The use of machine learning to optimize treatment of rheumatic diseases
In the past three decades, we have experienced an explosion of knowledge in all areas of medicine. The increasing use of “wearables” in everyday life in recent years, decreasing prices of biotechnical methods, and improved understanding of biomarkers and genetic sequencing methods, are creating unprecedented new opportunities, especially within rheumatology.
With the accelerating flow of scientific data, our knowledge grows rapidly, but also shows an immense growth in complexity.
Current challenges in diagnosis and management of rheumatic diseases
The field of rheumatology has seen enormous progress in recent years due to a more thorough understanding of the pathophysiology of immunological processes and the development of targeted drugs. Nevertheless, with few exceptions, rheumatic diseases are still not curable.
Many rheumatic diseases are heterogeneous in their manifestation. Disease progression, response rates to treatment, and the development of comorbidities frequently differ between individual patients. The treat-to-target approach is still not pursued in all patients.
The use of artificial intelligence and machine learning may be a game changer in this context. They are a driving force behind the expansion of knowledge in various fields of medicine (see figure 1). For example, artificial intelligence has already been successfully used for the analysis of big data in other disciplines such as radiology, ophthalmology, and dermatology.
Figure 1: Results on machine learning studies (rheumatology, ophthalmology, dermatology and radiology) on Pubmed.org (between 1999 and 15th October 2019)
Machine learning in rheumatology
The most frequent focus of artificial intelligence in rheumatic and musculoskeletal diseases is machine learning. Machine learning is a field of computer science that uses algorithms to develop predictive models from large data sets. Therefore, machine learning could serve as a sharp scalpel to dissect big data, uncover new patterns, and hence improve therapeutic options in rheumatology.
One subtype of machine learning, “deep learning”, has become more and more important in recent years. This method uses computationally complex algorithms to identify relationships within data sets.
A recently published meta-analysis by Liu and colleagues showed that deep learning diagnostic performance of medical imaging is equivalent to that of healthcare professionals. However, it seems likely to be a matter of time before deep learning will be superior to humans in the field of image recognition.
Disease activity prediction
There are also the first promising approaches of machine learning in disease activity prediction within rheumatology. Norgeot and colleagues were able to predict disease activity at the time of the next rheumatology visit in a test cohort of 116 patients with rheumatoid arthritis. This deep learning model was based on the electronic healthcare records of 820 patients from two different rheumatology clinics. The study showed that electronic health records can be used to compile accurate models to forecast complex disease outcomes with the help of deep learning models.
Identification of clinical subgroups
Orange and colleagues applied machine learning to even more complex data sets. By combining clinical, histologic and gene expression data they identified three distinct synovial subtypes in patients with rheumatoid arthritis.
For this study, 20 histologic features were assessed in 129 synovial tissue samples. Machine learning applied to histologic features, with gene expression subtypes serving as labels, generated an algorithm for the scoring of histologic features.
Patients with the high inflammatory synovial subtype exhibited higher levels of C-reactive protein (CRP), erythrocyte sedimentation rate, rheumatoid factor titer, and anti-cyclic citrullinated peptide antibodies. In addition, CRP levels were significantly correlated with the severity of pain in the high inflammatory subgroup.
Although this study also has some limitations (eg, the model was not validated using an independent data set, and the data sample was also very small) it demonstrates the potential of machine learning. It is a first indicator on how machine learning can help us to identify new subgroups in rheumatoid arthritis in the future. These subgroups could benefit from different therapeutic approaches or provide indications of the disease course.
Evaluation of clinical response to therapy
In particular, the evaluation of therapeutic responses by machine learning is the subject of new promising studies. This year, Guan and colleagues presented the best-performing model in predicting anti-TNF response in the DREAM Rheumatoid Arthritis Responder Challenge.
For the DREAM challenge, 1892 patients with rheumatoid arthritis from across 13 cohorts of patients were randomly selected. These data were used as a training data set and then validated from a data set of 680 individuals from the CORRONA registry. The authors used a Gaussian regression model which included demographics, baseline disease assessment, treatment, and single-nucleotide polymorphism array data.
The model proved to be useful to identify sub-cohorts of non-responders to TNF inhibitor therapy. In this model, 78% of individuals’ response statuses were correctly classified in the training cohort. The study shows promise as a first approach in guiding drug selections in clinical practice based on primarily clinical profiles with additional genetic information.
Machine learning has the potential to facilitate a better differentiation between individual disease severities and outcomes. Depending on the predicted disease outcome, patients could benefit from more intensive treatment and care. On the other hand, patients with a better disease outcome could benefit from earlier tapering of their medication.
Challenges of machine learning
Despite the promising prospects of these results, there are still some challenges to be overcome. A crucial factor for the success of machine learning is the quality and size of the data: The fuel for machine learning is large and correct data sets.
Nevertheless, a large heterogeneity of data sets and their quality or methodical processing are still apparent.
A limitation of patient data can be upcoding or downcoding in electronic health registries. In order to facilitate the use of machine learning in rheumatology, the development of global, harmonized and comprehensive standards should be promoted to enhance the interoperability of large amounts of data.
Furthermore, interdisciplinary training on large data methods in rheumatic diseases for clinicians, healthcare professionals, health and life scientists, and data scientists should be promoted.
There is huge potential for machine learning to reduce uncertainties in medicine. It will be capable of recognizing associations that were previously unknown.
This will allow us to detect many new facets in certain diseases. In particular, subgroup analysis via machine learning will enable us to design clinical trials more intelligently and to develop more targeted treatments for patient groups, and ultimately for individual patients.
Undoubtedly, machine learning will improve knowledge and opportunities within medicine, including rheumatology. However, it cannot make any statement as to why certain associations exist.
The interpretation and verification for plausibility and the assessment of causality will have to be performed by the physician. Machine learning will therefore not be able to replace the physician, but rather help us to understand diseases better and to optimize treatment.
About the authors