A Clinical Risk Model to Predict Rapidly Progressive Interstitial Lung Disease Incidence in Dermatomyositis

Background: Rapidly progressive interstitial lung disease (RP-ILD) is a fatal complication of dermatomyositis (DM) and clinically amyopathic DM (CADM). The objective of this study was to evaluate risk markers associated with RP-ILD incidence in patients with DM/CADM and to develop a RP-ILD risk prediction (RRP) model. Methods: The clinical records of 229 patients with DM/CADM from Peking University People's Hospital, and 97 patients from four other independent clinical centers were retrospectively reviewed. Univariate and multivariate logistic regression analyses were performed to identify independent risk factors associated with later RP-ILD incidence to build a risk score model. The concordance index (C-index) and calibration curve were calculated to evaluate the predictive accuracy of the RRP model. Results: A multiparametric RRP model was established based on weighted clinical features, including fever (yes, 5; no, 0), periungual erythema (yes, 6; no, 0), elevated CRP (yes, 5; no, 0), anti-MDA5 antibody (positive, 8; negative, 0), and anti-Ro-52 antibody (positive, 6; negative, 0). Patients were divided into three risk groups according to the RRP total score: low, 0–9; medium, 10–19; high, 20–30. The C-index and calibration curve of the RRP model showed a promising predictive accuracy on the incidence of RP-ILD. Conclusion: The RRP model might promisingly predict the incidence of RP-ILD in DM/CADM patients to guide early individual treatment and further improve the prognosis of DM/CADM patients.


INTRODUCTION
Dermatomyositis is an autoimmune disease characterized by skin and muscle damage caused by muscular involvement and frequent extramuscular symptoms such as Raynaud's phenomenon, arthritis, and interstitial lung disease (ILD) (1). Clinically amyopathic dermatomyositis (CADM) is a combination of hypomyopathic DM (HDM) and amyopathic DM (ADM), with characteristic skin-predominant lesions (2)(3)(4)(5). ILD is one of the most severe complications of DM/CADM. Despite aggressive treatments, respiratory failure following rapidly progressive interstitial lung disease (RP-ILD) remains the main cause of death in more than 50% of DM/CADM patients (6)(7)(8)(9)(10). Consequently, the prediction and timely identification of RP-ILD symptom onset is vital for effective treatment during early disease development stages and might lead to improved prognosis with significantly reduced mortality rates in patients with DM/CADM (11)(12)(13).
A variety of studies have explored baseline parameters associated with RP-ILD in patients with DM/CADM but few practical quantitative methods were established and validated in clinical practice to predict the incidence of RP-ILD. Antimelanoma differentiation-associated gene 5 (MDA5) antibody, lymphocytes in peripheral blood, C-reactive protein (CRP), skin ulceration, and ferritin were reported as predictive factors for disease onset and poor prognosis of RP-ILD (7,8,14,15). However, previous studies were generally supported by limited cohort sizes or scattered case reports. Moreover, it seems reductive and inefficient to use a single clinical factor to predict the risk of an extremely heterogeneous disease as RP-ILD. In contrast, a holistic approach, based on multiple factors comprehensively evaluating personalized clinical characteristics, might provide a better predictive model for RP-ILD. Combining clinical and immunological factors might be valuable to evaluate disease severity, predict outcomes, and guide individualized treatment. Therefore, it is clinically significant to explore the RP-ILD-associated parameters at the onset of DM/CADM to establish a reliable early-stage RP-ILD risk prediction (RRP) model.
In this study, a practical score model to predict the incidence of RP-ILD was established through the identification and quantification of specific clinical parameters associated with later RP-ILD incidence in a patient cohort with DM/CADM. Moreover, this model was validated in different cohorts, aiming to guide personalized treatment during early RP-ILD development stages.  (16,17). The inclusion criteria include all the patients who had a definite DM or CADM diagnosis. DM was diagnosed according to the criteria of 2017 EULAR/ACR (17) and CADM was diagnosed according to criteria proposed by Sontheimer (5). By using incidence data of RP-ILD in DM/CADM of Peking University People's Hospital (33.6%, 77/229) and the other four independent clinical centers (9.3%, 9/97), the sample size was calculated in Power and Sample Size Free Calculators (http:// www.powerandsamplesize.com/). According to the calculation results, the study sample size should include 41 patients in each cohort, with a one-sided α of 5%, and a power of 80%. In this study, the sample size collected was larger than 41. To improve the accuracy of the model, it was ensured that each group of samples was larger than 41. Finally, grouped by the time period, there were 165 cases in the development cohort from 2010 to 2015, 64 cases in the internal validation cohort from 2016 to 2019, and 97 cases in the external validation cohort from 2010 to 2019. This study was approved by the ethics committee of Peking University People's Hospital according to the declaration of Helsinki. The waiver of consent was agreed upon by the institutional ethics committee due to the retrospective nature of the study.

Patients
Patient exclusion criteria included recent acute infection, pulmonary infarction, presence of heart failure, history of neoplasm, other connective tissue diseases concomitantly, or insufficient demographic, clinical, and laboratory test data. ILD was diagnosed by the findings of high-resolution computed tomography (HRCT), according to the International Consensus Statement of Idiopathic Pulmonary Fibrosis of the American Thoracic Society (18) and defined as previously described (19). Chest HRCT patterns were classified into non-specific interstitial pneumonia (NSIP), organizing pneumonia (OP), NSIP combined with OP pattern. Pulmonary function tests (PFT) and Bronchoalveolar lavage (BAL) examinations were performed to evaluate ILD% predicted forced vital capacity (FVC), percent predicted diffusing lung capacity for carbon monoxide (DLco), and total lung capacity (TLC) at the initial diagnosis. Based on the radiological assessment of the chest HRCT results, RP-ILD was defined as a progressive deterioration of ILD in 3 months combining with rapidly progressive severe dyspnea and hypoxemia, requiring oxygen therapy or ventilator care (20,21).

Statistical Analyses
Demographic and clinical characteristics of the development cohort and validation cohort were compared. Categoric variables were reported as counts (%) and compared using the χ 2 test. Continuous variables were presented as the mean ± SD and compared by the ANOVA or the Kruskal-Wallis test. The least significant difference (LSD) of ANOVA and the chi-square test were used in a pair-wise post-hoc analysis. Logistic regression analysis and forward elimination process selection were used to explore independent risk factors with multivariate analyses. Results of the regression models were shown as the odds ratio (OR) with 95% confidence interval (95% CI). The performance of the RRP model was measured by the Harrel concordance index (C-index) and calibration curve in R version 3.6.1 (http://www. r-project.org/). The degree of agreement between the predicted probabilities with the actual outcomes numerically was measured by calibration curve, with a larger C-index value denoting better predictive accuracy. A value of p < 0.05 was considered statistically significant. Demographic, clinical, and laboratory characteristics of the development cohort and the internal validation cohort, the external validation cohort are summarized in Supplementary Table 1. The demographic, clinical, and laboratory manifestations of the three cohorts were compared with pair-wise post-hoc analyses in Supplementary Table 2. All patients underwent CT scans and none was diagnosed as RP-ILD at the baseline. At the outset, 84.8% (140/165) of the patients in the development cohort showed ILD onset on CT, but they were unable to be diagnosed as RP-ILD at that time. In the clinical practice, RP-ILD was diagnosed later, with the occurrence and development of the disease during 2 to 4 weeks of hospitalization, which was defined as a progressive deterioration of ILD in 3 months combining with rapidly progressive severe dyspnea and hypoxemia, requiring oxygen therapy or ventilator care (20,21). Among 326 patients with DM/CADM, RP-ILD was found in 35.8% (59/165) of the developed cohort, 28.1% (18/64) of the internal validation cohort, 9.3% (9/97) of the external validation cohort, respectively. The incidence of RP-ILD did not differ among the three groups (p = 0.273, Supplementary Table 1). Based on the results of Supplementary Tables 1 and 2, it could be known that there was no difference between developed cohort and internal validation cohort in demographic, clinical, and laboratory factors, which was not like in pairs of internal validation cohort vs. external validation cohort and developed cohort vs. external validation cohort. Different features including Gottron's sign/papules, heliotrope rash, V sign, shawl sign, skin ulceration, fever, arthralgia, anti-aminoacyl-tRNA synthetase (ARS) antibodies, ANA, FVC% predicted, TLC% predicted, AST, ESR, and CRP, indicating that characteristics of patients were various in different hospitals. Detailed results were shown in Supplementary Tables 1, 2.

Calculating the RRP Score
Univariate and multivariable logistic regression analyses were used to explore independent risk factors for RP-ILD incidence in DM/CADM patients. To accurately screen for risk factors, factors with p < 0.05 in the univariate analysis were included in multivariate analysis as covariates. Analyzed factors included mechanic's hands (yes vs. no), skin ulceration (yes vs. no), periungual erythema (yes vs. no), fever (yes vs. no), anti-MDA5 antibody (positive vs. negative), anti-Ro-52 antibody (positive vs. negative), and elevated CRP (yes vs. no) ( Table 1) Table 1].
The above five independent risk factors were then combined to establish an RRP model. Scores were based on five independent risk predictors weighted by regression coefficients and rounded to the nearest whole number ( Table 2). Total risk scores were calculated by adding the weighted values of all prognostic variables. According to the RRP scores, patients were categorized into three risk groups: low-risk group, with RRP scores ranging from 0 to 9; medium-risk group, with RRP scores ranging from

Validation of the RRP Score System
After classifying the development cohort according to the RRP scores, 40.6% of the patients were assigned to the low-risk group, 49.70% to the medium-risk group, and 9.69% to the high-risk group. The results of the internal validation cohort were similar: 50% of the patients were in the low-risk group, 40.63% in the medium-risk group, and 9.37% in the highrisk group. In the external validation cohort, 56.70% of the patients were assigned to the low-risk group, 37.11% to the medium-risk group, and 6.19% to the high-risk group ( Table 3).
The C-index for differences in RP-ILD incidence among risk groups from the development cohort was 0.849, with 95% CI 0.791 to 0.907. The mean predicted RP-ILD rates of the lowrisk, medium-risk and high-risk groups were 8.61%, 46.41%, and 94.84%, respectively, while the observed rates were 8.96%, 45.12%, and 100%, respectively (Table 3, Figure 1A). Similarly, the C-index for differences in RP-ILD incidence rates among risk groups from the internal validation cohort was 0.928, with 95% CI 0.866 to 0.990 and the mean predicted RP-ILD rates of the low-risk, medium-risk, and high-risk groups were 2.7%, 42.87%, and 99.82%, with observed rates were 3.13%, 42.31%, and 100%, respectively (Table 3, Figure 1B). Finally, the C-index for differences in RP-ILD incidence rates among different risk groups from the external validation cohort was 0.948, with 95% CI 0.862 to 1.000 with mean predicted RP-ILD rates of the low-risk, medium-risk, and high-risk groups at 0.38%, 25.07%, and 96.14%, and observed rates were 1.82%, 22.22%, and 100%, respectively (Table 3, Figure 1C). Moreover, the calibration plot based on bootstrap resampling validation demonstrated promising agreement between predicted and observed RP-ILD rates in the development cohort (Figure 2A), the internal validation cohort (Figure 2B), and the external validation cohort (Figure 2C).

DISCUSSION
Lung involvement is the most common reason for respiratory failure and death in DM/CADM. RP-ILD in DM patients is usually fatal within a few weeks or months (22). Therefore, predictive parameters for the onset of RP-ILD are important in the early treatment of DM/CADM patients. Previous studies mainly focused on short-term or small sample size studies to identify individual risk parameters of RP-ILD instead of exploring the multifactorial approach englobing the extensive clinical data from DM/CADM patients. Furthermore, none of the previous studies proposed a quantitative model to predict the risk of RP-ILD incidence in clinical practice. RP-ILD is an extremely heterogeneous disease. It is then likely, that multiple prognostic factors combined will better evaluate disease severity, predict outcomes, and guide individualized treatment than single-factor based models. Therefore, it was hypothesized that RP-ILD related risk parameters of DM/CADM could establish a dependable RRP model. In this study, an RRP model was built by combining independent predictive factors, including fever, periungual erythema, elevated CRP, anti-MDA5 antibody, and anti-Ro52 antibody. The early identification of RP-ILD progression is crucial for effective therapy, prognosis, and reduced mortality. In recent knowledge, this is the first score system available for clinical practice, which is able to predict the risk of RP-ILD incidence in DM/CADM patients during the early disease stages. Moreover, an independent external validation further confirmed the reliability of this model, with high accuracy in patients with a high risk of RP-ILD incidence. Predictors of poor outcome during DM-ILD have been previously reported (23)(24)(25), however, fewer studies assessed the prognostic factors for myositis-associated RP-ILD. This study confirmed many single clinical and laboratory prognostic factors previously reported, such as periungual erythema, fever, and CRP, as risk factors for RP-ILD in myositis (8,15,(26)(27)(28). However, the majority of predictive factors for ILD (8,15,26,27,29), e.g., skin ulceration and hands of a mechanic, which have been previously identified by univariate analysis, were not classified as independent risk factors by multivariate analysis in the study.
A dose-dependent relationship between anti-MDA5 antibody and RP-ILD has been recently reported, unveiling anti-MDA5 antibodies as a significant risk factor for poor prognosis (30). Similarly, patients with anti-Ro-52 antibodies during juvenile myositis are more likely to develop severe ILD with a poor prognosis (31). Consistent with these reports, this study also demonstrated that anti-MDA5 antibody and anti-Ro-52 antibody were both specific biomarkers for DM/CADM-associated RP-ILD (8,32).
Based on the results in Supplementary Tables 1 and 2, it was found that the characteristics of patients in different hospitals were not the same. Therefore, there was a need to establish an evaluation system with a universal applicability in different hospitals. In the RRP score model of the study, patients with a score between 0 and 9 are considered as low-risk subgroups for the development of RP-ILD, while scores between 10 and 19 and 20 or more are at medium-risk and high-risk of RP-ILD development, respectively. The results of summarized calibration and discrimination for the RRP model in three cohorts all presented a high predictive accuracy for estimating the risk of RP-ILD supported in clinical settings. Therefore, researchers of the study thought that this model has general applicability, which still needs to be proved in the future.
This study presented several limitations. First, this study was a retrospective study, suboptimal compared to the prospective study and included hospitalized patients only. Therefore, it might have shown biases in patient selection. Second, the prediction model was only tested in a Chinese population. The generalizability of this model to other ethnicities needs to be validated in future studies. Third, the validation cohort has a limited sample size because of the rarity of the disease; this model needs to be further validated by more multi-center studies with enlarged patient cohorts. Fourth, this risk score was mainly based on clinical characteristics. Although several studies have reported that some serum biomarkers, including ferritin, KL-6, and IL-18, were related to RP-ILD and might be used as potential biomarkers for predicting disease severity and prognosis (10,33); the researchers of the study were unable to include these markers in the model due to the lack of these data in the retrospective study, whether those markers could be used as a reliable alternative biomarker needs further validation.

CONCLUSIONS
The first RRP score model capable of predicting the risk of RP-ILD incidence during early disease stages was provided. It is expected to be a valuable clinical tool that could guide early personalized treatment, improve prognosis, and reduce morality for DM/CADM patients.

DATA AVAILABILITY STATEMENT
The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

ETHICS STATEMENT
This study was approved by the Ethic Committee of Peking University People's Hospital according to the declaration of Helsinki. Waiver of consent was agreed upon by the institutional ethics committee due to the retrospective nature of the study.