We conducted a multicenter cohort study in seven reference centers within the Myosit’EST network in northeastern France. Patients were recruited from rheumatology, clinical immunology, and pulmonology departments. Inclusion criteria were: (1) age ≥18 years; (2) definite ASyS according to the proposed 2024 American College of Rheumatology/European Alliance of Associations for Rheumatology classification criteria [9]; (3) positivity for anti-ARS antibodies (anti-Jo-1, PL-7, PL-12, EJ, or OJ); (4) ILD documented by HRCT; and (5) availability of baseline FVC measurement at ILD diagnosis and at least one follow-up FVC measurement. Anti-Jo-1 antibodies were identified by immunoprecipitation or enzyme-linked immunosorbent assay (ELISA), whereas non-anti-Jo-1 antibodies were identified by line/dot immunoblot assays. Patients with substantial missing baseline data or with multiple or rare anti-ARS antibodies were excluded because of uncertain antibody profiles or limited assay reliability. Associated systemic autoimmune rheumatic diseases (SARDs) were defined according to international classification criteria [10–13].

In accordance with French legislation, written informed consent was not required; participants received an information notice and nonopposition form. The study was approved by the local ethics committee and registered at ClinicalTrials.gov (NCT04924465).

FVC was the primary endpoint and was expressed as percentage predicted according to the Global Lung Function Initiative (GLI) equations, enabling longitudinal comparisons [14]. Diffusing capacity of the lung for carbon monoxide (DLCO) was evaluated as a secondary endpoint.

Rapidly progressive ILD (RP-ILD) was defined as acute respiratory failure at diagnosis (PaO2 <60mmHg or oxygen requirement) without an alternative cause, or rapid worsening within 3 months, adapted from Wu et al. [15], defined by at least two of the following criteria: (1) increased ILD extent on HRCT, (2) worsening dyspnea, or (3) ≥5% relative decline in predicted FVC. ASyS-ILD relapse was defined as new bilateral ground-glass opacities on HRCT after initial improvement, without alternative causes, requiring intensification of immunosuppressive therapy [16]. Chronic respiratory failure was defined as PaO2 <70mmHg at rest on room air on at least two occasions separated by 2 weeks. Pulmonary hypertension (PH) was confirmed by right heart catheterization with mean pulmonary artery pressure (mPAP) ≥20mmHg [17].

Muscle involvement was defined by creatine kinase (CK) levels ≥2 times the upper limit of normal and/or active inflammatory lesions on magnetic resonance imaging (muscle edema and/or muscle contrast enhancement), myopathic electromyography findings, or biopsy-confirmed myositis. Severe muscle involvement was defined as CK ≥10 times the upper limit of normal, severe proximal muscle weakness (Medical Research Council score ≤3/5), or severe dysphagia [18]. Skin involvement included mechanic's hands or typical dermatomyositis lesions. Microvascular involvement included Raynaud phenomenon, distal ischemic lesions, or abnormal nailfold capillaroscopy findings. Cancer-associated myositis was defined as cancer occurring within 3 years before or after ASyS onset.

Data were retrospectively collected from ILD diagnosis to the last follow-up visit at the time of study inclusion. Clinical status, pulmonary function tests (PFTs), HRCT findings, treatments, extrarespiratory manifestations, and ILD outcomes (relapse, PH, chronic respiratory failure, and survival) were recorded longitudinally. ILD pattern and extent were based on radiologist reports when available.

Latent class mixed models (LCMMs) were used to identify patient subgroups according to longitudinal FVC and DLCO trajectories. LCMMs assume a finite number of latent classes with distinct average trajectories. The optimal number of classes was selected according to Akaike and Bayesian information criteria, posterior class-membership probabilities (>0.7–0.8), and clinical interpretability [19,20]. Baseline FVC at ILD diagnosis defined time 0. To ensure stable and interpretable trajectory estimation, follow-up was censored after visit 8 because beyond this time point more than 50% of observations were missing. LCMM-defined trajectory classes were subsequently used as exposure groups in survival analyses evaluating chronic respiratory failure using Kaplan–Meier estimates, log-rank tests, and Cox proportional hazards models. Quantitative variables are reported as medians (IQRs), and categorical variables as number (%). Groups were compared using the t test or Mann–Whitney U test for continuous variables and the chi-square or Fisher's exact test for categorical variables. To assess robustness of trajectory classification, a sensitivity analysis restricted LCMMs to patients with at least three FVC measurements (i.e., baseline plus at least two follow-up assessments).

Linear mixed-effects models with Gaussian errors were used to identify factors associated with ≥5% absolute change in predicted FVC over time in patients with at least three FVC measurements. Covariates were selected a priori according to clinical expertise. Time, candidate predictors, and time-by-covariate interactions were first assessed in univariable analyses. Variables with P<.10 for either the main effect or interaction term were retained in multivariable models. Main effects represent baseline differences between groups (or per-unit change for continuous covariates), whereas time-by-covariate interactions describe how the rate of FVC change over time varies according to each variable. Random intercepts and, when supported, random slopes with unstructured covariance matrices accounted for within-patient correlation. Competing models were compared using information criteria and residual diagnostics. A Gaussian mixed-effects model provided the best balance between goodness-of-fit and parsimony and was retained for inference (Appendices 1–4). Similar models were used for DLCO trajectories, defining significant change as ≥10% absolute variation over 5 years.

All tests were 2-sided with α=0.05. Analyses were performed using R version 4.3.0.

Among 208 patients with ASyS-ILD screened, 92 were included (Fig. 1). Most participants were women (72%), with a median age at ASyS diagnosis of 54 years (IQR, 46–65 years) (Table S1). Anti-Jo-1, anti-PL-7, anti-PL-12, and anti-EJ antibodies were present in 61%, 15%, 14%, and 10% of patients, respectively; no anti-OJ antibodies were identified.

Fig. 1.

Flowchart of the study. ACR, American College of Rheumatology; anti-ARS antibodies, anti-aminoacyl-transfer RNA synthetase antibodies; ASyS, antisynthetase syndrome; ASyS-ILD, antisynthetase syndrome-associated interstitial lung disease; EULAR, European Alliance of Associations for Rheumatology; FVC, forced vital capacity; ILD, interstitial lung disease.

ILD was present at ASyS diagnosis in 91% of patients. Among the 8 patients diagnosed during follow-up [5], were identified within 3 months after ASyS diagnosis. At ILD diagnosis, median FVC was 72% predicted (IQR, 61%–85%) and median DLCO was 50% predicted (IQR, 42%–63%). The most frequent HRCT pattern was nonspecific interstitial pneumonia (NSIP; 79%), and RP-ILD was present in 33% of patients.

Median follow-up duration was 43 months (IQR, 27–70 months), with 13 patients lost to follow-up (Table S2). During follow-up, 12% developed confirmed PH and 23% progressed to chronic respiratory failure. Patients received a median of three treatments (IQR, 2–4), most commonly glucocorticoids (93%), mycophenolate mofetil (48%), rituximab (41%), and methotrexate (36%).

Compared with the 40 excluded patients with missing serial PFT data, included patients had similar demographic, serological, and extrapulmonary characteristics (Table S3). In contrast, excluded patients had more severe respiratory involvement at ILD diagnosis.

Latent class mixed models (LCMMs) with 1–6 classes were compared (Appendix 5), and a 3-class model was selected according to information criteria, class separation, posterior probabilities (74.7%–83.1%; Appendix 6), and clinical interpretability. Class-specific comparison tables are presented in Tables S4–S6.

Class 1 (n=17) demonstrated progressive FVC decline (Fig. 2), with higher rates of ILD relapse, PH (47%), and chronic respiratory failure (53%). Anti-Ro52 antibodies tended to be more common (73%), and patients received a greater number of treatments (Table 1).

Fig. 2.

Three-class latent class mixed model (LCMM) trajectories of FVC over time. (A) Spaghetti plot of all individual trajectories and the average trend estimated using B-splines. (B) FVC trajectories over time according to the 3-class LCMM. The trajectories correspond to LCMM-predicted mean FVC values for each latent class, obtained from the fitted mixture model on a regular time grid. These curves represent model-based class-specific means under the parametric form assumed for the trajectories and therefore appear smoother and may differ slightly from the empirical smoothed curves. (C) Spaghetti plot of the 3-class LCMM with modeled trajectories estimated using B-splines. The trajectories result from spline-based smoothing of observed FVC measurements over time after plotting individual patient curves. These curves reflect the raw longitudinal data, including irregular timing of visits and measurement variability. Time 0 was defined as the date of baseline percentage predicted FVC at ILD diagnosis. %FVC, percentage predicted forced vital capacity; LCMM, latent class mixed model.

Class 2 (n=44) had preserved baseline FVC without clear long-term decline, younger age, and frequent joint involvement (66%) and mechanic's hands (39%). RP-ILD at diagnosis was less common than in Class 3.

Class 3 (n=31) had lower baseline FVC but showed marked improvement over time. RP-ILD at diagnosis was more frequent (45%), and an organizing pneumonia (OP) pattern was more common than in the other classes. No significant differences were observed in extrapulmonary manifestations or anti-ARS antibody profiles across classes.

Five-year chronic respiratory failure-free survival showed a trend toward poorer outcomes in Class 1 (Fig. S1). In Cox proportional hazards models, Classes 2 and 3 tended to have lower risk of chronic respiratory failure compared with Class 1 (reference group).

In the sensitivity analysis restricted to patients with at least three FVC measurements (N=79), the 3-class structure was preserved with similar trajectories (Fig. S2). However, the declining class was no longer observed as a distinct class and appeared to be partially redistributed into the stable group, whereas clinical characteristics remained consistent with the primary analysis (Tables S7 and S8).

DLCO trajectories were also best described by a 3-class model (Appendix 7), with posterior probabilities ranging from 72.4% to 81.0% (Appendix 8).

Among 78 patients, Class 1 (n=28) showed stable DLCO over time and received more immunosuppressive treatments than Class 3 (Fig. S3 and Table S9). Class 2 (n=9) demonstrated DLCO decline, more frequent associated SARDs, a tendency toward higher anti-Ro52 positivity, and increased rates of ILD relapse, chronic respiratory failure, and suspected or confirmed PH compared with Class 3. Class 3 (n=41) showed DLCO improvement and fewer ILD relapses. Class-specific comparison tables are presented in Tables S10–S12.

In univariable analyses, age, cancer-associated myositis, anti-ARS antibodies, RP-ILD at diagnosis, HRCT pattern, and number of treatments were associated with longitudinal FVC change over 5 years (n=74, Table 2).

In multivariable models (n=25), no baseline clinical or serological factor was associated with FVC at inclusion. Significant time-dependent effects were observed. NSIP (β=6.5; 95%CI, 1.8–11; P=.008) and OP patterns (β=7.7; 95%CI, 2.0–13; P=.009) were associated with a more favorable longitudinal FVC trajectory. RP-ILD at diagnosis showed a borderline interaction (β=3.0; 95%CI, −0.06 to 6.0; P=.055).

For DLCO, univariable analyses identified RP-ILD at diagnosis, usual interstitial pneumonia (UIP) pattern, antifibrotic therapy, and number of treatments as associated factors (n=77, Table S13). In multivariable models (n=47), DLCO improved over time overall (β=5.94; 95%CI, 0.8–11; P=.025), whereas UIP pattern was associated with DLCO decline (β=−8.5; 95%CI, −16 to 0.7; P=.032).