1 INTRODUCTION

Rheumatoid arthritis (RA) is a chronic autoimmune disease that primarily affects synovial joints, leading to inflammation and potential joint destruction. It affects approximately 1% of the global population and is more prevalent in women than in men. Furthermore, RA can affect systemic organs and systems, leading to comorbidities, such as myocardial infarction, stroke, and pulmonary fibrosis.¹ Early diagnosis and treatment are critical for preventing irreversible joint damage and improving long-term outcomes.²

Artificial intelligence (AI) refers to the simulation of human intelligence in machines designed to perform tasks that typically require human cognition, such as visual perception, speech recognition, decision-making, and language translation.³ Machine learning (ML) is a critical subcategory of AI that focuses on developing algorithms that enable computers to learn from data and improve their performance over time.⁴ The process of ML through which computers learn to generate the desired output is referred to as “training.” ML methods are generally classified into three main categories based on their training approach: supervised, unsupervised, and reinforcement learning. In supervised learning, models are trained to predict future outcomes through the identification of patterns from labeled input–output data. Popular techniques in this category include random forests, support vector machines (SVMs), neural networks, and natural language processing (NLP) models. NLP models are designed to interpret and analyze text and speech, making them useful for tasks such as the analysis of electronic health records (EHRs). Unlike supervised learning, unsupervised learning focuses on uncovering hidden structures and relationships within the input data, such as identifying clusters or reducing data dimensionality.

Another subset of ML is deep learning (DL), which utilizes neural networks with multiple layers (hence “deep”) to analyze various forms of data (Figure 1). DL uses artificial neural networks (ANNs) composed of multiple layers to extract complex patterns from large datasets. Among various types of ANNs, convolutional neural networks (CNNs) are particularly effective for image processing, making them invaluable for medical image analysis.⁵ In contrast, recurrent neural networks are designed to handle sequential data, making them suitable for tasks such as analyzing time-series medical data or predicting patient outcomes over time.^{6, 7} More recently, transformer models, which excel at processing sequential data through self-attention mechanisms, have gained prominence owing to their ability to provide context-aware predictions, thereby enhancing the integration and analysis of multimodal medical data^{8, 9} (Figure 2).

Recently, AI has emerged as a transformative tool in rheumatology.¹⁰ The capacity of AI to process and analyze vast heterogeneous datasets, including clinical records, imaging, and genomics, has established it as an indispensable tool for managing the complexity inherent in RA. By integrating diverse data modalities, AI facilitates the detection of previously unrecognized patterns, enabling early diagnosis, formulation of personalized treatment strategies, and identification of novel therapeutic targets. Moreover, AI-driven predictive models are increasingly being applied to forecast disease progression and evaluate the risk of comorbidities, such as cardiovascular complications, in patients with RA.

Despite the promise of AI in improving the early diagnosis and management of RA, it remains insufficiently mature for full integration into clinical practice. This review summarizes recent advances in the application of AI, particularly ML, DL, and NLP, in the field of RA, focusing on imaging techniques, multimodal data integration, and personalized medicine. The objective of this study was to highlight the progress made in the past 5 years and identify future research directions to enhance the management of RA.

2 RA IMAGING: RECENT ADVANCES

RA imaging primarily includes modalities such as radiography, ultrasound imaging, and magnetic resonance imaging (MRI). Each has specific applications in the diagnosis and monitoring of various diseases.¹¹ Various AI methods, such as CNNs, have been used in medical imaging to automatically segment joint regions, detect erosions and synovitis, and quantify joint space narrowing, offering improvements in consistency and efficiency, particularly in reducing observer variability, although not necessarily surpassing human accuracy in all cases (Tables 1–3).^12-30

3 INTEGRATION OF MULTIMODAL DATA IN RA

Multimodal data refer to the integration of diverse types of medical information, such as clinical notes, medical imaging, genomic data, and biosensor readings, all of which are used to provide a comprehensive understanding of patient health. In the context of RA, this could involve the fusion of text-based EHRs, medical images, such as radiographs and magnetic resonance images, and molecular profiles, such as genetic and biochemical data.

Over the past decade, substantial progress has been made in leveraging different AI methodologies for various types of medical data, each making distinct contributions to the diagnosis and treatment of complex RA conditions. DL models have demonstrated considerable efficacy in interpreting medical images and facilitating automated detection of joint erosion and inflammatory features.⁴² In 2023, Ahalya et al.⁴³ introduced RANet, a custom CNN model that detects RA from hand thermal images with up to 97% accuracy by integrating quantum techniques. Ojha et al.⁴⁴ utilized a CNN to predict RA from hand images and achieved an impressive accuracy of 97.5%. Similarly, Phatak et al.⁴⁵ demonstrated the effectiveness of a ResNet-101-based CNN for detecting inflammatory arthritis in hand joints using smartphone photographs, highlighting the robustness of the model across diverse skin tones without requiring feature engineering. Bai et al.⁴⁶ further enhanced diagnostic accuracy by developing an ANN for RA diagnosis, achieving an AUC of 0.951.

Likewise, ML algorithms have been extensively applied to genomic data to uncover genetic markers associated with disease susceptibility.^47-49 In PathME (Lemsara et al.,⁵⁰), a multimodal sparse autoencoder was employed to cluster patients using multi-omics data by integrating gene expression, miRNA, and methylation profiles to identify disease subgroups. Recently, Li et al.⁵¹ employed advanced ML algorithms, such as Boruta, Least Absolute Shrinkage and Selection Operator regression, and random forests, to identify PANoptosis-associated genes in RA, focusing on key biomarkers, such as SPP1, which were validated to be crucial for RA diagnosis and treatment response. Liu et al.⁵² integrated transcriptomic data using clustering techniques to identify novel biomarkers and RA subtypes based on immune signatures, highlighting the potential of combining supervised and unsupervised methods for biomarker discovery and patient stratification.

NLP has also been used to extract valuable insights from clinical narratives, thereby enhancing the analysis of large-scale EHR. Humbert Droz et al.⁵³ developed an NLP pipeline to extract RA disease activity measures from clinical notes across multiple EHR systems, thereby demonstrating the feasibility of integrating clinical narratives into multimodal frameworks for large-scale RA research. Similarly, England et al.⁵⁴ applied an NLP tool to extract forced vital capacity values from EHRs of patients with RA-associated interstitial lung disease, significantly enhancing the availability of these critical data for disease monitoring. These diverse applications illustrate the broad potential of AI across different data modalities, contributing to improved outcomes in patients with RA.

Studies employing AI in multimodal data in RA are rare. Wei et al.⁵⁵ employed large language models (LLMs), such as generative pre-trained transformer-3.5-Turbo, to investigate the molecular mechanisms and identify shared drugs for RA and arthrofibrosis. By analyzing over 370,000 publications and integrating bulk and single-cell sequencing data, they developed a multimodal framework highlighting the pathogenic similarities, key cell types, and therapeutic targets for both conditions. Although the integration of multimodal data using AI for RA remains in its nascent stages, preliminary trials have demonstrated promising results for other diseases. DeePaN, a graph convolutional network developed by Fang et al.,⁵⁵ integrates clinicogenomic data to stratify patients with non-small cell lung cancer, outperforming other unsupervised methods and enhancing personalized treatment strategies. Similarly, the holistic AI in medicine framework,⁵⁶ combines tabular, image, time-series, and textual data, resulting in a 6%–33% performance increase in tasks, such as chest pathology diagnosis and mortality prediction. In the same year, Tang et al.⁵⁷ introduced FusionM4Net, which integrates clinical and dermoscopic images with metadata for skin lesion classification, achieving 77.0% accuracy, surpassing single-modality models. However, some challenges remain, such as limitations in generalizing across datasets and the need for task-specific optimization. For RA, integration of diverse data types, such as imaging, clinical records, and genomics, offers substantial promise but requires further research. Future studies should focus on data harmonization and improving model generalizability to fully unlock the potential of AI in the diagnosis and management of RA.

Despite these advancements, several challenges still need to be overcome. High-dimensional data, such as medical imaging combined with genomics and clinical records, introduce what is known as the “curse of dimensionality,” a major obstacle to effective model training and generalization. The high complexity of multimodal data often results in overfitting, which limits the applicability of models to real-world settings. Techniques such as dimensionality reduction (e.g., using autoencoders) and feature selection are crucial for mitigating these effects.

4 PERSONALIZED MEDICINE IN RA

DL-based predictive models have been developed to predict patient responses to specific RA treatments. These models analyze EHRs, including past treatment outcomes, to identify patterns that suggest the most effective therapies for individual patients.⁵⁸ Recent studies have demonstrated the potential of AI-driven personalized medicine to improve outcomes in RA by tailoring treatment to the unique characteristics of each patient while still requiring clinical judgment.⁵⁹

Over the past 5 years, the application of ML in utilizing multi-omics and laboratory data to predict clinical outcomes has expanded significantly. In 2019, Kim et al.⁶⁰ applied a nonnegative matrix factorization approach to identify synovial transcriptional signatures associated with RA and used supervised learning models to predict the infliximab response, achieving an AUC of 0.92. Tao et al.⁶¹ applied random forest algorithms to gene expression and/or DNA methylation profiling to predict responses to adalimumab and etanercept therapies in patients with RA and achieved impressive accuracy rates of 85.9% and 79%, respectively. This demonstrates the application of supervised learning to enhance treatment personalization through high-dimensional biological data analysis. In 2021, Morid et al.⁶² explored the benefits of semi-supervised learning, specifically one-class SVM, over traditional supervised methods, such as extreme gradient boosting, for predicting step-up therapy in RA. In 2021, Jung et al.⁶³ applied k-means clustering, an unsupervised learning method, to classify 1103 patients with RA into four clinical phenotypes. In 2023, Ferreira et al.⁶⁴ used unsupervised k-means clustering to differentiate patients with RA based on their proteomic signatures. Anioke et al.⁶⁵ used hierarchical clustering to predict the progression to inflammatory arthritis based on lymphocyte subsets. Their findings underscored the utility of unsupervised ML for integrating multi-omics data, thereby contributing to a deeper molecular understanding of the pathogenesis of RA.

DL has also seen considerable advancement over the past 5 years, particularly in the context of RA. Xing et al.⁶⁶ integrated DNNs with ML models to predict the activity of inhibitors targeting key RA-related kinases, including spleen tyrosine kinase, Bruton's tyrosine kinase, and Janus kinase. This integration demonstrated the potential of DL in drug discovery for RA, in which targeting multiple pathways could lead to better therapeutic outcomes. Wang et al.⁶⁷ reviewed the use of DL in combination with traditional and Western medicines to enhance RA diagnosis and treatment. They employed CNNs and cloud computing to analyze multiple data types, such as imaging and clinical records, leading to more precise diagnoses and optimized treatment plans. Kalweit et al.⁶⁸ applied deep embedded clustering combined with an adaptive deep autoencoder to categorize patients with RA based on their response to biological disease-modifying antirheumatic drugs, identifying distinct patient clusters with different drug responses.

NLP has also become an important tool for improving personalized therapeutic strategies for RA, with LLMs being extensively utilized for complex data extraction and integrative analyses. Nashwan and AbuJaber⁶⁹ discussed how LLMs can optimize EHRs by automating data extraction, streamlining patient communication, and enhancing decision-making, underscoring the potential of these models to improve RA management. Furthermore, Wei et al.⁷⁰ utilized GPT-3.5-Turbo to explore the molecular mechanisms shared between RA and arthrofibrosis, demonstrating how LLMs can assist in drug repositioning by integrating transcriptomic and literature data, thereby supporting personalized therapeutic strategies. Advancements in LLMs have facilitated the extraction of critical clinical information including diagnoses, outcomes, and treatments, thereby enhancing the utilization of real-world clinical data for RA.

Furthermore, the integration of multimodal data may be key to advancing personalized medicine for RA.⁷¹ However, as model complexity increases, interpretability often diminishes, raising critical questions regarding balancing AI reliance with clinician expertise. The black-box nature of complex AI models necessitates their cautious deployment, particularly when weighing AI-generated recommendations against human clinical judgments. Nevertheless, the utility of AI in evaluating drug safety profiles and predicting optimal therapeutic pathways offers substantial promise for augmenting clinician decision-making and delivering individualized care. These findings underscore the transformative potential of AI for advancing personalized rheumatology.

5 CHALLENGES AND FUTURE DIRECTION

One of the major challenges in applying AI to RA research is the scarcity and quality of available data. Ensuring that the data used for the training models are comprehensive and representative of diverse RA patient populations is critical.⁷² Researchers are exploring methods, such as data augmentation, transfer learning, and the use of synthetic data, to enhance the model performance and generalizability.⁷³ Recently, federated learning has been explored as a method to enhance data utilization without compromising privacy.⁷⁴ This approach allows multiple sites to contribute to the model training by sharing model coefficients instead of raw data. However, federated learning involves significant challenges, such as the high costs of the central server setup and intersite communication. Moreover, it does not completely solve the data privacy issue because models can still potentially “memorize” the training data. To address this, techniques such as differential privacy have been employed to prevent data memorization with the aim of creating more secure and effective AI solutions in RA research.^{75, 76}

In addition to technical challenges, ethical considerations are crucial to addressing errors flagged by AI when applied in RA research. If AI models support clinical decisions, the responsibility for errors owing to data inaccuracies or model misinterpretations will raise significant ethical concerns.^77-79 Particularly in high-stakes contexts, such as RA diagnosis and treatment planning, any errors flagged by AI systems must be meticulously evaluated.⁷⁷ Errors identified using AI tools may result from biases inherent in the training datasets, which may not fully represent the diverse population of patients with RA, potentially leading to algorithmic bias.⁷⁷ Furthermore, AI-generated errors can affect patient outcomes if not managed transparently, making it essential to establish ethical guidelines for interpreting and rectifying these errors in the clinical workflow.^{78, 79}

To mitigate these risks, it is essential to incorporate robust auditing processes and error-handling mechanisms into the deployment of AI tools.⁸⁰ This approach ensures that clinicians are fully aware of the limitations of AI and can make informed decisions, balancing the recommendations from AI with clinical judgment.^{80, 81} The black-box nature of DL models has raised concerns regarding their interpretability in clinical settings. Efforts to develop explainable AI techniques, such as saliency maps and attention mechanisms, are critical to ensure that these models can be trusted and adopted by healthcare professionals.⁸² The complexity of the DL models poses a challenge in terms of their interpretability, which is crucial for their clinical adoption. Researchers are exploring various methods to improve the explainability of these models and ensure that their predictions are trusted by clinicians.⁴²

Ethical and regulatory challenges in deploying AI for RA include concerns regarding patient data privacy, the potential for algorithmic bias, and the need for regulatory frameworks that can keep pace with rapid technological advancements.⁴ These challenges are compounded by the complexity of integrating AI with existing clinical workflows and ensuring that AI-driven decisions are transparent and interpretable to both clinicians and patients.^{83, 84} Furthermore, ensuring equitable access to AI technologies across different healthcare settings is essential to prevent the exacerbation of existing disparities in RA care.⁸⁵

Future research in AI, especially DL for RA, is likely to focus on improving model accuracy, expanding the use of multimodal data, and addressing the challenges of data scarcity and model interpretability.^{4, 5, 42} The recent literature emphasizes the integration of clinical, genetic, and imaging data to refine prediction models and personalize treatment strategies.^{4, 5, 32, 42, 86} For instance, advancements, such as combining DL with multi-omics and imaging modalities, are expected to enhance predictive capabilities and provide real-time actionable insights for clinicians.³² Moreover, addressing data scarcity through transfer learning and synthetic data generation could overcome the current limitations, leading to more robust and generalizable models.

6 CONCLUSION

Over the past several years, significant advancements in AI, particularly in DL, have transformed the landscape of RA research. Traditional ML techniques have laid the groundwork for data-driven insights, and the application of DL has shown promise in contributing to various areas of RA management. These include enhancing diagnostic consistency through the analysis of complex imaging data, assisting in the identification of potential biomarkers for the prediction of treatment response, and supporting the integration of multimodal data sources to inform personalized treatment plans. However, these methods require robust clinical validation and illustration before they can be broadly adopted and accepted. Despite these advancements, challenges remain, particularly in model interpretability, the risk of overfitting, and the need for robust clinical validation across diverse patient populations. As AI techniques continue to evolve, their integration with clinical practice holds promise for the further refinement of RA diagnosis, prognosis, and treatment, thereby moving the field closer to the realization of precision medicine.

AUTHOR CONTRIBUTIONS

Yiduo Sun conceptualized the study, conducted the literature review, wrote the manuscript, and prepared the figures and tables. Jin Lin contributed to the literature review, assisted in the writing of sections related to AI methodologies and applications, and provided critical revisions. Weiqian Chen supervised the study, provided overall guidance, and finalized the manuscript. All authors read and approved the final manuscript.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflict of interest.

ETHICS STATEMENT

This review article does not involve human participants, animal studies, or any other ethical concerns.