the Ottawa Ankle Rule for diagnosing ankle fractures and Well's Rule for diagnosing deep vein thrombosis. On a final note, there's a very important aspect of diagnostic research that we haven't mentioned, and that's the validation of diagnostic models. You'll hear more about this in the context of prognostic research next week. But the same concept applies to diagnostic models. For now, it's sufficient to say that developing and assessing the accuracy of a diagnostic rule isn't the end of the story. You'll have to show that your diagnostic rule also works in future patients. So we've now seen, for our cow milk allergy research problem, how we could produce a tool for primary care practitioners to use in practice to help diagnose future patients. Hopefully this would be of help to efficiently diagnose or exclude cow milk allergy in future patients with similar complaints as the baby presented at the beginning of this week.
