Archetypal analysis of pain measures in a diverse population-based cohort: Exploring cross-sectional relationships with outcomes Authors: Parsa Pazooki, Liubov Arbeeva, Simoni Tomar, Carolina Alvarez, Tessa Walker, Yvonne M Golightly, B Duncan X Lascelles, Amanda E Nelson Published: 2026 ABSTRACT AbstractObjectivePain experiences are complex, multifactorial, and subjective. Pain phenotyping can provide a holistic view of pain experiences by identifying distinct pain profiles. This study aimed to characterize chronic pain phenotypes using multiple validated osteoarthritis (OA) pain measures and assess their associations with functional and psychological outcomes.DesignParticipants were from the Johnston County Health Study (JoCoHS), a population-based cohort of non-Hispanic White, non-Hispanic Black, and Hispanic adults aged 35–70 years. Fourteen validated pain measures were administered. Archetypal analysis, a soft clustering technique, was applied to identify pain archetypes. Associations between pain phenotypes and varying functional and psychological characteristics were examined using Dirichlet regression for modeling archetype membership proportions. Results were validated in an independent cohort.ResultsAmong 834 participants (33% male, 27% with symptomatic knee OA, mean age 55 ± 9.6 years, mean BMI 33 ± 8 kg/m2), five pain archetypes emerged: (1) Low Pain, (2) High Somatization and Catastrophizing, (3) Widespread Pain, Aching, and Stiffness (PAS), (4) Knee Pain, and (5) High Pain and Distress. The Low Pain archetype was associated with better physical performance, higher functional status, and lower anxiety and depression, while the High Pain and Distress and Knee Pain archetypes were associated with worse outcomes. Pain-related anxiety and depression were associated with higher membership in the High Somatization archetype.ConclusionsPain phenotyping using brief self-reported measures identified distinct profiles with varying functional and psychological characteristics. These findings highlight the potential of phenotyping for improved understanding of the pain experience.Keywords: Pain, Clustering, Function, Cohort Chronic pain is a burdensome and often debilitating condition that plagues many in the United States and worldwide [1]. Osteoarthritis (OA) is a leading cause of pain and disability among older adults, characterized by chronic joint degeneration and a complex, multifaceted pain experience [2]. The subjective and multifactorial nature of the pain experience, coupled with its physical, emotional, and psychological components, poses significant challenges for assessment and optimal treatment results. Previous studies as well as clinical practice have revealed that the experience of OA-associated pain is variable and influenced by numerous factors including the involved joint site, multiple versus single joint OA involvement, and potentially mechanistic subtypes of OA (e.g., atrophic, erosive, chondrocalcinosis) [3]. Other studies have also revealed that neurophysiologic differences in modulation of pain significantly alter the pain experience and even the response to common pain treatments [4]. Further, there is a growing body of evidence showing that a bidirectional relationship exists between a person's emotional responses and the pain experience [5]. In clinical practice, numerous validated tools exist to measure pain and functionality, yet the heterogeneity of individual experiences remains poorly understood, in part because pain itself is unobservable (i.e., a latent construct) and is clinically reflected in more than one of these surrogate measures. One way to help consolidate the diverse pain experience among patients is through the use of phenotyping. Pain phenotyping, broadly, is an emerging approach that seeks to categorize individuals into profiles based on their distinct pain experiences and associated characteristics. Pain phenotyping has the potential to identify subgroups of patients who may benefit from specific treatments [6]. Pain phenotyping in the disease of OA has previously involved development of profiles based on the mechanism of OA [7], pain sensitization [8], and prognostic indicators [7], among other factors. Given that the pain experience includes psychosocial as well as biological components, it is important to consider multiple dimensions (multidimensional data) when defining pain phenotypes. Modern phenotyping approaches leverage advanced statistical techniques, such as clustering analysis, to identify patterns within multidimensional data [7,9]. These methods are capable of accounting for a wide array of variables, including self-reported pain intensity, psychological factors like anxiety or catastrophizing, and/or structural features or physical performance measures, which may be measured on continuous, ordinal, or binary scales. Unlike latent class-based methods, clustering approaches can accommodate continuous pain measures (i.e., without discretization), preserving meaningful graduations in pain experience. Although this is a notable improvement on previous phenotyping methods that primarily focused on limited features of the pain experience, there has been little acknowledgement that individuals with chronic pain experience multiple pain components simultaneously, but to varying degrees. Therefore, identifying specific pain phenotypes with clearly defined boundaries (e.g., partitioning the participants into discrete, separate, clusters) remains challenging. To address this gap, we applied archetypal analysis, a soft-clustering technique that identifies extreme, data-driven profiles defining the edges of a multidimensional symptom distribution while allowing individuals to have partial membership across archetypes. Unlike centroid-based hard clustering approaches such as k-means that assign each person into a single group and their soft extensions (e.g., fuzzy k-means), which emphasize grouping around central tendencies, archetypal analysis focuses on boundary-defining profiles that characterize the full range of observed pain experiences. This distinction was especially relevant in this community-based cohort, which includes a relatively younger population and a smaller proportion of individuals reporting severe pain, where identifying archetypes at the extremes of the distribution (at least three were hypothesized) may be informative. By describing these archetypes and subsequently examining their relationships with and various functional assessments, this work sought to enhance the understanding of pain variability and contribute valuable insights to the growing body of literature on pain phenotyping. Methods Participants Data were from participants in the Johnston County Health Study (JoCoHS), a community-based cohort study that enrolled (n = 902, 2019–2024) residents of Johnston County, NC. All participants provided written informed consent, and the study was approved by the UNC IRB (#18–0438). Detailed information on data collection, measurements procedures are described elsewhere [10]. Participants provided socio-demographic and clinical information and completed symptomatic and functional questionnaires including multiple validated instruments related to pain, performance-based OARSI core measures, and Patient-Reported Outcomes Measurement Information System (PROMIS) assessments, detailed below and in Supplemental Table A. For external validation, we used data available from the Johnston County Osteoarthritis Project (JoCoOA) 2013–2015 follow-up visit (n = 878, 32% males, 34% Black, mean age 71 ± 7 years) [11]. JoCoOA participants are completely independent of JoCoHS participants (see additional information in the Supplement). Pain assessments (used for archetyping, Table 1) Table 1List of validated pain instruments collected, derived measures used in analysis, range and interpretation of values.Table 1Validated pain instrumentsDerived measures used in analysisRangeInterpretationIntermittent and constant osteoarthritis pain (ICOAP)1.Constant pain subscale0–100Lower is better2.Intermittent pain subscale3.Total scorePain Catastrophizing Scale (PCS)4.PCS total score0–52Lower is betterPatient Health Questionnaire (PHQ) somatic symptom severity scale from initial 15 items, excluding 2 items (one not asked and one sensitive question with high missing rate)5.PHQ-13 total score (13 items)0–26Lower is better“On MOST days of ANY ONE MONTH in the LAST 12 MONTHS did you have pain, aching, or stiffness (PAS) in any of the following?” Rated as none (0), mild (1), moderate (2), or severe (3) for cervical, thoracic, and lumbar spine and left and right jaw, heel, shoulder, elbow, wrist, hand, hip, knee, ankle, foot6.Total of number of joints, other than the knee, with any PAS0–21Lower is better7.Sum of severity (0–3) for all joints, other than the knee0–638.Knee laterality with any PAS0–29.Sum of severity for knees0–6Pressure pain thresholdDolorimeter, standard mechanical pressure-based, at the upper trapezius muscle on left and right side. Pressure applied at 1 kg/s until pain reported, up to 4 kg of pressure maximum10.Mean of two trials (within 0.4 kg) on each side; lowest (worst) pressure pain threshold from either side used0–4 kgHigher is betterKnee Injury and Osteoarthritis Outcome Score (KOOS)11.Pain (9 items)0–100Higher is better12.Symptoms (7 items)13.Function of daily and higher level activities (9 items)14.Quality of life (QOL, 4 items) List of validated pain instruments collected, derived measures used in analysis, range and interpretation of values. 1.Constant pain subscale Constant pain subscale 2.Intermittent pain subscale Intermittent pain subscale 3.Total score Total score 4.PCS total score PCS total score 5.PHQ-13 total score (13 items) PHQ-13 total score (13 items) 6.Total of number of joints, other than the knee, with any PAS Total of number of joints, other than the knee, with any PAS 7.Sum of severity (0–3) for all joints, other than the knee Sum of severity (0–3) for all joints, other than the knee 8.Knee laterality with any PAS Knee laterality with any PAS 9.Sum of severity for knees Sum of severity for knees 10.Mean of two trials (within 0.4 kg) on each side; lowest (worst) pressure pain threshold from either side used Mean of two trials (within 0.4 kg) on each side; lowest (worst) pressure pain threshold from either side used 11.Pain (9 items) Pain (9 items) 12.Symptoms (7 items) Symptoms (7 items) 13.Function of daily and higher level activities (9 items) Function of daily and higher level activities (9 items) 14.Quality of life (QOL, 4 items) Quality of life (QOL, 4 items) Presence and severity of pain, aching, or stiffness in multiple joint sites were obtained for 23 sites (i.e., cervical, thoracic, or lumbar spine, and left or right jaw, heel, shoulder, elbow, wrist, hand, hip, knee, ankle, and foot). Participants were asked if on most days of any one month in the last 12 months they had symptoms of pain, aching, or stiffness in each joint site (yes/no), and if yes, to rate their symptoms as mild (1), moderate (2), or severe (3). Symptom presence was defined as number of joints with symptoms (0–21) for joints other than the knees which were scored 0–2 for no, unilateral, or bilateral involvement. Symptom severity was defined using the sum of the symptom scores at each joint site (0–3, none-severe, for all joints, except for knees, with a range 0–63 for sites other than the knees and 0–6 for the knees alone). Intermittent and Constant osteoarthritis (OA) Pain (ICOAP) [12]. The ICOAP was used to assess pain in the knee by considering both constant and intermittent pain experiences. The ICOAP consists of 11 items each of which is scored from 0 (no pain) to 4 (extreme pain) to assess different aspects of pain within the past week. A score is calculated for the constant pain subscale (0–20), intermittent pain subscale (0–24), and total pain (0–44, calculated by summing the subscale scores). Knee Injury and Osteoarthritis Outcome Score (KOOS) [13]. KOOS was developed to determine the impact of knee OA and/or injury on daily function and quality of life. KOOS scores were computed for five subscales of pain (9 items), function in daily and higher-level activities (short version, 9 items), symptoms (7 items), and quality of life (4 items). Scores are calculated separately for each subscale and range from 0 (the worst possible knee symptoms) to 100 (indicating no problems). Patient Health Questionnaire-15 (PHQ-15) [14] somatic symptom scale. PHQ-15 typically has a range from 0 to 30. Two sensitive questions (on menstrual cramps and sexual relations) were excluded due to the high number of missing data, so the score range in this study was from 0 to 26. Pain Catastrophizing Scale (PCS) [15]. This 13-item instrument asks participants to describe their thoughts and feelings when in pain, assessing domains of rumination, magnification, and helplessness; all items are rated on a scale of 0–4, with higher scores indicating more pain catastrophizing for a total score ranging from 0 to 52. Pressure Pain Sensitivity [16]. Pressure pain threshold was measured for both the left and right sides with a mechanical pressure-based dolorimeter placed at upper trapezius muscle and applied by an examiner at a rate of 1 kg per second until the participant reported pain or until a maximum pressure of 4 kg was attained without pain. At least two trials within 0.4 kg of each other were recorded, and the mean of the earliest consecutive pair was computed to define each side; the lowest (worst) value from either side for each participant was used in analyses. Self-reported and objective measures used for association analyses Two-Minute Step Test [17]. Participants were asked to stand next to a wall and to lift each leg, bending at the knee, to a height at least half the distance between the patella than the iliac crest, stepping in place as many times as possible within 2 min. 30-Second Chair Stand Test (30s-CST) [18]. From a seated position in a chair of standard height, participants rose to a full standing position and returned to a seated position as many times as possible within 30-s. Patient-Reported Outcomes Measurement Information System (PROMIS) [19]. Items from six PROMIS domains were completed by each participant, including assessments for anxiety, depression, fatigue, pain Interference, physical function, and sleep disturbance. A separate question asking about how much pain interfered with sleep in the past week and ranging from 0 (none) to 10 (completely) was also assessed. Timed Up and Go (TUG) [20]. Participants were asked to rise from a chair, walk a distance of 3 m, turn around, walk back to the chair, and sit down; time to complete this task was recorded in seconds. Eight foot walk [21]. For participants able to walk alone or without holding on to another person, the time (seconds) to complete an eight foot (2.44 m) walk was recorded for two trials and the average time used to compute gait in meters/second. Statistical analysis Density plots were used to check marginal distributions of pain variables and outcomes. The rank-based inverse normal transformation of all pain variables was performed to improve skewed distributions and to convert each variable's scale into a common measurement scale. First, Principal Components Analysis (PCA) was used to visualize and explore the multidimensional (14 variables, Table 1) data. We then used an unsupervised “fuzzy” clustering approach, archetypal analysis (AA, R package archetypes [22]), to identify different pain phenotypes represented in the JoCoHS cohort. AA is a method that deals with data multidimensionality and can identify latent multidimensional profiles called archetypes. Unlike other clustering techniques, AA identifies extreme or “pure” archetypes that are not necessarily represented in the sample, quantifies the degree to which an individual's features fit these archetypes, and determines the contribution of each feature to each archetype (similar to loadings from a principal components analysis). At the initial stage, we tested models with 2–10 archetypes and chose the optimal number, N, based on the relative decrease in the residual sum of squares (RSS). In addition, the stability of 2–10 archetypes was assessed as follows. For each K = from 2 to 10, we randomly selected 90% of original data, applied the archetype algorithm for the corresponding K, and compared the result to the original assignment (e.g., with full data) using the Rand index (Supplemental Fig. A). The procedure was repeated 100 times, for different membership cut-offs. The final model assigned N scores, one for each archetype to each participant. Each score ranged from 0 to 1, and all scores added up to one for each person, which represents the compositional nature of the data. We used Dirichlet Regression models to evaluate the associations of pain archetypes with demographic and clinical characteristics (i.e., sex, race, age, or BMI), self-reported PROMIS measures and objective physical performance. This model implemented in the DirichletReg R package [23] allows us to analyze a set of dependent variables, which are the degrees to which an individual participates in each archetype. These N dependent variables lie in a bounded interval from 0 to 1 and sum to one, representing the compositional nature of the data. Dirichlet regression produces an intercept and slope for all independent variables separately for each archetype. To enhance interpretability of the intercepts and slopes, we used these coefficients to estimate the membership degrees in each pain archetype across all values of PROMIS measures and physical performance tests available in our data. A statistically significant slope for an archetype indicates that the probability of belonging to that specific trajectory changes with the predictor; it does not indicate differences between archetypes but rather reflects within-archetype changes in proportional membership. Analyses were repeated using all available similar measures in the JoCoOA external validation cohort using the archetypes identified in the JoCoHS cohort. Results Cohort demographics The sample comprised 834 of 902 total JoCoHS participants (33% males, 33% Black, 14% Hispanic; mean age 55 ± 10 years) who had complete data for all pain measures (demographics were similar for those included and excluded, Supplementary Table A). Of these, 246 participants (27%) had symptomatic knee OA (defined as a Kellgren-Lawrence grade [24] of 2 or more and symptoms in the same joint), with substantial variability observed across knee-specific pain measures, multi-joint pain burden, psychological symptoms, and functional performance. Cluster discovery Based on the relative decrease in the residual sum of squares, we determined that the optimal number of archetypes was five (Supplemental Fig. A). The final selection of five archetypes was supported by the scree plot based on percentage of explained variances in PCA analysis (Supplemental Figure B) and principal component (PC) loadings plot (Supplemental Figure C). More than 80% of the information contained in the data were retained by the first 5 PCs.; the stability of 5 archetypes solution was acceptable (Supplemental Figure D). For descriptive purposes, we assigned each person to the “primary” archetype with the highest membership degree. We used the first two PCs to visualize the data and colored each individual point according to its primary archetype, where the size of each point reflects the degree of membership in the corresponding primary archetype (Fig. 1). Although this 2-D representation cannot fully describe the distribution of pain archetypes in the cohort, it provides a good illustration of the advantage of the fuzzy clustering when the data has a complex structure, and no clear grouping is obvious. Two-dimensional representation following Principal Components Analysis of the five groups identified using the maximum membership across 5 archetypes for each participant, e.g., each individual's primary archetype. Each point represents a JoCoHS participant and is colored according to the primary archetype; the size of each point reflects the degree of membership in the corresponding primary archetype. The five archetypes (i.e., pain profiles, Fig. 2, Supplemental Figure E) can be interpreted as follows: 1) Low Pain, 2) High Somatization and Catastrophizing, 3) Widespread Pain, Aching, and Stiffness (PAS, from Table 1) in multiple joints, 4) Knee Pain (as primary source of pain burden), and 5) High Pain and Distress. The Low Pain archetype was characterized by low scores for pain, somatization, and catastrophizing. The High Somatization and Catastrophizing archetype had low pain and high function, but high scores on PCS and PHQ13. The Widespread PAS archetype had high pain burden, pain severity and worse KOOS scores than the first two archetypes. The Knee Pain archetype demonstrated higher scores on ICOAP, greater decrements in KOOS subscales (worse scores), and high pain burden and severity specifically at the knee. Finally, the High Pain and Distress group had the worst scores across all measures, including the worst pain burden in severity at all joints, and by far the lowest pressure pain threshold (Fig. 2, Supplemental Figure E). Radar chart showing the contribution of each of 14 pain variables, ranging from 0% to 100% in the five archetypes (i.e., Low Pain [red], High somatization and catastrophizing [blue], Widespread pain, aching, and stiffness [green], Knee Pain [purple], High pain and distress [black]). Axes for all 14 pain measures are equally scaled from 0 to 100. ∗indicates that higher is better (no pain) for the corresponding outcome (Pressure Pain threshold, KOOS QoL, KOOS Function, KOOS Symptoms, and KOOS Pain). The yellow line shows the values for a hypothetical phenotype where all scores are ideal (no pain, no distress). The archetypes are demonstrated separately in Supplemental Figure E. Overall, 326 participants (39%) demonstrated dominant membership in a single archetype (membership ≥0.6, a threshold chosen pragmatically to reflect clear predominance of a single archetype over others), whereas the majority (508 participants; 61%) exhibited mixed membership across multiple archetypes. Specifically, 138 (17%) participants had extreme values (defined as an archetype membership ≥0.6) for the Low Pain archetype, 32 (4%) participants had extreme values for the High Pain archetype, and 36 (4%), 62 (7%), and 58 (7%) met this criterion for archetypes 2, 3, and 4, respectively. Clinical description of clusters We did not observe major differences by sex and race across the five archetypes. The average membership degree in the Low Pain archetype was slightly higher in males compared to females, and females were more likely to be represented in the High Pain group (Supplemental Table B). We found no strong correlations between membership degree in each archetype and age and BMI. However, higher BMI was associated with membership in the Knee pain group (rho = 0.23, p < 0.001) and inversely with the Low Pain membership (rho = −0.25, p < 0.001, Supplemental Table C). Associations with outcomes Results from the Dirichlet regression models demonstrated associations among archetypes and cross-sectional outcome measures (Fig. 3, Supplementary Table D). Effect sizes are presented as model-based predicted membership differences across observed ranges of outcomes. Overall, better performance on physical tasks, higher functional status, and lower levels of pain-related anxiety and depression were associated with higher membership degree in the Low Pain archetype (red line in figures). Specifically, higher PROMIS Physical Function T-scores were associated with a higher membership in Low Pain archetype (p < 0.001, Supplementary Table D), reflected by an increase in predictive probability from 0.07 at the lowest observed score to 0.56 at the highest score (Fig. 3A). In stark contrast, worse physical performance, lower functional status, and higher degrees of anxiety and depression were associated with higher membership in the High Pain and Distress (black line) and Knee Pain (purple line) archetypes. Membership in the Knee Pain group (purple line) is characterized by poor self-reported (Fig. 3A) and performance-based physical function (Fig. 3C–D). For example, the predictive probability of being in Knee Pain archetype was 0.37 for those with the lowest (worse) PROMIS Physical Function score, and 0.49 for participants with the highest Timed Up and Go in the cohort. These probabilities were lower for those with the highest Physical Function Score (probability = 0.08, p < 0.001) and the lowest Timed Up and Go in the cohort (probability = 0.15, p < 0.001). The widespread PAS group (green lines) did not have strong associations with reported or observed outcomes (flat in all figures). Pain-related anxiety and depression were associated with higher membership degree in the High somatization archetype (Supplemental Figure F). Estimated (from Dirichlet Regression) membership degrees in each pain archetype across different values of PROMIS measures and physical performance tests. External validation in data from the independent JoCoOA (N = 878, mean age 71.4 (SD = 7.7), mean BMI 31.0 (6.4), 34% Black, 32% males) was based on replication of archetypal structure, rather than exact quantitative equivalence, given the older age and the absence of ICOAP. We found similar archetypes and associations (Fig. 4 and Supplemental Figure G–H). Radar chart in the external validation cohort (JoCoOA) demonstrating the contribution of 11 available pain variables to the five archetypes found in the JoCoHS (for comparison to Fig. 3; Low Pain [red], High somatization and catastrophizing [blue], Widespread pain, aching, and stiffness [green], Knee Pain [purple], High pain and distress [black]). The yellow line shows the values for a hypothetical phenotype where all scores are ideal (no pain, no distress). The archetypes are demonstrated separately in Supplemental Figure G. Axes for all 11 pain measures are equally scaled from 0 to 100. ∗indicates that higher is better (no pain) for the corresponding outcome (Pressure Pain threshold, KOOS QoL, KOOS Function, KOOS Symptoms, and KOOS Pain). Discussion Our study identified five pain profiles using archetypal analysis, with each profile characterized by unique combinations of pain intensity, psychological distress, and functional limitations. In contrast to hard clustering techniques such as K-means, where individuals are assumed to be a member of one and only one class, archetypal analysis is a “fuzzy” clustering method, which allows each person to partially belong to multiple pain phenotypes, assuming individual heterogeneity in pain experience. The external validation of these profiles in an independent cohort reinforces the robustness of our findings. Our outcome analysis revealed that the ‘Low Pain’ profile was associated with better physical performance and lower levels of anxiety and depression, while the ‘High Pain and Distress’ group exhibited significant functional and psychological impairments. This aligns with the accepted notion that higher degrees of psychological and physical pain burdens can result in functional limitations while lower degrees of both result in better functionality. Murphy et al. also demonstrated this by identifying three clusters of pain phenotypes in older adults with hip or knee OA, associating the highest pain levels with depression, fatigue, poor sleep, and a high comorbidity burden [25]. This bears clinical significance when recognizing and addressing the ways in which chronic pain can inhibit patients in ways that extend beyond mobility. Interestingly, the ‘Widespread PAS’ and ‘High Somatization and Catastrophizing’ archetypes were not strongly associated with functional outcomes. This suggests that neither physical pain burden nor catastrophizing alone are primary drivers of functional status, a finding that adds nuance to previous research emphasizing the role of psychological factors in pain perception and disability [25,26]. While not strongly associated with outcomes, these archetypes did exhibit similar performances on outcome measures, as seen in Fig. 3. This apparent similarity may be due to some shared psychological characteristics, such as elevated distress and pain-related cognitive patterns, that contribute to both profiles. Future research should explore whether refinements in classification could yield clearer distinctions between these two pain profiles. The ‘Knee Pain’ archetype was associated—though not as strongly as the ‘High Pain and Distress’ group—with worse functional outcomes, including poorer performance on the 2-min step test and greater reported sleep interference. This suggests that localized pain burden, particularly in weight-bearing joints like the knee, may be a stronger predictor of functional impairment than widespread pain. It must be noted that the observed associations between pain phenotypes and functional outcomes may, in part, be influenced by the nature of the functional assessments used, which primarily involved lower-extremity tasks. It is possible that assessments focusing on upper-body function could yield different patterns of association, particularly for phenotypes characterized by primarily lower extremity pain. Future studies should explore whether pain phenotypes predict functional impairments in a broader range of physical activities. In addition, given the well-documented structure-symptom discordance in OA, where radiographic severity does not always align with symptom burden [27,28], future research could be directed at examining the degree to which patients in the ‘Knee Pain’ group exhibit radiographic evidence of knee damage. The relevance of pain phenotyping extends beyond OA and into broader pain management strategies. The emergence of ‘precision pain medicine’ emphasizes a tailored approach to pain treatment based on individual characteristics, including psychological and functional factors [1]. The Initiative on Methods, Measurement, and Pain Assessment in Clinical Trials (IMMPACT) has played a key role in synthesizing evidence-based recommendations for pain management, highlighting the importance of psychosocial factors, sleep, and patient-reported measures in treatment selection. For example, IMMPACT findings have shown that patients with significant psychosocial contributions to their pain tend to have worse sleep and may respond differently to analgesics, with less opioid benefit and a better response to non-opioid therapies. Similarly, in diabetic neuropathy, patients with lower neuropathic pain severity were found to respond better to duloxetine than pregabalin, demonstrating the importance of phenotype-specific treatment selection [1]. Our study contributes to this evolving field by providing a unique and nuanced methodology for pain phenotyping that has established associations with outcomes that are relevant to chronic pain in OA. The IMMPACT trial and the overall growing body of research on pain phenotyping underscores the importance of capturing the heterogeneity of the pain experience to predict patient outcomes and guide treatment. Prior studies have used various classification methods, including symptom-based stratification [29], chronicity-based grouping [26], and somatosensory profiling [30], to identify potentially clinically meaningful subgroups and assess response to treatments. For example, Lapin et al. stratified patients with low back pain into mild, mixed, and significant symptom groups based on PROMIS scores, finding that those with greater symptom burden required more specialist interventions but also demonstrated greater long-term improvement with appropriate treatment [29]. Our findings are broadly consistent with prior pain phenotyping studies, which have identified clinically meaningful subgroups characterized by differing burdens of pain, psychological distress, and functional limitation. However, prior studies have relied on traditional clustering approaches that assign individuals to discrete groups or have focused on more selected symptomatic populations. By contrast, our use of archetypal analysis allows graded membership across overlapping profiles, which may better reflect the multidimensional and heterogeneous nature of the OA pain experience. For example, Murphy et al. identified three clusters among 129 older adults with symptomatic knee or hip OA, including a high symptom burden group characterized by pain, fatigue, sleep, and mood disturbance; a lower symptom burden group; and an intermediate group with prominent sleep impairment. These patterns parallel our ‘High Pain and Distress’ and ‘Low Pain’ archetypes, supporting the reproducibility of clinically meaningful pain subgroups. Our study extends these findings by examining a larger, community-based cohort that included individuals with and without diagnosed OA and by employing archetypal analysis rather than hierarchical clustering, allowing participants to demonstrate partial membership across multiple pain profiles. Given the significant associations between pain phenotypes and functional outcomes in our study as well as the phenotype-dependent treatment responses demonstrated in the studies discussed above, future research should investigate whether treatment responses vary between the phenotypes derived from our study. We anticipate that individuals in the ‘High Pain and Distress’ group, who exhibit widespread impairments, may benefit from more robust and multidisciplinary interventions targeting both pain and psychological distress, while those in the ‘Knee Pain’ phenotype may obtain relief from more localized treatments. Our study has several strengths. The instruments we utilized for analysis were brief and largely patient-reported, making them easy to implement in clinical practice. In practice, these instruments could be administered during routine visits and entered into a clinical decision-support tool based on the archetypal model to generate each patient's degree of membership across the five pain profiles. This information could help identify patients with predominant pain-distress, localized joint pain, or low symptom burden patterns and guide referrals or treatment selection. Many of the phenotyping approaches discussed above carry a time and cost burden that preclude them from being readily used in most clinical settings. In addition, these are established measures, which means most pain or OA clinics are likely already familiar with them, further easing their implementation. The AA approach allows an individual to share features across more than one archetype, which is more reflective of complex and heterogeneous nature of pain. Limitations of this analysis include its cross-sectional design, which prevents the assessment of temporal relationships between the patient-reported measures and functional outcomes. As a result, we cannot determine causality or the directionality of the observed associations between pain profiles and outcomes. Longitudinal studies are needed to explore whether changes in pain profiles over time influence functional outcomes and treatment responses. In addition, while our study demonstrates that distinct pain phenotypes exist, the approach to phenotyping used here may not be entirely comprehensive. The selected assessments, though practical and widely used, may not capture all relevant dimensions of the pain experience, potentially introducing biases in phenotype classification. The lack of ICOAP and the older age of the validation cohort limited exact replication, highlighting the need for future validation studies using harmonized pain instruments. Future studies should explore additional dimensions, such as sensory profiles, inflammatory markers, or broader functional assessments beyond lower-extremity tasks, in addition to other cohorts and populations, to refine pain phenotyping approaches and enhance their clinical utility. In conclusion, the experience of pain is variable, influenced by joint involvement, mechanistic subtypes, neurophysiologic differences and psychological overlays, rendering effective pain management a persistent challenge. Our study identified five distinct pain profiles using AA, highlighting this heterogeneity in pain experience. We found that those with lower overall levels of pain exhibited better function and lower psychological distress, whereas those with higher overall pain had significant impairments. Interestingly, pain-related catastrophizing alone did not strongly predict functional limitations, while localized knee pain was associated with worse performance in the functional tests selected in our study. In practice, these findings provide a potential framework for stratifying patients using readily available, multidimensional pain assessments to better inform prognosis and guide targeted treatment strategies. Incorporating such phenotype-based approaches into clinical workflows and research design may improve patient selection for interventions and advance the development of more precise, individualized pain management strategies in OA. Future research should explore whether treatment responses differ across these phenotypes, potentially paving the way for precision pain medicine approaches. Author contributions 1)Conception and design: PP, LA, YMG, BDXL, AEN; Analysis: LA, CA; Acquisition of data: TW, YMG, AEN; Interpretation of data: all authors Conception and design: PP, LA, YMG, BDXL, AEN; Analysis: LA, CA; Acquisition of data: TW, YMG, AEN; Interpretation of data: all authors 2)Drafting the article: PP, LA, ST, AEN; Critical revisions: all authors Drafting the article: PP, LA, ST, AEN; Critical revisions: all authors 3)Final approval of the article: all authors Final approval of the article: all authors Funding Funding provided in part by the Centers for Disease Control and Prevention U01DP006266, NIH/NIAMS K24AR081368, P30AR072580, and R01AR080733-03S1. The funders had no role in the work or decision to publish. Conflict of interest Dr. Nelson has grants not related to this work from NIH/NIAMS, has received honoraria from MedScape Education and Novo Nordisk, and honoraria and travel support from the American College of Rheumatology and Osteoarthritis Research Society International. Drs. Nelson and Golightly are Associate Editors of Osteoarthritis and Cartilage. Acknowledgments We would like to thank the current and former staff, collaborators, and all participants in the Johnston County studies, without whom this work would not be possible. Handling Editor: Professor H Madry Footnotes Supplementary data to this article can be found online at https://doi.org/10.1016/j.ocarto.2026.100823. Contributor Information Parsa Pazooki, Email: parsa_pazooki@med.unc.edu. Liubov Arbeeva, Email: liubov_arbeeva@med.unc.edu. Simoni Tomar, Email: tomar.simoni@gmail.com. Carolina Alvarez, Email: alvarec@live.unc.edu. Tessa Walker, Email: twalkerinspired@gmail.com. Yvonne M. Golightly, Email: ygolightly@unmc.edu. B. Duncan X. Lascelles, Email: dxlascel@ncsu.edu. Amanda E. Nelson, Email: aenelson@med.unc.edu. Supplementary data The following are the Supplementary data to this article: Multimedia component 1 Multimedia component 2 References Associated Data Supplementary Materials Multimedia component 1 Multimedia component 2