Discussion

This review has brought to light important information on the current status of big data in RMDs and in other medical fields. Only a few authors clearly defined what they meant by ‘big data’, and the provided definitions were quite different. There were also disparities in the number of data found in the articles. Data sources were varied, and were mostly clinical in RMDs; whereas they were equally clinical, biological or radiological in the very recent publications outside RMDs. Both traditional and AI methods were used to analyse big data, and among machine learning methods, the most represented was artificial neural networks, independently of medical field and data source.

This study has strengths and weaknesses. First, only one reviewer performed the screening of the articles and extracted information from the selected articles. However, support was provided by coauthors, especially when data scientist skills were needed. Second, the research was only performed on PubMed MEDLINE with a language restriction. Ideally, multiple bibliographical databases would have been searched. The key words used to perform the SLR were restrictive. To assess the impact of the choice of key words, a sensitivity analysis showed that the use of additional key words referring to RMDs by name and/or to specific AI methods would have brought less than 10% more articles. We believe this indicates the validity of the present findings. Moreover, the aim was to obtain an overview rather than an exhaustive view of the topic. In the future, SLR in individual fields such as imaging, computational biology or clinical research, could be considered. Indeed, imaging and the other big data sources are very wide fields, and specific researches in each of them would have certainly found additional articles. We did not perform a SLR outside of the field of RMDs, however, this part of the work was only meant to be used as a comparison with RMDs and not to be exhaustive; furthermore, giving the fact that 33 794 articles related to big data out of RMDs were found using our key words, performing a SLR would not have been feasible. Choosing the right key words for this research was an issue, and some references in the field of RMDs were not picked up, such as the ActConnect Study25; however, we used the best keywords available at that time, and even if a MeSH term was created in 2019 for ‘big data’, all articles using this concept are not referenced yet under this MeSH term. Finally, classification of statistical methods, in particular machine learning methods, may be discussed. However, it was based on accepted classifications, which are considered reliable references in this field.22 23

A key finding from this review is that there is no consensual definition of big data. First defined as data sets too large or complex for traditional analysis methods,11 this concept has evolved and the ‘5 V’ paradigm (for volume, velocity, veracity, variety and value) is more and more used.26–28 The definition provided in recent EMA recommendations may be considered as a synthesis of all these notions.12 Although all the authors of the selected articles agree that big data refers to a very large number of data points, there is also no consensual ‘cut-off’ to define what is meant by ‘very large’. Some authors proposed log(n×p) superior or equal to 7 (n being the number of units of observation and p the number of variables),29 however, giving the rapid growth of datasets in the last decade, some authors rather propose to think in terms of terabytes (10¹²) or petabytes (10¹⁵).9 10 Nevertheless, even terabytes and petabytes will be soon too restrictive, since according to an International Data Corporation report prediction, the global data volume will grow exponentially from 4.4 zettabytes to 44 zettabytes (10²¹) between 2013 and 2020.30 This issue shows that the definition of big data is beyond the scope of the characteristics of data type and cannot be restricted to the size or volume of those data.31 It confirms also the disparity in the number of data reported in the studies. Thus, the amount of data is not the same when considering clinical or imaging data, since a single radiological exam can contain millions of pixels, and some imaging techniques such as MRI can also contain several images or sequences; this point makes complex the estimation of the number of datapoints in imaging. However, beyond their volume, what makes big data a challenge is their complexity based on heterogeneity, multidimensionality and the fact that they are dynamic—in other terms, all the previous single dimensions are dynamically connected. None of the selected articles addressed clearly these issues, despite studying complex connexions between heterogeneous, multidimensional and dynamic data offer unparalleled opportunities to personalise medicine.

In this review, clinical data were the most frequent source of big data in RMDs, whereas the distribution was more spread out outside of RMDs. This could be explained by the fact that, except clinical and radiological data, other data sources may not be so well implemented in rheumatology, whereas outside rheumatology, the literature review only picked up extremely recent articles (due to the retrochronological approach) and omics are a rapidly evolving field. With the increasing amount of information collected by registries, Electronic Health Records and the increasing use of sensors collecting in real time patients’ data, clinical research must evolve to take advantage of these new sources of information and implement them in routine practice.2 32 Given the exhaustive nature of clinical big data, they could be particularly interesting in the future to study rare diseases, rare outcomes and evaluate the efficacy of treatments in non-selected populations, which are difficult to assess in usual clinical trials.2 33 Omics is a growing field, particularly promising for personalised medicine as it supports the discovery of predictive biomarkers and therapeutic targets.3 4 34 Imaging is also a very interesting application of big data for diagnosis and clinical decision making. Since any single radiological exam compiles a huge amount of data, medical imaging is particularly conducive to the use of AI and notably machine learning methods.35 36 Examples of applications of big data in medical imaging are numerous and varied, from diagnosis and follow-up of cancers,37 38 to scoliosis39 or diabetic retinopathy.40 As social networks and Internet-driven data are exponentially growing, text mining is becoming a relevant source for health information: recent applications were the prediction of influenza and pertussis epidemics thanks to Google searches41 42 and prediction of depression thanks to Facebook statuses.43 In the field of RMDs, only one paper included in the SLR was based on Google, Wikipedia and Youtube searches concerning inflammatory vasculitis,44 contrasting with the variety of examples provided in other medical fields. This could be explained by the fact that rheumatic diseases and symptoms are less common than the examples cited above or that RMD specialists are less aware of these methods at this time. However, it is probable that this novel source of information will play an important role in health research in the years to come, particularly given the increasing focus on patient-driven and community-driven research.45

This work revealed the use of various statistical methods, traditional or AI-related, to analyse big data. The use of usual statistical methods for big data, such as χ², Student’s t-test or logistic regression may seem paradoxical, because big data may be too complex to be analysed with these tools.11 However, we found out that more than 20% of articles related to big data used traditional statistical methods (27% in RMDs and 22% in other medical fields). This may be because of lack of knowledge of AI methods, or because traditional statistics allowed to answer the clinical questions.46 Thus, in cohorts or registries, the number of patients is higher than the number of variables collected for each of them, so even if it is big data by ‘the number’, it may not be too complex for usual methods. Another possibility is that more specific methods such as AI are not yet well implemented in every research unit. Most of the selected articles used AI, and in particular machine learning methods; indeed, these methods seem more relevant to analyse huge and complex data, such as genomic or imaging-driven data.35 47 Moreover, these methods are tolerant of poor quality of underlying data,48 which is a common issue in big registries, were missing data are frequent. However, the risk of an inappropriate use of these methods is creating quantitative fallacy and over fitting models which could not be generalisable in clinical practice.49 That is why AI and machine learning algorithms should be validated and regulated to be integrated in medical practice.50 In the present review, between 5% and 10% of the selected articles did not mention explicitly the kind of machine learning algorithm which was used, indicating a need for better reporting.

In conclusion, this work gives an overview of the current status of big data in RMDs, and in medicine in general. Data sources and types are varied, and methods used to analyse them are heterogenous and not always well reported. This variety of sources and methods hold promises for potential applications of big data in rheumatology and in other medical fields, and may lead to a major change in health research for the years to come.