The promise of machine learning to inform the management of juvenile idiopathic arthritis

1. Introduction

Juvenile idiopathic arthritis (JIA) encompasses a heterogeneous group of inflammatory disorders linked by chronic joint inflammation [1]. Stratifying patients in a logical manner would improve the understanding and management of disease. Such a stratification would also enable the selection of patients with similar characteristics across clinical trials and epidemiological studies, enabling a comparison of results [2] and helping to design targeted and more effective management plans.

JIA is classified under the consensus-driven International League of Associations for Rheumatology (ILAR) system comprising seven subtypes [1]. These subtypes are a good first attempt at reconciling heterogeneous labels for JIA and have provided a framework for further studying JIA [3–6]. However, they only apply to patients <16 years of age, and debate continues as to whether they represent paediatric versions of adult phenotypes [7]. Subsequently, the Paediatric Rheumatology International Trials Organization (PRINTO) proposed a new system comprising six new subtypes designed to better align with adult phenotypes [8]. However, both systems contain undifferentiated subtypes that describe patients who do not satisfy criteria for one of the other subtypes, suggesting unexplored population structures and matching clinical recognition that some subtypes are still heterogeneous [9]. This is echoed in heterogeneous response and clinical success of biologic therapies against targets such as interleukin (IL)-1, IL-6, and tumour necrosis factor in limited subgroups of children from the different clinical classifications [10]. Therefore, other unexplored facets of disease, such as biology, may help in further stratifying JIA. However, the sheer number of measurements, spread across novel data domains, makes a purely consensus-driven approach challenging in terms of identifying patterns given the limited clinical experience with these domains.

2. The potential of machine learning (ML) for stratifying patients

Thanks to advances in genomic technologies and improvements in data management, we should consider a comprehensive, data-driven, ML-based approach to identify new disease signatures and patient groups in JIA. An obvious advantage of this approach is it would produce a stratification that reflects both the range of clinical phenotypes observed and the underlying biology of disease. Two categories of ML approaches are relevant to this review: supervised learning, which predicts labels from data (e.g., their ILAR or PRINTO subtype), and unsupervised learning, which identifies structures, such as patterns and groupings, independently of existing labels [11,12]. Several factors, beyond the limitations of consensus-based approaches, influence which category to pursue, including the breadth of heterogeneous data yet to be considered in classification. In a rare disease, measurements can outnumber patients such that we can stratify each patient into their own group defined by a distinct subset of measurements, the extreme outcome of overfitting and truly personalized medicine. Moreover, measurements can differ in units and scales and domains can differ in the number of measurements. For these reasons, unsupervised learning approaches may be preferred. We can broadly divide unsupervised learning into two approaches relevant to this review: dimensionality reduction and clustering. Dimensionality reduction groups related measurements together into signatures, and include methods such as principal component analysis (PCA) and non-negative matrix factorization (NMF) [13,14]. As the number of patients increases, we can consider deep learning techniques such as autoencoders and generative adversarial networks (GANs) [12]. Cluster analysis groups patients sharing similar characteristics together, such as how patients score on these signatures, and include methods such as hierarchical clustering, k-means clustering, Gaussian mixture models, and similarity network fusion followed by spectral clustering [15,16]. These two approaches are not mutually exclusive.

3. Prior applications of ML in JIA

Several studies have hinted at the usefulness of using unsupervised ML to stratify JIA. In more JIA-related applications, Martini et al. clustered antinuclear antibody (ANA)-associated measures using multiple correspondence analysis to suggest that ANAs are important for classification in four ILAR subtypes [3]. Van den Ham et al., using hierarchical clustering on plasma and synovial fluid from patients spanning four ILAR subtypes, found that patients with systemic arthritis differed in plasma cytokine expression from patients with oligoarthritis (≤4 inflamed joints) [4]. Jarvis et al., using pathway analysis on gene expression from blood from patients with polyarthritis (≥5 inflamed joints), identified networks of differentially expressed genes in neutrophils in patients compared to healthy individuals [5]. Van Nieuwenhove et al. identified an innate natural killer T cell signature common across all ILAR subtypes using random forests [6]. This last study highlights the importance of including as many ILAR subtypes as reasonable to fully take advantage of ML and construct a broadly applicable stratification without pre-selecting the subtypes to study.

To explore the usefulness of expanding the types of data to explore, we conducted several proof-of-principle studies in JIA. In the first study, we sought to identify disease signatures and phenotypes from clinical and cytokine expression data using probabilistic PCA [17] – explicitly chosen for its ability to handle missing data, which arise in clinical data due to deviations from study protocols and assay failures. PCA identified four signatures associated with various aspects of disease activity, demographics, and immune cell subsets. GMMs on signature scores produced five groups with distinct clinical and biological phenotypes. In particular, one group, with low clinical disease activity but high levels of circulating pro-inflammatory cytokines, experienced worse outcomes. In a second study, we applied multilayer NMF to complete joint involvement data in JIA and identified seven joint signatures related to distinct areas of the body, which then corresponded with seven phenotypes [18]. These phenotypes were each defined by few joints (i.e., sparse) and remained stable throughout disease course as patients experienced progressively fewer inflamed joints, building upon the idea of “indicator joints” considered by clinicians as indicators of poor outcome [19]. We then identified a novel clinical feature, the degree of localization, that measured the agreement of a patient’s joint involvement profile with their phenotype. Patients with non-localized joint presentation experienced worse outcomes than those with localized presentation regardless of treatment protocol, mirroring clinical intuition about the disease outcomes of less differentiated patients [18]. These proof-of-concept studies demonstrated that unsupervised ML can identify clinically and biologically sensible patterns and stratifications. In both studies, signatures were more reproducible than groups, suggesting that unsupervised analyses should emphasize the discovery of signatures over patient groups.

4. Challenges to stratification

Important challenges remain in terms of how to construct a better unsupervised stratification for JIA. Given the availability of more powerful yet data-intensive methods such as deep autoencoders and GANs, the first challenge involves the collection of not just more data, but also a consistent set of measurements. To fully realize the potential of these approaches, the Understanding Childhood Arthritis Network (UCAN) is uniting international JIA research consortia under a common framework, which will enable the characterization of the full set of JIA phenotypes observed globally. We can complement this harmonized approach through transfer learning by adapting pre-trained models from other diseases to more effectively make use of the collected data, although an open question is the relevance of pre-trained signatures compared to completely unsupervised ones. As more data are collected, refinements should be made to the data-driven stratification. Self-supervised learning may provide one such means, which can be further augmented by introducing expert knowledge between individual stages of refinement. Broadly, we could use unsupervised learning to identify an initial set of patient groupings, incorporate adjustments to these groupings from experts, and then apply self-supervised learning to refine these groupings. As more data are collected, the possibility of encountering free-text fields increases. These fields may contain useful information that are obvious to clinicians but would normally require numerous person-hours to extract into a usable form for ML. Natural language processing may be a potential avenue for using this information to its maximum potential. Once a stratification is generated, another challenge related to naming all entities in this stratification arises, especially as medication indications are linked to names, pointing to the need for common nomenclature. Another challenge is ensuring that the resulting stratification is generalizable among different geographic areas and ethnicities. Although many groups have put in significant effort in validating proof-of-concept stratifications with validation cohorts, the reality is that JIA is a rare disease, thus increasing the sensitivity of a data-driven stratification to more extreme phenotypes.

5. Conclusion

Expert-driven classifications arise from decades of experience with widely used disease activity measures. Experience with new genomic measures continues to be developed, so as such, in the short term, data-driven classifications are likely to result in more homogeneous patient groupings. The promise of integrating expert guidance on how best to use data-driven classifications is a stronger link to response to medications and clinical outcomes. Data-driven stratifications will identify links between biological findings and clinical phenotypes, thus providing evidence to guide therapeutic decision-making and usher in a new era in precision child health.

Abbreviations and acronyms:

ANA

antinuclear antibody

CV

cross-validation

ERA

enthesitis-related arthritis

GAN

generative adversarial network

GMM

Gaussian mixture model

HLA

human leukocyte antigen

ILAR

International League of Associations for Rheumatology

JIA

juvenile idiopathic arthritis

ML

machine learning

PRINTO

Paediatric Rheumatology International Trials Organization

RF

rheumatoid factor

Th

T-helper

UCAN

Understanding Childhood Arthritis Network