A Deep Learning System for Diagnosis of Rheumatoid Arthritis on Digital Hand Photographs

A Deep Learning System for Diagnosis of Rheumatoid Arthritis on Digital Hand Photographs | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A Deep Learning System for Diagnosis of Rheumatoid Arthritis on Digital Hand Photographs Ryosuke Hanaoka, Satoshi Shinohara This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7564397/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 24 Mar, 2026 Read the published version in BMC Rheumatology → Version 1 posted 11 You are reading this latest preprint version Abstract Objectives: To develop and evaluate a deep learning model for diagnosing untreated rheumatoid arthritis (RA) using digital camera images of bilateral dorsal hands, benchmarking its performance against the widely-used 2010 ACR/EULAR criteria as a clinical reference standard. Methods: This pilot study included 170 participants (86 RA, 84 non-RA) who presented with joint symptoms at participating medical institutions. Digital images of both dorsal hands were captured under standardized conditions and processed using a deep learning-based background removal algorithm. A Swin Transformer-based model was developed and trained on these images. Model performance was evaluated using area under the receiver operating characteristics curve (AUROC), sensitivity, specificity, and calibration metrics. Gradient-Weighted Class Activation Mapping (Grad-CAM) was employed to visualize the model’s decision-making process. Results: The deep learning model achieved an AUROC of 0.870 (95% CI: 0.708-0.988), compared with 0.981 (95% CI: 0.953-1.010) for the ACR/EULAR criteria, with the difference not reaching statistical significance (p=0.131). While demonstrating comparable sensitivity to the ACR/EULAR criteria, the model showed lower specificity, accuracy, and F1-score. Post-Platt scaling calibration analysis revealed good alignment with ideal calibration in the 0.4–0.6 probability range. Grad-CAM visualization confirmed that the model focused on clinically relevant joint regions, particularly the metacarpophalangeal and proximal interphalangeal joints. Conclusion: Our deep learning-based approach for RA diagnosis using standard digital camera images demonstrated clinically viable performance, albeit with lower specificity than the ACR/EULAR criteria. This accessible screening tool could potentially expedite early RA detection, particularly in resource-limited settings. Larger multi-centre studies are needed to validate our findings and establish broader clinical applicability. (Up to 10) rheumatoid arthritis diagnosis deep learning ACR/EULAR criteria Figures Figure 1 Figure 2 Figure 3 Figure 4 Key messages A deep learning model using regular digital camera images of hands can diagnose rheumatoid arthritis with clinically viable performance. The model focuses on clinically relevant joint features, particularly in metacarpophalangeal and proximal interphalangeal joints. This accessible screening tool could expedite early RA detection in resource-limited healthcare settings. Introduction Rheumatoid arthritis (RA) is a chronic inflammatory disorder of unknown aetiology characterized by synovial infiltration of various inflammatory cells through autoimmune mechanisms, leading to irreversible destruction of synovial joints and resulting in severe physical disability and social disadvantage for patients [ 1 ]. Joint destruction in RA progresses more rapidly in the early stages after onset. Without appropriate treatment, significant bone destruction is often observed within two years of onset [ 2 ]. Therefore, to prevent irreversible joint damage, it is critical to establish a diagnosis as early as possible and initiate specific treatment [ 3 ]. With the availability of numerous effective drugs, the prevention of joint destruction has become an achievable goal when adequate treatment is administered by specialists during the early stages of the disease [ 4 ]. However, many cases still exist where the optimal timing for treatment is missed because of delays in care-seeking by patients, failure by physicians to diagnose RA, or insufficient treatment being provided, leading to severe and irreversible physical disabilities [ 5 ]. To address these challenges, it is essential to develop accessible resources for the public that enable patients to self-assess their RA status and monitor their condition [ 6 ]. We hypothesized that an RA diagnostic model using deep learning could achieve diagnostic performance comparable to the 2010 ACR/EULAR criteria, which serve as the current standard for RA assessment in clinical practice and research settings without the need for laboratory examination results [ 7 ]. To evaluate the feasibility of this approach, we conducted a preliminary pilot study targeting patients presenting with chief complaints of joint pain, joint swelling, and joint stiffness at participating medical institutions, as a precursor to a prospective cohort study. If the utility of this RA diagnostic model that applies deep learning to digital images of the dorsal aspects of both hands is validated, we anticipate that delays in seeking medical care for RA could be significantly reduced. Preventing declines in physical function in many cases may positively influence medical economics and the broader socioeconomic environment [ 8 ]. Methods Participants This study used data from 180 individuals aged between 20 and 79 years who visited participating medical institutions with joint pain, joint swelling, or joint stiffness, and who had not previously received disease-modifying antirheumatic drugs or corticosteroid therapy. The cohort included 97 patients diagnosed with RA by two board-certified rheumatologists of the Japan College of Rheumatology with over 10 years of clinical experience, and 84 control subjects in whom RA and other collagen-vascular diseases were ruled out. Consistent with the well-established epidemiology of RA, which shows a strong female predominance (approximately 3:1 female-to-male ratio), our initial cohort included 15 male and 82 female RA patients compared with 4 male and 80 female non-RA subjects. To prevent potential gender-based confounding that could bias model training toward gender features rather than disease-specific pathological changes, we performed age-matched sampling among male patients, resulting in the exclusion of 11 male RA cases to achieve balanced gender distribution between groups. This study was conducted as a retrospective observational analysis of routinely collected clinical data. Written informed consent was waived by the ethics committee in accordance with national ethical guidelines. Participants were informed of the study via opt-out notices posted in the outpatient waiting area. Data Collection Data extracted from medical records included age, sex, disease duration, and the presence or absence of tenderness or swelling in the following bilateral joints: proximal interphalangeal (PIP), metacarpophalangeal, first interphalangeal, wrist, elbow, shoulder, hip, knee, ankle, and metatarsophalangeal. Additionally, we extracted information on serum C-reactive protein (CRP) levels, erythrocyte sedimentation rate (ESR), rheumatoid factor (RF), and anti-cyclic citrullinated peptide antibodies (ACPA). Image Acquisition and Processing Bilateral dorsal hand images were captured against a green surgical drape background by trained nurses using a C40Z digital camera (Olympus Optical Co., Ltd). Images were taken under standardized lighting conditions without flash in a designated room to minimize variation. The digital images were saved in JPEG format and processed using a deep learning-based background removal algorithm (GitHub: https://github.com/danielgatis/rembg ). Subsequently, the files were processed in Pixelmater (Pixelmator Team, https://www.pixelmator.com/mac/ ) to manually separate the left- and right-hand images. Using Pillow ( https://pypi.org/project/Pillow/ .) and OpenCV libraries (Open Source Computer Vision Library, https://opencv.org/ ), we applied padding to create square images and horizontally flipped the left-hand images. While most image processing steps were automated using rembg and Pillow libraries, the separation of left- and right-hand images required manual intervention. Preliminary analysis confirmed that automated background removal significantly improved model focus on anatomical features rather than environmental artifacts. Neural Network Architecture and Training Prior to any model development, we first performed a strict data split to ensure complete independence of the test set. The dataset was divided into training (60%), validation (30%), and test (10%) sets, while maintaining the same ratio of RA to non-RA cases across all sets through stratified sampling. To ensure unbiased evaluation of the model’s generalization ability, the test set was immediately sequestered and was not accessed during any part of the model development process, including architecture selection, hyperparameter tuning, and performance optimization, with model development and optimization being performed exclusively using the training and validation sets. We adopted a neural network architecture based on the Swin Transformer for the image classification task of distinguishing RA patients from non-RA individuals. The Swin Transformer is a hierarchical vision transformer utilizing shifted windows, which efficiently captures both local and global features within images [ 9 ]. Transfer learning from ImageNet-1K pretraining has been shown to provide robust feature representations that can be effectively adapted to medical imaging tasks, particularly when training data is limited [ 10 , 11 ]. The Swin Transformer base model (swin_base_patch4_window7_224) pretrained on ImageNet-1K was used as the backbone architecture. This variant processes input images with a patch size of 4×4 pixels and employs a window size of 7×7 for self-attention computation, with input images resized to 224×224 pixels. Regularization techniques including dropout (rate = 0.6), attention dropout (rate = 0.5), and stochastic depth (rate = 0.5) were applied to enhance generalization performance. Cross-entropy loss with label smoothing (0.15) was used as the loss function, and the AdamW optimizer (learning rate = 2e − 5 , weight decay = 2.5e − 2 ) with a OneCycleLR scheduler (max_lr = 2.2e − 5 , pct_start = 0.3, div_factor = 15, final_div_factor = 2000, anneal_strategy='cos') was employed for optimization. Data preprocessing and augmentation techniques were applied, including random rotations (± 15°), affine transformations (translation: ± 0.08, scale: 0.92–1.08), colour jittering (brightness, contrast, saturation: ± 0.2, hue: ± 0.05), and random perspective transformations (distortion scale: 0.15, probability: 0.4). The model was trained with a batch size of 64 over 300 epochs with gamma = 0.9 and seed = 42. Mixup (alpha = 0.8) was used as an additional data augmentation technique to further improve model generalization. Scoring of ACR/EULAR Classification Criteria The 2010 ACR/EULAR rheumatoid arthritis classification criteria evaluate four domains to calculate a total score ranging from 0 to 10 points: joint involvement (0–5 points), serology (0–3 points), acute-phase reactants (0–1 point), and symptom duration (0–1 point). A total score of 6 or greater classifies a patient as having RA. Joint involvement is scored based on the number and size of involved joints: 0 points for 1 large joint, 1 point for 2–10 large joints, 2 points for 1–3 small joints, 3 points for 4–10 small joints, and 5 points for more than 10 joints including at least 1 small joint. Serology scores are based on RF and ACPA levels: 0 points for negative results, 2 points for low-positive, and 3 points for high-positive results. Acute-phase reactants (ESR and CRP) are scored as 1 point if elevated, and symptom duration is scored as 1 point if symptoms have persisted for 6 weeks or longer [ 7 ]. Model Calibration and Interpretability To ensure reliable probability estimates, we performed probability calibration using Platt scaling [ 12 ]. This method, which has been shown to be effective for calibrating deep neural networks [ 13 ], involves training a logistic regression model for which the independent variables are the raw model outputs (logits) and the dependent variable is the reference standard RA diagnosis established by rheumatologists. The calibration model was trained using the validation set logits and their corresponding true diagnostic labels, and was then applied to transform both validation and test set predictions into well-calibrated probabilities. This calibration step is crucial for obtaining reliable probability estimates, especially for clinical applications where decision thresholds are important. The calibrated probabilities were then used for all subsequent analyses, including receiver operating characteristics (ROC) curve generation and performance metric calculations. The optimal classification threshold was determined by maximizing Youden’s index for the validation set. This threshold was then applied to the test set for final performance evaluation. Gradient-Weighted Class Activation Mapping (Grad-CAM) was applied to interpret the model’s predictions and identify regions of input images influential in the decision-making process. Forward and backward hooks were registered on the last block of the Swin Transformer to capture feature maps and gradients during forward and backward passes. The Grad-CAM algorithm computed class activation maps by weighting the captured feature maps with the gradients’ global average pooling, followed by ReLU activation and normalization. The resulting heatmaps were overlaid on the original images to highlight salient regions [ 14 ]. Outcomes The primary outcome was the area under the ROC curve (AUC) of the deep learning model in comparison with the ACR/EULAR classification criteria when applied to the test set, with the reference standard diagnosis being established by experienced rheumatologists. Secondary outcomes on the test set included sensitivity, specificity, accuracy, F1-score, and Brier score to assess calibration performance, all evaluated using the optimal threshold determined from the validation set. While the 2010 ACR/EULAR criteria were originally developed for classification purposes in research settings, they have become the de facto standard for clinical RA assessment and have been widely used as benchmarks in diagnostic studies [ 15 – 17 ]. Multiple validation studies have evaluated their diagnostic performance and recommended their use for uniformity and comparability in diagnostic research [ 15 ]. In this context, we used these criteria as our primary comparator to evaluate the clinical utility of our deep learning approach. Statistical Analysis Statistical analyses were conducted using Python with the packages scikit-learn and statsmodels. Continuous variables were compared using the Wilcoxon rank-sum test, and categorical variables were analysed using Pearson’s chi-square test. Model performance was evaluated using sensitivity, specificity, accuracy, and F1 score. ROC curves were generated and the AUCs were calculated with 95% confidence intervals. Comparison of AUCs between the deep learning model and ACR/EULAR criteria was performed using DeLong’s test. Model calibration was assessed using Brier scores and calibration plots, with Platt scaling applied for probability calibration. The optimal probability threshold was determined using Youden’s index. Statistical significance was defined as P < 0.05. Ethical Considerations This study adhered to the Declaration of Helsinki and the Ethical Guidelines for Medical and Health Research Involving Human Subjects. The study protocol was reviewed and approved by the Ethics Committee of National Hospital Organization Tochigi Medical Centre (approval number: 2025-02). Clinical trial number: not applicable. Written informed consent was waived by the ethics committee as this was a retrospective observational study using existing clinical data with minimal risk to participants. Information about the study was disclosed to participants via the postings in the outpatient waiting area, providing them the opportunity to opt-out of participation in accordance with Japanese ethical guidelines for medical research. Because this observational study used photographs taken with standard digital cameras and data extracted from routine medical records, there was no additional burden, expected risks, or direct benefits to the participants. Results Patient Characteristics Initially, 97 patients with RA and 84 subjects without RA were enrolled. After age-based pair matching of male subjects to address gender imbalance, the final analysis included 86 RA cases and 84 non-RA cases (Fig. 1 ). The gender distribution in our final cohort (4 males, 162 females) reflects the established epidemiological pattern of RA, which predominantly affects women with a typical female-to-male ratio of 2–4:1. Our decision to match by gender was driven by the need to ensure that the deep learning model would focus on disease-specific morphological features rather than gender-related anatomical differences in hand appearance. Table 1 shows the characteristics of these matched groups. There were no significant differences in age, sex, or symptom duration between the two groups. Inflammatory markers such as ESR and CRP, and immunological markers such as RF and ACPA, were significantly higher in the RA group (p < 0.05 for all). Table 1 Baseline characteristics and ACR/EULAR classification criteria scores of RA and non-RA patients nonRA (n = 84) RA (n = 86) difference (95% C.I.) age, years (SD) 61.7 (11.6) 62.6 (12.3) 0.9 (-2.7 to 4.4) sex, female (%) 80 (95.2) 82 (95.3) 0.1 (-6.3 to 6.5) symptom duration, week (SD) 78.7 (145.0) 66.6 (171.9) -12.1 (-59.9 to 35.6) ESR, mm/hr (SD) 12.3 (11.7) 39.0 (32.1) 26.6 (19.4 to 33.8) CRP, mg/dl (SD) 0.19 (0.66) 1.29 (2.25) 1.10 (0.61 to 1.60) RF, IU/ml (SD) 9.9 (18.8) 127.5 (195.0) 117.6 (76.2 to 159.0) ACPA, U/ml (SD) 1.1 (2.8) 182.3 (298.9) 181.2 (118.0 to 244.3) joint score 0 (%) 51 (60.7) 14 (16.3) -43.3 (-57.5 to -31.4) 1 (%) 2 (2.4) 1 (1.2) -1.2 (-5.2 to 2.8) 2 (%) 19 (22.6) 37 (43.0) 19.3 (6.6 to 34.2) 3 (%) 11 (13.1) 26 (30.2) 17.1 (5.0 to 29.2) 5 (%) 1 (1.2) 8 (9.3) 8.1 (1.5 to 14.7) inflammation score 0 (%) 45 (53.6) 5 (5.8) -47.8 (-59.5 to -36.0) 1 (%) 39 (46.4) 81 (94.2) 47.8 (36.0 to 59.5) immunological score 0 (%) 70 (83.3) 6 (7.0) -76.3 (-86.0 to -66.7) 2 (%) 6 (7.1) 2 (2.3) -4.8 (-11.2 to 1.5) 3 (%) 8 (9.5) 78 (90.7) 81.2 (72.4 to 90.0) duration score 0 (%) 4 (4.8) 7 (8.1) 3.4 (-4.0 to 10.7) 1 (%) 80 (95.2) 79 (91.9) -3.4 (-10.7 to 4.0) overall score 1 (%) 22 (26.2) 0 (0.0) -26.2 (35.6 to -16.8) 2 (%) 19 (22.6) 2 (2.3) -20.3 (-29.8 to -10.8) 3 (%) 18 (21.4) 0 (0.0) -21.4 (-30.2 to -12.7) 4 (%) 13 (15.5) 3 (3.5) -10.8 (-20.6 to -3.3) 5 (%) 8 (9.5) 11 (12.8) 3.3 (-6.2 to 12.7) 6 (%) 2 (2.4) 8 (9.3) 6.9 (-0.0 to 13.9) 7 (%) 1 (1.2) 36 (41.9) 39.5 (30.0 to 51.4) 8 (%) 1 (1.2) 20 (23.3) 22.1 (12.8 to 31.3) 10 (%) 0 (0.0) 6 (7.0) 7.0 (1.6 to 12.4) Continuous variables are presented as mean (standard deviation), and categorical variables are presented as number (%). Differences between groups are shown with their 95% confidence intervals. ACPA, anti-citrullinated peptide antibody; CRP, C-reactive protein; ESR, erythrocyte sedimentation rate; nonRA, non-rheumatoid arthritis; RA, rheumatoid arthritis; RF, rheumatoid factor; SD, standard deviation; 95% C.I., 95% confidence interval The most common diagnoses were normal findings or transient conditions, followed by osteoarthritis and fibromyalgia (Supplementary Table 1). Five patients presented with extra-articular inflammatory conditions, and two patients were diagnosed with hyperuricemia. Supplementary Fig. 1 shows the distribution of CRP, ESR, RF, and ACPA levels in the non-RA group. Some patients exhibited elevated CRP and ESR levels, but most of these elevations were mild. Excluding one case of pneumonia and two cases of mycobacterial infection, the elevated inflammatory markers were considered to be of limited clinical significance. In our cohort, the joint involvement scores showed clear differences between the groups, with the RA group having a higher proportion of moderate to severe scores (2 points or higher) and the non-RA group predominantly showing normal scores (0 points). Regarding acute-phase reactant scores, 94.2% of the RA group showed abnormal values (1 point), whereas 46.4% of the non-RA group had elevated inflammatory markers despite non-inflammatory clinical diagnoses. For the serology scores, 90.7% of the RA group had high-positive results (3 points), while 83.3% of the non-RA group had negative results (0 points). There was no difference in symptom duration scores between the groups. The total ACR/EULAR classification criteria scores showed distinct distributions in the two groups. The RA group demonstrated higher total scores, with 64% scoring 7 or above, whereas only 3.6% of the non-RA group reached this threshold. This distribution aligns with the expected performance of these classification criteria in distinguishing RA from other forms of arthritis and joint conditions. Model Performance The performance of the trained model was evaluated using the test set, and the results of this evaluation are shown in Table 2 . The model demonstrated high sensitivity, but relatively low specificity. After implementing Platt scaling calibration using the validation set and applying the threshold of 0.564 determined by Youden index analysis, we observed improved specificity without substantial deterioration in accuracy or F1 score, despite a slight decrease in sensitivity. In comparison, the ACR/EULAR classification criteria with a cutoff score of 6 points showed comparable sensitivity to the deep learning model but superior specificity, accuracy, and F1 score. Table 2 Diagnostic performance metrics comparing ACR/EULAR RA classification criteria and deep learning model with and without probability calibration ACR/EULAR RA criteria† Uncalibrated Deep Learning RA score Calibrated Deep Learning RA score‡ Sensitivity 0.778 0.944 0.778 Specificity 1.000 0.667 0.833 Accuracy 0.889 0.806 0.806 F1-score 0.875 0.829 0.800 †Using the established cutoff score of 6 points in the ACR/EULAR criteria; ‡Using a cutoff value of 0.564; ACR/EULAR criteria, American College of Rheumatology and European Alliance of Associations for Rheumatology Rheumatoid Arthritis classification criteria; RA, rheumatoid arthritis Figure 2 presents the calibration plots after Platt scaling. The Brier scores were 0.150 for the deep learning model and 0.112 for the ACR/EULAR classification criteria. Although the ACR/EULAR criteria performed marginally better, both methods demonstrated substantially better performance than random prediction (0.250). The calibration plots showed good alignment with ideal calibration in the 0.4–0.6 probability range. However, both methods exhibited strong binary tendencies in their probability distributions, with limited intermediate probability predictions. Notably, the deep learning score showed an abrupt transition around the 0.8 probability threshold. Figure 3 presents ROC curves comparing the overall performance of the ACR/EULAR RA classification criteria and the deep learning RA score. The ACR/EULAR RA criteria achieved an AUC of 0.981 (95% CI: 0.953–1.010), exceeding the deep learning RA score’s AUC of 0.870 (95% CI: 0.708–0.988) by 0.111 (95% CI: −0.033–0.255). However, this difference did not reach statistical significance (p = 0.131). Grad-CAM visualizations revealed that the model appropriately focused on clinically relevant joint regions typically affected in early RA, particularly the metacarpophalangeal and PIP joints (Fig. 4 ). The model paid special attention to the joint margins and surrounding soft tissues, aligning with conventional RA assessment. Discussion In this pilot study, we developed a deep learning model for diagnosing untreated RA using digital camera images of bilateral dorsal hands and evaluated its performance against the 2010 ACR/EULAR RA classification criteria. Analysis of the ROC curve demonstrated that the deep learning model achieved clinically viable performance. While the AUC of the 2010 ACR/EULAR RA classification criteria surpassed that of the deep learning model, this difference did not reach statistical significance. Using the cutoff value determined through validation set calibration, the deep learning model’s sensitivity for detecting RA was comparable to the 2010 ACR/EULAR classification criteria. However, the 2010 criteria demonstrated superior specificity, accuracy, and F1-score compared with the deep learning model. Post-Platt scaling calibration analysis revealed good alignment, with ideal calibration for both methods in the 0.4–0.6 probability range, although both exhibited a strong binary tendency in their probability distributions. Notably, visualization of decision rationale using Grad-CAM confirmed that the model appropriately focused on clinically relevant joint regions, particularly the morphology and skin creases of the metacarpophalangeal and PIP joints, which are typically affected in early RA. Our choice to benchmark against the ACR/EULAR criteria reflects their widespread adoption in clinical practice for RA assessment. While originally designed for research classification, these criteria have evolved into the most commonly used standardized approach for RA evaluation in real-world clinical settings, making them an appropriate reference standard for diagnostic performance comparison. Current clinical methods for evaluating RA include joint inspection and palpation, acute phase reactants such as CRP and ESR from blood samples, immunological markers such as RF and ACPA, plain radiography, joint ultrasonography, and magnetic resonance imaging. In today’s expanding field of machine learning and deep learning research, attempts to apply deep learning techniques to these modalities are emerging worldwide. However, there are relatively few reports on the use of digital camera images to evaluate RA, as employed in our study, despite the fact that this is arguably the most accessible imaging modality for patients. To our knowledge, Hügle et al. were the first to evaluate RA using digital camera images of the dorsal hand [18]. They focused on dorsal skin creases of PIP joints, training ResNet34 to differentiate between swollen and non-swollen joints using cropped local images, suggesting local findings in joints as potential digital biomarkers for RA. Our study similarly found that the deep learning model focused on finger creases for RA discrimination, aligning with their findings. However, while we successfully analysed the entire hand image, their study remained at a more fundamental stage, analysing only the local features of the PIP joints. Phatak et al. acquired smartphone images of both dorsal hands from 200 patients with inflammatory arthritis and 200 healthy individuals [19]. Using ResNet-based analysis, they successfully detected arthritis from whole-hand images with high accuracy. Their success can likely be attributed to the availability of a relatively large sample size and the fact that they identified hands with physician-confirmed arthritis. In contrast, our study aimed to explore whether a model could detect finger joint inflammation that is difficult for human observers to identify. We attempted to detect RA cases—diagnosed comprehensively through clinical examinations, laboratory tests, and radiographic findings—using only hand photographs. The seemingly lower diagnostic performance of our study compared to theirs may be explained by the inclusion of many RA patients who had been clinically diagnosed but did not exhibit detectable joint inflammation upon physical examination. Despite tackling a more challenging task, our study achieved a certain degree of success. We hypothesize that this may be due to the difference in model architectures: while they employed a CNN-based approach, we utilized a Swin Transformer. CNNs respond more strongly to fine-grained texture patterns in images, whereas Swin Transformers, due to their hierarchical nature, can capture both subtle local changes and broader global variations. This characteristic of Swin Transformers may be particularly advantageous in analyzing medical images, which often present diverse and complex patterns of change. A notable advantage of our approach is its reliance solely on digital camera images of dorsal hands, which can be easily captured by almost everyone with everyday equipment, rather than clinical tests or advanced imaging. Our system, which requires neither expensive imaging equipment nor blood biomarkers, could particularly enhance the detection of early RA in primary care settings and regions with limited rheumatology resources. More rapid and accessible screening could expedite referrals to rheumatologists, reduce diagnostic delays, and enable timelier initiation of disease-modifying therapy—essential factors for minimizing joint damage and maintaining patient function. Despite promising results, several limitations warrant acknowledgment. First, our analysis is based on a limited sample size from clinical settings, potentially resulting in insufficient model training and performance acquisition; however, increasing the sample size should lead to improved training and performance, not less. We must also acknowledge that the insufficient size of the test set sample may have compromised stable performance evaluation. Larger-scale multi-centre studies are essential to confirm generalizability and establish external validity. Second, while the background removal procedures improved uniformity, data reliability may still show variations due to differences in lighting, patient positioning, or camera angles. Successful real-world deployment will require rigorous training procedures and prospective validation across various image capture conditions. Our gender matching approach, while necessary to prevent model bias toward gender-specific features, may limit the generalizability of our findings to the broader RA population. However, this approach ensures that the model's diagnostic performance is based on pathological changes rather than demographic characteristics, which is essential for developing a clinically relevant diagnostic tool. In conclusion, our deep learning-based RA diagnostic approach, enhanced by advanced architecture, demonstrated the potential to achieve clinically viable accuracy. Future large-scale multi-centre prospective studies are needed to establish more stable and reliable results. Declarations Acknowledgements (if applicable): We thank the patients and staff at the participating medical institutions for their cooperation. We thank Edanz (https://jp.edanz.com/ac) for editing a draft of this manuscript. Group authorship list (if applicable and all group members fulfil ICMJE criteria for authorship): Not applicable Author contributions: RH designed the study, analysed the data, and drafted the manuscript. SS collected the data. All authors critically revised the manuscript and approved the final version. Funding (if specific to this study): This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Conflict of interest statement (in addition to author disclosure form): The authors declare no conflict of interest. Data availability statement: The data underlying this article will be shared upon reasonable request to the corresponding author. Clinical trial number: not applicable. Human Ethics and Consent to Participate declarations: This study was approved by the Ethics Committee of National Hospital Organization Tochigi Medical Centre (approval number: 2025-02). Written informed consent was waived by the ethics committee in accordance with national ethical guidelines for retrospective studies. Information regarding the study was disclosed to participants via opt-out postings in the outpatient waiting area. References Firestein GS, McInnes IB. Immunopathogenesis of rheumatoid arthritis. Immunity 2017;46:183-196. van der Heijde DM. Joint erosions and patients with early rheumatoid arthritis. Br J Rheumatol 1995;34 Suppl 2:74-8. van der Linden M, le Cessie S, Raza K, van der Woude D, Knevel R, Huizinga TW, et al. Long-term impact of delay in assessment of patients with early rheumatoid arthritis. Arthritis Rheum 2010;62:3537-46. Smolen JS, Landewé RBM, Bijlsma JWJ, Burmester GR, Dougados M, Kerschbaumer A, et al. EULAR recommendations for the management of rheumatoid arthritis with synthetic and biological disease-modifying antirheumatic drugs. Ann Rheum Dis 2020;79:685-99. Stack RJ, Nightingale P, Jinks C, Shaw K, Herron-Marx S, Horne R, et al. Delays between the onset of symptoms and first rheumatology consultation in patients with rheumatoid arthritis in the UK: an observational study. BMJ Open 2019;9:e024361. Connelly K, Segan J, Lu A, Saini M, Cicuttini FM, Chou L, et al. Patients’ perceived health information needs in inflammatory arthritis: A systematic review. Semin Arthritis Rheum 2019;48:900-10. Aletaha D, Neogi T, Silman AJ, Funovits J, Felson DT, Bingham CO 3rd, et al. 2010 Rheumatoid arthritis classification criteria. Arthritis Rheum 2010;62:2569-81. Feldman DE, Bernatsky S, Houde M, Beauchamp ME, Abrahamowicz M. Early consultation with a rheumatologist for RA: does it reduce subsequent use of orthopaedic surgery? Rheumatology 2013;52:452-59. Liu Z, Ning J, Cao Y, Hu H, Wei Y, Zhang Z, et al. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. Int Conf Comput Vis 2021:10012-22. Tajbakhsh N, Shin JY, Gurudu SR, et al. Convolutional neural networks for medical image analysis: full training or fine tuning? IEEE Trans Med Imaging. 2016;35(5):1299-1312. Kather JN, Pearson AT, Halama N, et al. Transfer learning for medical image classification: a literature review. BMC Med Imaging. 2022;22:69. Platt JC. Probabilistic Outputs for Support Vector Machines and Comparisons to Regularized Likelihood Methods. Smola AJ, Bartlett PL, Schölkopf B, Schuurmans D, editors. Advances in Large Margin Classifiers. Cambridge (MA): MIT Press; 2000. p.61-74. Guo C, Pleiss G, Sun Y, Weinberger KQ. On Calibration of Modern Neural Networks. Proceedings of the 34th International Conference on Machine Learning. 2017:1321-30. Selvaraju RR, Cogswell M, Das A Vedantam R, Parikh D, Batra D. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. Int Conf Comput Vis 2017:618-26. Alves C, Luime JJ, van Zeben D, et al. Diagnostic performance of the ACR/EULAR 2010 criteria for rheumatoid arthritis and two diagnostic algorithms in an early arthritis clinic (REACH). Ann Rheum Dis. 2011;70(9):1645-1647. Varache S, Cornec D, Morvan J, et al. Diagnostic accuracy of ACR/EULAR 2010 criteria for rheumatoid arthritis in a 2-year cohort. J Rheumatol. 2011;38(7):1250-1257. Britsemmer K, Ursum J, Gerritsen M, et al. Validation of the 2010 ACR/EULAR classification criteria for rheumatoid arthritis: slight improvement over the 1987 ACR criteria. Ann Rheum Dis. 2011;70(8):1468-1470. Hügle T, Caratsch L, Caorsi M, Maglione J, Dan D, Dumusc A, et al. Dorsal Finger Fold Recognition by Convolutional Neural Networks for the Detection and Monitoring of Joint Swelling in Patients with Rheumatoid Arthritis. Digit Biomark 2022;6:31-5. Phatak S, Saptarshi R, Sharma V, Shah R, Zanwar A, Hegde P, et al. Incorporating computer vision on smart phone photographs into screening for inflammatory arthritis: results from an Indian patient cohort. Rheumatology (Oxford). 2024 16:keae678. Additional Declarations No competing interests reported. Supplementary Files SupplementaryFigure1.png Supplementary Figure 1. Distribution of laboratory markers in RA and non-RA patients. Histograms showing the distribution of C-reactive protein (CRP), erythrocyte sedimentation rate (ESR), rheumatoid factor (RF), and anti-citrullinated protein antibodies (ACPA) levels in both RA (red) and non-RA (blue) groups. While some non-RA patients exhibited elevated inflammatory markers (CRP and ESR), most elevations were mild. The RF and ACPA levels show a clear distinction between the two groups, with significantly higher values in the RA population. supplementarytable1.docx Supplementary material Supplementary Table 1. Final diagnoses of patients in the non-RA group. The table shows the distribution of final diagnoses among the 84 patients classified as non-RA. Normal findings or transient conditions were the most common (42.86%), followed by osteoarthritis (29.76%) and fibromyalgia (11.90%). Five patients presented with extra-articular inflammatory conditions (organizing pneumonia, pneumonia, psoriasis, tuberculosis, and nontuberculous mycobacteriosis). Cite Share Download PDF Status: Published Journal Publication published 24 Mar, 2026 Read the published version in BMC Rheumatology → Version 1 posted Editorial decision: Revision requested 27 Jan, 2026 Reviews received at journal 23 Jan, 2026 Reviewers agreed at journal 02 Jan, 2026 Reviews received at journal 01 Nov, 2025 Reviewers agreed at journal 21 Oct, 2025 Reviewers agreed at journal 18 Oct, 2025 Reviewers invited by journal 14 Sep, 2025 Editor invited by journal 12 Sep, 2025 Editor assigned by journal 10 Sep, 2025 Submission checks completed at journal 10 Sep, 2025 First submitted to journal 08 Sep, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7564397","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":515564154,"identity":"a1bfc0bf-9341-4336-a8cc-d6ecda6d418e","order_by":0,"name":"Ryosuke Hanaoka","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA70lEQVRIiWNgGAWjYBACAzB5gEGOgZmBDSEsQYQWY9K1JDYwIGvBB8zZzx788OGMXfqG4+zPHjBU2OUxSDcfYLDcgVuLZU9esuSMG8m5Gw7zmBswnEkuZpA5lsAgeQaPww7kGEjzfGDO3XaYh02Cse1AYoNEjgGDZBseLeffGP/m+VCfbnaY/RlUS/4H/Fpu5JhJ89w4nGB2mMEMZgsDAS1vzCxnnDluuP8wj5lEwpnkxDaJNIMDeP1yPsf4xodj1fKS/cefSXyosEvsl0h++FgST4ihggQGSOwclmwgVgsMMH4kWcsoGAWjYBQMYwAAy2FQ9LByVt0AAAAASUVORK5CYII=","orcid":"","institution":"National Hospital Organization Tochigi Medical Centre","correspondingAuthor":true,"prefix":"","firstName":"Ryosuke","middleName":"","lastName":"Hanaoka","suffix":""},{"id":515564155,"identity":"bf04a53d-e210-449f-971e-77c98adc5ec0","order_by":1,"name":"Satoshi Shinohara","email":"","orcid":"","institution":"Tochigi Rheumatology Clinic","correspondingAuthor":false,"prefix":"","firstName":"Satoshi","middleName":"","lastName":"Shinohara","suffix":""}],"badges":[],"createdAt":"2025-09-08 12:53:19","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7564397/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7564397/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s41927-026-00639-7","type":"published","date":"2026-03-24T16:10:55+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":91956579,"identity":"e268be14-5179-46dd-8a5c-4b13ec639573","added_by":"auto","created_at":"2025-09-23 07:17:20","extension":"jpg","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":235327,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/f258d32249e9b3e8afff3855.jpg"},{"id":91955388,"identity":"c4955d55-b5fe-4833-b717-0d32442d00a9","added_by":"auto","created_at":"2025-09-23 07:09:20","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":42184,"visible":true,"origin":"","legend":"","description":"","filename":"Originalarticle0904.docx","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/333af506e336d8bd49197513.docx"},{"id":91955402,"identity":"6d61b5c1-4688-404a-93fb-9acca213b798","added_by":"auto","created_at":"2025-09-23 07:09:20","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":328502,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/5203f1d08a7c0e030c0372f1.png"},{"id":91956586,"identity":"9e7b7dc1-b2fa-48b0-9dda-17aa02568c2f","added_by":"auto","created_at":"2025-09-23 07:17:20","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":17610,"visible":true,"origin":"","legend":"","description":"","filename":"Table1.docx","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/8e24abd9569b39723aae4418.docx"},{"id":91955393,"identity":"e21ad46e-048f-4f7b-8564-71b013e2a296","added_by":"auto","created_at":"2025-09-23 07:09:20","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":272620,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/29527a907ac9078d3f492904.png"},{"id":91955401,"identity":"18327e9f-fee5-482f-8ab9-15e17ee54dee","added_by":"auto","created_at":"2025-09-23 07:09:20","extension":"jpg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":96164,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/bcd2b468d04dbf6da971c241.jpg"},{"id":91958158,"identity":"ccef96b0-f949-4c9c-a9ff-77b002babff6","added_by":"auto","created_at":"2025-09-23 07:33:20","extension":"docx","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":14789,"visible":true,"origin":"","legend":"","description":"","filename":"Table2.docx","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/9b208dd190c32baba29f8bf2.docx"},{"id":91955405,"identity":"d45a8028-1929-43af-880a-99fd3bfa632e","added_by":"auto","created_at":"2025-09-23 07:09:20","extension":"json","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4844,"visible":true,"origin":"","legend":"","description":"","filename":"492d60199dad4399b5eb671fca620f1a.json","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/2ed663c8c93584273793101b.json"},{"id":91956583,"identity":"11d66e90-938d-468c-8182-b53b4b587f8a","added_by":"auto","created_at":"2025-09-23 07:17:20","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":288633,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/dcc8f3b6595121e79f3c05c4.png"},{"id":91955403,"identity":"b23edde3-2da9-455f-8f25-71961df307e1","added_by":"auto","created_at":"2025-09-23 07:09:20","extension":"docx","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":14800,"visible":true,"origin":"","legend":"","description":"","filename":"supplementarytable1.docx","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/5b1d769199805382a22e59ff.docx"},{"id":91955416,"identity":"1bc15056-c4a0-4d25-be02-e0c1cb74de30","added_by":"auto","created_at":"2025-09-23 07:09:20","extension":"xml","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":92327,"visible":true,"origin":"","legend":"","description":"","filename":"492d60199dad4399b5eb671fca620f1a1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/e0d27d23ec272202ebd74a7c.xml"},{"id":91956587,"identity":"106af93e-2416-4e75-ba48-dbea84c4b2c8","added_by":"auto","created_at":"2025-09-23 07:17:20","extension":"jpg","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":235327,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/dd09127e50e3fd25d5b57ed8.jpg"},{"id":91957757,"identity":"21a063b3-5bc7-48d6-842f-d46ed95ef80a","added_by":"auto","created_at":"2025-09-23 07:25:20","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":328502,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/90918c6c62de0e78d2cefbee.png"},{"id":91955406,"identity":"8f23a320-2e98-4dcb-8104-79d45061b22e","added_by":"auto","created_at":"2025-09-23 07:09:20","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":272620,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/2457ab43be8707eea4f29425.png"},{"id":91955412,"identity":"8442a924-ea23-4e6c-b654-aa9de129dfb9","added_by":"auto","created_at":"2025-09-23 07:09:20","extension":"jpg","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":96164,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/091d6e21f0ede3560d5017f2.jpg"},{"id":91956591,"identity":"f3b330d5-4c3c-42d6-ad1e-f021356eb4c7","added_by":"auto","created_at":"2025-09-23 07:17:20","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":52746,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/ee90c3bc36567c6f0a1fa65e.png"},{"id":91957758,"identity":"386269ca-a376-4cbc-8271-c6fe6f1e9e7f","added_by":"auto","created_at":"2025-09-23 07:25:20","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":126598,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/c4e2acbfbd53a009df8c4f9c.png"},{"id":91956589,"identity":"e399aa2e-4a06-45e9-808a-5b8466835516","added_by":"auto","created_at":"2025-09-23 07:17:20","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":101672,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/71303689d04fccf0c26dd7f4.png"},{"id":91955410,"identity":"bf8702de-4c8f-43da-9efd-96eba0833126","added_by":"auto","created_at":"2025-09-23 07:09:20","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":89268,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/b39cc6d99ad51125b8c9fb09.png"},{"id":91955414,"identity":"3acdeccf-af1f-45fa-8a42-794206231e60","added_by":"auto","created_at":"2025-09-23 07:09:20","extension":"xml","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":87006,"visible":true,"origin":"","legend":"","description":"","filename":"492d60199dad4399b5eb671fca620f1a1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/4a8cc212ef7630ee839e1d6f.xml"},{"id":91955415,"identity":"701fbbb2-5e1c-47af-b504-2473895bc8f1","added_by":"auto","created_at":"2025-09-23 07:09:20","extension":"html","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":97938,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/39ca6289d9d84426baac6de7.html"},{"id":91957754,"identity":"356fdf5b-6f12-464a-b88a-c8b7bcbd87a7","added_by":"auto","created_at":"2025-09-23 07:25:20","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":235327,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFlow diagram of patient selection and image processing for the deep learning model development.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eInitial enrolment included 97 RA patients (15 males, 82 females) and 84 non-RA subjects (4 males, 80 females). To address gender imbalance, age-based 1:1 matching was performed among the male patients, resulting in the exclusion of 11 male RA cases. This yielded a final cohort of 86 RA patients (4 males, 82 females) and 84 non-RA subjects (4 males, 80 females). Digital images of bilateral dorsal hands underwent preprocessing steps including background removal, bilateral hand separation, and horizontal flipping of left-hand images, resulting in 172 RA images and 168 non-RA images. The processed image dataset was then divided into training, validation, and test sets in a 6:3:1 ratio while maintaining stratification of RA and non-RA cases.\u003c/p\u003e","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/5d72f9914c8eb2af524bbc17.jpg"},{"id":91956581,"identity":"567d4473-ac3a-4355-9824-d202c29f040e","added_by":"auto","created_at":"2025-09-23 07:17:20","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":328502,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCalibration plots comparing the performance of the deep learning model and ACR/EULAR RA classification criteria.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCalibration curves comparing the ACR/EULAR RA criteria (blue line) and Deep Learning RA Score (red line) against perfect calibration (dashed diagonal line). The plot shows the relationship between predicted probabilities and observed frequencies, with both methods showing different calibration patterns across probability ranges.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/97ccc6e10ce8dc2780f71ab0.png"},{"id":91955397,"identity":"abf4236c-c64e-43f1-a187-5c69bda888dc","added_by":"auto","created_at":"2025-09-23 07:09:20","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":272620,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eReceiver Operating Characteristics (ROC) curves comparing the diagnostic performance of the deep learning model and ACR/EULAR RA classification criteria.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe ROC curves demonstrate the diagnostic performance of the ACR/EULAR RA classification criteria (blue line) and the deep learning RA score (red line) in discriminating between RA and non-RA cases. The ACR/EULAR criteria achieved an AUC of 0.981 (95% CI: 0.953–1.010), while the deep learning model achieved an AUC of 0.870 (95% CI: 0.730–1.010). Although the ACR/EULAR criteria showed numerically superior performance with a difference in AUC of 0.111, this difference did not reach statistical significance (p = 0.131). The diagonal dashed line represents random classification (AUC = 0.5). Both methods demonstrated performance substantially better than random classification, with the deep learning model showing particularly high sensitivity at the expense of specificity, as evidenced by the steep initial rise in its ROC curve.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/4c5de63e630709508586e728.png"},{"id":91956584,"identity":"11792322-b531-4463-97a0-c169edd5816b","added_by":"auto","created_at":"2025-09-23 07:17:20","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":96164,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eVisualization of model attention using Gradient-Weighted Class Activation Mapping (Grad-CAM).\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eRepresentative Grad-CAM visualizations demonstrating the regions of interest that influenced the deep learning model's decision-making process in diagnosing RA. (A) A patient with RA correctly classified by the model. The heat maps (shown in blue-purple overlay) highlight the areas where the model focused its attention, particularly the skin creases on the extensor surface of the finger joints (white arrows). (B) A patient without RA correctly classified as non-RA by the model. The model appropriately focused on clinically relevant anatomical features, particularly the metacarpophalangeal and proximal interphalangeal joints and their surrounding soft tissues. These regions are typically affected in early RA, and this analysis confirms that the model learned to identify clinically meaningful features. The varying intensity of the activation patterns across different hand regions suggests the model's ability to detect subtle morphological changes characteristic of RA involvement.\u003c/p\u003e","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/1fa0e767326f93890133354c.jpg"},{"id":105755487,"identity":"96a372d9-5eff-4087-a227-1344ce0f41a7","added_by":"auto","created_at":"2026-03-30 16:27:32","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1763098,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/ef45c101-345e-49a1-852c-a2b3be450274.pdf"},{"id":91955399,"identity":"189a7e62-f61a-4df0-867f-5426e4160d3b","added_by":"auto","created_at":"2025-09-23 07:09:20","extension":"png","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":288633,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSupplementary Figure 1. Distribution of laboratory markers in RA and non-RA patients.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHistograms showing the distribution of C-reactive protein (CRP), erythrocyte sedimentation rate (ESR), rheumatoid factor (RF), and anti-citrullinated protein antibodies (ACPA) levels in both RA (red) and non-RA (blue) groups. While some non-RA patients exhibited elevated inflammatory markers (CRP and ESR), most elevations were mild. The RF and ACPA levels show a clear distinction between the two groups, with significantly higher values in the RA population.\u003c/p\u003e","description":"","filename":"SupplementaryFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/d3bc40ba7188f330f15d585d.png"},{"id":91957755,"identity":"845e1a3b-5965-468e-84e5-09281a3f581a","added_by":"auto","created_at":"2025-09-23 07:25:20","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":14800,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSupplementary material\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSupplementary Table 1. Final diagnoses of patients in the non-RA group.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe table shows the distribution of final diagnoses among the 84 patients classified as non-RA. Normal findings or transient conditions were the most common (42.86%), followed by osteoarthritis (29.76%) and fibromyalgia (11.90%). Five patients presented with extra-articular inflammatory conditions (organizing pneumonia, pneumonia, psoriasis, tuberculosis, and nontuberculous mycobacteriosis).\u003c/p\u003e","description":"","filename":"supplementarytable1.docx","url":"https://assets-eu.researchsquare.com/files/rs-7564397/v1/3540c7fd7834e41040d82cd7.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Deep Learning System for Diagnosis of Rheumatoid Arthritis on Digital Hand Photographs","fulltext":[{"header":"Key messages","content":"\u003cul\u003e\n \u003cli\u003eA deep learning model using regular digital camera images of hands can diagnose rheumatoid arthritis with clinically viable performance.\u003c/li\u003e\n \u003cli\u003eThe model focuses on clinically relevant joint features, particularly in metacarpophalangeal and proximal interphalangeal joints.\u003c/li\u003e\n \u003cli\u003eThis accessible screening tool could expedite early RA detection in resource-limited healthcare settings.\u003c/li\u003e\n\u003c/ul\u003e"},{"header":"Introduction","content":"\u003cp\u003eRheumatoid arthritis (RA) is a chronic inflammatory disorder of unknown aetiology characterized by synovial infiltration of various inflammatory cells through autoimmune mechanisms, leading to irreversible destruction of synovial joints and resulting in severe physical disability and social disadvantage for patients [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Joint destruction in RA progresses more rapidly in the early stages after onset. Without appropriate treatment, significant bone destruction is often observed within two years of onset [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Therefore, to prevent irreversible joint damage, it is critical to establish a diagnosis as early as possible and initiate specific treatment [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. With the availability of numerous effective drugs, the prevention of joint destruction has become an achievable goal when adequate treatment is administered by specialists during the early stages of the disease [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eHowever, many cases still exist where the optimal timing for treatment is missed because of delays in care-seeking by patients, failure by physicians to diagnose RA, or insufficient treatment being provided, leading to severe and irreversible physical disabilities [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. To address these challenges, it is essential to develop accessible resources for the public that enable patients to self-assess their RA status and monitor their condition [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eWe hypothesized that an RA diagnostic model using deep learning could achieve diagnostic performance comparable to the 2010 ACR/EULAR criteria, which serve as the current standard for RA assessment in clinical practice and research settings without the need for laboratory examination results [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. To evaluate the feasibility of this approach, we conducted a preliminary pilot study targeting patients presenting with chief complaints of joint pain, joint swelling, and joint stiffness at participating medical institutions, as a precursor to a prospective cohort study.\u003c/p\u003e\u003cp\u003eIf the utility of this RA diagnostic model that applies deep learning to digital images of the dorsal aspects of both hands is validated, we anticipate that delays in seeking medical care for RA could be significantly reduced. Preventing declines in physical function in many cases may positively influence medical economics and the broader socioeconomic environment [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eParticipants\u003c/p\u003e\u003cp\u003eThis study used data from 180 individuals aged between 20 and 79 years who visited participating medical institutions with joint pain, joint swelling, or joint stiffness, and who had not previously received disease-modifying antirheumatic drugs or corticosteroid therapy. The cohort included 97 patients diagnosed with RA by two board-certified rheumatologists of the Japan College of Rheumatology with over 10 years of clinical experience, and 84 control subjects in whom RA and other collagen-vascular diseases were ruled out. Consistent with the well-established epidemiology of RA, which shows a strong female predominance (approximately 3:1 female-to-male ratio), our initial cohort included 15 male and 82 female RA patients compared with 4 male and 80 female non-RA subjects. To prevent potential gender-based confounding that could bias model training toward gender features rather than disease-specific pathological changes, we performed age-matched sampling among male patients, resulting in the exclusion of 11 male RA cases to achieve balanced gender distribution between groups. This study was conducted as a retrospective observational analysis of routinely collected clinical data. Written informed consent was waived by the ethics committee in accordance with national ethical guidelines. Participants were informed of the study via opt-out notices posted in the outpatient waiting area.\u003c/p\u003e\u003cp\u003eData Collection\u003c/p\u003e\u003cp\u003eData extracted from medical records included age, sex, disease duration, and the presence or absence of tenderness or swelling in the following bilateral joints: proximal interphalangeal (PIP), metacarpophalangeal, first interphalangeal, wrist, elbow, shoulder, hip, knee, ankle, and metatarsophalangeal. Additionally, we extracted information on serum C-reactive protein (CRP) levels, erythrocyte sedimentation rate (ESR), rheumatoid factor (RF), and anti-cyclic citrullinated peptide antibodies (ACPA).\u003c/p\u003e\u003cp\u003eImage Acquisition and Processing\u003c/p\u003e\u003cp\u003eBilateral dorsal hand images were captured against a green surgical drape background by trained nurses using a C40Z digital camera (Olympus Optical Co., Ltd). Images were taken under standardized lighting conditions without flash in a designated room to minimize variation.\u003c/p\u003e\u003cp\u003eThe digital images were saved in JPEG format and processed using a deep learning-based background removal algorithm (GitHub: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/danielgatis/rembg\u003c/span\u003e\u003cspan address=\"https://github.com/danielgatis/rembg\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). Subsequently, the files were processed in Pixelmater (Pixelmator Team, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.pixelmator.com/mac/\u003c/span\u003e\u003cspan address=\"https://www.pixelmator.com/mac/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) to manually separate the left- and right-hand images. Using Pillow (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://pypi.org/project/Pillow/\u003c/span\u003e\u003cspan address=\"https://pypi.org/project/Pillow/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.) and OpenCV libraries (Open Source Computer Vision Library, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://opencv.org/\u003c/span\u003e\u003cspan address=\"https://opencv.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), we applied padding to create square images and horizontally flipped the left-hand images. While most image processing steps were automated using rembg and Pillow libraries, the separation of left- and right-hand images required manual intervention. Preliminary analysis confirmed that automated background removal significantly improved model focus on anatomical features rather than environmental artifacts.\u003c/p\u003e\u003cp\u003eNeural Network Architecture and Training\u003c/p\u003e\u003cp\u003ePrior to any model development, we first performed a strict data split to ensure complete independence of the test set. The dataset was divided into training (60%), validation (30%), and test (10%) sets, while maintaining the same ratio of RA to non-RA cases across all sets through stratified sampling. To ensure unbiased evaluation of the model’s generalization ability, the test set was immediately sequestered and was not accessed during any part of the model development process, including architecture selection, hyperparameter tuning, and performance optimization, with model development and optimization being performed exclusively using the training and validation sets.\u003c/p\u003e\u003cp\u003eWe adopted a neural network architecture based on the Swin Transformer for the image classification task of distinguishing RA patients from non-RA individuals. The Swin Transformer is a hierarchical vision transformer utilizing shifted windows, which efficiently captures both local and global features within images [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Transfer learning from ImageNet-1K pretraining has been shown to provide robust feature representations that can be effectively adapted to medical imaging tasks, particularly when training data is limited [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe Swin Transformer base model (swin_base_patch4_window7_224) pretrained on ImageNet-1K was used as the backbone architecture. This variant processes input images with a patch size of 4×4 pixels and employs a window size of 7×7 for self-attention computation, with input images resized to 224×224 pixels.\u003c/p\u003e\u003cp\u003eRegularization techniques including dropout (rate = 0.6), attention dropout (rate = 0.5), and stochastic depth (rate = 0.5) were applied to enhance generalization performance. Cross-entropy loss with label smoothing (0.15) was used as the loss function, and the AdamW optimizer (learning rate = 2e\u003csup\u003e− 5\u003c/sup\u003e, weight decay = 2.5e\u003csup\u003e− 2\u003c/sup\u003e) with a OneCycleLR scheduler (max_lr = 2.2e\u003csup\u003e− 5\u003c/sup\u003e, pct_start = 0.3, div_factor = 15, final_div_factor = 2000, anneal_strategy='cos') was employed for optimization.\u003c/p\u003e\u003cp\u003eData preprocessing and augmentation techniques were applied, including random rotations (± 15°), affine transformations (translation: ± 0.08, scale: 0.92–1.08), colour jittering (brightness, contrast, saturation: ± 0.2, hue: ± 0.05), and random perspective transformations (distortion scale: 0.15, probability: 0.4). The model was trained with a batch size of 64 over 300 epochs with gamma = 0.9 and seed = 42. Mixup (alpha = 0.8) was used as an additional data augmentation technique to further improve model generalization.\u003c/p\u003e\u003cp\u003eScoring of ACR/EULAR Classification Criteria\u003c/p\u003e\u003cp\u003eThe 2010 ACR/EULAR rheumatoid arthritis classification criteria evaluate four domains to calculate a total score ranging from 0 to 10 points: joint involvement (0–5 points), serology (0–3 points), acute-phase reactants (0–1 point), and symptom duration (0–1 point). A total score of 6 or greater classifies a patient as having RA. Joint involvement is scored based on the number and size of involved joints: 0 points for 1 large joint, 1 point for 2–10 large joints, 2 points for 1–3 small joints, 3 points for 4–10 small joints, and 5 points for more than 10 joints including at least 1 small joint. Serology scores are based on RF and ACPA levels: 0 points for negative results, 2 points for low-positive, and 3 points for high-positive results. Acute-phase reactants (ESR and CRP) are scored as 1 point if elevated, and symptom duration is scored as 1 point if symptoms have persisted for 6 weeks or longer [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eModel Calibration and Interpretability\u003c/p\u003e\u003cp\u003eTo ensure reliable probability estimates, we performed probability calibration using Platt scaling [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. This method, which has been shown to be effective for calibrating deep neural networks [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], involves training a logistic regression model for which the independent variables are the raw model outputs (logits) and the dependent variable is the reference standard RA diagnosis established by rheumatologists. The calibration model was trained using the validation set logits and their corresponding true diagnostic labels, and was then applied to transform both validation and test set predictions into well-calibrated probabilities. This calibration step is crucial for obtaining reliable probability estimates, especially for clinical applications where decision thresholds are important. The calibrated probabilities were then used for all subsequent analyses, including receiver operating characteristics (ROC) curve generation and performance metric calculations.\u003c/p\u003e\u003cp\u003eThe optimal classification threshold was determined by maximizing Youden’s index for the validation set. This threshold was then applied to the test set for final performance evaluation.\u003c/p\u003e\u003cp\u003eGradient-Weighted Class Activation Mapping (Grad-CAM) was applied to interpret the model’s predictions and identify regions of input images influential in the decision-making process. Forward and backward hooks were registered on the last block of the Swin Transformer to capture feature maps and gradients during forward and backward passes. The Grad-CAM algorithm computed class activation maps by weighting the captured feature maps with the gradients’ global average pooling, followed by ReLU activation and normalization. The resulting heatmaps were overlaid on the original images to highlight salient regions [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eOutcomes\u003c/p\u003e\u003cp\u003eThe primary outcome was the area under the ROC curve (AUC) of the deep learning model in comparison with the ACR/EULAR classification criteria when applied to the test set, with the reference standard diagnosis being established by experienced rheumatologists. Secondary outcomes on the test set included sensitivity, specificity, accuracy, F1-score, and Brier score to assess calibration performance, all evaluated using the optimal threshold determined from the validation set. While the 2010 ACR/EULAR criteria were originally developed for classification purposes in research settings, they have become the de facto standard for clinical RA assessment and have been widely used as benchmarks in diagnostic studies [\u003cspan additionalcitationids=\"CR16\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e–\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Multiple validation studies have evaluated their diagnostic performance and recommended their use for uniformity and comparability in diagnostic research [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. In this context, we used these criteria as our primary comparator to evaluate the clinical utility of our deep learning approach.\u003c/p\u003e\u003ch2\u003eStatistical Analysis\u003c/h2\u003e\u003cp\u003eStatistical analyses were conducted using Python with the packages scikit-learn and statsmodels. Continuous variables were compared using the Wilcoxon rank-sum test, and categorical variables were analysed using Pearson’s chi-square test. Model performance was evaluated using sensitivity, specificity, accuracy, and F1 score. ROC curves were generated and the AUCs were calculated with 95% confidence intervals. Comparison of AUCs between the deep learning model and ACR/EULAR criteria was performed using DeLong’s test. Model calibration was assessed using Brier scores and calibration plots, with Platt scaling applied for probability calibration. The optimal probability threshold was determined using Youden’s index. Statistical significance was defined as P \u0026lt; 0.05.\u003c/p\u003e\u003cp\u003eEthical Considerations\u003c/p\u003e\u003cp\u003e This study adhered to the Declaration of Helsinki and the Ethical Guidelines for Medical and Health Research Involving Human Subjects. The study protocol was reviewed and approved by the Ethics Committee of National Hospital Organization Tochigi Medical Centre (approval number: 2025-02). Clinical trial number: not applicable. Written informed consent was waived by the ethics committee as this was a retrospective observational study using existing clinical data with minimal risk to participants. Information about the study was disclosed to participants via the postings in the outpatient waiting area, providing them the opportunity to opt-out of participation in accordance with Japanese ethical guidelines for medical research. Because this observational study used photographs taken with standard digital cameras and data extracted from routine medical records, there was no additional burden, expected risks, or direct benefits to the participants.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003ePatient Characteristics\u003c/p\u003e\u003cp\u003eInitially, 97 patients with RA and 84 subjects without RA were enrolled. After age-based pair matching of male subjects to address gender imbalance, the final analysis included 86 RA cases and 84 non-RA cases (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe gender distribution in our final cohort (4 males, 162 females) reflects the established epidemiological pattern of RA, which predominantly affects women with a typical female-to-male ratio of 2\u0026ndash;4:1. Our decision to match by gender was driven by the need to ensure that the deep learning model would focus on disease-specific morphological features rather than gender-related anatomical differences in hand appearance.\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the characteristics of these matched groups. There were no significant differences in age, sex, or symptom duration between the two groups. Inflammatory markers such as ESR and CRP, and immunological markers such as RF and ACPA, were significantly higher in the RA group (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05 for all).\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eBaseline characteristics and ACR/EULAR classification criteria scores of RA and non-RA patients\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003enonRA (n\u0026thinsp;=\u0026thinsp;84)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRA (n\u0026thinsp;=\u0026thinsp;86)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003edifference (95% C.I.)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eage, years (SD)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e61.7 (11.6)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e62.6 (12.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.9 (-2.7 to 4.4)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003esex, female (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e80 (95.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e82 (95.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.1 (-6.3 to 6.5)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003esymptom duration, week (SD)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e78.7 (145.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e66.6 (171.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-12.1 (-59.9 to 35.6)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eESR, mm/hr (SD)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e12.3 (11.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e39.0 (32.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e26.6 (19.4 to 33.8)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCRP, mg/dl (SD)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.19 (0.66)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1.29 (2.25)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1.10 (0.61 to 1.60)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRF, IU/ml (SD)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e9.9 (18.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e127.5 (195.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e117.6 (76.2 to 159.0)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eACPA, U/ml (SD)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1.1 (2.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e182.3 (298.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e181.2 (118.0 to 244.3)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ejoint score\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e0 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e51 (60.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e14 (16.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-43.3 (-57.5 to -31.4)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e2 (2.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1 (1.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-1.2 (-5.2 to 2.8)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e2 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e19 (22.6)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e37 (43.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e19.3 (6.6 to 34.2)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e11 (13.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e26 (30.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e17.1 (5.0 to 29.2)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e5 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1 (1.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e8 (9.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e8.1 (1.5 to 14.7)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003einflammation score\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e0 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e45 (53.6)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e5 (5.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-47.8 (-59.5 to -36.0)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e39 (46.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e81 (94.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e47.8 (36.0 to 59.5)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eimmunological score\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e0 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e70 (83.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e6 (7.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-76.3 (-86.0 to -66.7)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e2 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e6 (7.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e2 (2.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-4.8 (-11.2 to 1.5)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e8 (9.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e78 (90.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e81.2 (72.4 to 90.0)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eduration score\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e0 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e4 (4.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e7 (8.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e3.4 (-4.0 to 10.7)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e80 (95.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e79 (91.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-3.4 (-10.7 to 4.0)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eoverall score\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e22 (26.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0 (0.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-26.2 (35.6 to -16.8)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e2 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e19 (22.6)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e2 (2.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-20.3 (-29.8 to -10.8)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e18 (21.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0 (0.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-21.4 (-30.2 to -12.7)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e4 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e13 (15.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e3 (3.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-10.8 (-20.6 to -3.3)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e5 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e8 (9.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e11 (12.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e3.3 (-6.2 to 12.7)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e6 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e2 (2.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e8 (9.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e6.9 (-0.0 to 13.9)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e7 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1 (1.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e36 (41.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e39.5 (30.0 to 51.4)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e8 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1 (1.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e20 (23.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e22.1 (12.8 to 31.3)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e10 (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0 (0.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e6 (7.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e7.0 (1.6 to 12.4)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e\u003cp\u003eContinuous variables are presented as mean (standard deviation), and categorical variables are presented as number (%). Differences between groups are shown with their 95% confidence intervals. ACPA, anti-citrullinated peptide antibody; CRP, C-reactive protein; ESR, erythrocyte sedimentation rate; nonRA, non-rheumatoid arthritis; RA, rheumatoid arthritis; RF, rheumatoid factor; SD, standard deviation; 95% C.I., 95% confidence interval\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe most common diagnoses were normal findings or transient conditions, followed by osteoarthritis and fibromyalgia (Supplementary Table\u0026nbsp;1). Five patients presented with extra-articular inflammatory conditions, and two patients were diagnosed with hyperuricemia. Supplementary Fig.\u0026nbsp;1 shows the distribution of CRP, ESR, RF, and ACPA levels in the non-RA group. Some patients exhibited elevated CRP and ESR levels, but most of these elevations were mild. Excluding one case of pneumonia and two cases of mycobacterial infection, the elevated inflammatory markers were considered to be of limited clinical significance.\u003c/p\u003e\u003cp\u003eIn our cohort, the joint involvement scores showed clear differences between the groups, with the RA group having a higher proportion of moderate to severe scores (2 points or higher) and the non-RA group predominantly showing normal scores (0 points). Regarding acute-phase reactant scores, 94.2% of the RA group showed abnormal values (1 point), whereas 46.4% of the non-RA group had elevated inflammatory markers despite non-inflammatory clinical diagnoses. For the serology scores, 90.7% of the RA group had high-positive results (3 points), while 83.3% of the non-RA group had negative results (0 points). There was no difference in symptom duration scores between the groups.\u003c/p\u003e\u003cp\u003eThe total ACR/EULAR classification criteria scores showed distinct distributions in the two groups. The RA group demonstrated higher total scores, with 64% scoring 7 or above, whereas only 3.6% of the non-RA group reached this threshold. This distribution aligns with the expected performance of these classification criteria in distinguishing RA from other forms of arthritis and joint conditions.\u003c/p\u003e\u003cp\u003eModel Performance\u003c/p\u003e\u003cp\u003eThe performance of the trained model was evaluated using the test set, and the results of this evaluation are shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The model demonstrated high sensitivity, but relatively low specificity. After implementing Platt scaling calibration using the validation set and applying the threshold of 0.564 determined by Youden index analysis, we observed improved specificity without substantial deterioration in accuracy or F1 score, despite a slight decrease in sensitivity. In comparison, the ACR/EULAR classification criteria with a cutoff score of 6 points showed comparable sensitivity to the deep learning model but superior specificity, accuracy, and F1 score.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eDiagnostic performance metrics comparing ACR/EULAR RA classification criteria and deep learning model with and without probability calibration\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eACR/EULAR RA criteria\u0026dagger;\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eUncalibrated Deep Learning RA score\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eCalibrated Deep Learning RA score\u0026Dagger;\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSensitivity\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.778\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.944\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.778\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1.000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.667\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.833\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.889\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.806\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.806\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eF1-score\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.875\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.829\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.800\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e\u003cp\u003e\u0026dagger;Using the established cutoff score of 6 points in the ACR/EULAR criteria; \u0026Dagger;Using a cutoff value of 0.564; ACR/EULAR criteria, American College of Rheumatology and European Alliance of Associations for Rheumatology Rheumatoid Arthritis classification criteria; RA, rheumatoid arthritis\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents the calibration plots after Platt scaling. The Brier scores were 0.150 for the deep learning model and 0.112 for the ACR/EULAR classification criteria. Although the ACR/EULAR criteria performed marginally better, both methods demonstrated substantially better performance than random prediction (0.250). The calibration plots showed good alignment with ideal calibration in the 0.4\u0026ndash;0.6 probability range. However, both methods exhibited strong binary tendencies in their probability distributions, with limited intermediate probability predictions. Notably, the deep learning score showed an abrupt transition around the 0.8 probability threshold.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents ROC curves comparing the overall performance of the ACR/EULAR RA classification criteria and the deep learning RA score. The ACR/EULAR RA criteria achieved an AUC of 0.981 (95% CI: 0.953\u0026ndash;1.010), exceeding the deep learning RA score\u0026rsquo;s AUC of 0.870 (95% CI: 0.708\u0026ndash;0.988) by 0.111 (95% CI: \u0026minus;0.033\u0026ndash;0.255). However, this difference did not reach statistical significance (p\u0026thinsp;=\u0026thinsp;0.131).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eGrad-CAM visualizations revealed that the model appropriately focused on clinically relevant joint regions typically affected in early RA, particularly the metacarpophalangeal and PIP joints (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). The model paid special attention to the joint margins and surrounding soft tissues, aligning with conventional RA assessment.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this pilot study, we developed a deep learning model for diagnosing untreated RA using digital camera images of bilateral dorsal hands and evaluated its performance against the 2010 ACR/EULAR RA classification criteria. Analysis of the ROC curve demonstrated that the deep learning model achieved clinically viable performance. While the AUC of the 2010 ACR/EULAR RA classification criteria surpassed that of the deep learning model, this difference did not reach statistical significance. Using the cutoff value determined through validation set calibration, the deep learning model’s sensitivity for detecting RA was comparable to the 2010 ACR/EULAR classification criteria. However, the 2010 criteria demonstrated superior specificity, accuracy, and F1-score compared with the deep learning model. Post-Platt scaling calibration analysis revealed good alignment, with ideal calibration for both methods in the 0.4–0.6 probability range, although both exhibited a strong binary tendency in their probability distributions. Notably, visualization of decision rationale using Grad-CAM confirmed that the model appropriately focused on clinically relevant joint regions, particularly the morphology and skin creases of the metacarpophalangeal and PIP joints, which are typically affected in early RA.\u003c/p\u003e\n\u003cp\u003eOur choice to benchmark against the ACR/EULAR criteria reflects their widespread adoption in clinical practice for RA assessment. While originally designed for research classification, these criteria have evolved into the most commonly used standardized approach for RA evaluation in real-world clinical settings, making them an appropriate reference standard for diagnostic performance comparison.\u003c/p\u003e\n\u003cp\u003eCurrent clinical methods for evaluating RA include joint inspection and palpation, acute phase reactants such as CRP and ESR from blood samples, immunological markers such as RF and ACPA, plain radiography, joint ultrasonography, and magnetic resonance imaging. In today’s expanding field of machine learning and deep learning research, attempts to apply deep learning techniques to these modalities are emerging worldwide. However, there are relatively few reports on the use of digital camera images to evaluate RA, as employed in our study, despite the fact that this is arguably the most accessible imaging modality for patients.\u003c/p\u003e\n\u003cp\u003eTo our knowledge, Hügle et al. were the first to evaluate RA using digital camera images of the dorsal hand [18]. They focused on dorsal skin creases of PIP joints, training ResNet34 to differentiate between swollen and non-swollen joints using cropped local images, suggesting local findings in joints as potential digital biomarkers for RA. Our study similarly found that the deep learning model focused on finger creases for RA discrimination, aligning with their findings. However, while we successfully analysed the entire hand image, their study remained at a more fundamental stage, analysing only the local features of the PIP joints.\u003c/p\u003e\n\u003cp\u003ePhatak et al. acquired smartphone images of both dorsal hands from 200 patients with inflammatory arthritis and 200 healthy individuals [19]. Using ResNet-based analysis, they successfully detected arthritis from whole-hand images with high accuracy. Their success can likely be attributed to the availability of a relatively large sample size and the fact that they identified hands with physician-confirmed arthritis. In contrast, our study aimed to explore whether a model could detect finger joint inflammation that is difficult for human observers to identify. We attempted to detect RA cases—diagnosed comprehensively through clinical examinations, laboratory tests, and radiographic findings—using only hand photographs. The seemingly lower diagnostic performance of our study compared to theirs may be explained by the inclusion of many RA patients who had been clinically diagnosed but did not exhibit detectable joint inflammation upon physical examination.\u003c/p\u003e\n\u003cp\u003eDespite tackling a more challenging task, our study achieved a certain degree of success. We hypothesize that this may be due to the difference in model architectures: while they employed a CNN-based approach, we utilized a Swin Transformer. CNNs respond more strongly to fine-grained texture patterns in images, whereas Swin Transformers, due to their hierarchical nature, can capture both subtle local changes and broader global variations. This characteristic of Swin Transformers may be particularly advantageous in analyzing medical images, which often present diverse and complex patterns of change.\u003c/p\u003e\n\u003cp\u003eA notable advantage of our approach is its reliance solely on digital camera images of dorsal hands, which can be easily captured by almost everyone with everyday equipment, rather than clinical tests or advanced imaging. Our system, which requires neither expensive imaging equipment nor blood biomarkers, could particularly enhance the detection of early RA in primary care settings and regions with limited rheumatology resources. More rapid and accessible screening could expedite referrals to rheumatologists, reduce diagnostic delays, and enable timelier initiation of disease-modifying therapy—essential factors for minimizing joint damage and maintaining patient function.\u003c/p\u003e\n\u003cp\u003eDespite promising results, several limitations warrant acknowledgment. First, our analysis is based on a limited sample size from clinical settings, potentially resulting in insufficient model training and performance acquisition; however, increasing the sample size should lead to improved training and performance, not less. We must also acknowledge that the insufficient size of the test set sample may have compromised stable performance evaluation. Larger-scale multi-centre studies are essential to confirm generalizability and establish external validity. Second, while the background removal procedures improved uniformity, data reliability may still show variations due to differences in lighting, patient positioning, or camera angles. Successful real-world deployment will require rigorous training procedures and prospective validation across various image capture conditions.\u003c/p\u003e\n\u003cp\u003eOur gender matching approach, while necessary to prevent model bias toward gender-specific features, may limit the generalizability of our findings to the broader RA population. However, this approach ensures that the model's diagnostic performance is based on pathological changes rather than demographic characteristics, which is essential for developing a clinically relevant diagnostic tool.\u003c/p\u003e\n\u003cp\u003eIn conclusion, our deep learning-based RA diagnostic approach, enhanced by advanced architecture, demonstrated the potential to achieve clinically viable accuracy. Future large-scale multi-centre prospective studies are needed to establish more stable and reliable results.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements (if applicable):\u003c/strong\u003e We thank the patients and staff at the participating medical institutions for their cooperation. We thank Edanz (https://jp.edanz.com/ac) for editing a draft of this manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eGroup authorship list (if applicable and all group members fulfil ICMJE criteria for authorship):\u0026nbsp;\u003c/strong\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions:\u0026nbsp;\u003c/strong\u003eRH designed the study, analysed the data, and drafted the manuscript. SS collected the data. All authors critically revised the manuscript and approved the final version.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding (if specific to this study):\u0026nbsp;\u003c/strong\u003eThis research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflict of interest statement (in addition to author disclosure form):\u0026nbsp;\u003c/strong\u003eThe authors declare no conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability statement:\u0026nbsp;\u003c/strong\u003eThe data underlying this article will be shared upon reasonable request to the corresponding author.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical trial number:\u003c/strong\u003e not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eHuman Ethics and Consent to Participate declarations:\u003c/strong\u003e This study was approved by the Ethics Committee of National Hospital Organization Tochigi Medical Centre (approval number: 2025-02). Written informed consent was waived by the ethics committee in accordance with national ethical guidelines for retrospective studies. Information regarding the study was disclosed to participants via opt-out postings in the outpatient waiting area.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eFirestein GS, McInnes IB. Immunopathogenesis of rheumatoid arthritis. \u003cem\u003eImmunity\u003c/em\u003e 2017;46:183-196.\u003c/li\u003e\n\u003cli\u003evan der Heijde DM. Joint erosions and patients with early rheumatoid arthritis. \u003cem\u003eBr J Rheumatol\u003c/em\u003e 1995;34 Suppl 2:74-8.\u003c/li\u003e\n\u003cli\u003evan der Linden M, le Cessie S, Raza K, van der Woude D, Knevel R, Huizinga TW, et al. Long-term impact of delay in assessment of patients with early rheumatoid arthritis. \u003cem\u003eArthritis Rheum\u003c/em\u003e 2010;62:3537-46.\u003c/li\u003e\n\u003cli\u003eSmolen JS, Landew\u0026eacute; RBM, Bijlsma JWJ, Burmester GR, Dougados M, Kerschbaumer A, et al. EULAR recommendations for the management of rheumatoid arthritis with synthetic and biological disease-modifying antirheumatic drugs. \u003cem\u003eAnn Rheum Dis \u003c/em\u003e2020;79:685-99.\u003c/li\u003e\n\u003cli\u003eStack RJ, Nightingale P, Jinks C, Shaw K, Herron-Marx S, Horne R, et al. Delays between the onset of symptoms and first rheumatology consultation in patients with rheumatoid arthritis in the UK: an observational study. \u003cem\u003eBMJ Open\u003c/em\u003e 2019;9:e024361.\u003c/li\u003e\n\u003cli\u003eConnelly K, Segan J, Lu A, Saini M, Cicuttini FM, Chou L, et al. Patients\u0026rsquo; perceived health information needs in inflammatory arthritis: A systematic review. \u003cem\u003eSemin Arthritis Rheum\u003c/em\u003e 2019;48:900-10.\u003c/li\u003e\n\u003cli\u003eAletaha D, Neogi T, Silman AJ, Funovits J, Felson DT, Bingham CO 3rd, et al. 2010 Rheumatoid arthritis classification criteria. \u003cem\u003eArthritis Rheum\u003c/em\u003e 2010;62:2569-81.\u003c/li\u003e\n\u003cli\u003eFeldman DE, Bernatsky S, Houde M, Beauchamp ME, Abrahamowicz M. Early consultation with a rheumatologist for RA: does it reduce subsequent use of orthopaedic surgery? \u003cem\u003eRheumatology\u003c/em\u003e 2013;52:452-59.\u003c/li\u003e\n\u003cli\u003eLiu Z, Ning J, Cao Y, Hu H, Wei Y, Zhang Z, et al. Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows. \u003cem\u003eInt Conf Comput Vis\u003c/em\u003e 2021:10012-22.\u003c/li\u003e\n\u003cli\u003eTajbakhsh N, Shin JY, Gurudu SR, et al. Convolutional neural networks for medical image analysis: full training or fine tuning? IEEE Trans Med Imaging. 2016;35(5):1299-1312. \u003c/li\u003e\n\u003cli\u003eKather JN, Pearson AT, Halama N, et al. Transfer learning for medical image classification: a literature review. BMC Med Imaging. 2022;22:69.\u003c/li\u003e\n\u003cli\u003ePlatt JC. Probabilistic Outputs for Support Vector Machines and Comparisons to Regularized Likelihood Methods. Smola AJ, Bartlett PL, Sch\u0026ouml;lkopf B, Schuurmans D, editors. Advances in Large Margin Classifiers. Cambridge (MA): MIT Press; 2000. p.61-74.\u003c/li\u003e\n\u003cli\u003eGuo C, Pleiss G, Sun Y, Weinberger KQ. On Calibration of Modern Neural Networks. \u003cem\u003eProceedings of the 34th International Conference on Machine Learning.\u003c/em\u003e 2017:1321-30.\u003c/li\u003e\n\u003cli\u003eSelvaraju RR, Cogswell M, Das A Vedantam R, Parikh D, Batra D. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. \u003cem\u003eInt Conf Comput Vis\u003c/em\u003e 2017:618-26.\u003c/li\u003e\n\u003cli\u003eAlves C, Luime JJ, van Zeben D, et al. Diagnostic performance of the ACR/EULAR 2010 criteria for rheumatoid arthritis and two diagnostic algorithms in an early arthritis clinic (REACH). Ann Rheum Dis. 2011;70(9):1645-1647. \u003c/li\u003e\n\u003cli\u003eVarache S, Cornec D, Morvan J, et al. Diagnostic accuracy of ACR/EULAR 2010 criteria for rheumatoid arthritis in a 2-year cohort. J Rheumatol. 2011;38(7):1250-1257. \u003c/li\u003e\n\u003cli\u003eBritsemmer K, Ursum J, Gerritsen M, et al. Validation of the 2010 ACR/EULAR classification criteria for rheumatoid arthritis: slight improvement over the 1987 ACR criteria. Ann Rheum Dis. 2011;70(8):1468-1470.\u003c/li\u003e\n\u003cli\u003eH\u0026uuml;gle T, Caratsch L, Caorsi M, Maglione J, Dan D, Dumusc A, et al. Dorsal Finger Fold Recognition by Convolutional Neural Networks for the Detection and Monitoring of Joint Swelling in Patients with Rheumatoid Arthritis. \u003cem\u003eDigit Biomark\u003c/em\u003e 2022;6:31-5.\u003c/li\u003e\n\u003cli\u003ePhatak S, Saptarshi R, Sharma V, Shah R, Zanwar A, Hegde P, et al. Incorporating computer vision on smart phone photographs into screening for inflammatory arthritis: results from an Indian patient cohort. Rheumatology (Oxford). 2024 16:keae678.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-rheumatology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"brhm","sideBox":"Learn more about [BMC Rheumatology](http://bmcrheumatol.biomedcentral.com)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/brhm/default.aspx","title":"BMC Rheumatology","twitterHandle":"@BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"(Up to 10) rheumatoid arthritis, diagnosis, deep learning, ACR/EULAR criteria","lastPublishedDoi":"10.21203/rs.3.rs-7564397/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7564397/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eObjectives:\u003c/strong\u003e To develop and evaluate a deep learning model for diagnosing untreated rheumatoid arthritis (RA) using digital camera images of bilateral dorsal hands, benchmarking its performance against the widely-used 2010 ACR/EULAR criteria as a clinical reference standard.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods:\u003c/strong\u003e This pilot study included 170 participants (86 RA, 84 non-RA) who presented with joint symptoms at participating medical institutions. Digital images of both dorsal hands were captured under standardized conditions and processed using a deep learning-based background removal algorithm. A Swin Transformer-based model was developed and trained on these images. Model performance was evaluated using area under the receiver operating characteristics curve (AUROC), sensitivity, specificity, and calibration metrics. Gradient-Weighted Class Activation Mapping (Grad-CAM) was employed to visualize the model’s decision-making process.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003e The deep learning model achieved an AUROC of 0.870 (95% CI: 0.708-0.988), compared with 0.981 (95% CI: 0.953-1.010) for the ACR/EULAR criteria, with the difference not reaching statistical significance (p=0.131). While demonstrating comparable sensitivity to the ACR/EULAR criteria, the model showed lower specificity, accuracy, and F1-score. Post-Platt scaling calibration analysis revealed good alignment with ideal calibration in the 0.4–0.6 probability range. Grad-CAM visualization confirmed that the model focused on clinically relevant joint regions, particularly the metacarpophalangeal and proximal interphalangeal joints.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion:\u003c/strong\u003e Our deep learning-based approach for RA diagnosis using standard digital camera images demonstrated clinically viable performance, albeit with lower specificity than the ACR/EULAR criteria. This accessible screening tool could potentially expedite early RA detection, particularly in resource-limited settings. Larger multi-centre studies are needed to validate our findings and establish broader clinical applicability.\u003c/p\u003e","manuscriptTitle":"A Deep Learning System for Diagnosis of Rheumatoid Arthritis on Digital Hand Photographs","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-23 07:09:15","doi":"10.21203/rs.3.rs-7564397/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-01-27T13:42:14+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-23T16:47:37+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"200454160449124654979496035941255903364","date":"2026-01-02T11:28:41+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-01T08:30:02+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"112052957756926316710230600913004502330","date":"2025-10-21T21:16:16+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"61940530174067170533002365534140461008","date":"2025-10-19T01:08:02+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-09-14T08:25:23+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-09-12T18:59:54+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-09-10T05:21:48+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-09-10T05:20:59+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Rheumatology","date":"2025-09-08T12:43:54+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-rheumatology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"brhm","sideBox":"Learn more about [BMC Rheumatology](http://bmcrheumatol.biomedcentral.com)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/brhm/default.aspx","title":"BMC Rheumatology","twitterHandle":"@BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"56ffc65f-b7ac-4767-9368-417c1afffc87","owner":[],"postedDate":"September 23rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-03-30T16:21:41+00:00","versionOfRecord":{"articleIdentity":"rs-7564397","link":"https://doi.org/10.1186/s41927-026-00639-7","journal":{"identity":"bmc-rheumatology","isVorOnly":false,"title":"BMC Rheumatology"},"publishedOn":"2026-03-24 16:10:55","publishedOnDateReadable":"March 24th, 2026"},"versionCreatedAt":"2025-09-23 07:09:15","video":"","vorDoi":"10.1186/s41927-026-00639-7","vorDoiUrl":"https://doi.org/10.1186/s41927-026-00639-7","workflowStages":[]},"version":"v1","identity":"rs-7564397","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7564397","identity":"rs-7564397","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}