Ground Zero Diagnosis

Longstanding challenges of effective diagnosis of UTIs point to the need for a more fundamental rethink in practitioner decision making

At the heart of medicine is the requirement of accurate diagnosis. Without this it is immensely difficult to know what the suitable treatments should be. So far, so familiar: the medical profession has been doing this for a very long time and have had remarkable success in combatting disease of the many of the world’s populations. 

New health challenges (e.g. new variants of COVID and other infectious diseases) highlight the importance of effective diagnosis, but there are also longstanding ones that need addressing such as urinary tract infections (UTIs). This is a widespread issue with the prevalence in women over 65 years of age approximately 20% (compared with approximately 11% in the overall population). Between 50% and 60% of adult women will have at least one UTI in their life, and close to 10% of postmenopausal women indicate that they had a UTI in the previous year.  

One of the challenges of UTIs is obtaining an accurate diagnosis. This is clearly an important issue in its own right but also acts as a starting point for considering the challenges facing practitioners needing to make real-world decisions in changeable environments.

Limits of testing

One of the key challenges for accurate diagnosis of UTIs is that conventional biomarkers involving lab test results have weak accuracy for diagnosis and differential diagnosis for UTIs . Urinary dipsticks, used by GPs as a first line UTI diagnostic tool are, according to some reports, grossly insensitive missing up to 70% of urinary infections. The tests are especially unsuitable for detecting embedded UTIs, which have relatively few free-floating bacteria.

The danger here is that, as is widely understood in the psychology literature, doctors can sometimes struggle to factor in the underlying probabilities of a false negative from these tests into their diagnosis, as Gerd Gigerenzer has explored. 

However, there is a wider issue that also merits attention. Received wisdom in medicine, but also more widely, point to the use of base rates for accurate prediction. Generally speaking, the current concept of risk prediction / diagnosis is based on a reference population to compare new patients. This focus is often discussed with regard to base rate fallacy, a type of bias in which people tend to ignore the base rate in favour of the individuating information. 

Without a measure of the base rate it is difficult to interpret the meaningfulness of test results – we may have naturally occurring levels of the substance that is being measured and need to establish the point at which it merits status as a ‘condition’. The basic assumption is that the reference population is similar to any one individual: and so variance from the measures occurring naturally in the population allow us to make an effective diagnosis.

But there are problems here: In his book Cystitis Unmasked, James Malone-Lee writes that the “diagnostic threshold” for identifying a UTI is based on a 1950s study of 74 pregnant women with pyelonephritis (a serious kidney infection). Somehow, Malone-Lee suggests, the method used in the study became adopted worldwide to diagnose UTI in non-pregnant people. While there is a case for evaluation against the wider population, a non-comparable base-rate is clearly problematic.

But there is also part of a much wider challenge.  Fiedler (2000) and Krynski and Tenenbaum (2007) consider that while base rates are large samples of a broad population, real-world decision making generally requires predictions to be made from a limited sample. In this way base rate information is inevitably often irrelevant – based on factors such as geographic location (e.g. the same base rate may not apply between countries) or age of data (e.g. due to the way base rates change over time).

So is the answer to get effective base rate information?  On the one hand it will be possible to more accurately assess the probability of the presence of a UTI if we can know the incidence a relevant population and use that as a basis to interpret the test results (notwithstanding how flawed they reportedly are).  However, one can quickly see the practicalities of doing this being difficult; not only does defining the ‘relevant’ population create difficulties given the range of dimensions that could be used, and the limits of data for each of these combinations make it not a viable solution. Moreover, health is not a closed system – external environmental factors may impact individual’s health as well, which may fluctuate and change across time. Therefore, we are never comparing like with like – there will always be individuating and external factors that makes it very difficult to compare against the base rates of certain tests.

Added to this, most of the variables used to predict and diagnose health capture the state of the body (such as cortisol level or blood pressure).  But cortisol is newly formed constantly so that the measure we take today, is not the same as the measure from last week. Biological units of measurement are not stable and rigid.  Indeed, biological mechanisms are dependent on the processes of their host organ or organism. These life processes (such as aging) are notoriously hard to capture and account for in any measurement. In fact, there are not many instances in medicine where processes are included in the prediction model– despite medicine agreeing that processes such as aging are relevant (and interacting) processes.

What do we do about it?

It might be easy to assume then that when we have very tangible data points then medicine allows us to be accurate.  But reality is not quite so straightforward. While medical practitioners rely on data from diagnostic tests, applying the associated population probabilities to the individual can be problematic for the variety of reasons we have outlined.

We therefore call for a more Bayesian approach – the population statistics are the ‘prior probabilities’ which are then adapted based on the reports of the patient.  Patient knowledge is critical as the individual’s profile can render the general population as not relevant. But it is also related to processes (which the patient is able to self-report on to a large degree) rather than states alone. 

This would make the case for relational diagnosis where the healthcare practitioner spends time with the patient, as it introduces adaptive mechanisms into the process.  Without this the healthcare practitioner is thrown back onto the the use of a probability based diagnosis:  the ongoing challenges of women’s experience of UTIs surely illustrate the problematic nature of this. 

Indeed,  the danger here is alienation from the medical establishment, which can lead to seeking explanations and solutions that are potentially problematic, as indicated by a study of Reddit discussion on the topic which suggests there is an abundance of misinformation on this issue.

Additional thoughts

Science historian Lorraine Dalston suggests that COVID means we are living in a period of radical novelty and radical uncertainty which means we are thrown into a state of ‘ground-zero empiricism.’  As we struggle to make sense of a rapidly changing environment, we simply do not have the data to make sense of the world in the ways we have in the past.  While that quote comes from a blog post at the outset of COVID, it remains highly relevant, specifically related to diagnosis of long COVID but also the alarming new variants of a range of new viruses. 

And more broadly we are seeing the way in which the underlying ‘processes’ are changing through a combination of climate change, COVID and technology, reshaping the ‘states’ that we use for measurement. Practitioners therefore may need to rethink ways in which population data serves as a helpful reference point, with new more agile and flexible approaches starting to come to the fore.