Answers are grounded in this paper and cite their sources.
Big data sets and novel analytical methods present us with an opportunity and a challenge to push the boundaries of medical research to move towards more targeted, personalised management strategies 1. The concept of ‘big data’ and the promises made concerning the potential of big data to generate novel hypotheses and provide most of the solutions in health care is becoming everyday parlance, with increasing belief that the data must and will speak for themselves. In childhood allergic diseases, discovering latent structure and patterns within large data sets using unbiased machine learning techniques has been used to identify distinct subgroups (classes or clusters) under the umbrella diagnoses of asthma, eczema and rhinitis 2, 3. The study by Goksör et al. 4 in the current issue of the journal is a timely reminder that in addition to disaggregating the structure in the big data sets using ‘unbiased’ analytical techniques, we need to look at the detailed associations of the temporality and comorbidity of symptoms via more traditional epidemiological approaches to fully understand the development of allergic diseases. This study from the Swedish longitudinal cohort presents a valuable epidemiological analysis which may contribute to a better understanding of the development of allergic symptoms during childhood. With the advent of the computational revolution combined with the amplified scale of biological, genetic and phenotypic healthcare data which has become available, the horizons of a data-driven hypothesis-generating approach to understanding disease have somewhat overshadowed the more traditional epidemiological hypotheses-testing approach based on carefully constructed scientific questions and observations. The ever-increasing quantity of data which is generated has made it impossible at times to know what we are looking for, and what questions need to be asked. One way of tackling this has been by taking an agnostic approach towards understanding the structure of the data by using unsupervised machine learning algorithms. Machine learning searches through data to look for patterns, combining mathematical modelling for analysing high-dimensional data and computational statistics to identify model-based patterns or structure within the data. However, nothing replaces and indeed nothing can replace carefully constructed epidemiological studies which ask specific questions and interrogate data sets to test specific hypotheses based on scientific intuition and prior clinical knowledge, and which act as a sanity check to validate the findings of hypothesis-generating studies. In his classical textbook, Rothman describes epidemiology as the study of the distribution and determinants of disease, as well as its frequency and occurrence 5. Epidemiology is concerned with understanding associations and risk factors of diseases, by using carefully constructed frameworks to understand disease causality 5. These concepts, which seem so simple on the surface, are fundamental if we are to extract meaningful clinical interpretations of the vast amount of medical data with which we are constantly presented. Goksör et al. 4 use this epidemiological approach, presenting a carefully structured question to investigate whether allergic manifestations in early life are associated with an increased risk in allergic manifestations in later life. The way in which they go about answering this question is likewise carefully constructed, looking at a range of symptoms and definitions of eczema, asthma, allergic rhinitis and food allergy. This study takes us back to the fundamental principles of epidemiology, which is understanding causality in order to influence policy. This was the aim of the ‘first epidemiologist’, Sir John Snow, who established the cause of the 1854 cholera epidemic in London was contaminated water, rather than bacteria from the surrounding area, which led to the closing down of the water pumps in Soho. We need to go back to the basics of epidemiology so that we do not miss the opportunity to capitalise on what epidemiology is all about: understanding causality and distinguishing causality from confounding. Confounding, an important albeit too often overlooked concept in epidemiology, is the study of effects which may explain away apparently causal associations observed in a statistical model. Confounding is a complex scenario and the existence of latent or unobserved confounders is an area which is acknowledged, but in which much development needs to be encouraged on the more practical level 6. The thorough analysis by Goksör et al. 4 has made an attempt to disaggregate all possible explanations of the development of, and the relationship between allergic manifestations during childhood, through carefully addressing confounding. The results challenge the view that development of allergic manifestations in childhood is a progressive development of one disease. Traditionally, the term ‘atopic march’ has been used to describe the progression of allergic disease from eczema to asthma and allergic rhinitis during childhood 7, and under this conceptual framework, eczema is seen as the first step in a march of symptoms leading to the subsequent development of asthma and allergic rhinitis. The concept of atopic march has gained traction both in clinical guidelines and in the design of prevention studies. Some guidelines recommend that clinicians inform parents that children with eczema may later develop asthma, and general practice trainees are often taught that the effective control of eczema not only improves eczema symptoms and quality of life, but also may prevent the atopic march 8. This paradigm has also been used as a basis for asthma prevention strategies, with the rationale that effective treatment of eczema in early life could interrupt the atopic march and decrease the risk of subsequent asthma 9. However, the investments by pharmaceutical industry in such studies have thus far resulted in failure 10. We have recently questioned the paradigm that atopic march adequately describes the natural history of allergic diseases during childhood by studying the development of allergic diseases in two large UK birth cohorts using a machine learning approach, to show that less than 7% of children with any symptoms of allergic disease follow this presumed trajectory 2. Similar to the findings by Goksör et al. 4, on a population level we observed a pattern similar to that of the atopic march. However, on an individual level, such a profile existed only in a small number of children. This apparent contradiction between what can be observed on a population level with the observed pattern on an individual level suggests that allergic disease is not a binary trait, but that there are heterogeneous patterns of symptom manifestation over time. Indeed, taken together, these findings indicate that the development of eczema, asthma and rhinitis has heterogeneous manifestations during childhood and that these different patterns of symptom manifestation may have different underlying biological mechanisms. This emphasises the need to carefully deconstruct different individual profiles and understand their unique pathophysiological mechanisms. Atopic march has traditionally been described within a causal framework 7, 8. Goksör et al. 4 have pushed this area of research by trying to better understand the association between allergic disease in infancy with allergic disease in later childhood (age eight years). They found that allergic disease at the age of eight years was related to the number of allergic manifestations in infancy. As in our previous study which used machine learning approach 2, the authors pointed out that such a finding does not imply causality (or, as they phrase it, the ‘progressive development of one underlying disease’), but rather a co-occurrence of allergic diseases which starts in early life. The authors reiterate that there is little evidence to infer causality, and when causality has indeed been inferred in previous studies 4, 7, 8, 10, the premises upon which this inference has been made have been erroneous. The authors state that ‘Children with allergic manifestations in infancy, for example either a recurrent wheeze, eczema or food allergy, had a higher prevalence of allergic disease at eight years of age, namely either doctor-diagnosed asthma, allergic rhinitis, eczema or food allergy ‘4. However, the authors illustrate that this association is by no means causal. It should be noted that although the co-occurrence of symptoms within a particular individual is greater than that which is expected to occur by chance, this does not necessarily extrapolate to the general population and to the linked progression of symptoms. The co-occurrence of eczema, wheeze and rhinitis can be a consequence of an underlying latent (nonobserved) and as yet unexplained disease status within the population, with distinct mechanism underpinning different symptom profiles. Good science is about merging different schools of thought for developing the bigger picture. The two approaches of machine learning (which helps us to identify structure in complex data sets) and carefully principled epidemiological examination to test hypotheses need to go hand in hand. These two disciplines present different aspects and perspectives of the same problem with different biases. There is a need to foster collaborative research more deeply in order to achieve a holistic ‘Team Science’ approach to understanding disease mechanisms 1. The data analytical process should never become disengaged from the thinking scientist process. In fact, complex statistical methods based on a pattern-recognition approach to understand the latent structure within the data need to work together with carefully constructed epidemiology in order to find mutually enriching solutions, so that data really does ultimately help save lives. Dr Danielle Belgrave is supported by the Medical Research Council Career Development Fellowship MR/M015181/1.
No takes yet. Share an insight, caveat, or question.
Belgrave et al. (2016) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: