Current application,possibilities,and challenges of artificial intelligence in the management of rheumatoid arthritis,axial spondyloarthritis,and psoriatic arthritis

Artificial intelligence

AI can be explained as machines that mimic human intelligence and make decisions similar to humans. ⁷ Unlike traditional programming methods, AI learns from data to uncover relationships and derive insights, combining computer science and statistics. ML, natural language processing (NLP), speech recognition, expert systems, robotics, computer vision, and several other branches work harmoniously to serve AI. All the branches of AI, especially ML and NLP, appear to be more prominent in this area of medicine.

Machine learning

ML, a specialized field within AI, enables systems to recognize patterns and generate predictions from data without explicit programming. ML models identify relationships within data, improving decision-making processes used in diagnosis, prognosis, and treatment planning in medicine.^7,8 Models generalize patterns learned from datasets to predict new outcomes.^7,8 Primary ML methods include supervised learning (learning input-output relationships from labeled datasets, such as linking patient symptoms to diagnoses), unsupervised learning (identifying hidden structures or patient subgroups from unlabeled data), semi-supervised learning (combining limited labeled data with extensive unlabeled data, particularly useful in rare diseases), and reinforcement learning (learning sequential decision-making through feedback, potentially beneficial for developing treatment strategies).^7,8

Deep learning

DL is a subset of ML, using neural networks with multiple layers to automatically learn representations from large and complex datasets. ⁹ Unlike traditional ML algorithms, such as the Support Vector Machine (SVM), Random Forests (RnFr), Decision Trees, and Gradient Boosting Machines, which normally require manual feature extraction, DL models learn relevant features themselves from raw data. ³ Therefore, they are particularly suited to high-dimensional and unstructured data represented by images, audio, and text. DL models have proved particularly powerful, as their depth enables them to capture non-linear, complex relationships among data. The common architectures are convolutional neural networks (CNNs) for image data and recurrent neural networks (RNNs) for sequential data like time series or natural language. The transformer architecture is a DL model designed primarily for NLP and sequence-related tasks. Transformers were proposed by researchers at Google in 2017. They utilized an “attention” mechanism that freed the model to process relationships across the entire input sequence in parallel rather than sequentially, as in RNNs. ¹⁰ This makes transformers very efficient, especially for dealing with long data sequences. Popular models include BERT, generative pre-trained transformer (GPT), and T5—all based on transformer architecture. ¹¹ In-depth analysis of electronic health records (EHR), radiology reports analysis, genomic data analysis, and medical literature review are readily available. Vision transformers (ViTs) are DL models that adapt the transformer architecture, initially designed for NLP, to image analysis tasks. By dividing images into fixed-size patches and processing them as sequences, they effectively capture long-range dependencies, achieving state-of-the-art performance in image classification. This enables them to be applied in medical image segmentation and classification.^12,13

Radiomics

Radiomics is a technique responsible for extracting and analyzing complex features from medical images, such as X-rays, computed tomography (CT), and magnetic resonance imaging (MRI) scans. Radiomics can characterize tissue properties by examining details within images that may not be visible to the human eye, providing valuable insights for diagnosis, treatment planning, and disease monitoring. ¹⁴ Although initially popular in cancer diagnostics, radiomics has also become relevant for rheumatology, helping to evaluate joint inflammation, bone erosion, and synovial thickness, which are critical information for disease activity and prognosis.

Traditional radiomics uses feature extraction algorithms to identify structural, textural, and shape-based characteristics within the images, and ML algorithms then classify these features. More recently, DL techniques, especially CNNs ViTs, have also been applied in radiomics to automatically learn and classify features from images, further enhancing diagnostic capabilities and predictive modeling.^12–14

Natural language processing

NLP is a sub-field of AI that enables computers to understand, interpret, and generate human language.^15,16

In medicine, NLP can potentially improve the quality of patient care, research, and administrative activities by extracting relevant information from the unstructured clinical notes of EHRs, medical literature, and patient-generated data.^15,17 In addition, NLP can support mining scientific literature and clinical trial data for drug candidates and help describe the mechanism of action, including adverse effects. ¹⁷ One of the most notable advancements in NLP is the development of large language models (LLMs), such as GPT and BERT, which are trained on vast amounts of text data to perform complex language tasks. These models can summarize clinical notes, extract disease-relevant features, assist in automated diagnosis, and support clinical decision-making by understanding the context and semantics of medical language. ¹⁵

Summary of AI and related concepts was in Table 1

EHRs, medical imaging data, genomic and other omics data, clinical trial data, patient-generated data, and medical literature and research data are key data sources that form the backbone of AI applications in medical practice and research. These sources may help to an accurate diagnosis, treatment, and continuous patient monitoring.

OPEN IN VIEWER

Table 1.

Overview of AI methodologies, algorithms, and their applications in medicine.

Methodology	Hierarchical relation to AI	Definition	Example algorithms	Potential medical applications
AI	Core concept	Machines capable of mimicking human intelligence and decision-making by learning from data	–	Diagnostic assistance, decision-making, treatment planning, and patient monitoring
ML	Subfield of AI	Enables systems to recognize patterns and generate predictions from data without programming.	SVM, RnFr, KNN, neural networks	Diagnosis, prognosis, and treatment strategy formulation
Supervised learning	Subtype of ML	Learns relationships from labeled datasets linking inputs to defined outputs.	Linear regression, decision trees, neural networks	Diagnosis, risk prediction based on patient data
Unsupervised learning	Subtype of ML	Identifies hidden patterns or groups from unlabeled data, useful for clustering.	K-means clustering, hierarchical clustering	Patient subgrouping, discovery of new disease subtypes
Semi-supervised learning	Subtype of ML	Combines small amounts of labeled data with extensive unlabeled data, improving accuracy in situations where labeled data is scarce.	Semi-supervised SVM, label propagation	Analysis of rare or complex diseases
Reinforcement learning	Subtype of ML	Models sequential decision-making through feedback received from previous actions.	Q-learning, deep Q-networks	Development of sequential treatment strategies
DL	Subfield of ML	Uses multi-layered neural networks to automatically learn representations from large, complex datasets.	CNNs, RNNs, transformer architectures (BERT, GPT, T5)	Medical image analysis, genomic data analysis, and literature review
ViTs	Subtype of DL	Adapts transformer architecture to image data, processing images as sequences of fixed-size patches to capture complex patterns.	ViT	Medical image segmentation, image classification
Radiomics	Application field utilizing ML and DL	Extracts complex features from medical imaging modalities to characterize tissue properties not visible to the human eye.	Feature extraction algorithms, CNNs, ViTs	Evaluation of joint inflammation, bone erosion, tumor diagnostics, and disease monitoring
NLP	Subfield of AI	Enables computers to understand, interpret, and generate human languages, extracting insights from textual data.	Transformer models (BERT, GPT), text classification algorithms	Clinical note analysis, patient monitoring, clinical coding, drug mechanism research, and literature analysis
LLM	Subfield of DL and NLP	LLMs are DL models trained on massive text datasets to understand and generate human language with context-awareness and reasoning ability.	GPT (OpenAI), BERT (Google), T5, LLaMA (Meta), PaLM (Google)	Summarization of clinical notes, automated report generation, medical question answering, clinical decision support, literature mining, conversational agents for patient interaction

AI, artificial intelligence; CNN, convolutional neural networks; DL, deep learning; GPT, generative pre-trained transformer; KNN, K-nearest neighbors; LMM, large language model; ML, machine learning; NLP, natural language processing; RNN, recurrent neural networks; SVM, support vector machines; ViT, vision transformer.

Predicting the risk of RA

After recognizing that early diagnosis and detection can dramatically change the course of RA, interest in the disease’s early phase and the preclinical phase has grown. ¹⁸ In this regard, predicting the risk of developing RA plays an important role. Thus, identifying these people allows for close monitoring that may help reduce the disease burden from the beginning. The recognition and follow-up of at-risk individuals could also be significant in gaining valuable molecular and clinical data regarding the preclinical phase of RA. As they represent an important group for future potential RA prevention studies, they will likely provide further insight into the RA development process. AI applications may benefit this area, where predictive factors are limited to serological and clinical data. AI models have been utilized to predict RA risk by analyzing serum proteomics, genomics, SNPs, clinical and laboratory data, EHRs, and imaging data.^19–28

A study by Chung et al. ²³ using ML algorithms based on human leukocyte antigen (HLA) reported that alleles HLADQA103:03, DQB10401, and DRB1*0405 could predict the risk of RA development. The following non-HLA genes were studied for their association with RA development: PTPN22, STAT4, TRAF1, CD40, and PADI4; decision tree model showed the best performance, with sensitivity and specificity of around 70%. ²⁴ Besides, IRF5 gene polymorphisms have been shown to influence RA susceptibility, particularly in anti-citrullinated protein antibody (ACPA)-negative and shared epitope-positive patients. ²⁹ A model based on 23 proteins, regardless of ACPA status, predicted the risk of developing RA with 91.2% accuracy. ²⁵ Another study using ML and weighted gene co-expression network analysis identified mitochondrial oxidative stress-related genes with high predictive value in RA: CDKN1A, GADD45B, and MAFF. ²⁶

Fujii et al. developed a Feedforward Neural Network model to predict the progression from seronegative undifferentiated arthritis (UA) to RA using clinical and laboratory data. Utilizing the KURAMA cohort (training dataset, 210 patients with seronegative UA, 27.1% progressed to RA) and the ANSWER cohort (validation dataset, 140 patients, 32.1% progressed to RA), the model incorporated demographic data, acute phase reactants, autoantibodies, and physical examination findings, achieving an area under the curve (AUC) of 0.924 in the training set and 0.777 in the validation set. ²⁷ Salehi et al. aimed to estimate the time and risk of RA onset in 154 preclinical RA patients using various survival ML models, with the Random Survival Forest (RSF) model performing best (C-index: 0.798). By employing SHapley Additive Explanations (SHAP), the model identified baseline rheumatoid factor (RF) and anti-cyclic citrullinated peptide (anti-CCP) antibody levels as key predictors, facilitating patient stratification into low-, medium-, and high-risk groups; however, the model needs further validation. ²⁸

When combined with ML algorithms, ultrasonography may help predict RA development.^30,31 In the study by Daskareh et al., 326 patients with arthralgia were followed up with regular ultrasound imaging over 2 years, and 123 patients developed RA by the end of the follow-up. ML data analysis showed that, even with a need for further validation, radiocarpal synovial thickness, PIP/MCP synovitis, wrist effusion, and RF and anti-CCP levels were associated with the development of RA. ³⁰ Hu et al. aimed to improve early identification of RA progression in 432 UA patients using ML models, particularly a RnFr model integrating clinical data with the 18-joint ultrasound scoring system (US18). The RnFr model demonstrated and validated the highest accuracy and sensitivity, with SHAP analysis highlighting US18 Grade 2 joint counts, total US18 score, and swollen joint count as key predictive variables. ³¹

Diagnosis and differential diagnosis of RA

The diagnosis of RA is clinical, but specific diagnostic and classification criteria assist clinicians in identifying the disease and distinguishing it from other conditions. ³² Besides clinical signs and symptoms, serological markers, especially rheumatoid factor and anti-cyclic citrullinated peptide, ease the diagnostic process. Given the crucial role of early diagnosis and treatment in managing the disease, tools that facilitate and confirm diagnosis effectively have become increasingly important. ³² Data from EHRs, sensor and monitoring data, imaging and biopsy data, and omics data—such as proteomics, genomics, metagenomics, transcriptomics, and metabolomics—have been used in AI applications developed for RA diagnosis and classification.^19,32–34

Predicting the joint damage and disease flares

Assessment of disease activity and prediction of disease activity scores

Accurate assessment of disease activity in RA is essential for guiding treatment decisions, yet conventional indices like DAS28 or Clinical Disease Activity Index (CDAI) can be time-consuming. AI can streamline this process by predicting disease activity using diverse data sources, including EHRs, clinical notes (via NLP algorithms, named entity recognition, and sentiment analysis), laboratory values, imaging, and patient-reported outcomes (PROs), NLP techniques.^115–123

Norgeot et al. developed a longitudinal DL model to predict RA disease activity at the next clinical visit using structured EHR data from two distinct health systems. Despite differing patient populations and treatment patterns, the model achieved high predictive performance at the university hospital (AUROC = 0.91) and maintained reasonable accuracy at the safety-net hospital (AUROC = 0.74). ¹¹⁶ Feldman et al. evaluated whether supplementing Medicare claims data with EHR data improves the estimation of RA disease activity, using DAS28-C-reactive protein (CRP) as the reference standard. Among 300 patients from the BRASS cohort, adding EHR-derived medications and laboratory data to claims data improved model performance, with C-statistics increasing from 0.61 (claims-only) to 0.76 (claims + medications + labs) for classifying high/moderate versus low disease activity. While continuous DAS28-CRP estimation remained suboptimal (adjusted R ²  = 0.18 at best), combining claims and EHR data enabled reasonable discrimination between binary disease activity categories. ¹¹⁷ Spencer et al. reported a ML model to estimate CDAI scores from clinical notes using data from the OM1 RA Registry. The model demonstrated strong performance with an AUC of 0.88, PPV of 0.80, NPV of 0.84, and Spearman’s R of 0.72. ¹¹⁸ Curtis et al. assessed whether longitudinal PROs could serve as surrogates for the CDAI in predicting LDA among RA patients starting new biologics. ML techniques, primarily RnFrs, yielded an accuracy of approximately 80%, highlighting the potential of PRO data for real-world disease activity classification when physician-derived measures are unavailable. ¹¹⁹ However, the main barriers to daily usage include the heterogeneity of EHR data, variability in clinical documentation practices, and difficulty distinguishing between disease symptoms and treatment-related side effects. ¹¹⁵ Moreover, NLP models may struggle with domain-specific language, abbreviations, and inconsistent use of clinical terminology. Incorporating multimodal data integration approaches, expanding training datasets to include diverse patient populations, and enhancing model interpretability through tools like SHAP may overcome the abovementioned issues. ¹¹⁵

Mobile activity trackers may be useful for predicting disease activity and PROs in the RA.^124,125 Rao et al. developed ML models to predict PRO scores in RA patients. They suggested that activity tracker data can effectively monitor health status over time while underlining the need for external validation. ¹²⁴ The weaRAble-PRO study explored how digital health technologies, including smartphones and wearables, could enable the development of innovative RA assessment methods by combining PRO (PRO-RAPID3 Score) with sensor-based data. Over 14 days, health status, mobility, fatigue, and RA-specific symptoms measured with smartwatches and iPhones accurately differentiated between RA patients and healthy controls (F1: 0.807) and improved the detection of RA severity and flare risk when paired with PROs (F1: 0.833 compared to 0.759 for PROs alone). ¹²⁶

Besides NLP algorithms, AI-based MRI and/or ultrasonography imaging can potentially assess RA disease activity.^127–130 Mao et al. developed an AI-based model using dynamic contrast-enhanced MRI to quantify synovitis in RA patients, achieving high disease activity assessment performance with AUCs ranging from 0.941 to 0.965 and Dice scores between 0.557 and 0.615. The model correlated with expert assessments (Spearman’s 0.884–0.927 for joints and 0.736–0.831 for whole hands). ¹²⁷ The RATING system, a DL-based tool using multimodal ultrasound and self-supervised pretraining, improves RA activity scoring accuracy to 86.1% and enhances radiologists’ diagnostic performance from 41.4% to 64.0%. ¹²⁸ However, external validation is needed for both studies.

Two novel composite remote disease activity indexes, ThermoDAI and ThermoDAI-CRP, are based on thermal imaging and ML to assess RA activity without formal joint counts. ThermoDAI and ThermoDAI-CRP showed strong correlations with ultrasound-determined synovitis (GS = 0.52–0.58; PD = 0.56–0.61) and high sensitivity for detecting active synovitis, outperforming PGA and PGA + CRP in relation to CDAI, SDAI, and DAS28-CRP scores (ρ > 0.81).^131,132

Predicting treatment response

In RA treatment, the vast armamentarium of therapeutic options, glucocorticoids, conventional and targeted-synthetic DMARDs (tsDMARDs), and bDMARDs, presents challenges in selecting optimal therapies due to the lack of solid data considering different patient populations, variability in drug responses, and wisdom in the trial-and-error approach. AI applications may lead to better prediction of treatment response of DMARDs and adverse event risk.^133–143

For the prediction of methotrexate response from fecal RNA using RnFr, an AUC of 0.84 was achieved. ¹⁴⁴ To predict adverse events via EHRs, Surendran et al. developed a machine-learning model using RnFr to forecast liver enzyme elevation in RA patients treated with methotrexate. This model demonstrated high accuracy (F1 score: 0.87), with baseline high-normal transaminase levels and elevated lymphocyte and neutrophil counts identified as the primary predictors. ¹⁴⁵ Furthermore, other genome, exome, and biologic material sequencing models predicted methotrexate response with an AUC ranging from 0.7 to 0.8. ¹⁹ Li et al. developed a diagnostic xgboost model using circRNA biomarkers (hsa-circ0002715, hsa-circ0001946) identified through ML techniques for predicting methotrexate-insufficient response and optimizing TNFi treatment in RA patients. ¹⁴⁶ Li et al. aimed to predict 6-month non-remission in 222 DMARD-naïve RA patients starting methotrexate monotherapy, using baseline characteristics from the ARCTIC trial through a ML model. A super learner algorithm combining elastic net, RnFr, and SVM achieved AUC-receiver operating characteristic (ROC) scores of 0.75–0.76 with sensitivities of 0.77–0.81. The model identified the baseline Rheumatoid Arthritis Impact of Disease score as the most powerful predictor. ¹⁴⁷

ANNs and RnFr algorithms have been used to predict responses to TNFi like infliximab, adalimumab, and etanercept in RA patients.^19,141 Based on clinical and molecular data, these models reached a high degree of accuracy, sensitivity, and specificity, demonstrating the capability of ML to guide better treatment decisions. Sonomoto et al. developed inexpensive ML models, explicitly using lasso logistic regression, for predicting CDAI remission 6 months after initiating TNFi in patients with RA. The models, which relied exclusively on routine clinical data input without subjective physician contribution, demonstrated promising accuracy and underlined the feasibility of regional or institution-specific precision medicine approaches. ¹⁴⁸ A study integrating clinical and molecular data attained an AUC of 0.91 in predicting the response to anti-TNF therapy. The investigation examined the impact of RETN polymorphisms on remission in patients with RA treated with TNF-α inhibitors, utilizing ML to forecast remission, revealing the elastic net algorithm showing the best performance in predictions. ¹⁴⁹

Besides the TNFi, several prediction models were published for different biologics. Johansson et al. presented the development and validation of a remission prediction score for tocilizumab monotherapy in RA using RCT and real-world data. The models’ performance in predicting remission was good, with improved discrimination achieved by retraining on real-world data and expanding the variable set. ¹⁵⁰ Rehberg et al. applied a ML-based decision tree (GUIDE) to identify a clinically actionable rule for predicting response to sarilumab in RA. The rule—anti-CCP positivity combined with CRP >12.3 mg/L—was derived from the MOBILITY trial and validated across multiple trials (MONARCH, ASCERTAIN), showing strong predictive value for ACR and DAS28-CRP responses. Although the performance was limited in TNFi-refractory patients (TARGET), the approach demonstrated potential for guiding individualized biologic selection between sarilumab and adalimumab. ¹⁵¹ Kalweit et al. applied DL to cluster RA patients and investigate their response to b/tsDMARDs. Results showed that patients with at least two conventional synthetic DMARDs (csDMARDs) and prednisone at the start of b/tsDMARD, male patients, and those with lower disease burden had a significantly better response to tocilizumab compared with adalimumab, with hazard ratios (HRs) ranging from 3.64 to 8.44. In contrast, seronegative women not using prednisone at initiation and seropositive women with greater disease burden and longer disease duration had a higher risk of non-response to golimumab compared with adalimumab: HRs 2.36 and 5.27, respectively. ¹⁵² Alten et al. applied ML models, particularly a gradient-boosting classifier, to predict 12-month abatacept retention in 5320 RA patients from the ACTION and ASCORE trials, achieving an AUC of 0.620 and an F1 score of 0.659. Key predictive factors identified through SHAP analysis included low BMI, ACPA positivity, low patient global assessment, younger age, and better functional status. ¹⁵³

A recent scoping review by Benavent et al. ¹⁵⁴ aimed to evaluate the application of AI in predicting treatment response to DMARDs in patients with RA. Out of 66 studies included, 32 specifically focused on RA. The review assessed various AI techniques, including ML models such as RnFrs, SVMs, NNs, and DL approaches. These methods analyzed diverse data types, including clinical parameters, imaging, genetic, and transcriptomic data. ¹⁵⁴ Overall, the included models demonstrated moderate to high predictive performance, with reported AUC values typically ranging between 0.70 and 0.90 and some models exceeding 0.90. However, the review also highlights the need for further validation and standardization of models before widespread clinical adoption. ¹⁵⁴ Messelink et al. performed text mining and feature weighting on real-world data of RA patients to identify those with D2T RA. They also developed a predictive model assessing the risk of developing D2T RA before starting the first biological or tsDMARD. It correctly identified 79% of the D2T cases, with an AUROC of 0.73. ¹⁵⁵ Besides all these studies, a systematic review has been conducted to appraise the performance and transparency of ML models that predict RA treatment response. Among 29 analyzed studies, the adherence to reporting standards was moderate, as assessed by TRIPOD (42.9%–45.6%), and most of the studies had unclear or high risk of bias (79.3%). Although ML methods are increasingly applied in RA, the findings underline the need for improved transparency, calibration, and better handling of missing data to enhance the reliability of these models. ¹⁵⁶

Drug development and repurposing

AI, particularly ML, may play a significant role in drug repurposing for RA by facilitating the identification of novel therapeutic candidates.^157–161 Various ML techniques, including text mining, structure-based, signature-based, and network-based approaches, are employed to analyze vast datasets, extract patterns, and predict potential drug candidates.^{157,160,162–167} For instance, ML models are used to prioritize genes, predict drug-target interactions, and identify effective inhibitors against inflammatory targets, even for the production of virtual twin synovial joints.^{157,167–170} Integrating ML and NLP with conventional methodologies enhances the efficiency of the drug discovery process by rapidly identifying promising candidates and minimizing experimental costs. ¹⁷¹ Moreover, ML-based approaches improve the accuracy of predictions by combining multi-omics data, which is essential for addressing the complexity of RA pathogenesis. ¹⁵⁷

Prabha et al. utilized ML to identify novel and potential TNFi, addressing the limitations of existing treatments for RA, such as cost, administration challenges, and side effects. The best performance among all the tested algorithms was exhibited by the RnFr model, which yielded an accuracy of 87.96% and a sensitivity of 86.17% for the combination of 1D, 2D, and fingerprint features; this is a novel application of ML in predicting TNFi and offers a cost-effective and efficient approach to drug discovery for RA. ¹⁷² Tariq et al. integrated a multi-omics approach combining transcriptomics and epigenomics data with bioinformatics and ML techniques to identify 18 novel multi-evidence genes (MEGs) associated with RA. Twelve of these MEGs, previously unlinked to RA, were found to be part of critical protein-protein interaction (PPI) networks and RA-related pathways, exploring the possibility of AI-driven analysis for discovering new therapeutic markers. ¹⁷³

However, it should be underlined that effective implementation of AI requires comprehensive, connected datasets and validation at the individual patient level. ¹⁷⁴

Assessment of co-existing conditions

RA is associated with several comorbidities, making the early identification of at-risk patients critical for preventing and managing these complications. ML and DL algorithms have been applied to identify and stratify comorbidities. Crowson et al. analyzed comorbidity patterns in RA patients using four clustering methods and compared them to non-RA individuals. Although RA patients had a higher overall comorbidity burden, the comorbidity patterns were similar across groups, and the instability of clustering methods underscored the need for caution when interpreting results from a single approach. ¹⁷⁵ Solomon et al. used multiple ML algorithms (K-mode, K-mean, regression-based, and hierarchical clustering) to cluster 24 comorbidities in 11,883 RA patients from the CorEvitas registry over 6 years. While ML-derived comorbidity clusters produced similarly strong predictive models for CDAI and health assessment questionnaire (HAQ)-DI as those using individual comorbidities, the clusters were often clinically difficult to interpret. ¹⁷⁶

The complexity of the multimorbidity network in RA has only recently been described as involving a complex interplay between various organs and systems affected by systemic inflammation, which is crucial to consider within the network concept. ¹⁷⁷ Key nodes in this network include cardiovascular diseases (CVDs), metabolic comorbidities, musculoskeletal system complications, infections, malignancies, mental health disorders, gastrointestinal and hepatobiliary disorders, and pulmonary complications. ¹⁷⁷ These interrelated conditions complicate disease management and significantly impact prognosis and quality of life; therefore, a personalized and multidisciplinary approach is necessary. ML and DL approaches can elucidate the underlying patterns and possibly the key factors in understanding this phenomenon. Utilizing ML, England et al. identified patterns of multimorbidity among patients with RA related to cardiopulmonary, cardiometabolic, and mental health and chronic pain disorders. Compared to individuals without RA, patients with RA had significantly higher odds of these multimorbidity patterns, particularly concerning mental health and chronic pain disorders. ¹⁷⁸

Important note to readers

Whether AI or classical statistics is used to generate prediction models, the generalizability of these models remains an important research question. The performance of these models should be evaluated in prospective designs across groups with varying ethnic and social backgrounds.

Predicting the risk and detecting axSpA

From a basic science perspective, Han et al. aimed to identify ankylosing spondylitis (AS)-specific mRNA biomarkers from whole blood using an ML-based feature selection method, support vector machine-recursive feature elimination (SVM-RFE), applied to the mRNA expression profile (GSE73754) from the GEO. ML identified 13 key mRNAs, with IL17RA, Sqstm1, Picalm, Eif4e, Srrt, Lrrfip1, Synj1, and Cxcr6 validated as significant for AS diagnosis. Notably, Cxcr6, IL17RA, and Lrrfip1 were associated with AS severity. ²⁰⁸ In another study, Alber et al. used single-cell CITE-seq technology and ML models to analyze peripheral blood mononuclear cells in AS patients, achieving an AUROC of >0.95 for AS classification. Overexpression of CD52, TNFSF10, IL-18Rα, and cytotoxic genes in specific cell subsets was identified. ²⁰⁹

Novel AI models are being developed to detect axSpA early, integrating various data sources to strengthen diagnostic accuracy. Jia et al. proposed the improved optimization algorithm SCJAYA (Salp Chain JAYA Algorithm), which integrates the salp swarm’s foraging behavior with cooperative predation strategies to improve the diagnosis of AS. The binary version of SCJAYA was applied to feature selection in the bSCJAYA-FKNN (fuzzy K-nearest neighbor) classifier, which showed excellent performance metrics: 99.23% accuracy and 99.52% specificity on both AS-specific and public datasets. ²¹⁰ Deodhar et al. and Zhang et al. developed models investigating medical and pharmacy claims data and routine blood tests for high-risk disease identification. At the same time, Ye et al. have extended this approach by incorporating MRI findings with clinical risk factors into a predictive clinical-radiomics nomogram. These models utilize a wide range of biomarkers to facilitate diagnostics and suggest efficient strategies for early diagnosis of axSpA.^211,212 Using a machine-learning approach, Zhu et al. developed a diagnostic model for AS using routine blood tests, liver function, and kidney function: seven factors were identified, including ESR, red blood cell count, mean platelet volume, albumin, aspartate transaminase, and creatinine. These were integrated into a nomogram that demonstrated high accuracy, with an AUC of 0.878 in the training and 0.823 in the validation cohort. ²¹³ To address the clinical heterogeneity of axSpA, Sun et al. identified two distinct patient subtypes with unsupervised ML. Their predictive model integrated biomarkers like CRP, neutrophils, and monocytes and reached a high AUC of 0.983. This stratification may help design treatment strategies for patient profiles. ²¹⁴ Besides, Wen et al. used ML algorithms, including LASSO regression, SVM-RFE, and RnFr, to analyze gene expression profiles from the GEO database. They identified IL2RB and ZDHHC18 as potential diagnostic biomarkers for AS. The diagnostic model achieved AUC values exceeding 0.84 across multiple validation datasets. ²¹⁵ From the same database, Li et al. found that pyroptosis-related genes may be a potential biomarker and drug target for AS. ²¹⁶

Kennedy et al. aimed to develop an ML model to predict AS using primary care health records from the Secure Anonymised Information Linkage databank. The model was trained on clinical, demographic, and laboratory data, including symptoms, comorbidities, medication use, diagnostic codes, and blood test results. Using decision trees built on principal component analysis, the model was trained on a cohort of 543 male and 250 female AS patients, along with 2900 male and 2900 female controls, with additional tests on 1559 AS patients and 4500 controls. The model showed high prediction accuracy for males, with a PPV of 76.69%, and for females, 78.3%. However, applying the model to a general population with a low prevalence of AS reduced the PPV to 0.33% for males and 0.25% for females. ²¹⁷

NLP may aid in improving the classification of axSpA by using EHRs.^218–220 Benavent et al. used NLP to extract clinical data from the EHRs of 4337 patients diagnosed with spondyloarthritis at a large hospital. The NLP system extracted information on demographics, SpA subtypes, comorbidities, and treatment patterns, including methotrexate and adalimumab as the most commonly used csDMARD and bDMARD, respectively. The technology was highly reliable, with F-1 scores above 0.80 for data extraction. These findings illustrate the potential for NLP to enhance SpA patient profiling and improve clinical management by efficiently analyzing unstructured EHR data. ²¹⁸ However, we as clinicians need to do better to help AI(!). In the SpAINET study, NLP of EHRs from axSpA and PsA patients treated with bDMARDs in Spain was performed. ²²¹ Findings suggest a critical gap in real-world data for both diseases concerning the assessment and management of disease activity. Notably, there was a very low record of any activity score concerning the disease at the start of biologic therapy, pointing out the need for enhanced clinical documentation practices. The SpAINET study underlines the potential of AI and sophisticated data analysis to cover existing shortcomings in axSpA and PsA management for more active and specialized patient management. ²²¹

An important problem in daily clinical practice is differentiating axSpA from other causes of chronic back pain. Redeker et al. reported a RnFr-based ML model to distinguish axSpA from non-axSpA in patients with chronic back pain using clinical data from 939 patients (659 axSpA, 280 non-axSpA). The model achieved high performance with an accuracy of 0.9234, sensitivity of 0.9586, specificity of 0.8438, and ROC-AUC of 0.9717, with HLA-B27, insidious onset of back pain, and SIJ erosions being the most important predictors. ²²²

Insights into imaging techniques for diagnosis

Radiography, or X-ray, remains a major modality for diagnosing axSpA due to its wide availability and affordability. ²⁰⁷ The modified New York criteria for diagnosing AS, one of the subtypes of axSpA, have traditionally relied on the presence of radiographic sacroiliitis, which is typically defined as grade 2 bilaterally or grade 3 unilaterally. Plain radiographs have several limitations, the most significant of which is their insensitivity early in the disease process. ²⁰⁷

AI models can enhance the sensitivity of radiographic analysis.^{13,103,223–225} A CNN model based on the ResNet architecture, trained to differentiate between normal and sacroiliitis-affected images, showed a very high AUC of 0.97 on a validation dataset, comparable to expert radiologist performance. ²²³ Another study applied transfer learning techniques on networks like VGG-16, ResNet-101, and Inception-v3, achieving high sensitivity and specificity. ²²⁴ A semi-supervised DL model diagnosed AS using pelvic radiographs, with an expert-level performance with only 10% labeled data. Validated on an independent test set of 982 images, the model reached 89.1% accuracy, 86.5% recall, and 85.9% precision, while interpretability analysis confirmed its clinical reliability. ²²⁶ Koo et al. developed a DL model to automatically grade vertebral corners on spinal radiographs for Modified Stoke Ankylosing Spondylitis Spinal Score (mSASSS) scoring in AS. The model achieved high accuracy (91.6%); however, further validation is needed. ²²⁷ Similarly, Chen et al. proposed a DL model to automate mSASSS using lateral X-ray images of the cervical and lumbar spine in patients with AS. The model was trained on data from 554 patients and reached an accuracy of 86.5% in detecting vertebral structural damage and scoring individual vertebral corners. Both studies need further validation. ²²⁸ Further, integrating the anatomical landmark knowledge gathered by the X-ray into the model enhanced the capacity of the AI model. ¹⁰³

Ultrasound plays a role in assessing peripheral inflammation, including enthesitis and synovitis, which are common manifestations of SpA. The limited use of ultrasound in axSpA diagnosis is attributed to the dependence on the interpreter’s expertise; therefore, it is not easily standardizable. An AI-driven regression model for calculating the Madrid Sonographic Enthesis Index, an indicator of enthesitis severity, demonstrates that AI may bring consistency to ultrasound interpretation, reducing variability and improving reliability in the clinical setting. ²²⁹ However, readers should be cautious as this study has not been published in full-text, and results should be interpreted cautiously. The role of ultrasound in axSpA remains complementary, and its wider application is pending more extensive validation studies.

CT imaging offers superior sensitivity and specificity compared to radiography, especially for identifying bone erosions, sclerosis, and ankylosis in the SIJs. However, the images produced by CT are complex and require expertise for interpretation. AI-driven models have further advanced CT-based axSpA analysis, allowing for a more detailed assessment of structural changes. Liu et al. ²³⁰ proposed a model that combines U-Net segmentation with radiomics, predicting an accuracy of 87.3% in diagnosing sacroiliitis from CT images, nearly 10% higher than that achieved with traditional methods. Castro-Zunti et al. ²³¹ examined additional applications of DL using an InceptionV3 architecture for detecting sacroiliac erosions and structural damage, surpassing musculoskeletal radiologists in sensitivity and specificity. In a multi-center study, Van Den Berghe et al. trained a DL model to identify erosions and ankylosis on CT scans by concentrating on the cortical edges of the SIJs. Utilizing Grad-CAM++ visualization, this model attained high sensitivity and specificity, showcasing AI’s potential to establish standardized, reproducible CT-based diagnostic processes in axSpA. ²³²

MRI provides unique insights into bone and soft tissue, which are imperative for the early diagnosis of axSpA, especially when radiographic findings are inconclusive. ²⁰⁷ Early inflammatory changes, such as BME, can be identified on MRI and are considered important markers for axSpA. ²³³ However, BME assessment is subject to significant intra- and inter-observer variability, highlighting the need for standardized approaches.

AI has significantly improved the detection of BME and/or structural lesions and synovitis on MRI.^234–246 Recently, Nicolaes et al. evaluated a previously trained deep-learning algorithm for detecting inflammation in SIJ MRIs from 731 axSpA patients (326 nr-axSpA, 405 r-axSpA) across two trials (RAPID-axSpA, C-OPTIMISE). The algorithm, blinded to clinical data, achieved a sensitivity of 70%, specificity of 81%, PPV of 84%, NPV of 64%, Cohen’s kappa of 0.49, and absolute agreement of 74% compared to expert readings. ²³⁴ A DL model proposed and tested by Bordner et al. utilized MRI data to identify BME in SIJs, facilitating the prediction of active sacroiliitis according to the ASAS criteria in patients with chronic inflammatory back pain. The model demonstrated excellent performance, as evidenced by an AUC of up to 0.98 and a specificity of 100% in external validation using data from the DESIR cohort. ²⁴⁵ Tenorio et al. reported that radiomic features extracted from fat-suppressed MRI sequences (SPAIR and STIR) are significantly associated with active sacroiliitis and clinical markers of SpA. The models were validated against expert musculoskeletal radiologists, showing strong concordance. ²⁴⁷ Ozga et al. ²⁴⁸ created a model with a Dice coefficient of 0.98 for segmenting lesions, achieving near-perfect agreement with expert radiologists. Other studies identified active sacroiliitis using DL networks and combined models with the Spondyloarthritis Research Consortium of Canada (SPARCC) scoring system to ensure accurate grading. Zhang et al. merged radiomics from MRI with clinical data, achieving excellent diagnostic performance in distinguishing axSpA from mimics. For this purpose, they focused on structural markers of erosions, subchondral lesions, and sclerosis. Structural lesions such as erosions and subchondral abnormalities are crucial features of axSpA diagnosis on MRI but can be challenging to detect. ²⁴⁹

Radiomics and clinical data have effectively distinguished between axSpA and non-axSpA, and inflammatory and structural changes have been combined to enhance early diagnosis using AI models. ²³⁹ For instance, Zhang et al. developed and validated a prediction model for distinguishing axSpA from non-axSpA using SIJ-MRI imaging features and clinical risk factors, applied to 942 patients (707 axSpA, 235 non-axSpA). The combined clinical-imaging model using tabular neural network (TabNet) achieved the highest diagnostic performance with an AUC of 0.93, outperforming other ML models, with key predictors including joint space changes, erosions, HLA-B27 positivity, and CRP values. ²⁵⁰ AI tools may also provide benefits in diagnosing axSpA in HLA-B27-negative patients. Lu et al. reported NegSpA-AI, a DL tool using SI MRI and clinical data, which achieved AUCs of 0.878, 0.870, and 0.815 across internal, external, and prospective test sets, respectively. It significantly improved diagnostic accuracy, sensitivity, and specificity for junior radiologists and rheumatologists by up to 11.5%, 13.3%, and 20.6% in tests involving 454 HLA-B27-negative axSpA patients. ²⁵¹ In addition to BME, AI applications may provide valuable assessments of structural lesions. Li et al. reported a UNet-based DL model for segmenting fat metaplasia on SIJ MRI and a classification model to distinguish axSpA from non-axSpA using data from 706 patients. The segmentation model achieved Dice Similarity Coefficients of 81.86% (internal) and 85.44% (external), while the classification model reached AUCs of 0.876 (internal) and 0.799 (external). Model assistance also improved radiologists’ performance. ²⁵²

An attention-deep neural network has also assessed spinal inflammation in STIR MRI images of axSpA patients, achieving an AUC-ROC of 0.87, sensitivity of 0.80, and specificity of 0.88, comparable to radiologists’ assessments. ²⁵³ For hip involvement, integrating clinical data with MRI features like BME and effusion, AI models demonstrate expert-level accuracy, improving early detection and grading. ²⁵⁴ Additionally, AI models have employed denoising and intensity standardization algorithms to address heterogeneity in MRI protocols, improving diagnostic reliability across various imaging platforms. This harmonization is essential for enabling accurate, cross-center diagnostic comparisons and establishing a new standard for MRI in axSpA care.

Positron emission tomography (PET)-CT-based low-dose computed tomography (LDCT) may give insights into the pathogenesis of axSpA. Recently, van der Heijden developed and validated two automated segmentation methods using AI techniques—morphological operations and a multi-atlas-based approach—for detecting lesions in posterior spinal joints, SIJs, and disco vertebral units in SpA patients using Na[18F]F PET/LDCT. The atlas-based method, employing ML for enhanced segmentation accuracy, achieved superior performance with a maximum AUC of 0.90, making it the preferred approach for PET-based lesion analysis. ²⁵⁵

Advanced models integrating MRI radiomics, other imaging techniques, clinical biomarkers, and predictive analytics represent a modern, comprehensive approach to managing axSpA. For instance, merging structural and inflammatory changes from MRI with clinical data enables better disease monitoring and classification, allowing for early detection and treatment planning. However, further validation in diverse patient populations is necessary for these AI-driven tools to become dependable components of routine care in axSpA.

Predicting radiographic progression, treatment response and follow-up

The potential of AI is not limited to imaging since predictive models showed promise in predicting the disease course and treatment strategies in axSpA.^256–259 ML algorithms have pointed out prognostic factors, including baseline inflammatory markers and structural damage, which may help the clinician guide the treatment approach.

Recently, Dorfner et al. studied the role of anatomy-centered DL models for detecting radiographic sacroiliitis and predicting disease progression in axSpA using pelvic radiographs from four cohorts (total n = 2422). Comparing standard and anatomy-centered models, the latter achieved higher AUCs (0.899, 0.846, 0.957) and accuracies (0.821, 0.744, 0.906) across three test datasets. High-risk patients identified by the anatomy-centered model had an odds ratio of 2.16 for radiographic sacroiliitis progression within 2 years, indicating improved generalizability and predictive capability. ²⁶⁰ Koo et al. utilized the RnFr algorithm to analyze time-series clinical features to predict radiographic progression AS. Their model yielded a mean accuracy of 73.73% and an AUC of 0.79, reflecting robust predictive performance. ²⁵⁷ Baek et al. ²⁵⁸ developed a machine-learning model using clinical data to predict radiographic progression in axSpA patients; the model performed better than the generalized linear or input model to predict MSA score (RMSE = 2.83). Joo et al. used agglomerative hierarchical clustering and ML models to classify 412 axSpA patients into three phenogroups with distinct clinical characteristics and radiographic progression rates. Phenogroup 2 (male smokers) had the worst progression, while Phenogroup 3 (uveitis-only) had the least. ²⁵⁹ However, Garofoli et al. ²⁶¹ compared ML algorithms with traditional statistical models, and comparable predictive accuracies were found between the two approaches. This would mean that though AI offers promising tools, its superiority over conventional methods of early disease progress is not yet well established.

ML can be a valuable tool in predicting treatment responses for axSpA. Using baseline demographic and laboratory data, Lee et al. developed an ANN model to forecast early TNFi use in patients with AS. Compared to logistic regression, SVM, RnFr, and XGBoost models, the ANN model reached an AUC of 0.783. The strong predictors identified were CRP and ESR levels. ²⁶² Fernández-Carballido et al. assessed sex differences and other factors influencing TNFi response in 969 axial SpA patients (315 females, 654 males) using statistical and AI-based analyses from the BIOBADASER registry. Analyses highlighted age at treatment initiation as the primary factor linked to poor TNFi response, especially when combined with female sex or cardiovascular risk factors. Females showed lower BASDAI50 responses and lesser reductions in ASDAS-CRP by the second year, with older age also contributing to unfavorable outcomes. ²⁶³

A recent scoping review by Benavent et al. ¹⁵⁴ also evaluated the application of AI in predicting treatment response (anti-TNF and IL-17 blockers) in patients with SpA, including axial and peripheral subtypes. Various AI methods such as SVMs, RnFr, logistic regression, and DL models were employed, analyzing data from clinical, imaging (particularly MRI), and molecular sources. ¹⁵⁴ Overall, the performance of AI models in SpA studies was promising, with reported AUC values generally ranging from 0.70 to 0.90, indicating moderate to high predictive accuracy. ¹⁵⁴ However, the review emphasized that more robust validation in independent cohorts and improved data standardization are needed before these models can be reliably integrated into clinical practice. ¹⁵⁴

Studies on the extra-articular involvements, comorbidities, and long-term follow-up of axSpA are largely missing. Bioinformatic analysis using ML algorithms found ST8SIA4 and lysosomal pathways as a common pathogenic gene and mechanism in AS and atherosclerosis. ²⁶⁴ Regarding cardiovascular risk assessment in axSpA, recent data suggested that the combination of QRISK3 and SCORE2 yielded the most accurate predictive performance. ²⁶⁵ Navarini et al. evaluated the performance of seven cardiovascular risk algorithms and three ML techniques for predicting CV events in AS patients. Among the traditional algorithms, SCORE and RRS performed best (C-statistic: 0.71 and 0.72), whereas RF was the best among ML techniques tested, with an AUC of 0.73. Feature analysis identified CRP as the most important predictor—a different profile from traditional CV risk factors. ²⁶⁶ Although AI can enhance cardiovascular risk assessment by integrating clinical, imaging, and biomarker data to develop more accurate prediction models beyond traditional risk calculators, the current literature has no solid data on this issue.

AI can support mortality risk assessment by analyzing complex, longitudinal datasets—including disease activity scores, comorbidities, treatment patterns, and imaging—to identify high-risk profiles with greater precision. This could facilitate early intervention strategies and better management plans to reduce long-term morbidity and mortality. However, currently, there is no robust evidence to explore the usage of AI in this setting.

Mobile applications and clinical decision support systems

Mobile applications may facilitate monitoring and enhancing physical activity,^267,268 mobility,^269,270 self-reporting symptoms, disease activity, and real-time communication with healthcare providers. Gossec et al. applied ML to activity tracker data to predict patient-reported RA and axial SpA flares. Over 3 months, physical activity data (steps per minute) and weekly flare assessments from 155 patients were analyzed. The ML model exhibited high accuracy in detecting flares, with sensitivity at 96%, specificity at 97%, and positive and NPV at 91% and 99%, respectively. ¹¹² A proof of concept was established in one study using a smartphone application to self-monitor disease activity in axSpA patients, showcasing how this might improve patient engagement and facilitate tight control strategies. ²⁷¹

The CDSS integrates patient data, clinical guidelines, and predictive analytics to help providers make informed decisions. For instance, in axSpA, CDSS will analyze PROs, imaging findings, and laboratory results to provide treatment recommendations for personalized care. Using AI-enhanced CDSS might improve diagnostic accuracy, optimize treatment strategy, and reduce variability in clinical practice.

Looking ahead, AI can facilitate the management of axSpA by continuous patient monitoring and CDSS. These platforms may also promote patient engagement, leading to better adherence to treatment protocols and proactive health management.

Important note to readers

Whether AI or classical statistics is used to generate prediction models, the generalizability of these models remains an important research question. The performance of these models should be evaluated in prospective designs across groups with varying ethnic and social backgrounds.

Diagnosis and progression from psoriasis to PsA

For screening purposes for patients applying to primary care centers, Reed et al. aimed to generate and validate a smartphone application using ML algorithms to screen for different forms of hand arthritis, including PsA, across multiple rheumatology practices. The app utilized a CNN model trained on 1577 hand images and integrated survey data, achieving 95.2% accuracy, 76.9% precision, 90.9% recall, and 95.8% specificity for PsA detection; however, results need further external validation. ²⁷² McArdle et al. aimed to identify serum protein biomarkers that differentiate early inflammatory arthritis patients with PsA from those with RA using nano-LC-MS/MS, aptamer-based assays, and multiplexed antibody assays. ML analysis achieved an AUC of 0.94 for nano-LC-MS/MS, 0.69 for bead-based immunoassays, and 0.73 for aptamer-based analysis, while validation using RnFr models reached AUCs of 0.79 and 0.85, underlining the need for external validation. ²⁷³ Xu et al. ²⁷⁴ develop ML models for PsA diagnosis and progression risk prediction using multimodal clinical data from 3961 patients, including demographic, laboratory, clinical, and treatment-related variables. ML algorithms, particularly XGBoost and AdaBoost, were applied to integrate these diverse data sources, achieving an AUC of 0.87 for PsA diagnosis and 0.80 for predicting PsA progression risk. Although these findings demonstrate the potential of multimodal AI approaches, further validation using larger and more diverse datasets is necessary. ²⁷⁴ Love et al. developed an AI-based algorithm that combined structured data with unstructured clinical notes using NLP. This strategy identified 31 PsA-related predictors and achieved a PPV of 90%–93% with a sensitivity of 87%, outperforming the PPV of 57% obtained using a single PsA code alone. ²⁷⁵ The PredictAI™ ML tool was evaluated for the early identification of undiagnosed PsA patients using data from Maccabi Healthcare Service spanning 2008–2020. The model, developed using data from 1 to 3 years before the diagnosis of PsA, displayed 90% specificity within the psoriasis cohort, with sensitivities of 51% 1 year and 38% 4 years prior and a PPV of up to 36.1%. This demonstrated the model’s utility in the early identification of PsA and allowing timely interventions to prevent joint damage and enhance patient outcomes. ²⁷⁶ Recently, Rudge et al. developed a ML model to detect PsA early and identify important clinical indicators in primary care. Using EHRs from the Clinical Practice Research Datalink, models were created employing Bayesian Networks (BN) and RnFr algorithms. ²⁷⁷ A cohort of 122,330 psoriasis patients was examined, with 2460 developing PsA; the best BN model achieved an AUC of 0.823 and a Precision-Recall Area Under the Curve (PRAUC) of 0.221, while the RnFr model reached an AUC of 0.851 and a PRAUC of 0.261. The effectiveness of the models may be limited by incomplete primary care data, delayed PsA diagnosis, and challenges in model calibration, highlighting the need for further validation and refinement. ²⁷⁷

A computational automated prediction tool for PsA in psoriasis patients was developed. ²⁷⁸ Researchers defined nine new loci influencing psoriasis and its clinical subtypes by analyzing data from six cohorts comprising more than 7000 genotyped PsA and psoriasis patients. Integrating a molecular signature composed of 200 genetic markers resulted in a remarkable AUC of 0.82, reflecting good predictive performance. ²⁷⁸ Mulder et al. applied flow cytometry combined with ML in immune cell profiling to delineate PsO from PsA. Their best RnFr classification model exhibited robust performance, with an AUC of 0.95. ²⁷⁹ Key differences included enriching differentiated CD4+ T-cells and decreasing memory T-cells and monocytes in the peripheral blood of PsA patients. Moreover, some immune subsets showed an association with joint scores in PsA—a potential tool for early PsA diagnosis in PsO patients, aiding in the timely referral and treatment of PsA. ²⁷⁹ Jalali-Najafabadi et al. ²⁸⁰ developed feature selection and PsA risk prediction models using genetic data from 1462 PsA cases and 1132 cutaneous-only psoriasis cases, with validation in the UK Biobank dataset. Seven supervised ML methods were trained using stratified nested cross-validation, and the best model achieved an AUC of 0.61 for internal cross-validation, 0.57 for the internal holdout set, and 0.58 for the external dataset. While HLA_C_*06 was initially identified as the most informative genetic variant, mitigating confounding features revealed HLA-B27 as the most important. Still, individual features alone showed only moderate predictive accuracy, suggesting that integrating multiple HLA features improves prediction performance. ²⁸⁰ The performance of a CNN model, utilizing temporal diagnostic and prescription data from Taiwan’s National Health Insurance Research Database, in predicting a 6-month PsA risk in PsO patients was assessed. The overall performance based on AUC reached 0.70, with a sensitivity of 0.80 and an NPV of 0.93. This model has shown good performance in identifying PsO patients at higher risk of subsequently developing PsA. ²⁸¹ Another study from Switzerland investigated the predictability of PsA in patients with psoriasis. Swiss Dermatology Network on Targeted Therapies data was analyzed using gradient-boosted decision trees and mixed models, focusing on age, Psoriasis Area and Severity Index (PASI), physical well-being, and nail psoriasis severity. ²⁸² Despite the limitations due to modest sample size and generalizability problems, the models achieved AUROCs of up to 73.7%, indicating moderate predictive performance. ²⁸²

Disease activity assessment

Using ML models, Choksi et al. aimed to identify serum metabolomic markers associated with skin disease activity in PsA patients. Serum samples from 150 PsA patients were analyzed, and predictive models achieved an AUC of 0.813 for distinguishing low and high disease activity. Potential biomarkers identified include phospholipids, bile acids, and oxylipins, but further validation is required. ²⁸³ Another study analyzed serum metabolites associated with PsA disease activity using mass spectrometry and ML, achieving classification models with AUC up to 0.818. Key metabolites, such as lysophosphatidylcholine and sphingomyelin, were identified as potential biomarkers for PsA activity. ²⁸⁴ A recent study applied ML to explore the components of the DAS28-CRP score for patients with PsA. Tenderness in the right index finger metacarpophalangeal joint proved to be the most informative indicator for staging PsA activity. ²⁸⁵ In a multicenter observational study of 158 patients with recent-onset PsA, key factors associated with a psoriatic arthritis impact of disease (PsAID) score of ⩾4, indicating high disease impact, were identified via ML algorithms. They found that higher HAQ scores, increased pain levels, lower educational attainment, and higher physical activity were significantly associated with elevated PsAID scores, with the mean accuracy measure surpassing 85%. ²⁸⁶ The same study group identified severe disease predictors, including high global pain, CRP levels, hypertension, gluteal/perianal psoriasis, and the need for synthetic DMARDs. Using ML models such as XGBoost and RnFr achieved more than 80% accuracy of predictions for severe cases, thus pointing out the rigorous management of pain, inflammation, and comorbidities in the PsA. ²⁸⁷ As the other part of the study, Queiro et al. aimed to develop ML models to predict minimal disease activity (MDA) in the same patient population. A RnFr algorithm revealed the most predictive variables: global pain, PsAID score, patient global assessment of disease, and HAQ score, achieving a prediction accuracy of 85.94% for MDA. ²⁸⁸

For assessing the articular activity status remotely, Webster et al. aimed to validate a smartphone-based app (Psorcast) for assessing cutaneous and musculoskeletal signs of psoriatic disease using digital measures and ML models. Participants included 104 individuals with PsA, PsO, or healthy controls, and digital assessments were compared with clinical evaluations by rheumatologists and dermatologists. ML models analyzing hand photos achieved 76% accuracy in detecting nail PsO. At the same time, the Digital Jar Open test for physician-assessed upper extremity involvement (tender joint or enthesitis) reached an AUROC of 0.68. Although results are acceptable for this proof-of-concept study, further validation in larger cohorts is necessary before clinical application. ²⁸⁹ Nail psoriasis, which affects around 50% of psoriasis patients, is difficult to quantify because of heterogeneous nail involvement, making manual scoring with the Nail Psoriasis Severity Index (NAPSI) impractical in clinical settings. Using BEiT, the researchers developed a neural network-based system for fully automated modified NAPSI scoring. The system achieved an AUC of 88% and a Pearson correlation of 90% with human annotations. This openly available system enables efficient, observer-independent evaluation of nail psoriasis severity in clinical use. ²⁹⁰ Ricardo et al. ²⁹¹ aimed to assess the performance and inter-reader agreement between AI-determined NAPSI scores using a CNN and dermatologist-assigned scores for psoriatic fingernail images. In a dataset of 240 images, the AI model achieved excellent interclass correlation coefficients (ICCs) of 0.81 for overall NAPSI, 0.75 for matrix NAPSIm, and 0.81 for bed NAPSIb, outperforming dermatologist agreement which had ICCs of 0.43, 0.56, and 0.53, respectively; underlining the potential of CNNs to enhance accuracy and reliability in NAPSI scoring, though limitations include limited sample size and potential underrepresentation of racial/ethnic minorities. ²⁹¹

Prediction of treatment response

Using AutoML, ML models were developed to predict the change in therapy, treatment response, and disease progression in psoriasis vulgaris and PsA with high accuracy (AUC up to 0.92). The main predictors included baseline PASI scores, modification of pruritus, initial therapeutic agents, and psychological factors such as depression and anxiety. ²⁹² In a multimodal risk prediction study, treatment response analysis revealed that IL-17 inhibitors were effective for severe psoriasis, whereas methotrexate showed limited efficacy in preventing PsA progression (HR = 0.139; 95% CI, 0.040–0.474; p-value = 0.00024). ²⁷⁴ Gottlieb et al. combined ML with evidence-based medicine to investigate baseline data from 2148 PsA patients. They applied Bayesian elastic net regression on 275 predictors. They found variables associated with a good response to secukinumab, being anti-tumor necrosis factor-naïve, treated with one prior anti-tumor necrosis factor agent, not receiving methotrexate, with enthesitis at baseline, and with shorter PsA disease duration, allowing precise medicine strategies. ²⁹³ Post hoc ML analysis of 10 clinical trials of secukinumab (FUTURE, MAXIMISE, and MEASURE), Baraliakos et al. identified that PsA patients with certain characteristics (clusters) responded better to secukinumab 300 mg compared to 150 mg. Improved response to higher doses was observed, particularly in patients with axial involvement, higher articular burden, and those with pronounced psoriasis. Specifically, clusters representing overweight patients with extensive joint involvement (knees, shoulders, elbows, wrists) and psoriasis, as well as those with axial manifestations and oligoarthritis, demonstrated significantly better improvements in tender joint counts and swollen joint counts when treated with secukinumab 300 mg. ²⁹⁴ Richette et al. analyzed data from two phase III clinical trials (DISCOVER-1 and 2) of guselkumab in bio-naïve PsA patients to identify distinct PsA phenotypes using ML techniques. Non-negative matrix factorization was utilized to cluster 661 patients based on baseline clinical features, resulting in eight distinct clusters characterized by varying joint involvement, enthesitis, dactylitis, skin, and nail involvement. Guselkumab treatment improved MDA and Disease Activity index for PSoriatic Arthritis (DAPSA) scores across all clusters, with the best responses seen in clusters with high skin involvement and axial manifestations. However, clusters with small joint involvement showed lower initial responses, showing the heterogeneity of PsA and the need for patient-specific treatment approaches. ²⁹⁵

Prediction of comorbidities

Although the association and impact of comorbidities on PsA have been hot topics in recent years, studies of AI applications in the prediction, assessment, or follow-up of comorbidities on PsA are in their infancy. ²⁹⁶ Regarding cardiovascular risk assessment, recent evidence suggests that the combined use of QRISK3 and SCORE2 provided the best predictive accuracy. Although the literature is very scarce on this topic, Navarini et al. applied ML approaches, including SVMs and RnFr, to predict cardiovascular risk in PsA. Their best models outcompeted the standard risk calculators, yielding AUC estimates between 0.76 and 0.85, reflecting superior predictive performance. ²⁹⁷

AI can improve mortality risk assessment by combining information from clinical findings, test results, and comorbid conditions. This approach may help clinicians identify vulnerable patients earlier and guide more targeted and effective care plans. However, this review did not identify any study about the use of AI in mortality prediction.

Important note to readers

Whether AI or classical statistics is used to generate prediction models, the generalizability of these models remains an important research question. The performance of these models should be evaluated in prospective designs across groups with varying ethnic and social backgrounds.