Site-level cutaneous phenotyping framework for anti-MDA5-positive dermatomyositis: exploratory associations with systemic immune reference markers Authors: Ziyuan Tian, Lingbo Bi, Jiahui Han, Ke Xue, Yong Cui Published: 2026 ABSTRACT ObjectivesTo propose and evaluate a CDASI-informed, medical-record-based site-by-lesion cutaneous phenotyping framework for anti-MDA5-positive dermatomyositis, using established systemic and immunological markers as reference anchors to examine whether this framework provides additional phenotypic resolution beyond conventional binary skin assessment.MethodsThis single-centre retrospective cross-sectional study enrolled 339 anti-MDA5-positive DM patients. Cutaneous involvement was coded as binary ulceration, itch, and scale features across seven prespecified anatomical regions. The framework was CDASI-informed but did not use formal CDASI activity or damage scores. FLATCAN components, PAH, IgG, and IgM were used as systemic and immunological reference markers. Analyses included FDR-corrected univariate screening, hypothesis-driven multivariable association models, restricted cubic spline analysis, exploratory internal model-performance summaries, cross-correlation SVD (CC-SVD), and exploratory clustering.ResultsThe framework revealed complementary skin-systemic association patterns. First, the spatial extent of ulceration refined a conventional binary ulcer signal: any cutaneous ulceration was associated with CD8 + T-cell depletion (OR = 3.35, P < 0.001; FDR q = 0.026), and ulcer site count remained independently associated with CD8 + depletion (OR = 1.43 per site, P = 0.008). Second, anatomical location contributed distinct information: facial ulceration was associated with PAH (OR = 2.28, P = 0.003) more strongly than overall ulcer site count. Third, lesion-feature type separated different immunological patterns, with ulceration preferentially associated with CD8+/IgG-related signals and itch with IgM-related signals. CC-SVD organised these observations into exploratory dimensions, including a bootstrap-stable facial involvement-PAH/infection dimension and a statistically less stable acral ulcer-CD8+/IgG dimension.ConclusionSite-by-lesion cutaneous phenotyping may provide a useful framework for studying phenotypic heterogeneity in anti-MDA5-positive dermatomyositis. The observed associations are exploratory and require prospective validation using standardised skin assessment and clinical outcome follow-up.Graphical abstract Supplementary InformationThe online version contains supplementary material available at 10.1186/s13075-026-03834-6. Introduction Anti-melanoma differentiation-associated gene 5 (MDA5)-positive dermatomyositis (DM) is characterised by severe systemic manifestations, including vasculopathic skin lesions, rapidly progressive interstitial lung disease, and high short-term mortality [1–3]. Established prognostic biomarker frameworks such as FLATCAN, FLAIR, and CRAFT have identified laboratory features, including CD8 + T-cell depletion, hyperferritinaemia, hypoalbuminaemia, LDH elevation, CRP elevation, and age, as markers associated with adverse outcomes [4–6]. In the present study, variables highlighted by these models were used as established systemic reference markers for evaluating a cutaneous phenotyping framework. Skin involvement is clinically visible and biologically relevant in DM, but conventional descriptors often reduce cutaneous disease to the presence or absence of findings such as ulcers, mechanic’s hands, or Gottron papules/signs [7–9]. Such binary descriptors may lose information about anatomical location, lesion-feature type, and spatial extent. A structured observation framework may therefore help organise cutaneous heterogeneity and identify related statistical associations, without establishing a prediction model. The Cutaneous Dermatomyositis Disease Area and Severity Index (CDASI) provides a conceptual basis for structured regional assessment of dermatomyositis skin disease [10–12]. However, retrospective admission records often do not contain complete formal CDASI activity or damage scores. Therefore, rather than reconstructing or transposing CDASI scores, we used a CDASI-informed, medical-record-based site-by-lesion framework to organise available skin documentation into analysable regional descriptors. Accordingly, this study deconstructed cutaneous involvement into prespecified anatomical sites and lesion-feature types and evaluated whether this framework revealed interpretable skin-systemic association patterns when benchmarked against established systemic and immunological reference markers [13]. The aim was not to build a clinical risk model, but to assess whether structured cutaneous phenotyping could serve as a research framework for anti-MDA5-positive dermatomyositis. Methods Study design and patients This retrospective cross-sectional study was conducted at China-Japan Friendship Hospital, Beijing, in accordance with the Declaration of Helsinki. The study was approved by the Institutional Ethics Committee (approval numbers: 2023-KY-056); the requirement for informed consent was waived due to the retrospective nature of the study. We screened 513 patients first hospitalised with dermatomyositis between January 2015 and January 2024, reclassified cases per the 2017 EULAR/ACR criteria [14] or 2018 ENMC classification criteria [15], confirmed anti-MDA5 positivity by standardised line immunoassay, and excluded overlap syndromes, incomplete cutaneous documentation, and multi-antibody co-positivity, yielding 339 patients (Fig. 1). Fig. 1Study cohort and analytical framework. A Patient enrolment flow diagram: of 513 patients screened, 339 anti-MDA5-positive dermatomyositis patients with complete cutaneous documentation were included. B Site-by-lesion phenotyping framework: seven prespecified anatomical regions (face, neck, upper limbs, hands/fingers, trunk, back, lower limbs) were each independently coded for ulcer, itch, and scale (binary), generating a 7 × 3 site-by-lesion matrix per patient; mucosal ulcer, alopecia, and photosensitivity were recorded separately. The framework was CDASI-informed but did not use formal CDASI scoring. Each dimension was benchmarked against established systemic and immunological reference markers (FLATCAN components, PAH, IgG, IgM)/ Alt text: Two-panel figure: (A) enrolment flow diagram from 513 screened to 339 included; (B) anterior and posterior body diagrams with seven colour-coded anatomical regions and the site-by-lesion framework summary Study cohort and analytical framework. A Patient enrolment flow diagram: of 513 patients screened, 339 anti-MDA5-positive dermatomyositis patients with complete cutaneous documentation were included. B Site-by-lesion phenotyping framework: seven prespecified anatomical regions (face, neck, upper limbs, hands/fingers, trunk, back, lower limbs) were each independently coded for ulcer, itch, and scale (binary), generating a 7 × 3 site-by-lesion matrix per patient; mucosal ulcer, alopecia, and photosensitivity were recorded separately. The framework was CDASI-informed but did not use formal CDASI scoring. Each dimension was benchmarked against established systemic and immunological reference markers (FLATCAN components, PAH, IgG, IgM)/ Alt text: Two-panel figure: (A) enrolment flow diagram from 513 screened to 339 included; (B) anterior and posterior body diagrams with seven colour-coded anatomical regions and the site-by-lesion framework summary Data collection Cutaneous features were retrospectively extracted from first-hospitalisation medical records using a CDASI-informed site-by-lesion phenotyping framework. The CDASI served as a conceptual reference for regional skin assessment, but formal CDASI activity or damage scores were not available, were not reconstructed, and were not transposed into the seven anatomical regions used in this study. Cutaneous variables were extracted independently by two researchers from first-hospitalisation admission notes, which in our institutional practice contain the most detailed dermatological description and are temporally synchronous with the first complete laboratory panel. Discrepancies were resolved by joint review of the original record; cases with persistently ambiguous or incomplete cutaneous documentation were excluded at the screening stage. Seven prespecified anatomical regions were used: face, upper limbs excluding hands/fingers, hands/fingers, neck, back, trunk, and lower limbs. For each region, ulceration, itch, and scale were coded as binary features. Ulceration was coded as present when at least one ulcer was documented in that region; therefore, ulcer site count reflected the number of ulcer-involved regions rather than the number, size, depth, or severity of individual ulcers. Itch was coded as present only when a documented cutaneous lesion in that region was described as pruritic, and was not intended to represent isolated regional itch without corresponding skin involvement. In addition to site-specific lesion features, three composite variables–any ulcer, any itch, and any scale–were defined according to whether the corresponding feature was present in any anatomical region. Mucosal ulceration, alopecia, and photosensitivity were recorded separately because they were not anatomically compatible with the seven-region cutaneous matrix. Gottron papules/signs and periungual changes were not included in the site-by-lesion matrix because they were not consistently documented with sufficient regional granularity in retrospective records and were not part of the elemental ulcer/itch/scale feature set used for this framework. Systemic and immunological reference markers included the six FLATCAN components, PAH, IgG, and IgM. The six FLATCAN components were extracted as continuous values and dichotomised at thresholds prespecified from our centre’s prior FLATCAN cohort study [4] (age ≥ 50 years, albumin < 31 g/L, LDH ≥ 341 U/L, CD8 + < 80 cells/µL, ferritin ≥ 887 ng/mL, CRP ≥ 2.14 mg/dL); both forms were used in analysis. CD8 + T-cell depletion was prespecified as the primary systemic reference marker because of its established association with mortality risk in prior anti-MDA5-positive dermatomyositis cohorts. PAH was defined as estimated systolic pulmonary artery pressure ≥ 40 mmHg on transthoracic echocardiography. Serum immunoglobulin levels (IgG, IgM) were recorded. Missing data (0.33% overall; ≤2.9% per variable) were handled by multiple imputation (MICE, five datasets, pooled by Rubin’s rules). Statistical analysis Univariate screening For univariate screening, 14 skin variables, including the composite skin variables, were compared against 42 systemic and immunological reference variables (14 × 42 = 588 pairwise comparisons). These 14 variables comprised 3 composite variables and 11 site-specific variables retained on feature-specific event-count thresholds (ulcer ≥ 34, itch ≥ 30, scale ≥ 20 events), with thresholds calibrated to the underlying prevalence of each lesion feature to ensure stable association estimation. Categorical comparisons used χ² or Fisher exact tests; continuous variables used Mann-Whitney U tests. Benjamini-Hochberg FDR correction was applied across the prespecified screening matrix, and q values were used for hypothesis prioritisation. Correlation structure Pairwise Spearman correlation matrices were computed to assess collinearity among skin features and to identify clustering among systemic and immunological reference variables, informing variable selection for multivariable association models (Supplementary Figure S2). Multivariable regression Five exploratory association models were constructed: logistic regression for CD8 + T-cell depletion, hypoalbuminaemia, and PAH, and ordinary least squares regression for log-transformed IgG and IgM. All models were adjusted for age and sex. Incremental ROC analysis, restricted cubic splines, and decision curve analysis were used as exploratory internal model-performance summaries rather than evidence of clinical decision rules. Variable selection was hypothesis-driven, with L1-penalised regression used as a sensitivity analysis (Supplementary Table S7). Under-documentation sensitivity analyses simulated additional ulcer events at low-prevalence ulcer sites using random and worst-case allocation strategies (Supplementary Table S8). Continuous CD8⁺ count analysis To complement the binary threshold analysis, log-transformed CD8 + count was modelled as a continuous reference variable using OLS regression. A minimally adjusted model included any cutaneous ulceration, age, and sex; a fully adjusted model additionally included PAH and log-transformed ferritin to account for haemodynamic and inflammatory confounding. Lymphocyte count was not included because total lymphocytes contain CD8 + T cells as a subset, creating potential collider bias. To characterise where in the CD8 + distribution the ulcer extent signal operated, logistic regression was repeated across multiple CD8 + thresholds (< 80, < 100, <150, < 200, <300 cells/µL). Cross-correlation SVD and exploratory clustering For CC-SVD, we used 14 binary skin/mucocutaneous features (13 site × feature combinations with ≥ 5% prevalence plus mucosal ulcer; composite variables excluded to avoid redundancy) and 10 systemic reference variables: CD8 + depletion, hypoalbuminaemia, hyperferritinaemia, elevated LDH, elevated CRP, PAH, pulmonary infection, fungal infection, IgG, and IgM. Composite skin variables were excluded to avoid redundancy, and age was excluded because it was treated as a regression adjustment variable. After z-score standardisation, each matrix entry represented a Pearson correlation between one skin/mucocutaneous feature and one systemic reference variable. Dimension significance was assessed by permutation testing (1,000 iterations), loading stability by bootstrap resampling (1,000 iterations), and sensitivity analyses by removing infection variables or scale features. As a complementary patient-level analysis, Ward’s hierarchical clustering was applied to the same 14 prevalence-filtered, non-composite binary skin/mucocutaneous features. Cluster number was selected using silhouette and Calinski-Harabasz indices; K-means sensitivity analysis (k = 4, 50 random initialisations) evaluated concordance with Ward clustering; and bootstrap stability was evaluated over 200 iterations(Supplementary Figure S3). All analyses were performed in Python 3.12 using pandas, NumPy, SciPy (v1.15.2), scikit-learn (v1.8.0), and matplotlib. For the high-dimensional screening step, FDR-adjusted q values were used for hypothesis prioritisation. For hypothesis-driven multivariable association models, nominal two-sided P values, effect sizes, confidence intervals, and consistency with the framework-level analyses were interpreted exploratorily. Results Baseline characteristics and construction of the site-by-lesion matrix Of 339 patients (median age 51, IQR 40–58; 65.5% female; median disease duration 3 months), 328 (96.8%) had ILD and 101 (29.8%) had PAH. Ulceration was present in 118 (34.8%), pruritus in 178 (52.5%), and scaling in 114 (33.6%). Facial involvement was the most common across all three lesion features (ulcer 28.6%, itch 46.9%, scale 19.8%); finger and upper limb were the next most affected sites for ulcer and scale, while upper limb and neck/back ranked second for itch (Table 1). Among FLATCAN components, 45 (13.3%) had CD8 + depletion (< 80 cells/µL), 36 (10.6%) had hypoalbuminaemia (< 31 g/L), 103 (30.4%) had hyperferritinaemia, 93 (27.4%) had elevated CRP, and 86 (25.4%) had elevated LDH. Median CD8 + count was 225 cells/µL (IQR 124.5-410.5). Full baseline data are presented in Table 1. This distribution provided the basis for analysing three complementary dimensions of skin involvement: spatial extent, anatomical location, and lesion-feature type. Table 1Baseline characteristics of study patients (N = 339)VariableValueStatistic typeDemographics & General Information Age (years)51.00 (40.00–58.00)Median (IQR) Male117 (34.5%)n (%) Disease duration (months)3.00 (2.00–6.00)Median (IQR)Medications Glucocorticoids331 (97.6%)n (%) Glucocorticoid dose (mg/day)40.00 (30.00–45.00)Median (IQR) Immunosuppressants321 (94.7%)n (%) Biologics79 (23.3%)n (%) Symptomatic treatment330 (97.3%)n (%) JAK inhibitor66 (19.5%)n (%) Combination therapy114 (33.6%)n (%)Skin Involvement Facial lesion311 (91.7%)n (%) Itch159 (46.9%)n (%) Scale67 (19.8%)n (%) Ulcer97 (28.6%)n (%) Site feature count2.00 (1.00–3.00)Median (IQR) Upper limb lesion314 (92.6%)n (%) Itch72 (21.2%)n (%) Scale37 (10.9%)n (%) Ulcer44 (13.0%)n (%) Site feature count1.00 (1.00–2.00)Median (IQR) Lower limb lesion218 (64.3%)n (%) Itch24 (7.1%)n (%) Scale13 (3.8%)n (%) Ulcer14 (4.1%)n (%) Site feature count1.00 (0.00–1.00)Median (IQR) Back lesion208 (61.4%)n (%) Itch32 (9.4%)n (%) Scale6 (1.8%)n (%) Ulcer3 (0.9%)n (%) Site feature count1.00 (0.00–1.00)Median (IQR) Neck lesion239 (70.5%)n (%) Itch32 (9.4%)n (%) Scale9 (2.7%)n (%) Ulcer2 (0.6%)n (%) Site feature count1.00 (0.00–1.00)Median (IQR) Finger lesion307 (90.6%)n (%) Itch31 (9.1%)n (%) Scale58 (17.1%)n (%) Ulcer51 (15.0%)n (%) Site feature count1.00 (1.00–2.00)Median (IQR) Trunk lesion69 (20.4%)n (%) Itch17 (5.0%)n (%) Scale8 (2.4%)n (%) Ulcer7 (2.1%)n (%) Site feature count0.00 (0.00–0.00)Median (IQR) Total feature count6.00 (4.00–8.00)Median (IQR) Total ulcer count0.00 (0.00–1.00)Median (IQR) Total itch count1.00 (0.00–1.00)Median (IQR) Total scale count0.00 (0.00–1.00)Median (IQR)Pain Pain (any)307 (90.6%)n (%) Joint pain219 (64.6%)n (%) Muscle pain257 (75.8%)n (%) Nail pain9 (2.7%)n (%) Skin pain score11 (3.2%)n (%) Total pain score2.00 (1.00–2.00)Median (IQR)Clinical Manifestations Mucosal ulcer66 (19.5%)n (%) Interstitial lung disease (ILD)328 (96.8%)n (%) Pulmonary infection290 (85.5%)n (%) Pulmonary arterial hypertension (PAH)101 (29.8%)n (%) Bone/joint damage39 (11.5%)n (%) Heart damage122 (36.0%)n (%) Abnormal liver function192 (56.6%)n (%) Osteoporosis15 (4.4%)n (%) Viral infection200 (59.0%)n (%) Bacterial infection188 (55.5%)n (%) Fungal infection257 (75.8%)n (%) Alopecia61 (18.0%)n (%) Dysphagia101 (29.8%)n (%) Photosensitivity47 (13.9%)n (%) Raynaud’s phenomenon13 (3.8%)n (%) Muscle atrophy18 (5.3%)n (%) Muscle edema64 (18.9%)n (%) Dryness51 (15.0%)n (%)Laboratory Tests Lymphocyte count (×10⁹/L) *0.90 (0.57–1.33)Median (IQR) Lymphocyte ratio (%) *15.90 (10.95–22.50)Median (IQR) ESR (mm/h) *18.00 (10.00–33.00)Median (IQR) Creatine kinase (U/L)44.00 (27.00–90.00)Median (IQR) ALT (U/L) *38.00 (24.00–67.53)Median (IQR) AST (U/L) *37.00 (25.00–64.00)Median (IQR) Creatinine (µmol/L)48.70 (40.10–58.00)Median (IQR) D-dimer (mg/L) *0.53 (0.30–1.02)Median (IQR) IgG (mg/dL) *1240.00 (1025.00–1550.00)Median (IQR) IgM (mg/dL) *109.00 (81.50–152.50)Median (IQR) Anti-MDA5 antibody reactivity grade1.00 (1.00–2.00)Median (IQR) Ferritin (ng/mL) *495.70 (197.60–996.30)Median (IQR) CD8⁺ T-cell count (cells/µL)225.00 (124.50–410.50)Median (IQR) LDH (U/L)281.00 (234.50–345.50)Median (IQR) Albumin (g/L)35.80 (32.70–38.90)Median (IQR) CRP (mg/dL)0.66 (0.28–2.59)Median (IQR) Ro-52 antibody positive172 (50.7%)n (%)FLATCAN Score High age181 (53.4%)n (%) Low albumin36 (10.6%)n (%) High LDH86 (25.4%)n (%) High ferritin103 (30.4%)n (%) Low CD8⁺ count45 (13.3%)n (%) High CRP93 (27.4%)n (%)Abbreviations: IQR Interquartile range, ILD Interstitial lung disease, PAH Pulmonary arterial hypertension, ESR Erythrocyte sedimentation rate, CK Creatine kinase, ALT Alanine aminotransferase, AST Aspartate aminotransferase, CRP C-reactive protein, LDH Lactate dehydrogenaseAnti-MDA5 antibody results were recorded as a semi-quantitative laboratory grade: 1 = +, 2 = ++, and 3 = +++. Values are presented as median of the ordinal reactivity grade rather than true endpoint titresContinuous variables: median (Q1–Q3). Categorical variables: n (%), where n is the number of patients with the feature present (value = 1) out of N = 339Missing data (superscript indicates number of missing values):Total: 31 missing values across 9 variables (31/9,492 data cells = 0.33%). All handled by multiple imputation prior to analysisLymphocyte ratio (%): 10 value(s) missingD-dimer (mg/L): 8 value(s) missingESR (mm/h): 3 value(s) missingAST (U/L): 3 value(s) missingIgG (mg/dL): 2 value(s) missingIgM (mg/dL): 2 value(s) missingFerritin (ng/mL): 1 value(s) missingLymphocyte count (×10⁹/L): 1 value(s) missingALT (U/L): 1 value(s) missing Table 1Baseline characteristics of study patients (N = 339)VariableValueStatistic typeDemographics & General Information Age (years)51.00 (40.00–58.00)Median (IQR) Male117 (34.5%)n (%) Disease duration (months)3.00 (2.00–6.00)Median (IQR)Medications Glucocorticoids331 (97.6%)n (%) Glucocorticoid dose (mg/day)40.00 (30.00–45.00)Median (IQR) Immunosuppressants321 (94.7%)n (%) Biologics79 (23.3%)n (%) Symptomatic treatment330 (97.3%)n (%) JAK inhibitor66 (19.5%)n (%) Combination therapy114 (33.6%)n (%)Skin Involvement Facial lesion311 (91.7%)n (%) Itch159 (46.9%)n (%) Scale67 (19.8%)n (%) Ulcer97 (28.6%)n (%) Site feature count2.00 (1.00–3.00)Median (IQR) Upper limb lesion314 (92.6%)n (%) Itch72 (21.2%)n (%) Scale37 (10.9%)n (%) Ulcer44 (13.0%)n (%) Site feature count1.00 (1.00–2.00)Median (IQR) Lower limb lesion218 (64.3%)n (%) Itch24 (7.1%)n (%) Scale13 (3.8%)n (%) Ulcer14 (4.1%)n (%) Site feature count1.00 (0.00–1.00)Median (IQR) Back lesion208 (61.4%)n (%) Itch32 (9.4%)n (%) Scale6 (1.8%)n (%) Ulcer3 (0.9%)n (%) Site feature count1.00 (0.00–1.00)Median (IQR) Neck lesion239 (70.5%)n (%) Itch32 (9.4%)n (%) Scale9 (2.7%)n (%) Ulcer2 (0.6%)n (%) Site feature count1.00 (0.00–1.00)Median (IQR) Finger lesion307 (90.6%)n (%) Itch31 (9.1%)n (%) Scale58 (17.1%)n (%) Ulcer51 (15.0%)n (%) Site feature count1.00 (1.00–2.00)Median (IQR) Trunk lesion69 (20.4%)n (%) Itch17 (5.0%)n (%) Scale8 (2.4%)n (%) Ulcer7 (2.1%)n (%) Site feature count0.00 (0.00–0.00)Median (IQR) Total feature count6.00 (4.00–8.00)Median (IQR) Total ulcer count0.00 (0.00–1.00)Median (IQR) Total itch count1.00 (0.00–1.00)Median (IQR) Total scale count0.00 (0.00–1.00)Median (IQR)Pain Pain (any)307 (90.6%)n (%) Joint pain219 (64.6%)n (%) Muscle pain257 (75.8%)n (%) Nail pain9 (2.7%)n (%) Skin pain score11 (3.2%)n (%) Total pain score2.00 (1.00–2.00)Median (IQR)Clinical Manifestations Mucosal ulcer66 (19.5%)n (%) Interstitial lung disease (ILD)328 (96.8%)n (%) Pulmonary infection290 (85.5%)n (%) Pulmonary arterial hypertension (PAH)101 (29.8%)n (%) Bone/joint damage39 (11.5%)n (%) Heart damage122 (36.0%)n (%) Abnormal liver function192 (56.6%)n (%) Osteoporosis15 (4.4%)n (%) Viral infection200 (59.0%)n (%) Bacterial infection188 (55.5%)n (%) Fungal infection257 (75.8%)n (%) Alopecia61 (18.0%)n (%) Dysphagia101 (29.8%)n (%) Photosensitivity47 (13.9%)n (%) Raynaud’s phenomenon13 (3.8%)n (%) Muscle atrophy18 (5.3%)n (%) Muscle edema64 (18.9%)n (%) Dryness51 (15.0%)n (%)Laboratory Tests Lymphocyte count (×10⁹/L) *0.90 (0.57–1.33)Median (IQR) Lymphocyte ratio (%) *15.90 (10.95–22.50)Median (IQR) ESR (mm/h) *18.00 (10.00–33.00)Median (IQR) Creatine kinase (U/L)44.00 (27.00–90.00)Median (IQR) ALT (U/L) *38.00 (24.00–67.53)Median (IQR) AST (U/L) *37.00 (25.00–64.00)Median (IQR) Creatinine (µmol/L)48.70 (40.10–58.00)Median (IQR) D-dimer (mg/L) *0.53 (0.30–1.02)Median (IQR) IgG (mg/dL) *1240.00 (1025.00–1550.00)Median (IQR) IgM (mg/dL) *109.00 (81.50–152.50)Median (IQR) Anti-MDA5 antibody reactivity grade1.00 (1.00–2.00)Median (IQR) Ferritin (ng/mL) *495.70 (197.60–996.30)Median (IQR) CD8⁺ T-cell count (cells/µL)225.00 (124.50–410.50)Median (IQR) LDH (U/L)281.00 (234.50–345.50)Median (IQR) Albumin (g/L)35.80 (32.70–38.90)Median (IQR) CRP (mg/dL)0.66 (0.28–2.59)Median (IQR) Ro-52 antibody positive172 (50.7%)n (%)FLATCAN Score High age181 (53.4%)n (%) Low albumin36 (10.6%)n (%) High LDH86 (25.4%)n (%) High ferritin103 (30.4%)n (%) Low CD8⁺ count45 (13.3%)n (%) High CRP93 (27.4%)n (%)Abbreviations: IQR Interquartile range, ILD Interstitial lung disease, PAH Pulmonary arterial hypertension, ESR Erythrocyte sedimentation rate, CK Creatine kinase, ALT Alanine aminotransferase, AST Aspartate aminotransferase, CRP C-reactive protein, LDH Lactate dehydrogenaseAnti-MDA5 antibody results were recorded as a semi-quantitative laboratory grade: 1 = +, 2 = ++, and 3 = +++. Values are presented as median of the ordinal reactivity grade rather than true endpoint titresContinuous variables: median (Q1–Q3). Categorical variables: n (%), where n is the number of patients with the feature present (value = 1) out of N = 339Missing data (superscript indicates number of missing values):Total: 31 missing values across 9 variables (31/9,492 data cells = 0.33%). All handled by multiple imputation prior to analysisLymphocyte ratio (%): 10 value(s) missingD-dimer (mg/L): 8 value(s) missingESR (mm/h): 3 value(s) missingAST (U/L): 3 value(s) missingIgG (mg/dL): 2 value(s) missingIgM (mg/dL): 2 value(s) missingFerritin (ng/mL): 1 value(s) missingLymphocyte count (×10⁹/L): 1 value(s) missingALT (U/L): 1 value(s) missing Baseline characteristics of study patients (N = 339) Abbreviations: IQR Interquartile range, ILD Interstitial lung disease, PAH Pulmonary arterial hypertension, ESR Erythrocyte sedimentation rate, CK Creatine kinase, ALT Alanine aminotransferase, AST Aspartate aminotransferase, CRP C-reactive protein, LDH Lactate dehydrogenase Anti-MDA5 antibody results were recorded as a semi-quantitative laboratory grade: 1 = +, 2 = ++, and 3 = +++. Values are presented as median of the ordinal reactivity grade rather than true endpoint titres Continuous variables: median (Q1–Q3). Categorical variables: n (%), where n is the number of patients with the feature present (value = 1) out of N = 339 Missing data (superscript indicates number of missing values): Total: 31 missing values across 9 variables (31/9,492 data cells = 0.33%). All handled by multiple imputation prior to analysis Lymphocyte ratio (%): 10 value(s) missing D-dimer (mg/L): 8 value(s) missing ESR (mm/h): 3 value(s) missing AST (U/L): 3 value(s) missing IgG (mg/dL): 2 value(s) missing IgM (mg/dL): 2 value(s) missing Ferritin (ng/mL): 1 value(s) missing Lymphocyte count (×10⁹/L): 1 value(s) missing ALT (U/L): 1 value(s) missing Extent dimension: ulcer involvement and CD8 + T-cell depletion Compared with patients without cutaneous ulceration, those with ulceration at any site had a higher rate of CD8 + depletion (22.9% vs. 8.1%; OR = 3.35, 95% CI 1.75–6.38, P < 0.001; FDR q = 0.026). After FDR correction, this remained among the strongest skin-laboratory associations across the 588 comparisons, showing that the association between ulceration and CD8 + depletion was detectable even with a conventional binary skin variable. Site-specific analysis showed that this association was also related to ulcer extent. Ulcer site count remained independently associated with CD8 + depletion (OR = 1.43 per additional site, 95% CI 1.10–1.86, P = 0.008; Table 2, Panel A), with modest internal information gain beyond age, sex, and PAH (ΔAUC = 0.060, 95% CI 0.022–0.123, P < 0.001; Fig. 2A). Exploratory decision-curve analysis favoured the model incorporating cutaneous features across threshold probabilities of approximately 1%-49% (Fig. 2C), but did not establish a clinical decision rule. RCS analysis suggested a nonlinear dose-response relationship (overall P = 0.004; nonlinearity P = 0.002), with a descriptive inflection at approximately 2 involved sites (Fig. 2B). Table 2Multivariable regression analyses: predictors of FLATCAN severity components, serum immunoglobulin levels, and pulmonary arterial hypertension (N = 339)PanelPredictorEstimate (95% CI)P valueΔAUCPanel A. CD8⁺ T-cell depletion (CD8⁺ <80 cells/µL) Binary logistic; n = 339; Nagelkerke R² = 0.194 No. of ulcer sites (0–7) OR 1.43 (1.10–1.86) 0.008 — Age (years) OR 1.09 (1.05–1.13) < 0.001 —Sex (female)OR 0.57 (0.27–1.21)0.142—PAHOR 1.57 (0.79–3.14)0.200—Panel B. Hypoalbuminaemia (albumin < 31 g/L) Binary logistic; n = 339; Nagelkerke R² = 0.094 Facial itch OR 2.20 (1.01–4.79) 0.047 — Age (years) OR 1.04 (1.01–1.07) 0.016 —Facial ulcerOR 1.51 (0.63–3.62)0.351—Upper limb ulcerOR 1.30 (0.46–3.63)0.622—Sex (female)OR 1.18 (0.52–2.70)0.694—Panel C. Serum IgG (log-transformed) OLS linear; n = 339; R² = 0.066 Any ulcer β 0.089 (0.012–0.166) 0.024 — Ferritin (log) β 0.049 (0.020–0.078) < 0.001 —Any itchβ 0.046 (− 0.027–0.120)0.216—Sex (female)β 0.043 (− 0.036–0.123)0.284—Age (years)β − 0.001 (− 0.004–0.002)0.608—Panel D. Serum IgM (log-transformed) OLS linear; n = 339; R² = 0.089 Any itch β 0.186 (0.057–0.314) 0.005 — Sex (female) β 0.275 (0.143–0.407) < 0.001 —Any ulcerβ 0.052 (− 0.083–0.186)0.452—Age (years)β − 0.004 (− 0.009–0.001)0.147—Panel E. Pulmonary arterial hypertension Binary logistic; n = 339; Nagelkerke R² = 0.115; AUC = 0.671 (0.604–0.736) Facial ulcer OR 2.28 (1.33–3.97) 0.003 + 0.073 Facial itch OR 1.78 (1.06–3.05) 0.035 + 0.026Hyperferritinaemia (≥ 887)OR 1.69 (0.99–2.92)0.051+ 0.015Age (years)OR 1.01 (0.99–1.04)0.201—Sex (female)OR 1.09 (0.62–1.99)0.762—OR Odds ratio, β Unstandardised regression coefficient, CI Confidence interval, PAH Pulmonary arterial hypertension, AUC Area under the receiver operating characteristic curve, ΔAUC Incremental AUC from nested ROC analysis. Bold rows indicate statistically significant predictors (P < 0.05). Panels A–B: binary logistic regression; Panels C–D: ordinary least squares linear regression; Panel E: binary logistic regression with incremental ROC. All models adjusted for age and sex. The ΔAUC column applies only to Panel E; dashes indicate not applicableBolded entries in the table body indicate statistically significant predictors (P < 0.05); bold panel headings are used for table structure only Table 2Multivariable regression analyses: predictors of FLATCAN severity components, serum immunoglobulin levels, and pulmonary arterial hypertension (N = 339)PanelPredictorEstimate (95% CI)P valueΔAUCPanel A. CD8⁺ T-cell depletion (CD8⁺ <80 cells/µL) Binary logistic; n = 339; Nagelkerke R² = 0.194 No. of ulcer sites (0–7) OR 1.43 (1.10–1.86) 0.008 — Age (years) OR 1.09 (1.05–1.13) < 0.001 —Sex (female)OR 0.57 (0.27–1.21)0.142—PAHOR 1.57 (0.79–3.14)0.200—Panel B. Hypoalbuminaemia (albumin < 31 g/L) Binary logistic; n = 339; Nagelkerke R² = 0.094 Facial itch OR 2.20 (1.01–4.79) 0.047 — Age (years) OR 1.04 (1.01–1.07) 0.016 —Facial ulcerOR 1.51 (0.63–3.62)0.351—Upper limb ulcerOR 1.30 (0.46–3.63)0.622—Sex (female)OR 1.18 (0.52–2.70)0.694—Panel C. Serum IgG (log-transformed) OLS linear; n = 339; R² = 0.066 Any ulcer β 0.089 (0.012–0.166) 0.024 — Ferritin (log) β 0.049 (0.020–0.078) < 0.001 —Any itchβ 0.046 (− 0.027–0.120)0.216—Sex (female)β 0.043 (− 0.036–0.123)0.284—Age (years)β − 0.001 (− 0.004–0.002)0.608—Panel D. Serum IgM (log-transformed) OLS linear; n = 339; R² = 0.089 Any itch β 0.186 (0.057–0.314) 0.005 — Sex (female) β 0.275 (0.143–0.407) < 0.001 —Any ulcerβ 0.052 (− 0.083–0.186)0.452—Age (years)β − 0.004 (− 0.009–0.001)0.147—Panel E. Pulmonary arterial hypertension Binary logistic; n = 339; Nagelkerke R² = 0.115; AUC = 0.671 (0.604–0.736) Facial ulcer OR 2.28 (1.33–3.97) 0.003 + 0.073 Facial itch OR 1.78 (1.06–3.05) 0.035 + 0.026Hyperferritinaemia (≥ 887)OR 1.69 (0.99–2.92)0.051+ 0.015Age (years)OR 1.01 (0.99–1.04)0.201—Sex (female)OR 1.09 (0.62–1.99)0.762—OR Odds ratio, β Unstandardised regression coefficient, CI Confidence interval, PAH Pulmonary arterial hypertension, AUC Area under the receiver operating characteristic curve, ΔAUC Incremental AUC from nested ROC analysis. Bold rows indicate statistically significant predictors (P < 0.05). Panels A–B: binary logistic regression; Panels C–D: ordinary least squares linear regression; Panel E: binary logistic regression with incremental ROC. All models adjusted for age and sex. The ΔAUC column applies only to Panel E; dashes indicate not applicableBolded entries in the table body indicate statistically significant predictors (P < 0.05); bold panel headings are used for table structure only Multivariable regression analyses: predictors of FLATCAN severity components, serum immunoglobulin levels, and pulmonary arterial hypertension (N = 339) OR Odds ratio, β Unstandardised regression coefficient, CI Confidence interval, PAH Pulmonary arterial hypertension, AUC Area under the receiver operating characteristic curve, ΔAUC Incremental AUC from nested ROC analysis. Bold rows indicate statistically significant predictors (P < 0.05). Panels A–B: binary logistic regression; Panels C–D: ordinary least squares linear regression; Panel E: binary logistic regression with incremental ROC. All models adjusted for age and sex. The ΔAUC column applies only to Panel E; dashes indicate not applicable Bolded entries in the table body indicate statistically significant predictors (P < 0.05); bold panel headings are used for table structure only Fig. 2Ulcer burden and CD8⁺ T-cell depletion: discriminative value, dose–response, and clinical decision analysis (N = 339). A Incremental ROC analysis with four nested models: Model 1 (age + sex + PAH), Model 2 (+ any skin ulcer), Model 3 (+ mucosal ulcer), Model 4 (+ facial itch). Total ΔAUC = 0.060 (95% CI 0.022–0.123, P < 0.001). B Restricted cubic spline regression showing the dose–response relationship between ulcer site count and CD8⁺ depletion risk (overall P = 0.004; nonlinearity P = 0.002), adjusted for age and sex only (linear reference OR = 1.46 per site; differs from Table 2 Panel A [OR = 1.43], which additionally adjusted for PAH). C Decision curve analysis comparing the base model versus the model incorporating cutaneous features. Alt text: Three-panel figure showing ROC curves, restricted cubic spline dose-response curve, and decision curve analysis for the association between cutaneous ulcer features and CD8-positive T-cell depletion Ulcer burden and CD8⁺ T-cell depletion: discriminative value, dose–response, and clinical decision analysis (N = 339). A Incremental ROC analysis with four nested models: Model 1 (age + sex + PAH), Model 2 (+ any skin ulcer), Model 3 (+ mucosal ulcer), Model 4 (+ facial itch). Total ΔAUC = 0.060 (95% CI 0.022–0.123, P < 0.001). B Restricted cubic spline regression showing the dose–response relationship between ulcer site count and CD8⁺ depletion risk (overall P = 0.004; nonlinearity P = 0.002), adjusted for age and sex only (linear reference OR = 1.46 per site; differs from Table 2 Panel A [OR = 1.43], which additionally adjusted for PAH). C Decision curve analysis comparing the base model versus the model incorporating cutaneous features. Alt text: Three-panel figure showing ROC curves, restricted cubic spline dose-response curve, and decision curve analysis for the association between cutaneous ulcer features and CD8-positive T-cell depletion Reverse-stratified descriptive analysis using FLATCAN components as grouping variables was directionally consistent with this pattern: patients with CD8 + depletion had significantly higher prevalence of any ulcer (P < 0.001), facial ulcer (P = 0.037), finger ulcer (P = 0.019), and upper limb ulcer (P = 0.017), together with elevated ferritin, D-dimer, and depressed albumin (all P < 0.001). Notably, high CRP (≥ 2.14 mg/dL) showed no significant association with any cutaneous feature, suggesting that the ulcer-CD8 + association was not paralleled by a comparable CRP-defined acute-phase signal in this dataset. Together, these findings illustrate the extent dimension of the framework: ulceration was not only present or absent, but its regional distribution provided additional phenotypic information for severe CD8 + depletion within this cohort. Low-prevalence ulcer-site under-documentation simulations supported the relative stability of the any-ulcer-CD8 + signal, while showing greater sensitivity of ulcer site count under extreme worst-case allocation (Supplementary Table S8). Full univariate comparisons for ulcer, itch, and scale features are provided in Supplementary Tables S1–S3 and Supplementary Figure S1. Continuous CD8 + count analysis within the extent dimension To evaluate whether the ulcer-CD8 + association extends beyond the binary < 80 cells/µL threshold, log-transformed CD8 + count was modelled as a continuous reference variable. In a minimally adjusted model (age, sex), any cutaneous ulceration was associated with a 22% reduction in geometric mean CD8 + count (β = −0.253, P = 0.006). After further adjustment for PAH and log-transformed ferritin, the association was attenuated but remained significant (β = −0.182, P = 0.047), corresponding to approximately 17% lower geometric mean. Ulcer site count showed a similar direction (β = −0.069 per site) but did not reach significance in the fully adjusted model (P = 0.089), and the RCS nonlinearity term was not significant for continuous CD8+ (P = 0.235), in contrast to the binary threshold analysis. To reconcile this discrepancy, the ulcer site count odds ratio was examined across multiple CD8 + thresholds (Supplementary Table S6). The site count OR was strongest at CD8 + < 80 cells/µL (OR = 1.48, P = 0.006) and remained significant through < 200 cells/µL (OR = 1.31, P = 0.015), but was non-significant at < 300 cells/µL (OR = 1.14, P = 0.262). This gradient indicates that binary ulceration captures the overall ulcer-CD8 + association across the full distribution, whereas site counting provides additional stratification mainly in the lower CD8 + range, especially around thresholds used to define severe depletion. Location dimension: facial involvement and PAH In multivariable analysis, facial ulcer was the strongest independent correlate of PAH (OR = 2.28, 95% CI 1.33–3.97, P = 0.003; Table 2, Panel E), followed by facial itch (OR = 1.78, 95% CI 1.06–3.05, P = 0.035). The total internal information gain from adding cutaneous features and hyperferritinaemia to a demographic baseline model was ΔAUC = 0.115 (95% CI 0.049–0.188, P < 0.001). In a parallel model replacing facial ulcer with ulcer site count, the site-count association was weaker (OR = 1.21 per site, P = 0.087). This pattern suggests that PAH aligned more closely with facial involvement than with generalized ulcer extent in this cohort. Lesion-feature dimension: ulceration and itch map to different immunological patterns The analyses above examined ulcer extent and its site-specific associations with systemic reference markers. We next examined whether lesion-feature type provided additional information by comparing the systemic correlates of ulceration and pruritus. Compared with patients without pruritus, those with pruritus at any site had higher rates of hypoalbuminaemia (14.0% vs. 6.8%; OR = 2.23, P = 0.031) and higher IgM levels (median 118 vs. 96.8 mg/dL, P < 0.001; FDR q = 0.034) but showed no significant association with CD8 + depletion (q = 0.805). In multivariable regression, facial itch was the only cutaneous feature associated with hypoalbuminaemia, although this was only borderline significant (OR = 2.20, 95% CI 1.01–4.79, P = 0.047; Table 2, Panel B); in contrast, any itch remained independently associated with IgM elevation (β = 0.186, P = 0.005; Table 2, Panel D). Ulcer and itch showed different immunoglobulin associations after multivariable adjustment: any ulcer was associated with IgG (β = 0.089, P = 0.024; Table 2, Panel C) but not IgM, whereas any itch was associated with IgM (β = 0.186, P = 0.005; Table 2, Panel D) but not IgG. Cross-pathway signals were present at the univariate level but did not survive FDR correction, supporting preferential rather than absolute associations. Scale features showed fewer associations with the selected reference markers. Site-specific immunoglobulin associations are detailed in Supplementary Table S4. The ulcer-IgG association was also sensitive to worst-case under-documentation assumptions and should therefore remain exploratory (Supplementary Table S8). Integrated framework map: cross-correlation decomposition and exploratory clustering Having examined the extent, location, and lesion-feature dimensions separately, we used CC-SVD to summarise exploratory skin-systemic covariation. SVD was applied to the cross-correlation matrix between the 14 prevalence-filtered, non-composite skin/mucocutaneous features and 10 systemic reference variables; two dimensions captured 76% of the total skin-systemic cross-correlation (Fig. 3A-B). Fig. 3Cross-correlation decomposition and cluster alignment (N = 339). A Patient clusters projected onto the CC-SVD two-dimensional space; ellipses denote 95% confidence regions; large markers indicate cluster centroids. B Variable loadings from the cross-correlation SVD; numbered markers correspond to the 10 highest-magnitude variables (see inset legend); solid arrows denote skin features, dashed arrows denote systemic outcomes. C Skin feature prevalence heatmap by cluster. D Clinical outcomes by cluster; laboratory medians annotated below. Alt text: Four-panel figure showing patient scatter in CC-SVD space with cluster ellipses, variable loading biplot with numbered arrows, skin feature prevalence heatmap, and clinical outcome bar chart across four Ward hierarchical clusters Cross-correlation decomposition and cluster alignment (N = 339). A Patient clusters projected onto the CC-SVD two-dimensional space; ellipses denote 95% confidence regions; large markers indicate cluster centroids. B Variable loadings from the cross-correlation SVD; numbered markers correspond to the 10 highest-magnitude variables (see inset legend); solid arrows denote skin features, dashed arrows denote systemic outcomes. C Skin feature prevalence heatmap by cluster. D Clinical outcomes by cluster; laboratory medians annotated below. Alt text: Four-panel figure showing patient scatter in CC-SVD space with cluster ellipses, variable loading biplot with numbered arrows, skin feature prevalence heatmap, and clinical outcome bar chart across four Ward hierarchical clusters Dimension 1 (59%, permutation P < 0.001) linked facial ulcer (+ 0.56) and facial itch (+ 0.51) with PAH (+ 0.47), pulmonary infection (+ 0.44), and fungal infection (+ 0.51). Loadings were bootstrap-stable within this cohort, and removing infection variables increased the PAH loading from + 0.47 to + 0.66. Dimension 2 (17%, permutation P = 0.016) linked acral ulcer and scale features with IgG (+ 0.54) and CD8 + depletion (+ 0.44). Ulcer features carried the largest loadings after scale removal, but individual Dim 2 bootstrap CIs crossed zero and permutation significance weakened in sensitivity analyses (P = 0.092–0.257), so this dimension should be interpreted cautiously. Ward clustering on the same 14 features identified four exploratory clusters (Table 3) (Fig. 3A, C-D). C1 (“diffuse”, n = 65) scored high on both dimensions, with the highest PAH (40%) but low CD8 + depletion (9%). C2 (“minimal”, n = 113) anchored the low end of both dimensions. C3 (“itch-dominant”, n = 114) aligned with Dim 1 and high infection burden. C4 (“ulcer-dominant”, n = 47) aligned with Dim 2 and had the highest CD8 + depletion (28%) and IgG (median 1425 mg/dL). Comprehensive baseline characteristics by cluster are provided in Supplementary Table S5. Table 3Cutaneous phenotype clusters and their clinical–immunological profiles (N = 339)Skin features (%)C1 Diffuse(n = 65)C2 Minimal(n = 113)C3 Itch-dom.(n = 114)C4 Ulcer-dom.(n = 47)Facial ulcer71%0%32%32%Finger ulcer26%0%5%60%UL ulcer35%0%2%40%Facial itch95%0%81%11%UL itch92%0%9%4%Clinical outcomes (%) CD8⁺ <80 cells/µL9.2%8.0%14.9%27.7% Low albumin16.9%7.1%11.4%8.5% PAH40.0%19.5%35.1%27.7% Pulm. infection89.2%81.4%89.5%80.9% Fungal infection76.9%72.6%84.2%61.7%Laboratory medians CD8⁺ count (cells/µL)251249214137 IgG (mg/dL)1280119012201425 IgM (mg/dL)12998111116 Albumin (g/L)36373635 Ferritin (ng/mL)481458512568 Anti-MDA5 reactivity grade1111Values are prevalence (%) for binary variables and median values for continuous variables. Ward hierarchical clustering on 14 binary skin features. K-means sensitivity analysis (k = 4) confirmed algorithm independence. Bootstrap stability varied across clusters. UL, upper limb; PAH, pulmonary arterial hypertension Table 3Cutaneous phenotype clusters and their clinical–immunological profiles (N = 339)Skin features (%)C1 Diffuse(n = 65)C2 Minimal(n = 113)C3 Itch-dom.(n = 114)C4 Ulcer-dom.(n = 47)Facial ulcer71%0%32%32%Finger ulcer26%0%5%60%UL ulcer35%0%2%40%Facial itch95%0%81%11%UL itch92%0%9%4%Clinical outcomes (%) CD8⁺ <80 cells/µL9.2%8.0%14.9%27.7% Low albumin16.9%7.1%11.4%8.5% PAH40.0%19.5%35.1%27.7% Pulm. infection89.2%81.4%89.5%80.9% Fungal infection76.9%72.6%84.2%61.7%Laboratory medians CD8⁺ count (cells/µL)251249214137 IgG (mg/dL)1280119012201425 IgM (mg/dL)12998111116 Albumin (g/L)36373635 Ferritin (ng/mL)481458512568 Anti-MDA5 reactivity grade1111Values are prevalence (%) for binary variables and median values for continuous variables. Ward hierarchical clustering on 14 binary skin features. K-means sensitivity analysis (k = 4) confirmed algorithm independence. Bootstrap stability varied across clusters. UL, upper limb; PAH, pulmonary arterial hypertension Cutaneous phenotype clusters and their clinical–immunological profiles (N = 339) Values are prevalence (%) for binary variables and median values for continuous variables. Ward hierarchical clustering on 14 binary skin features. K-means sensitivity analysis (k = 4) confirmed algorithm independence. Bootstrap stability varied across clusters. UL, upper limb; PAH, pulmonary arterial hypertension K-means sensitivity analysis showed moderate concordance with Ward clustering (adjusted Rand index = 0.60), and bootstrap stability varied across clusters (Jaccard range 0.58–0.90). The four-cluster solution should be regarded as an exploratory discretisation of the continuous CC-SVD space; future replication should focus on the cross-correlation dimensions rather than cluster assignments. Discussion The principal finding of this study is not any single skin-systemic association, but the demonstration that cutaneous involvement in anti-MDA5-positive dermatomyositis can be organised into an analysable site-by-lesion framework. When benchmarked against established systemic and immunological reference markers, this framework revealed extent-, location-, and lesion-feature-dependent association patterns that are not readily captured by conventional binary skin descriptors. The individual associations reported here should therefore be interpreted as framework-derived observations rather than as validated prognostic or predictive markers. Across the framework, we adopt an explicit hierarchy of evidence: binary any-ulcer–CD8⁺ depletion is the most stable extent-axis signal, retained statistical support under both random and worst-case allocation up to a 3-fold inflation multiplier (Supplementary Table S8); ulcer site count provides additional spatial resolution but with greater susceptibility to documentation error; and ulcer–IgG and dimension-2 patterns remain exploratory observations requiring prospective replication. This hierarchy guides the interpretive language used throughout the Discussion. Extent dimension: ulcer involvement and CD8 + depletion The ulcer-CD8 + depletion association illustrates how the site-by-lesion framework adds resolution beyond conventional binary assessment. Binary ulceration was significantly associated with CD8 + depletion, and site count suggested dose-response characteristics in the binary model (RCS nonlinearity P = 0.002), although this pattern was not reproduced when CD8 + was analysed continuously (P = 0.235). Ulcer site counting therefore adds spatial information to a binary ulcer descriptor, but its internal information gain was modest (ΔAUC = 0.060). Ulcers in anti-MDA5-positive DM have been linked to complement-mediated microvascular injury [16–18]. One speculative interpretation is that peripheral CD8 + depletion may reflect recruitment, exhaustion, or redistribution of cytotoxic T lymphocytes into damaged tissues [19–23], a hypothesis requiring dedicated immunophenotyping. The co-occurrence of facial ulceration and PAH (OR = 2.28) is also consistent with potentially shared microvascular pathology [18, 24, 25], but should remain a cautious location-dimension observation because it was derived from retrospective site-specific documentation and lacks prospective validation. Continuous CD8 + count analysis further contextualised the extent axis. Binary ulceration captured an association across the CD8 + distribution (adjusted β = −0.182, P = 0.047), whereas ulcer site counting showed a threshold-concentrated pattern below 200 cells/µL. This supports interpreting the approximately 2-site pattern as a descriptive inflection rather than an actionable cutoff. The low-prevalence ulcer-site simulations support binary ulceration as the most stable extent-axis signal, while indicating that site counts add regional resolution but depend more on complete documentation. Lesion-feature dimension: different lesion features, different immune-marker patterns Site-specific analysis suggested that ulceration and pruritus were associated with different systemic and immunological reference-marker patterns (CD8+/IgG vs. albumin/IgM). This preferential dissociation has been little explored in anti-MDA5-positive DM and suggests that deconstruction by lesion type and anatomical site may reveal research-relevant heterogeneity not captured by binary categorisation. Pruritus in DM has been linked to overexpression of IL-31 in lesional skin [26]. One speculative hypothesis is that inflammatory mediators could contribute to cutaneous itch, capillary protein loss, and polyclonal B-cell stimulation, providing a possible framework for the itch-hypoalbuminaemia and itch-IgM signals observed here. The facial itch-hypoalbuminaemia association was borderline (P = 0.047), whereas the itch-IgM association was more consistent within this dataset (P = 0.005, FDR q = 0.034). Cluster-level patterns did not map cleanly onto the regression-level itch-IgM signal. Whether IgG elevation with ulceration reflects antigen-driven humoral responses [3, 26], or IgM elevation with pruritus reflects polyclonal B-cell activation [27], remains speculative and warrants B-cell subset analysis [28]. The sensitivity of the ulcer-IgG model to worst-case under-documentation reinforces its hypothesis-generating status; the itch-IgM pattern likewise remains exploratory because of the retrospective and subjective nature of pruritus recording. Integrated framework structure: cross-correlation decomposition CC-SVD provided a framework-level summary of the regression findings. Dimension 1 (facial involvement-PAH/infection, 59% of cross-correlation) was bootstrap-stable within this cohort and persisted after infection variables were removed. Dimension 2 (acral ulcer-CD8+/IgG, 17%) was directionally consistent with the ulcer-CD8 + regression findings but statistically less stable. Regression analyses suggested lesion-feature separation (ulcer-CD8+/IgG; itch-albumin/IgM; cross-pathway associations not surviving FDR correction), whereas CC-SVD added location structure. Dim 1 was driven by facial ulcer and facial itch and associated with PAH and infection; Dim 2 was driven by acral ulcer features and associated with CD8 + depletion and IgG. Together, these findings indicate that both location and lesion type contributed to the observed skin-systemic patterns. The four clusters broadly corresponded to this space but should be viewed only as exploratory groupings: C4 aligned with Dim 2 and had the highest CD8 + depletion and IgG; C3 aligned with Dim 1 and high infection burden; and C1 had the highest PAH but low CD8 + depletion. Together, the regression models, CC-SVD, and low-prevalence ulcer-site simulations are mutually consistent with the evidence hierarchy outlined at the start of the Discussion, with binary ulceration–CD8⁺ depletion remaining the most robust extent-axis observation. Research implications and future validation The immediate implication is not that ulcer count, facial ulceration, or itch should be used as standalone clinical decision tools. Rather, future dermatomyositis studies may benefit from structured site-by-lesion recording to generate testable hypotheses about skin-systemic heterogeneity. Prospective cohorts with standardised skin assessment, direct clinical outcomes [29], and external validation are needed to determine whether this framework improves disease stratification. Comparison with existing literature Previous studies have noted associations between ulceration and anti-MDA5 antibody levels [27, 30–33], but have not evaluated site-specific associations, dose-response relationships with CD8 + depletion, or the preferential immunoglobulin association pattern. To our knowledge, few previous studies have examined cutaneous involvement in anti-MDA5-positive DM jointly by lesion type and anatomical site in relation to systemic and immunological reference markers. Limitations This study has several limitations. First, as a single-centre retrospective analysis, selection and information biases cannot be avoided. The framework was CDASI-informed but not equivalent to formal CDASI scoring: complete CDASI activity or damage scores were unavailable in retrospective records, and attempting to reconstruct them would have introduced greater missingness and inter-record heterogeneity than the simplified site-by-lesion coding used here. This trade-off prioritised data consistency over scoring granularity. Cutaneous features were extracted from medical records rather than prospective standardised skin assessment, and documentation may have varied by physician and anatomical visibility. Second, binary site-feature coding could not distinguish multiple ulcers within one region from a single ulcer; ulcer site count therefore reflects spatial extent across regions rather than lesion-level ulcer burden. The principal sources of measurement error differ between lesion features. Ulcer documentation is most likely affected by unidirectional under-recording at low-prevalence, low-visibility sites (lower limb 4.1%, trunk 2.1%); we therefore quantified its impact through simulation-based sensitivity analyses (Supplementary Table S8). Pruritus, by contrast, is patient-reported and subject to bidirectional misclassification, for which directional simulation is not appropriate; the itch-related findings are accordingly framed as exploratory throughout the Discussion. The inpatient first-hospitalisation cohort may overrepresent severe disease and may not generalise to milder outpatient or early-stage populations. Third, the cross-sectional design cannot establish temporal sequence or causality. The study did not directly evaluate mortality, ICU admission, mechanical ventilation, RP-ILD progression, or other serious clinical outcomes, and no external validation cohort was available. CC-SVD captures only linear associations; Dim 2 and cluster assignments remain exploratory and require replication. Conclusion In this retrospective cohort of anti-MDA5-positive dermatomyositis patients, a CDASI-informed site-by-lesion framework provided a structured way to examine cutaneous heterogeneity. Using established systemic and immunological markers as reference anchors, the framework revealed exploratory extent-, location-, and lesion-feature-dependent skin-systemic association patterns. These findings support further prospective evaluation of structured cutaneous phenotyping in dermatomyositis research, but do not establish a validated prognostic or predictive tool. Supplementary Information Additional file 1. Supplementary Table S1. Univariate comparison: Any Ulcer present vs. absent (N = 339). Supplementary Table S2. Univariate comparison: Any Itch present vs. absent (N = 339). Supplementary Table S3. Univariate comparison: Any Scale present vs. absent (N = 339). Supplementary Table S4. Site-specific associations between cutaneous features and serum immunoglobulin levels (N = 339). Supplementary Table S5. Comprehensive baseline characteristics by cutaneous phenotype cluster (N = 339). Supplementary Table S6. Threshold-specific association between ulcer site count and CD8⁺ T-cell count: adjusted odds ratios across multiple CD8⁺ thresholds (N = 339). Supplementary Table S7. Regularised variable selection across three LASSO approaches: methodological rationale for hypothesis-driven predictor selection (N = 339). Supplementary Table S8. Sensitivity analyses for retrospective under-documentation at low-prevalence ulcer sites: random and worst-case allocation strategies (N = 339). Supplementary Figure S1. Layer 1 univariate association analysis. Supplementary Figure S2. Spearman correlation matrices. Supplementary Figure S3. Cluster quality metrics and Ward dendrogram Additional file 1. Supplementary Table S1. Univariate comparison: Any Ulcer present vs. absent (N = 339). Supplementary Table S2. Univariate comparison: Any Itch present vs. absent (N = 339). Supplementary Table S3. Univariate comparison: Any Scale present vs. absent (N = 339). Supplementary Table S4. Site-specific associations between cutaneous features and serum immunoglobulin levels (N = 339). Supplementary Table S5. Comprehensive baseline characteristics by cutaneous phenotype cluster (N = 339). Supplementary Table S6. Threshold-specific association between ulcer site count and CD8⁺ T-cell count: adjusted odds ratios across multiple CD8⁺ thresholds (N = 339). Supplementary Table S7. Regularised variable selection across three LASSO approaches: methodological rationale for hypothesis-driven predictor selection (N = 339). Supplementary Table S8. Sensitivity analyses for retrospective under-documentation at low-prevalence ulcer sites: random and worst-case allocation strategies (N = 339). Supplementary Figure S1. Layer 1 univariate association analysis. Supplementary Figure S2. Spearman correlation matrices. Supplementary Figure S3. Cluster quality metrics and Ward dendrogram