Introduction
Psoriatic arthritis (PsA) is a chronic autoimmune disease characterised by peripheral arthritis, enthesitis, dactylitis, axial involvement, psoriasis and nail disease.1 The disease course is unpredictable and marked by fluctuating levels of disease activity and substantial individual variation in symptom burden and quality of life (QoL).23 Discerning how PsA symptoms evolve and how they differ between patients is relevant to identify opportunities to improve disease management and tailor treatment approaches.
Current treatment approaches aim to achieve predefined states of disease control, with consideration of the various disease domains.45 Composite indices such as the Disease Activity in Psoriatic Arthritis (DAPSA), Psoriatic Arthritis Disease Activity Score (PASDAS), Minimal Disease Activity (MDA) have improved standardised assessment of disease activity in clinical trials, yet are not routinely and uniformly used in everyday PsA care.67 Moreover, concerning the defined PsA disease activity states, despite low to moderate disease activity, individuals with PsA often experience persistent symptoms such as pain, fatigue, reduced mobility and sleep difficulties. This indicates a discordance between the physician-observed and lived experience of disease activity.8–12
Recent expert consensus advocates that the heterogeneity of PsA requires the patient perspective to be at the centre of the disease assessment, goal setting and decision-making.7 The 12-item Psoriatic Arthritis Impact of Disease questionnaire (PsAID-12) is a patient-reported outcome measure (PROM), specifically designed to capture the impact of PsA on daily life.713–15 Beyond its original purpose as a health-related quality of life (HRQoL) estimator, studies demonstrate that the PsAID-12 provides valuable predictive information regarding disease activity, flare risk and treatment response.16–19
Clustering, an unsupervised machine learning technique, is commonly used for exploratory data analysis.20 It opts to identify meaningful subgroups by uncovering hidden patterns without relying on predefined labels.2122 Previous studies have demonstrated the utility of clustering in identifying clinically meaningful patient subgroups based on baseline composite clinical data, revealing distinct phenotypes with differing disease burdens and treatment responses.23–27 This study introduces a novel perspective by focusing on symptom and impact severity profiles and by employing a data-driven, time-independent approach to capture the dynamic and heterogeneous nature of PsA. Using PsAID-12 data from the Dutch southwest Early Psoriatic Arthritis (DEPAR) cohort, the aim was to identify distinct states that reflect lived experience of PsA, examine their relationship to established disease activity measures and explore their stability and transitions over time.
Methods
Patient representatives were actively involved in the DEPAR study design. One patient representative contributed as coauthor and provided input on study interpretation and dissemination of the findings.
Dataset
This study used data from DEPAR, an ongoing, multicentre, prospective cohort study initiated in July 2013. Patient representatives were actively involved in shaping the study design. Details on the DEPAR cohort have been explained in detail by Kasiem et al.28 The cohort comprises patients who were newly diagnosed with PsA and were disease-modifying antirheumatic drug-naïve for musculoskeletal symptoms at baseline.
Inclusion was based on clinical diagnosis by the treating rheumatologist without requiring fulfilment of the classification criteria for psoriatic arthritis (CASPAR criteria).29 Participants received standard of care treatment.
Clinical assessments were collected by trained research nurses at baseline, every 3 months during the first year, every 6 months in the second year and annually thereafter. Prior to each visit, patients completed multiple PROMs. From DEPAR, data were selected up to March 2023, based on the availability of the PsAID-12,30 which was introduced to the cohort in 2015.
Instead of aggregating outcomes by patient (n=880), each clinical assessment was treated as a distinct observational unit, yielding 5099 data points. Table 1 summarises baseline participant characteristics; table 2 reports disease activity characteristics pooled across all observational units. A breakdown of disease activity by time point is provided in online supplemental material S1.
Variables and preprocessing
Clustering variables
The PsAID-12 is a PROM developed with patients to assess the symptoms and impact of PsA. It is the first patient-reported outcome (PRO) to capture HRQoL in PsA that is provisionally endorsed by both the Outcome Measurements in Rheumatology (OMERACT) initiative and the Group for Research and Assessment of Psoriasis and Psoriatic Arthritis (GRAPPA) for use in clinical trials.13 Comprising 12 items rated on a 0–10 scale, it reflects domains most relevant to patients’ lived experience.30 Symptom severity (physical and psychosocial) is captured by the components: pain, fatigue, skin problems, discomfort, sleep disturbance, anxiety/fear/uncertainty and depression. On the other hand, impact severity is captured by the components: work/leisure activities, functional capacity, coping, shame/embarrassment and social participation.13 The PsAID-12 has demonstrated robust psychometric properties in a large sample of patients with active PsA, supporting its use as a fit-for-purpose PRO.
It was validated in moderate-scale observational studies.14 Items cover key aspects of PsA, with pain, fatigue and skin problems identified as particularly burdensome. Within this study, all 12 domains were used as input variables for clustering analyses.
Data inspection and preprocessing
Prior to analysis, data inspection was performed to ensure data quality and to assess the suitability of PsAID-12 variables for clustering. The PsAID-12 data were evaluated for structure, distributions, potential outliers and relationships among key clinical variables. Spearman’s correlation coefficients were computed to examine associations between the clustering variables, as highly correlated variables may disproportionately influence clustering results.
Descriptive statistics and visualisations (eg, histograms, box plots) were used to quantify missingness and identify outliers and potential data entry errors. Skewness and kurtosis were assessed to characterise distributional properties and guide subsequent analytical decisions. The dataset predominantly reflected low disease activity, with several variables exhibiting floor effects, consistent with a treated observational cohort with prolonged follow-up (median 36 months).
Since all PsAID-12 variables stem from the same questionnaire with identical numeric answering scales, there were no missing items, nor was there a need for standardisation. Additional data inspection details are provided in online supplemental material S1.
Clustering
To derive data-driven states of patient-reported disease impact, various clustering solutions were computed using Python and/or STATA. Algorithms included partition-based methods (k-means++, spectral clustering), agglomerative hierarchical clustering and model-based methods (generalised structural equation models (GSEM) based on Gaussian, Poisson and Weibull distributions) for a range of 2–10 clusters (k). The included algorithmic families are commonly applied unsupervised learning techniques for disease phenotyping.
Their differences in mathematical principles allow for comprehensive evaluation of different structural assumptions within the data. Each algorithm was tuned for relevant hyperparameters and used the individual PsAID-12 domains as input variables. Candidate clustering solutions were evaluated for quality, stability and clinical interpretability.
More details on the clustering methods are reported in online supplemental material S2.
Clustering quality was defined in terms of internal cohesion and separation between clusters. It was primarily assessed following the bootstrapped smooth quadratic score (BQS) with 1000 bootstrap samples, as proposed by Coraggio and Coretto.31 Other method-independent cluster quality metrics, including silhouette scores, the Davies-Bouldin index and the Calinski-Harabasz index, were computed as complementary measures.20 Silhouette and two-dimensional T-distributed Stochastic Neighbour Embedding (t-SNE) plots were generated to visualise the identified clusters. Additionally, relevant method-specific quality metrics were computed for each clustering algorithm.
Clustering stability referred to the consistency of cluster assignments under data perturbations. To evaluate stability across resampling, Adjusted Rand Index scores were computed for the bootstrapped clustering analysis (ARIbootstrap). For each bootstrap sample, the original dataset was projected onto the clusters obtained from the bootstrap sample, using the highest membership probabilities.
The ARIbootstrap was then computed between the original cluster assignments and these projected assignments. The average was taken over all bootstrap samples to compare the cluster solutions. To further assess robustness, a leave-one-variable-out sensitivity analysis was performed.
For each variable, clustering was re-ran after removing the respective variable and the resulting clusters were compared with the original clusters using the ARI (ARIexcl.var).
Clustering solutions were first filtered based on cluster quality and stability thresholds: silhouette score >0.2, ARI scores >0.5 and cluster size <50. The best clustering solution for each candidate number of clusters (k) was then selected based on the highest BQS score.
Clinical interpretability was also considered to ensure both statistical robustness and practical relevance. The number of clusters (impact of disease states) was thus determined based on their clinical value as discussed within our multidisciplinary research team of researchers, rheumatologists and patient representatives.
Cluster interpretation
Symptom/Impact severity
Descriptive statistics were used to summarise the data. Kruskal-Wallis tests accompanied by a Dunn’s post hoc analysis using the Bonferroni correction were performed to examine differences between clusters. Significance was defined as p<0.05 for all tests, unless stated otherwise in the text.
Each cluster was labelled empirically, based on the clinical constellation of symptom/impact severity present.
Clinical comparators
Identified clusters were additionally evaluated using clinical assessments of peripheral arthritis (swollen joint count of 66 joints/tender joint count of 68 joints), enthesitis (Leeds Enthesitis Index and Maastricht Ankylosing Spondylitis Enthesitis Score, 0–17), psoriasis (body surface area (BSA), 0%–100%) and global and individual clinical assessments of arthritis and psoriasis (visual analogue scale (VAS), 0–100 mm). Inflammatory blood markers, including C reactive protein (CRP; mg/L) were excluded from the analysis because median CRP concentrations were below the 3–5 mg/L lower detection limit of the traditional CRP immunoassays available at the participating research centres.32 Dactylitis (count, 0–10) was excluded due to minimal variation in the observations, with 88% reporting a count of 0 and only ~1% reporting a count ≥3.
To assess alignment with established disease activity states, observational units were categorised according to DAPSA, PASDAS, MDA and PsAID-12 thresholds33 (online supplemental material S1), and cross-tabulations were used to examine the correspondence between the data-driven clusters and established disease activity states.
State transitions
A Sankey diagram was computed to review the trajectories of patients between symptom impact/severity states over time. The probabilities of transferring from one state to another at the next clinical visit were calculated and illustrated with a Chord diagram and state transition matrix.
Results
Identification of the PsA impact of disease states
Four algorithmic families (partition-based, spectral, hierarchical and model-based clustering) were implemented for the identification of data-driven impact of disease states. After tuning the relevant hyperparameters of each model, a set of 30 candidate clustering solutions were computed. The candidate solutions were first screened using predefined quality criteria.
For each potential number of clusters, the statistically strongest solution was retained. Model explanation and performances are reported in online supplemental material S2.
Following review and discussion within the research team, the Gaussian GSEM mixture model with six clusters was selected as the final solution (t-SNE projection and silhouette plot: online supplemental material S3). Further analysis in this paper is based on this final solution. This solution was preferred because it captured distinct profiles rather than broad severity categories, without introducing overly fragmented patterns.
The model demonstrated stable results when tested repeatedly using resampling techniques, with modest but consistent cluster separation (BQS1000=−20.11, average silhouette score=0.27, mean ARIbootstrap=0.76, ARIexcl.pain=0.91). Kruskal-Wallis tests indicated significant differences among the six clusters for all PsAID-12 domains (p<0.001). Dunn’s post hoc analyses further elicited significant inter-cluster differences of the PsAID-12 components, as reported in online supplemental material S3.
Characterisation of the PsA impact of disease states
The six GSEM clusters identified among the unique observational units (n=5099) are presented in figure 1. Cluster sizes varied considerably (table 3), ranging from 2143 observations (42%; 531 unique patients) to 357 observations (7%; 154 unique patients). As patients could occupy different states across visits, unique patient counts are not mutually exclusive.
Clusters were labelled empirically based on intercluster differences and interpretation of individual item scores guided by Gossec et al.14
Demographic and clinical characterisation of the PsA impact of disease states
When considering only the unique patients within each cluster, there were no meaningful differences in median age, BMI or symptom duration. Sex distribution differed significantly, with an over-representation of men in cluster 1, and a greater proportion of women in clusters 3 and 5 (pbf<0.05). Observed clinical and composite disease activity scores in the six GSEM clusters are presented in table 3.
Clusters differed significantly for clinical variables (Kruskal-Wallis, p<0.001). Pairwise differences are detailed in online supplemental material S3.
In cluster 1, representing very mild impact, the median composite indices correspond to DAPSA remission and PASDAS low disease activity. Minimal joint and skin involvement, absent enthesitis and a low Physician Global score are observed. From cluster 1 to cluster 2, representing very mild to mild impact, clinical measures rose with small but significant increases (pbf<0.001) for all variables except BSA.
In cluster 2, the composite indices indicated a state of low disease activity.
For clusters 3 and 4, representing substantial impact, median composite scores indicated DAPSA low disease activity and PASDAS moderate disease activity. Cluster 3 had more tender joints (pbf<0.01) and enthesitis (pbf<0.05), but a smaller BSA than cluster 4 (pbf<0.001). Differences in other variables, including composite indices, were non-significant.
Both clusters demonstrated higher clinical involvement than the mild impact state. Cluster 3 displayed a slight increase in all variables except skin measures. Cluster 4 was less distinct from cluster 2, as there were no significant differences in enthesitis, physician VAS Global or physician VAS Arthritis.
Cluster 5 and 6, representing severe impact states, exhibited the highest clinical burden related to tender joints and enthesitis. Composite scores indicated moderate disease activity in both clusters. Differences between cluster 5, characterised by lower skin and psychosocial burden, and cluster 6, characterised by pervasive impact, were clinically reflected by a significant difference in psoriasis severity (pbf<0.001), whereas musculoskeletal and composite indices were non-significant.
When comparing impact states with lower skin and psychosocial burden profiles, cluster 3, representing substantial impact, differed from cluster 5, representing severe impact, only in DAPSA (pbf<0.05). Comparisons between the higher-burden profiles revealed broader discrepancies. Although clusters 4 and 6, representing substantial and severe impact, respectively, had comparable psoriasis severity, they differed significantly in tender joints, enthesitis, Physician Global and composite scores.
Notably, cluster 3, representing substantial impact with lower skin and psychosocial burden, and cluster 6, representing pervasively severe impact, differed only in skin variables and DAPSA (pbf<0.001).
Know-groups comparison
Cross-classification, presented in table 4, demonstrates the correspondence between patient-reported impact of diseases and conventional disease activity states.
PASDAS, DAPSA and MDA
Remission and VLDA states were almost exclusively represented in cluster 1. Low disease activity states and MDA were distributed predominantly across clusters 1 and 2. In contrast, the no MDA, moderate and high disease activity categories exhibited substantial dispersion across all impact of disease clusters.
Observations classified as high disease activity were more frequently observed in clusters 5 and 6.
PsAID-12 thresholds by Gossec et al
In previous research by Gossec et al, PsAID-12 levels of symptom/impact severity were derived using receiver operating characteristic-based thresholds anchored to the DAPSA and PASDAS.14 Cross-classification with the data-driven PsAID-12 clusters displayed a clear ordinal pattern: remission and low disease activity states are confined to clusters 1 and 2. In contrast to broad dispersion observed for the composite moderate disease activity states, PsAID-12 moderate disease activity was concentrated primarily in cluster 2 with limited spill-over into cluster 3. High disease activity showed wide dispersion across clusters 3–6.
Cross-classification between disease activity states is available in online supplemental material S4 for reference.
Transitions between PsA impact of disease states
Longitudinal analysis of state transitions during the first 5 years of follow-up (figure 2) demonstrated that symptom/impact severity behaves dynamically rather than fixed, with patients with PsA transitioning between impact states over time.
Time-agnostic analysis based on the state transition matrix and chord diagram (figure 3) revealed that cluster stability varied by severity. The tendency to remain in the same cluster between consecutive visits was strongest in the extreme clusters: cluster 1 (very mild; p=0.82) and cluster 6 (severe; p=0.56). When transitions occur, they typically involve movement to adjacent clusters.
State transitions between clusters 3 and 4 were relatively infrequent (p≤0.10), while direct transitions from cluster 6 to clusters 1, 2 or 3 were rare (p≤0.01).
These results suggest that while PsA symptom/impact severity fluctuates over time, patients tend to progress or regress gradually along the spectrum of symptom/impact severity, with extreme impact states being more stable.
Discussion
Patients with PsA often experience their disease differently from what clinical assessments capture.8–12 In this study, we analysed PsAID-12 data from the DEPAR cohort, following a data-driven clustering approach. Earlier clustering studies in PsA have generally used baseline physician assessments and laboratory data to group patients and relate them to treatment response.23262734 Our study differs in two ways: it focuses on PROs and includes data from all follow-up visits rather than using baseline snapshots. The pooling of longitudinal data allowed us to capture states that are not bound to visit timing, an approach that parallels the work of Steinz et al.35
Six distinct ‘impact of disease’ states ranging from ‘very mild’ to ‘pervasively severe’ were observed. Patients in these extreme states tended to remain more often in the same state at the next visit. Those in intermediate states showed more movement, typically transitioning to neighbouring clusters.
Among states with similar overall impact, skin problems and psychosocial burden were the main distinguishing features. Higher skin burden coincided with greater anxiety, shame, depression, difficulties with social participation and poorer coping. This is consistent with previous studies showing that even mild skin problems can strongly affect QoL.36–39 Importantly, all states remained present over time, underscoring the need to routinely assess dermatological and psychosocial impact, alongside musculoskeletal inflammation throughout the entire disease course.
The impact of disease states complement traditional disease activity measures. This is suggested across both low and high levels of clinically measured disease activity: patients in the very mild cluster report residual pain, fatigue and skin problems, whereas patients with high disease activity do not always exhibit a severe impact state. These findings suggest that the relationship between clinically measured disease activity and experienced burden is modulated by additional factors.4041 Consequently, intensifying pharmacological treatment alone may be insufficient for patients who consistently experience severe disease impact despite moderate levels of clinically measured disease activity.
Non-pharmacological approaches, such as diet, cognitive-behavioural therapy, cessation programmes, sleep therapy, occupational therapy or self-management education could help address specific modulators of experienced burden.4243 Integrating PsAID-12 or other patient-reported impact assessments into routine care can be one way to bring these patients’ needs to light.
Findings need to be considered in the context of clinical decision-making. The ASSIST study reports physician assessment as the core driver of treatment escalation.44 PsAID-12 total scoring was not reported by physicians as influential, although it was associated with treatment escalation. However, most clinicians did report potential value of PsAID-12 scoring in boosting confidence for tapering.
Our very mild impact state may help serve this purpose. We argue that the clinical utility of the impact of disease states could extend beyond tapering, as they reveal areas of disease impact that may be insufficiently captured by traditional composite measures. In a heterogeneous condition such as PsA, individual PsAID-12 domains provide detailed information on specific aspects of disease impact, whereas the total PsAID-12 score and its associated thresholds offer a rapid overall assessment.45 The impact of disease clusters may offer an intermediate level of interpretation that unites multidimensional insight with clinical efficiency.
Strengths and limitations
DEPAR created a unique opportunity to study long-term PsAID-12 scores in everyday practice rather than in a selected trial population. All PsAID-12 domains were retained as clustering variables despite moderate-to-strong intercorrelations, because each is a unique patient-prioritised facet of disease impact. The shared variance may have weighted the solution towards overall severity rather than individual symptom patterns.
Even so, skin and psychosocial burden emerged as consistent differentiators, suggesting that meaningful information was retained beyond severity alone. Moreover, the predominance of low disease activity in DEPAR may have reduced the contrast between latent states and could partially explain the modest silhouette scores. As different algorithms and validity metrics can favour different structures within the same dataset, multiple algorithms and criteria were applied.
The final solution was selected carefully by triangulating quality, stability and clinical interpretability, the latter appraised by a multidisciplinary team of researchers, rheumatologists and patient representatives.
The pooling of repeated observations allowed us to identify patient-reported impact states across the disease course rather than fixed patient phenotypes. As a consequence, the clustering solution ignores within-patient correlation and may overestimate the clustering results. The late introduction of the PsAID-12 and unequal follow-up intervals may have distorted state prevalence and transition probabilities.
The Sankey plot was added to provide more information about the transitions over time. The reporting of the results was intentionally kept descriptive. Confirmation in an independent cohort is required before inferential modelling of determinants or trajectories.
Conclusion
Our impact of disease states demonstrates that patients with similar musculoskeletal disease activity can be stratified into distinct subgroups with markedly different skin and psychosocial burden. This stratification emphasises the importance of assessing dermatological, emotional and social determinants of well-being. Routine use of the PsAID-12 domains or similar measures may reveal impact of disease not adequately captured by composite scores and encourage non-pharmacological approaches to improve patient outcomes.
The impact of disease states could offer a practical framework for quick interpretation in clinical practice.
Supplementary material