Artificial intelligence in rheumatology research: what is it good for?

Introduction

Artificial intelligence (AI) has emerged as a transformative technology in medicine, providing rheumatology with innovative tools for research. AI, known as the capability of computational systems to perform tasks that typically require human intelligence, include learning patterns from prior data, understanding natural language, perception, reasoning, problem-solving.1 The impact of this technology in health sciences research is increasingly evident, with multiple applications gradually being integrated into the field of rheumatology.1 2 Indeed, different algorithms have led to the development of models for the diagnosis, evaluation, prognosis and prediction of disease.3 As AI has evolved, it has become increasingly important to differentiate between discriminative AI, widely used for studies on disease classification and prediction, and the more recently emergent generative AI, which holds promise for novel applications in research like hypothesis generation, clinical trial design, drug development, literature synthesis and writing support. Discriminative and generative AI differ in how they process data and apply their learning algorithms. While discriminative models focus on finding decision limits to predict labels, generative models analyse the underlying data distribution aiming to generate new data. Figure 1 summarises the main models used by these technologies and applications for research that we will explore in this review.

Discriminative AI includes a wide range of capabilities, such as distinguishing data to make classifications or predictions, as well as performing tasks like outlier detection and clustering. Radiology exemplifies the significant impact of discriminative AI.4 As a matter of fact, 723 of the 950 (76%) AI/ML (machine learning)-enabled medical devices approved by the Food and Drug Administration (FDA) as of August 2024 are related to this specialty.5 Other notable examples can be found in ophthalmology, where algorithms have shown the ability not only to diagnose ocular pathologies with greater accuracy than expert ophthalmologists but also to predict cardiovascular risk factors undetectable by fundus examination.6 Discriminative AI models have proven effective beyond image analysis, extending to fields like linguistics and other data types. For instance, a model used voice recordings and demographic data to predict dementia onset in patients with mild cognitive impairment, achieving around 80% accuracy.7 Although no FDA-approved AI/ML applications currently exist in rheumatology,5 numerous promising studies within the field of discriminative AI will be discussed in further detail.

Generative AI has recently transformed the AI landscape, particularly following the release of the chatbot ChatGPT in 2022, which has made AI more accessible to the general public.8 Generative AI can create new content based on existing data from various sources, including text generation, image or video creation. The tools based on this technology have shown promising applications in medicine, including demonstrating clinical knowledge by successfully achieving high accuracy in standardised examinations.9 Beyond knowledge-based tasks, ChatGPT’s responses to patient questions were shown to be often preferred to those by doctors for their quality and empathy.10 These results are remarkable given that the ChatGPT model is general-purpose and not specifically designed for medicine. The accuracy and reasoning skills of large language models (LLMs) have also been demonstrated in clinical examinations. Research has led to the development of specialised medical models like Med-Pathways Language Model (PaLM2), which achieved an accuracy similar to clinician answers (both exceeding 90%) in answering medical questions after being trained on six medical question-answering datasets.11

Rheumatology is a rapidly evolving specialty, thanks to the advent of advanced therapies and new technologies. Focusing on the management of chronic diseases with potential systemic involvement, it is a field rich in data and complex decision-making. Therefore, the use of AI tools holds the promise to transform clinical practice, leading to more informed decision making. Beyond the diagnostic level, there are promising predictive capabilities. In this regard, AI can assess disease activity, predict flares, determine optimal treatment dosages and anticipate patient responses based on clinical and serological biomarkers.12 Moreover, generative AI can function as a clinical decision support system, assist with administrative tasks, and enhance the quality of patient information and education.

The aim of this review is to highlight the current and future applications of AI in rheumatology, examine the mechanisms of AI, analyse state-of-the-art investigations and explore its integration into daily research practice. To achieve this, we conducted a narrative review including an electronic search in Medline and Embase for English-language sources from inception to September 2024. We employed a range of free-text terms including, but not limited to: “Artificial intelligence AND rheumatology”, “Machine learning AND rheumatology”, “Deep learning AND rheumatology”, “(Machine learning OR Deep learning) AND (rheumatoid arthritis OR spondyloarthritis OR psoriatic arthritis OR osteoarthritis OR lupus OR Sjogren)”, “Large language models”, “Natural language processing”, “(Predictive modeling OR Electronic medical records OR Risk stratification) AND rheumatology”, “(Large language models OR Natural language processing) AND rheumatology”, “ChatGPT AND rheumatology”. Furthermore, we conducted a manual search by examining the references cited in the included studies and technical computer science books. Priority was given to seminal references or those published within the last 2 years.

Key AI concepts for rheumatology research

The integration of AI into rheumatology research is becoming increasingly relevant, as the availability of complex datasets and advanced computation redefine how we approach and conduct scientific investigations.13 Given its capacity for use in research, AI-related concepts can help rheumatologists to effectively use these technologies in their work. Table 1 summarises the core principles in the most widely used AI algorithms, as well as their application in rheumatology.

An AI algorithm is a computational model designed to perform tasks by learning from data and identifying patterns, rather than relying solely on a predefined set of rules or instructions. These algorithms may improve their performance over time through experience, which is gained through an iterative process. These algorithms are typically used for classification or predictive purposes, such as diagnostic or prescriptive applications in medicine. This can be achieved by analysing large datasets, identifying relevant features and applying learnt patterns to new, unseen data.13

ML is a branch of AI that operates by feeding an algorithm with input data that reflects past observations, enabling it to construct a model to assess new, previously unseen observations. ML algorithms can be classified into four main types according to their training: supervised, unsupervised, self-supervised and reinforcement learning.13 Supervised algorithms are trained on a dataset where the output results are known and are used to label the outcomes. These have been the most used for clinical research. Unsupervised algorithms work with unlabelled data to identify patterns or clusters within datasets, making them useful for exploratory data analysis. There are two main types: clustering, which groups similar data points (eg, K-means, DBSCAN, hierarchical clustering), and dimensionality reduction, which simplifies data by reducing the number of features while preserving essential information (eg, PCA, t-SNE). These methods help uncover hidden structures without the need for labelled examples.14 Self-supervised learning creates internal labels within an unlabelled dataset, allowing models to learn without external annotation and guidance.15 Reinforcement learning adapts dynamically using reward-based feedback to maximise the performance of the algorithm.13

Deep learning (DL) is a subtype of ML that involves neural networks. A neural network is a particular ML algorithm based on successive layers of data transformation, inspired by the neural connections in the human brain. Neural networks are particularly effective with large volumes of data and demand significant processing power, which can be provided by processing units working in parallel. DL uses a high number of neuron layers, allowing for multiple levels of abstraction and has achieved noteworthy results in various applications, including text and image recognition.16 Transfer learning enables the adaptation of a DL model to specific imaging tasks (eg, rheumatological imaging classification) by leveraging pre-existing knowledge from extensive, non-specialised image datasets, enhancing model performance and reducing the need for large, specialised training datasets. Applications of DL include image recognition and natural language processing (NLP), which use images and text as input data, respectively.17 Indeed, one type of deep neural network algorithm gave birth to transformer technology.11 18 Transformers have revolutionised NLP with the so-called self-attention mechanism, which allows for capturing relations between words, allowing for efficient and accurate text generation. The seminal paper on this technology has garnered 140 000 citations by November 2024, reflecting the significant impact of language models on society.18

LLMs are advanced neural networks based on the transformer architecture.18 They are pretrained on vast amounts of unlabelled text data, typically sourced from the web, using self-supervised learning. This self-supervised learning involves predicting the next word in a sentence given the previous words (context), for which the model uses the surrounding context as the signal to learn and improve.17 These models are fine-tuned for specific tasks like question-answering and named entity recognition, showing their versatility and effectiveness in language understanding and generation. When models are able to process and integrate multiple types of data such as text, images and audio, this is known as multimodality.

Validation is a process that ensures a model’s generalisability and reliability by assessing its performance on unseen data before it is deployed in real-world applications.19 Evaluating discriminative AI models involves metrics familiar to rheumatologists, such as sensitivity (also known as recall in the field of ML) and specificity, which assess the ability to correctly identify true positives and true negatives.19 Precision, similar to positive predictive value (PPV), measures the proportion of true positives among all positive predictions, while the F1 score combines precision and recall into a single measure. Accuracy, which represents the overall correctness of the model’s predictions, reflects the proportion of true results (both true positives and true negatives) among the total cases. The area under the receiver operating characteristic curve (AUC-ROC) offers a visual summary of the model’s performance across thresholds, with the area reflecting the ability of the model to distinguish between classes.19

In evaluating generative AI models, additional metrics provide an understanding of model performance besides accuracy.20 Perplexity measures the model’s ability to predict the next word in a sequence, with lower scores indicating more precise predictions. Bilingual Evaluation Understudy assesses the similarity between the model-generated text and a human reference by comparing overlapping word sequences. Recall-Oriented Understudy for Gisting Evaluation measures the degree of n-gram overlap between generated and reference texts, emphasising recall. BERTScore further enhances evaluation by using Bidirectional Encoder Representations from Transformers (BERT) embeddings to compare the semantic similarity between generated and reference texts, capturing context-sensitive alignment. These complementary metrics allow for a comprehensive assessment of generative AI in clinical applications.21

Discriminative AI in rheumatology

Discriminative AI algorithms are focused on two main objectives. Classification analysis aims to evaluate previously described phenomena, attempting to describe features and ideally associations between risk factors (independent or predictor variables) and outcomes (dependent variables or events). On the other hand, predictive analysis, aims to forecast future events. Traditionally, this has been achieved using regression methodologies, including linear, logistic or Cox regression.19 Recently, AI has been employed to classify diseases and predict their progression using ML algorithms. These approaches may use various data types, including structured data, images and free-text information. Interpreting performances across studies is complex due to variations in datasets, patient cohorts and study outcomes, making direct comparison challenging. While metrics such as AUC and F1 metrics offer insights into model performance, their practical value depends on whether these AI advances lead to real-world clinical benefits. Validations against conventional models and interventions remain essential to establish AI’s utility and ensure its impact on patient care.

The application of AI in analysing structured data is advancing diagnostic accuracy, risk prediction and patient management in rheumatic and RMDs. In rheumatoid arthritis (RA), for example, a neural network model trained on demographic and laboratory data (including age, sex, rheumatoid factor, anti-citrullinated cyclic peptide and anti-carbamylated protein) achieved an F1 score of 0.92 in diagnosing RA,22 demonstrating accuracy comparable to, or exceeding, conventional diagnostic approaches. In predicting difficult-to-treat (D2T) RA, an extreme gradient boosting (XGBoost) model combined structured and unstructured data from 1873 patients, achieving an AUC-ROC of 0.88 for D2T identification and 0.73 for future D2T development prediction.23 In combination with structured data, integrating unstructured data—such as clinical notes and imaging—further enhances AI models’ ability to predict complex outcomes, as demonstrated in the previous study identifying RA subsets like D2T RA. Collectively, these applications underscore AI’s potential to exceed or complement standard statistical approaches by improving accuracy, sensitivity and specificity across RMD diagnostic and prognostic tasks.

For prognostic applications, structured data analyses have provided insights across various RMDs. In spondyloarthritis (SpA), K-means clustering applied to a longitudinal dataset identified two distinct disease activity trajectories—one with persistently high activity and another evolving to low activity—highlighting potential therapeutic approaches based on trajectory patterns.24 Similarly, in systemic lupus erythematosus (SLE), a random forest (RF) model predicted hospitalisations with an AUC-ROC of 0.75, using clinical markers such as dsDNA positivity, C3 levels, blood cell counts, inflammatory markers and albumin.25 Additionally, in pregnancy outcomes for women with SLE, a pre-pregnancy RF model achieved an AUC-ROC of 0.92, demonstrating high sensitivity (0.89) and specificity (0.94) in identifying adverse outcomes, a notable improvement compared with traditional models.26

The use of structured data and their analysis through AI has become an important axis in drug development and molecule generation, mainly based on ML and DL algorithms.27 Some of the use cases aim to identify drug targets and binding sites as well as to predict chemical properties (affinity, ability, lipophilicity, solubility, toxicity) of a compound. ML and DL algorithms may be used for efficacy evaluations of drugs through big data modelling and analysis.28 A relevant advancement in this regard has been conducted by AlphaFold, developed by DeepMind, in predicting structures of proteins. This enabled researchers to understand molecular targets more precisely, therefore supporting in identifying binding sites, refining drug designs and predicting protein interactions.29 Additionally, AI may assist clinical trial design and implementation by supporting selection of promising lead molecules based on patient-specific profiles, identifying suitable patient profiles and improving recruitment for clinical trials.

Predicting the suitability of treatments is crucial for improving research and clinical practice. One study used an ML model to predict methotrexate (MTX) response in RA patients using clinical data. A Least Absolute Shrinkage and Selection Operator algorithm, a method for fitting linear models, was employed in this project, achieving better performance than RF, with an AUC-ROC=0.79 (vs 0.68 in RF); this effective categorisation of patients into good and poor responders was achieved with baseline Disease Activity Score 28 (DAS-28), anti-citrullinated protein antibody and Health Assessment Questionnaire as top predictors.30 Combining clinical data with genomic biomarkers (single-nucleotide polymorphisms) and baseline DAS-28 has also shown promise in predicting MTX response in early RA; metrics of different supervised ML methods showed an AUC-ROC=0.84 in the training cohort, and a validation cohort accuracy of 0.76.31 Similarly, different ML models (linear regression, random forest, XGBoost and CatBoost) were evaluated for their ability to predict the probability of therapeutic response for bDMARDs in RA in the ESPOIR cohort, predicting response to tumorous necrosis factor inhibitors with an AUC-ROC of 0.72 (0.68 to 0.73), and yielding key predictors such as DAS28, lymphocytes, aspartate aminotransferase, neutrophils, age, weight and smoking status.32

Models aiming to predict the readmission risk of patients with RMDs after discharge, or the evolution of a given disease, have also been developed. By analysing data from electronic health records (EHRs), RF-based models aiming to predict patient’s return to the clinic achieved an AUC-ROC of 0.65, a sensitivity of 0.38 and a specificity of 0.79; follow-up duration, the prescription of DMARDs, corticosteroids, diagnosis of chronic polyarthritis, quality of life and patient occupation were identified as key variables.33 In the context of life-threatening illnesses such as systemic sclerosis, predictive modelling has been employed to estimate mortality rates drawing on clinical, demographic and spirometric data.34 The Naïve Bayes Classifier, a supervised ML algorithm, achieved an AUC-ROC=0.76 to predict 5-year mortality rates after internal cross-validation, which demonstrated superior predictive capability as compared with other algorithms, including RF (AUC-ROC=0.73), logistic regression (AUC-ROC=0.75) and Cox regression (AUC-ROC=0.724).34

Concerning imaging, studies have used various techniques from simple to complex. Using X-rays, ML models achieved up to 90.7% accuracy in distinguishing RA and OA from normal hand radiographs, though accuracy decreased (80.6%) when classifying all three classes together.35 Moving on to osteoarthritis (OA), a DL model was trained on knee radiographs to identify patients with and without pain progression, as measured by the Western Ontario and McMaster Universities Arthritis Index (WOMAC) pain score.36 The DL model achieved an AUC-ROC=0.80 in predicting pain progression, significantly higher (p<0.001) than a traditional model trained on demographic, clinical and radiographic risk factors. In axial imaging, a neural network based on 1553 pelvis X-rays evaluating the presence or absence of definite radiographic sacroiliitis as agreed in a central reading session, identified definite sacroiliitis with an AUC-ROC=0.94, a sensitivity of 0.92 and a specificity of 0.81 for the test dataset.37 CT has also benefited from AI, where neural networks trained on CT-derived 3D joint shapes distinguished hand joint patterns in RA with AUC-ROC=75%, psoriatic arthritis (PsA) with AUC-ROC=68% and healthy controls with AUC-ROC=82%. These models additionally identified disease-specific regions prone to erosions and bony spurs, contributing to classifying undifferentiated arthritis.38 Convolutional neural networks (CNNs) trained on sacroiliac joint images detected structural lesions such as erosion and ankylosis, achieving sensitivities of 0.95 and 0.82 and specificities of 0.85 and 0.97.39 Additionally, in Sjögren’s syndrome, a DL model using 500 CT images detected salivary gland damage in parotid glands with 96% accuracy, comparable to diagnosis of experienced radiologists.40

Regarding MRI, CNNs have also demonstrated the ability to differentiate between patients with RA and PsA based on patterns from hand MRIs, achieving AUC-ROC=0.75 for seropositive RA versus PsA, 0.74 for seronegative RA versus PsA and 0.67 for seropositive versus seronegative RA. Interestingly, adding demographic or clinical data to the networks did not provide improve classification.41 Non-articular MRI applications, such as brain MRI, have been used to evaluate fatigue in RA, showing that brain structural metrics were superior to clinical measures, with the highest prediction accuracy reaching 0.67.42 In SpA, MRI models have been developed to detect sacroiliac joint active damage. In fact, a deep neural network developed to detect MRI changes in sacroiliac joints indicative of axial SpA (axSpA) achieved a sensitivity of 0.88 and specificity of 0.71 for detecting inflammatory changes, and a sensitivity of 0.85 and specificity of 0.78 for structural changes in external validation.43 A multi-purpose MRI-based model using compound image transformations analysed knee cartilage in T2-weighted images to predict progression to symptomatic OA with an accuracy of 0.75, as defined by the WOMAC score 3 years post-baseline.44

Other imaging modalities, such as ultrasound, have demonstrated potential in predicting RA relapses and assessing joint conditions. A study comparing three ML classifiers found XGBoost to be the best-performing model (AUC-ROC=0.75), identifying 10 key features, including superb microvascular imaging scores of wrist and metatarsophalangeal joints.45 On vasculitis ultrasound, a study assessed the use of a CNN for detecting the halo sign in colour Doppler images for diagnosing giant cell arteritis, achieving an AUC-ROC=0.84 on the test set, with a 0.95 specificity and 0.60 sensitivity.46 For Sjögren’s syndrome, DL models used transfer learning to improve the automated segmentation of salivary gland ultrasonography, achieving a higher Intersection-over-Union (0.85) compared with both inter-observer agreement (0.76) and intra-observer agreement (0.84), indicating superior accuracy and consistency.47 Thermography, combined with AI, can detect RA activity by analysing temperature changes in hand joints. An ML-based method, ThermoJIS, for detecting joint inflammation in RA using hand thermography, correlated moderately with ultrasound scores and demonstrated with good diagnostic performance (AUC-ROC=0.78).48 Building on this, the study developed and validated two composite disease activity indices, ThermoDAI and ThermoDAI-CRP, which showed stronger correlations with ultrasound-determined synovitis (GS=0.52–0.58; PD=0.56–0.61) compared with patient global assessment (PGA) and PGA+CRP, and strong correlations with clinical indices (ρ>0.81).49

In the context of text analysis, discriminative AI using NLP has aided the analysis of vast amounts of EHRs, including tasks such as disease identification and clinical characteristics assessments. Several studies highlight NLP’s utility in rheumatology for disease detection. For instance, a validated ML pipeline identified RA patients with high performance, with support vector machines (AUC-ROC=0.98, F1 score 0.83) and gradient boosting (AUC-ROC=0.94, F1 score 0.82) outperforming simpler word-matching methods.50 Other study demonstrated the capability to identify axSpA through an unsupervised algorithm, incorporating both the NLP concept and ICD codes, with a sensitivity of 0.78, a specificity of 0.94 and an AUC-ROC of 0.93.51 This has also been explored in PsA, in which a sensitivity of 0.79 and a PPV of 0.93 were achieved when NLP was combined with billing codes.52 Further, a tool combining text mining with NLP-based exclusion accurately identified ANCA-associated vasculitis cases, achieving a PPV of 0.86 and outperforming traditional ICD-10 coding.53 This growing body of evidence supports the adoption of NLP technologies in accurately identifying RMDs.

Additional studies have focused on extracting clinical information beyond diagnoses from EHRs. For example, a recent study that included a dataset with around 64 million EHRs focused on the demographic and clinical characteristics of RA patients with interstitial lung disease (RA-ILD), yielding relevant information on prevalence, comorbidities and drug use in real life, with a high precision (F1 score over 0.7) for most of the assessed variables.54 Another algorithm extracted forced vital capacity from EHRs, strongly correlating (r=0.94) with pulmonary function test values.55 In RA, a study identified MTX-induced liver toxicity using NLP with a string-matching algorithm, achieving a PPV of 0.76.56 In another study, the analysis of structured and free-text EHR data from three hospitals showed limited disease activity evaluations in axSpA and PsA patients.57 For systemic autoimmune rheumatic diseases, an ML model predicted autoantibody testing needs and specialist referrals in systemic autoimmune diseases with AUC-ROC values from 0.91 to 0.94, enabling early detection up to 5 years before diagnosis.58 Another example illustrating the potential of AI in using large-scale real-world data is EPIC Cosmos, a vast inter-hospital database aggregating de-identified EHR from millions of patients across multiple health systems.59 EPIC Cosmos has enabled studies in different fields including rheumatology, such as recent work on SLE where researchers used Cosmos to enhance disease phenotyping and diagnosis. This study applied ICD codes to identify SLE patients and validated data quality against EULAR/ACR classification criteria, highlighting the need for integrating clinical notes to improve data completeness beyond structured EHR fields. While the study primarily relied on structured ICD codes for SLE phenotyping, the authors acknowledge plans to develop an NLP pipeline to analyse clinical notes, aiming to improve data completeness.60

Generative AI in rheumatology

Generative AI, the latest advancement in AI, is an emerging technology capable of creating new content in audio, image, video and text formats. It is based on foundation models—large-scale AI systems that acquire emergent capabilities across domains such as language, vision, robotics, reasoning and interaction. Their versatility allows them to adapt to diverse tasks, from NLP to computer vision and robotic control, by leveraging unlabelled data and self-supervised learning techniques. Among these, text-oriented applications have shown the most potentialities for research. At the core of generative AI are LLMs, which use transformer architecture to generate human-like responses based on input data.18 LLMs process and analyse input to generate outputs that mimic human reasoning based on statistical correlations, a capability that distinguishes them from discriminative AI, whose models produce a label or category based on the input, requiring explicit interpretation of the results.61 An example of interacting with these models is through widely recognised chatbots such as ChatGPT by OpenAI or Gemini by Google.62

Optimising the research process

AI, particularly through LLMs, has the opportunity to transform research by offering advanced tools that can support every stage of the process. Central to adopting these capabilities is the concept of prompting—the process of giving instructions to AI systems. Prompts can range from simple, direct queries to complex, structured inputs designed to elicit detailed responses.78 In research, prompting can be executed as ‘zero-shot’ learning, where the AI is given a task without any prior example or training; more refined prompts can guide AI to produce more focused and relevant information based on examples, such as few-shot prompting (giving examples) or chain-of-thought (providing step-by-step) prompting.

The role of AI in all aspects of research goes from idea generation and literature review to data analysis and manuscript preparation (figure 2). In the early stages of research, AI can significantly help via brainstorming sessions, in which a wide range of ideas and hypotheses can be explored.79 LLMs such as ChatGPT-4, Gemini, Perplexity and Claude are capable of generating diverse perspectives on a given research question, helping to promote the creativity that helps refine research objectives.62 These tools allow researchers to quickly iterate their ideas, explore conceivable investigative angles and develop a clear roadmap for studies. A recent study found that an AI model generated research ideas rated as more original and exciting than those of human scientists, though with slightly lower feasibility. Using the Claude 3.5 model, researchers produced 4000 ideas across several topics, and reviewers assessed these ideas without knowing their source. Despite AI’s high novelty scores, only about 200 ideas were genuinely unique, with creativity diminishing over time.80 The overwhelming abundance of options produced by AI challenges conventional creative processes, pushing researchers to shift from seeking single insights to generating numerous ideas for refinement. Rather than asking for one idea, AI enables researchers to request many altogether, allowing to sift through diverse suggestions and strategies. This surplus demands the skill of curating and discerning the best quality, which highlights a core value of AI in augmenting intellectual creativity and decision-making in research. As the research project progresses, AI tools can assist in refining the design and structure of the study. Beyond generating ideas, these systems can suggest detailed article structures, and help in drafting sections of a manuscript.

Figure 2

Request permission

Generative applications of artificial intelligence (AI) in research.

Conducting a literature review can also benefit from AI support. Tools such as Elicit or Research Rabbit provide curated bibliographies by searching with NLP.81 82 In addition, they can offer insights into the state-of-the-art of research concepts and visualise the relationships between key studies, potentially uncovering connections between research articles that may not be immediately apparent. They can also generate summaries from articles, extract precise data and even highlight emerging trends in the field, enabling researchers to stay ahead in their field.

AI tools can also take on a more active role during the data analysis phase. LLMs can assist in data analysis by guiding researchers through their analysis, performing statistical tests and generating insights from datasets. For instance, ChatGPT-4, can provide AI-generated code for statistical tools like SPSS, R or Python that facilitates the execution of complex analyses. Moreover, it can directly perform advanced computational calculations and data queries by inputting the prompt and the dataset to the system. As an additional support, it can also provide preliminary interpretations of data and give insights on the results. A recent study of 187 489 software developers using GitHub Copilot demonstrated how AI tools can reshape work by shifting focus from non-core management tasks to primary tasks, such as coding. This shift allowed developers to work more autonomously, explore new methods, and potentially reduce hierarchical dependencies.83

Finally, the writing phase—often the most time-consuming—can be facilitated using AI. Besides ChatGPT or Gemini, other tools such as Jenni AI can suggest article structures, and help draft coherent and well-organised manuscripts.84 These platforms can assist with translation, paraphrasing and ensuring that the text adheres to publication standards. As for the presentation of the results, image models can support on creating graphs, images or presentations. For example, platforms like Microsoft Copilot can create a presentation of the research.85

Conclusion

The combined strengths of discriminative and generative AI are revolutionising rheumatology research. Discriminative AI’s precise classification and prediction capabilities, paired with generative AI’s ability to synthesise and create content, may provide rheumatology researchers with powerful tools to enhance their work. These advancements can accelerate the research process and therefore contribute to the development of efficient processes in rheumatology. However, as we integrate these technologies into our research, we must proceed with caution, balancing innovation with responsibility to maximise their prospective impact on the field. The future trajectory of AI in rheumatology is within our hands, with its ultimate impact determined by our collective efforts and thoughtful application.