Unsupervised machine-learning algorithm identifies clinically distinct phenotypes across patients initiating belimumab in the Lupus Italian REgistry (LIRE)

1 Introduction

The biological anti-B Lymphocyte Stimulator (BLyS) belimumab is increasingly used as add-on therapy in systemic lupus erythematosus (SLE) refractory to standard of care. The eligibility of patients to belimumab is mostly based on physician judgement, as is the concept of SLE refractoriness. In fact, international recommendations for SLE management leave room to the treating clinician about the timing of belimumab allocation during disease course, defining refractoriness as the insufficient response to standard treatment, which encompasses glucocorticoids and hydroxychloroquine (HCQ), with or without the failure of a previous traditional immunosuppressant (1). Growing evidence from randomized trials, long-extension studies and real-life cohorts has suggested specific manifestations, namely mucocutaneous and articular, to display the best response to belimumab in the short term; conversely, active smoking, long-standing SLE, established damage and chronic lesions have deemed able to hinder belimumab effectiveness (2–5). Lately, belimumab has been approved as add-on treatment for active lupus nephritis (LN) (6), thus widening its potential applicability.

Although SLE patients candidate to belimumab may share some common clinical features, the heterogeneity of SLE, combined with the widening range of use of the drug according to physician judgement, suggest that the population of SLE patients undertaking belimumab entails an inner heterogeneity, which might at least partially account for the different reported effectiveness of belimumab across different cohorts and related satisfaction.

Unsupervised learning is a machine-learning-based approach widely used in several science branches including the medical field, which allows clustering and dimensionality reduction (7) from big data. Historically, Principal Component Analysis (PCA) has been used for such tasks (8), which obtains the components by standardizing the values, building the covariance matrix and, from there getting the eigenvalues and eigenvectors. PCA arrives at the final components which explain most of the variability in the dataset.

The unsupervised approach expresses the inertia of given variables in driving cluster separation in relation to other variables entailing non linear relationship with the chosen outcome. In fact, as opposed to the traditional supervised statistical models, unsupervised learning depicts the segregation of data devoid of their inclusion into a multivariable model, thus virtually overcoming any investigator-driven bias.

The aim of our study is to explore the existence of clustering across a wide patient population undertaking belimumab due to refractory SLE, in order to explore any clinical heterogeneity and the potential effects of clustering on response to belimumab.

The study was conducted with the support of the Italian Society of Rheumatology (SIR).

2 Patients and methods

We analyzed data from the multicentric Lupus Italian Registry (LIRE) which encompasses prospectively collected clinical and laboratory records of SLE patients initiating a new treatment due to refractory SLE (9). We selected patients exclusively starting belimumab. Exclusion criteria were incomplete records at baseline and missing data>8% at 6 and 12 months. An 8% threshold allowed to retain a sufficient sample while minimizing the need for imputation and preventing bias in the variance structure that drives clustering.

Data wrangling was applied to obtain a homogenous dataset suitable for machine-learning analysis aimed at investigation of baseline clustering of patients undertaking belimumab. One-hundred fifty baseline input variables from clinical and laboratory features were selected in the algorithm, encompassing demographics, disease manifestations by organ domain, serological markers, treatment history, comorbidities, and damage items; this comprehensive selection was consistent with the unsupervised nature of the analysis, which aims to identify data-driven groupings without a priori assumptions about which features may drive clustering. Among retained patients, residual missing values (<8% per variable) were imputed using mean substitution for continuous variables and most-frequent-category substitution for categorical variables. After imputation, the resulting complete dataset entered the Factor Analysis of Mixed Data (FAMD). After cluster visualization, characterization was carried out by comparing baseline demographics, clinical and laboratory features.

Disease activity was measured by the SLE-disease activity index 2000 (SLEDAI-2 K). Clinical response was defined as clinical SLEDAI-2 K (cSLEDAI-2 K) = 0 at 6 and 12 months (T6, T12) (10). Organ damage was evaluated with the SLICC damage index (SDI) (11). Clinical response and damage progression were assessed at 6 and 12 months from baseline.

2.1 Machine-learning methodology

Because both continuous and categorical variables were present, we applied FAMD, i.e., a generalization of PCA that can be deployed on large datasets with qualitative and quantitative variables. FAMD was implemented in a Python ver. 3.9 environment, using Prince ver. 0.7.1, scipy ver. 1.9.3, scikit-learn ver. 1.1.3 and numpy ver. 1.23.2 on a MacBook Pro with Apple Silicon M1MAX 64 GB RAM.

Continuous variables were inspected for outliers and their distributions; as FAMD internally standardizes the variables, no additional scaling was applied beyond the default implementation. Categorical variables were automatically one-hot encoded by the Prince library and divided by the square root of their probability, and all variables were centred before PCA. Only patients with complete baseline data were entered in the FAMD.

The number of principal components to retain was determined using the scree/elbow criterion applied to the proportion of variance explained, supported by visual inspection of the scree plot (Supplementary Figure S1) and the corresponding eigenvalues reported in Supplementary Table S1. Using the derived components, we performed Hierarchical Clustering on Principal Components (HCPC) with Ward's criterion and Euclidean distance (12). The optimal number of clusters was chosen by inspection of the dendrogram and by identifying the elbow in the plot of inter-cluster inertia gain vs. the number of clusters, which supported a four-cluster solution.

2.2 Statistics

Continuous variables are expressed as median (interquartile range, IQR). After testing for normality (Shapiro–Wilk), Kruskal–Wallis H test for multiple independent samples was used to perform inter-cluster comparisons of continuous variables, while the Chi-squared test was used for categorical variables. Over-time comparisons were assessed using repeated-measures ANOVA or the Friedman test, depending on data distribution. Longitudinal trajectories were also evaluated using linear mixed-effects models with random intercepts.

The association between baseline cluster membership and clinical response was evaluated using binary logistic regression models with cluster as the main independent variable and achievement of cSLEDAI-2K = 0 (yes/no) as the dependent variable at each timepoint. Covariates included baseline GC intake and baseline confounders i.e., age, sex, ethnicity. Statistical significance was set at a two-sided α of 0.05.

2.3 Ethic statement

All patients provided written informed consent at the time of enrolment. The study was conducted in compliance with the Helsinki Declaration and received approval from the Ethical Committee (EC) of the Coordinator Centre (code 3524/AO/15) and of local centres’ ECs.

3 Results

3.1 Patient population

Overall, 143 patients in the LIRE database started a biological agent as add-on treatment. Of these, 128 patients starting belimumab had complete baseline data and entered the clustering procedure. HCPC identified four main patient groups comprising overall 125 individuals; three additional patients formed small isolated terminal branches on the dendrogram at the selected cut-point and were therefore not assigned to any of the four main clusters.

Baseline demographic and clinical features of selected patients are reported in Table 1.

Table 1. Baseline demographic and clinical features across patient clusters.

across clusters. Kruskal–Wallis H test for multiple independent samples was used for continuous variables; chi-squared test for the categorical variable.

BMI, body mass index; C3/C4, complement fragment ¾; Hb, heamoglobin; WBC, white blood cells; SLEDAI-2 K, SLE-disease activity index 2000; SLICC-DI: SLICC damage index; PGA, physician global assessment; GC, glucocorticoids. Continuous variables are expressed as median (IQR); ethnicity as n (%).

3.2 Baseline clustering

The FAMD identified five principal components that were retained for clustering, explaining 68.37% of the total variance; additional components contributed to only marginal gains, as shown in Supplementary Table S1 and Supplementary Figure S1. Application of HCPC on these five components, together with inspection of the dendrogram and the elbow in the inter-cluster inertia gain curve, supported a four-cluster solution at baseline (Cluster 1: n = 55; Cluster 2: n = 28; Cluster 3: n = 14; Cluster 4: n = 28), as depicted in Figure 1. Three additional patients formed small isolated terminal branches at the chosen cut-point and were therefore not assigned to any of the four main clusters and excluded from cluster-level analyses.

Clusters were comparable in terms of age, gender and ethnic origin. Baseline demographics and clinical features across different clusters are shown in Table 1.

3.2.1 Disease activity

While baseline SLEDAI-2 K did not differ significantly across clusters, some differences were retrieved in the distribution of clinical items making up the score. Cluster 2 included the majority of patients with active renal involvement at baseline (p < 0.001). Cluster 3 was characterized by patients with severe SLE features, including central nervous system (CNS) and cardiovascular manifestations. Patients in Cluster 4 displayed more frequently fever and haematological abnormalities, though not reaching statistical significance. Haemolytic anemia, which is not in included among the SLEDAI-2 K items, was significantly more prevalent in this cluster.

No differences across other SLEDAI-2 K items emerged. The proportion of patients displaying articular, mucocutaneous and serosal manifestations was similar across all Clusters.

Clinical items of disease activity separating clusters at baseline are reported in Table 2.

Table 2. Baseline disease activity (SLEDAI-2 K) items across patient clusters.

SLEDAI-2 K, SLE disease activity index 2000; CNS, central nervous system; LN, lupus nephritis. Kruskal–Wallis H test for multiple independent samples.

3.2.2 Serology

All patients displayed positive anti-nuclear antibodies (ANA) and the majority were positive for anti-dsDNA antibodies. Patients in Cluster 2 most frequently displayed complement consumption and lower absolute values of serum C3 and C4. Antiphospholipid antibodies were more frequently positive among patients belonging to Cluster 3. No significant differences in other antibody specificity were observed between clusters (Table 3); no clustering of anti-ENA was observed.

Table 3. Baseline serology across patient clusters.

ANA, antinuclear antibodies; anti-dsDNA, anti-double stranded DNA antibodies; anti-Sm, anti-Smith antibodies; aCL, anticardiolipin; anti-B2GPI, anti-Beta2 glycoprotein I; LAC, lupus anticoagulant; ENA, extractable nuclear antigens.

3.2.3 Damage

Baseline SDI was significantly higher in patients within Cluster 3. SLICC items most frequently reported in this Cluster included organic brain damage, cognitive dysfunction and psychosis, and peripheral vascular disease. Osteoporotic fractures were also more frequent in this cluster. Patients in Cluster 4 displayed higher frequency of scarring alopecia and cutaneous ulcers.

Comorbidities were more common in Cluster 3 and included cerebrovascular disease and dementia. Hypertension was also more frequent in Cluster 3 (42.9% vs. 19.8%, p = 0.06).

SLICC-DI items and comorbidities separating clusters are reported in Table 4.

Table 4. Baseline treatment across patient clusters.

HCQ, hydroxychloroquine; GC, glucocortidoids; MMF, mycophenolate mofetil; AZA, azathioprine; MTX, methotrexate; CsA, cyclosporine A; CYC, cyclophosphamide; IS, immunosuppressants. Chi-squared test.

3.2.4 Treatment

Most patients across all clusters were receiving GC at baseline. Median baseline GC intake was higher within Cluster 3, though not reaching statistical significance. Cluster 3 displayed the lowest proportion of patients receiving baseline HCQ (78.6% vs. 95.2%, p = 0.05). The highest proportion of patients receiving an ongoing immunosuppressive treatment at baseline beyond HCQ was found in Cluster 2, among which mycophenolate mofetil was the most frequently used. Cluster 2 patients were as well more likely to have received cyclophosphamide in the past. Treatment across clusters is reported in Table 5.

Table 5. Damage and comorbidities at baseline.

Chi-squared test.

3.3 Clinical outcome at subsequent timepoints

Response rates and damage accrual were assessed at T6 and T12. Clustering was tested as independent variable to predict clinical response in a logistic regression model.

3.3.1 Disease activity

Global disease activity measured through the SLEDAI-2 K significantly decreased throughout month 12 in the whole cohort [SLEDAI-2 K, T0 vs. T6 vs. T12: 9 (6–12) vs. 4 (2–5) vs. 4 (2–6), p < 0.001] and across all groups (Figures 2A,B). Clinical response was defined as cSLEDAI-2K = 0 at T6 and T12. Proportion of responders was not significantly different between clusters at any timepoint (Supplementary Table S2); Cluster 2 showed a numerically higher rate of response at T12 as compared to other clusters, although the difference did not reach statistical significance (70.8% vs. 52.3%, p = 0.08). In logistic regression models with cluster membership as the main predictor, baseline clustering was not significantly associated with the probability of achieving cSLEDAI-2K = 0 at either T6 or T12.

3.3.2 Damage

The overall cohort accrued significant damage through month 12, which was mostly driven by SDI increase in Cluster 1 (Figures 3A,B). No pattern of repetitive SDI items was documented within this cluster. Cluster 3 patients displayed the higher SDI at T6 and T12 (p < 0.001), whose most represented items were consistent with those from baseline, namely cardiovascular and neurological. Median values of variations in disease activity and damage scores across clusters are reported in Supplementary Table S2.

4 Discussion

In this study, we investigated the heterogeneity of SLE patients initiating belimumab through an unsupervised approach of hierarchical Clustering on Principal Components. Our data show that patients undertaking belimumab due to refractory SLE are heterogeneous and can be grouped into data-driven clusters that largely reflect recognizable clinical patterns of disease expression, entailing different risk of damage progression through a 12-month observation.

The unsupervised machine-learning methodology is being increasingly widespread across several scientific fields, including medicine and health care (13–15). The method is capable of defining the intrinsic weight of any selected variable within a wide range of independent variables (7), hence allowing a virtually unbiased identification of clusters according to the autonomous segregation of specific variables.

Among autoimmune rheumatic diseases, recent data on rheumatoid arthritis submit the unsupervised hierarchical clustering approach as a useful tool to discriminate patient clusters according to the underlying inflammatory pathways, regardless of clinical similarities (16). Whether these findings alone might predict future disease course needs proof, yet the perspective is promising (17).

Growing evidence is available about unsupervised clustering approaches in SLE. While most concern diagnostic algorithms (18), fewer have pointed towards prediction of response which are mostly focused on LN (19), with currently limited clinical application. Our study explores the clinical heterogeneity of patients candidate to the same add-on biological treatment in a wide SLE cohort through an unbiased approach, thereby allowing investigation of drug effectiveness according to baseline independent clinical stratification. Our observations suggest that patients receiving belimumab are indeed heterogeneous despite comparable overall disease activity, and such heterogeneity does not solely depend on overt clinical manifestations leading to start belimumab. Rather, patients tend to cluster according to features including established damage, former treatment, concomitant comorbidities and serological features; moreover, specific clinical manifestations gather together contributing to complex patient phenotyping.

We identified four main clusters. Patients with predominantly active renal disease and a higher serological activity (Cluster 2) showed a numerically higher response rate at one year. Although this difference did not reach statistical significance, this observation is consistent with previous evidence supporting the use of belimumab as add-on therapy in renal SLE (20). Conversely, patients exhibiting activity in other major organ systems i.e., CNS and cardiovascular built up a cluster burdened with higher damage and comorbidities at baseline, with a history of higher daily GC dosage and lower HCQ intake, and displayed a higher damage burden throughout the observational period, in keeping with established damage triggering additional damage (21–23). Patients with constitutional symptoms and haematological manifestations grouped together, highlighting a real-life use of belimumab in refractory haemolytic anaemia. The last group included patients without prominent peculiar clinical manifestations, displaying lower baseline damage and enriched with adequate HCQ intake. Articular and mucocutaneous manifestations were comparable across different clusters, which may be due to widely accepted data on belimumab effectiveness in such domains (2, 24, 25). It should also be acknowledged that alternative biologic therapies such as rituximab are frequently considered in patients with severe or refractory SLE manifestations, including lupus nephritis, neuropsychiatric involvement, vasculitis and severe hematological disease. These clinical scenarios may partially overlap with the profiles identified in our clusters and highlight the complexity of treatment selection in real-world practice.

Disease activity decreased significantly in the whole cohort and across all groups and clinical response measured with the cSLEDAI-2 K was aligned across different clusters, suggesting that clustering solely based on clinical and laboratory features may not sufficiently discriminate belimumab responders from non-responders. However, it should be noted that patients deemed to belimumab in the Italian health care system need to fulfil established criteria including positive serology and/or a baseline SLEDAI-2K ≥ 10, which might have flattened subtle differences in patient profiling. Patients within Cluster 2 (active renal and serology) exhibited a numerically higher response rate at one year, thus supporting belimumab use in this subgroup. An additional aspect that was not specifically addressed in the present analysis is the time to response across clusters. Although response was evaluated at predefined timepoints (6 and 12 months), differences in the temporal dynamics of response may exist between clusters and could provide further insight into treatment effectiveness. This aspect warrants dedicated investigation in future studies.

Furthermore, it is likely that a clustering approach employing molecular rather than clinical phenotyping might define clusters according to the underlying immunological abnormalities, which may better pinpoint potential responders. Recent data have shown that underlying enrichment of different B cell populations and particularly transitional B cells in SLE patients are associated with a better response to belimumab, regardless of the clinical disease expression (26). Differences in treatment exposure across clusters should also be considered when interpreting the observed disease trajectories. In particular, patients in the renal/serological cluster more frequently received immunosuppressive therapy such as mycophenolate mofetil and had higher rates of prior cyclophosphamide exposure, reflecting the underlying severity of renal disease. In this context, the concomitant use of mycophenolate mofetil may have contributed to the observed outcomes and could potentially act synergistically with belimumab, particularly in patients with renal involvement. Therefore, the treatment response observed in this cluster should be interpreted considering both disease phenotype and background immunosuppressive therapy. Conversely, the cluster characterized by neuropsychiatric and cardiovascular manifestations displayed lower hydroxychloroquine use and higher glucocorticoid exposure, both factors previously associated with increased damage accrual in SLE. These differences likely reflect clinical decision-making and disease severity rather than intrinsic biological differences between clusters, and they may partly contribute to the observed outcome patterns.

Consistently, our data suggest that the integrated approach of clinical and immunological phenotyping might lead to a more precise stratification of patients, despite an apparent clinical overlap.

Damage worsening throughout the observational period was not significant across three out of four clusters, consistently with former data on belimumab halting damage progression (2, 4, 27, 28). However, the SDI score significantly increased within Cluster 1 throughout month 12. Despite a milder initial phenotype, a statistically significant yet clinically modest increase in SDI is expected throughout SLE course (22, 23) which is not hindered by the even optimized use of HCQ and may not be completely abolished by the addition of biological drugs, as suggested by the accrual of SDI items within Cluster 1. Additionally, it should be noted that the median disease duration at belimumab initiation in this cohort was around 10 years, which prevented early treatment and may have negatively impacted damage prevention (2). Nevertheless, it may not be ruled out that this cluster gathered patients vulnerable to damage progression despite an optimized treatment, possibly related to other unidentified factors. In this regard, Cluster 1 displayed a relatively high proportion of anti-ENA positive patients, which is a feature associated with interferon (IFN) activity, however whether this may be related to a more likely damage progression is not known and would require longitudinal studies on a larger patient group. Conversely, patients in Cluster 2 displayed a substantially stable SDI score notwithstanding the severe active manifestations at the time of belimumab initiation, which adds on well to a higher response rate in this group, thereby highlighting the advantage of an intensified treatment of renal manifestations in SLE. As expected, cluster 3 persistently displayed the highest damage score throughout the follow-up; patients in this cluster were given belimumab with several SLICC items and comorbidities established and while assuming the highest prednisone dosage and the lowest HCQ, which have been associated with damage progression and disease refractoriness (2, 25, 29). However, a higher baseline SDI did not prevent achievement of a clinically meaningful response among active domains across all clusters, thus underling the importance of optimizing treatment even in advanced disease stages.

It should be noted that because this analysis was exploratory and conducted in a relatively small cohort, the differences observed in damage trajectories across clusters should be interpreted cautiously. Factors such as baseline damage burden, treatment exposure, and comorbidities may have contributed to these patterns, warranting future studies in larger cohorts using fully adjusted multivariable models to explore modification of damage trajectory upon biological treatment and according to baseline clustering.

Our study has both strengths and limitations. Among the latter, the analysis was carried out on a relatively small number of patients, thus possibly masking the effect of other variables on clustering; data were retrieved from a real-world registry requiring imputation of missing data; finally, the study population derives from a national registry including patients initiating belimumab for refractory SLE, and therefore represents a selected subgroup of the overall SLE population. Clustering was performed only in patients with sufficiently complete baseline data, which may have introduced selection bias. Moreover, no independent external cohort with comparable data was available to validate the identified clusters. External validation will therefore be required in future studies to confirm the robustness and generalizability of these findings. On the other hand, this is the first study investigating the clinical heterogeneity of a wide SLE population candidate to an add-on biological treatment through an unbiased approach.

From a translational perspective, the identification of data-driven clinical clusters among patients initiating belimumab may help frame future strategies for patient stratification. Although the present clusters should not be interpreted as definitive therapeutic subgroups, they highlight the heterogeneity of patients receiving belimumab in clinical practice and suggest that different clinical trajectories may coexist within this population.

In particular, the cluster enriched in renal and serologically active disease showed a numerically higher response rate, whereas patients characterized by neuropsychiatric and cardiovascular involvement displayed a greater burden of established damage. Such patterns may help generate hypotheses for future studies exploring whether certain clinical profiles may benefit from earlier or more intensive biologic intervention.

More broadly, these findings support the concept that clinical phenotyping alone may only partially capture the complexity of SLE heterogeneity. Integration of clinical clustering with molecular and immunological profiling may represent a promising strategy to identify biologically meaningful patient subsets, improve treatment stratification and ultimately inform therapeutic decision-making (26, 30).

In conclusion, the unsupervised clustering approach highlighted the clinical heterogeneity of patients initiating belimumab in real-world practice and identified data-driven groupings that broadly reflect known patterns of disease manifestations, highlighting the clinical heterogeneity of patients initiating belimumab due to refractory SLE. Severe manifestations preceding belimumab treatment, increased baseline SDI and suboptimal hydroxychloroquine intake define a cluster of patients exhibiting higher damage burden despite addition of biological treatment. Although clustering did not clearly predict clinical outcome in this exploratory analysis, a greater response rate to belimumab seems likely in patients with active serology and renal involvement.

Because diverse immunological pathways may result in similar clinical manifestations, clinical and laboratory phenotyping cannot fully account for the complex variability meandering through an SLE population, thus demanding for an integrated molecular-driven clustering to enhance a tailored treatment approach. Unsupervised learning may complement clinical judgement leading to a deeper understanding of SLE phenotypes and individualized drug allocation.

Members of the LIRE Registry Study Group

The following collaborators contributed to data collection within the LIRE Registry but are not included as authors of this manuscript:

Alessandra Bortoluzzi, Section of Rheumatology, Department of Medical Sciences, University of Ferrara and Azienda Ospedaliera-Universitaria di Ferrara, Cona (Fe), Italy; Ettore Silvagni, Section of Rheumatology, Department of Medical Sciences, University of Ferrara and Azienda Ospedaliera-Universitaria di Ferrara, Cona (Fe), Italy; Elena Bartoloni, Rheumatology Unit, Department of Medicine and Surgery, University of Perugia, Perugia, Italy; Enrica Bozzolo, Internal Medicine and Division of Immunology, Transplantation and Infectious Diseases, IRCCS San Raffaele Hospital and Università Vita-Salute San Raffaele, Milan, Italy; Valentina Canti, Internal Medicine and Division of Immunology, Transplantation and Infectious Diseases, IRCCS San Raffaele Hospital and Università Vita-Salute San Raffaele, Milan, Italy; Paola Conigliaro, Rheumatology, Allergology and Clinical Immunology, Department of “Medicina dei Sistemi”, University of Rome “Tor Vergata”, Rome, Italy; Maria Chiara Ditto, Rheumatology Unit, Azienda Ospedaliera Città della Salute e della Scienza di Torino, Turin, Italy; Giacomo Emmi, Department of Medical, Surgical and Health Sciences, University of Trieste, Trieste, Italy; Clinical Medicine and Rheumatology Unit, Cattinara University Hospital, Trieste, Italy; Sara Ferrigno, Rheumatology, Allergology and Clinical Immunology, Department of “Medicina dei Sistemi”, University of Rome “Tor Vergata”, Rome, Italy; Franco Franceschini, Rheumatology and Clinical Immunology Unit, ASST Spedali Civili, Department of Clinical and Experimental Sciences, University of Brescia, Brescia, Italy; Micol Frassi, Rheumatology and Clinical Immunology Unit, ASST Spedali Civili, Department of Clinical and Experimental Sciences, University of Brescia, Brescia, Italy; Carlo Garaffoni, Section of Rheumatology, Department of Medical Sciences, University of Ferrara and Azienda Ospedaliera-Universitaria di Ferrara, Cona (Fe), Italy; Marcello Govoni, Section of Rheumatology, Department of Medical Sciences, University of Ferrara and Azienda Ospedaliera-Universitaria di Ferrara, Cona (Fe), Italy; Luca Iaccarino, Rheumatology Unit, Department of Medicine-DIMED, University of Padova, Padova, Italy; Angelo Manfredi, Internal Medicine and Division of Immunology, Transplantation and Infectious Diseases, IRCCS San Raffaele Hospital and Università Vita-Salute San Raffaele, Milan, Italy; Silvia Mancuso, Dipartimento di Scienze Cliniche, Internistiche, Anestesiologiche e Cardiovascolari (SCIAC), Sapienza University, Rome, Italy; Caterina Naclerio, Rheumatology Unit, Scarlato Hospital, Scafati (Sa), Italy; Valeria Orefice, Rheumatology Unit, San Camillo-Forlanini Hospital, Rome, Italy; Simone Parisi, Rheumatology Unit, Azienda Ospedaliera Città della Salute e della Scienza di Torino, Turin, Italy; Giulia Pazzola, Azienda USL-IRCCS di Reggio Emilia, Reggio Emilia, Italy; Immacolata Prevete, Rheumatology Unit, San Camillo-Forlanini Hospital, Rome, Italy; Giuseppe Alvise Ramirez, Internal Medicine and Division of Immunology, Transplantation and Infectious Diseases, IRCCS San Raffaele Hospital and Università Vita-Salute San Raffaele, Milan, Italy; Chiara Scirocco, Rheumatology Unit, San Camillo-Forlanini Hospital, Rome, Italy; Francesca Romana Spinelli, Dipartimento di Scienze Cliniche, Internistiche, Anestesiologiche e Cardiovascolari (SCIAC), Sapienza University, Rome, Italy; Margherita Zen, Rheumatology Unit, Department of Medicine-DIMED, University of Padova, Padova, Italy.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/flupu.2026.1757086/full#supplementary-material

Proportion of variance explained by PCA highlighting 5 principal components. Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Martina Cozzi, University Center of Excellence on Nephrologic, Italy

Reviewed by: Ioannis Parodis, Karolinska Institutet (KI), Sweden

Ioannis Tassiulas, Icahn School of Medicine at Mount Sinai, United States