Can Prediction Models in Primary Care Enable Earlier Diagnosis of Rare Rheumatic Diseases?

Fiona Pearce; Peter C. Lanyon; Richard A. Watts


Rheumatology. 2018;57(12):2065-2066. 

The impact of rare diseases (conditions affecting fewer than 1 in 50 000 people) is increasingly prominent in national healthcare policy development. European Reference Networks have been evolving since the early 2000s, to share information, research and knowledge in rare diseases. In 2009, a Recommendation of the Council of the European Union called for each EU State to put in place a rare diseases plan or strategy (2009/C 151/02). For example, the UK Strategy for Rare Diseases,[1] published in 2013, makes 51 commitments aimed at ensuring that people living with a rare disease have best-quality evidence-based care.

Recognizing that difficulty obtaining a timely diagnosis is a major problem, one of the five key areas of focus in the strategy is improving diagnosis and earlier intervention for those with a rare disease. Although the strategy appears focused on the majority of rare diseases that have childhood onset or an identified genetic cause, it is very important that the needs of non-genetic rare diseases, which often have clinical onset during adulthood, are not neglected.

The rare autoimmune rheumatic diseases, ANCA-associated vasculitis (AAV), SLE, myositis and scleroderma, form a significant component of this non-genetic, adult-onset group; and they frequently share similar clinical presentations, with non-specific symptoms in multiple organ systems. A theme that emerged at a Workshop for Rare Rheumatic diseases in November 2015 was the similarities of the patients' priorities expressed by charities that represent people living with these conditions, and a major concern was that many people experience delays of up to 5 years between the onset of their symptoms and diagnosis.[2]

In support of this need, one aspiration within the UK Strategy for Rare Diseases[1] is the development of 'effective IT support' for rare conditions presenting with common symptoms, specifically advocating use of computer prompts to alert primary care physicians to consider a rare disease diagnosis. Predictive models based on electronic algorithms have also been recommended by the National Institute for Health and Care Excellence for use in UK primary care to predict risk of cardiovascular disease[3] and fragility fractures,[4] and have been incorporated into one primary care software system to predict risk of cancer.[5]

Would such models work to enable earlier diagnosis in rare rheumatic diseases? This is important to consider, because the consequence of delayed diagnosis and treatment of these conditions, with potentially irreversible organ damage and high early mortality, could have an even greater impact on the individual than a delayed diagnosis of inflammatory arthritis.

To answer this, we need greater understanding of the health-seeking behaviour of patients in primary care prior to their diagnosis for people with rare autoimmune diseases. In other words, are there key predictive patterns (e.g. consultations with specific symptoms and frequency) that can be identified and which might enable earlier diagnosis of these conditions and lead to development of risk prediction models?

To our knowledge, this work has only been done and predictive models attempted in two rare rheumatic diseases. Both used a database of UK primary care records (United Kingdom Clinical Practice Research Datalink, UKCPRD) and a method similar to that used to develop a cancer prediction model.[5] The model for SLE had a sensitivity of 34% and a specificity of 90%,[6] which meant that it did not identify 66% of cases, and that if someone was flagged as at risk they had a 1.3% chance of having SLE (and 98.7% chance of not having it). Subsequently, we have recently reported that in the 5 years prior to a diagnosis of another rare autoimmune disease (granulomatosis with polyangiitis), although there is increased health care activity compared with those without this diagnosis, there are no specific symptoms, either alone or in combination, with sufficient discrimination to support model development.[7]

Why do the models seem not to work? There are two problems. The obvious, but more minor, problem is that as there is not one symptom that all people with each rare rheumatic condition share, the combinations of symptoms needed to build a model with a specificity as high as 90% reduce the sensitivity, and such models will only identify a small proportion of cases.

However, the less obvious, but crucial, consideration in prediction of rare disease is that the prevalence of the disease in the population being tested affects the predictive value of a model or test. The rarer the disease, the lower the positive predictive value of a test (i.e. the lower the probability that an individual has the disease if the prediction tool flags them as at risk). This may seem counter-intuitive. The conditions in which risk prediction tools are currently used are much more common than the rare rheumatic conditions: cardiovascular disease affects ~3% of the population, osteoporosis ~2% and cancers ~0.9%, compared with 0.0049% for SLE[8] and 0.002% for AAV, scleroderma and myositis.[9,10] If we compare simple predictive models (that flag people as high risk or not) that perform equally well at the model-building stage, and all have a sensitivity of 100% and a specificity of 90%, when applied to each condition they would have different positive predictive values. In cardiovascular disease 31% of people flagged as at risk would have cardiovascular disease, in osteoporosis 20% of people flagged as at risk would have osteoporosis, for cancers 8.6% of people flagged as at risk would have cancer, for the rare rheumatological diseases 0.47% of people flagged as at risk would have SLE and only 0.0002% of people flagged as at risk for AAV, scleroderma or myositis would have that disease.

Therefore, predictive models running in the background of primary care software, prompting consideration of rare disease diagnoses seem unlikely to be effective. We need to seek other avenues for improving the diagnostic odyssey for our patients, including those with conditions such as Behcet's disease and relapsing polychondritis that are rarer than those considered above. This will include continued support from rheumatologists in the assessment of acute presentations, secondary care of people with undiagnosed multisystem disease, and awareness that attendance at multiple different ambulatory specialities may indicate the need to search for one of these conditions.