This retrospective study involving human participants was reviewed and approved by the ethics committee of West China Hospital and the First Affiliated Hospital of Chongqing Medical University.

A total of 189 anti-MDA5 + DM-ILD patients who underwent CT examination from August 2014 to March 2022 in West China Hospital of Sichuan University were consecutively recruited. In addition, the anti-MDA5 + DM-ILD patients with HRCT from August 2019 to May 2022 in the First Affiliated Hospital of Chongqing Medical University were consecutively collected as the external validation dataset. The diagnosis of DM was made according to the 119th ENMC or 224th ENMC classification criteria, and clinically amyopathic DM was confirmed [19,20,21]. ILD was confirmed by typical radiological features in chest CT [2]. RP-ILD was defined as rapid progression of dyspnea symptoms, rapid worsening of HRCT findings, or decrease in partial pressure of oxygen > 1.33 kPa (10mmHg) within 3 months [11, 22]. Each patient was diagnosed with RP-ILD within three months after CT examination.

Inclusion criteria included adult-onset disease (age > 18 years), a diagnosis of DM, positive anti-MDA5 autoantibody, diagnosed with ILD for the first time on chest CT, without a history of drug-induced interstitial changes, and without a history of lobectomy. Exclusion criteria included inadequate image quality, lack of HRCT scan, and moderare-large pleural effusion. Finally, a total of 160 patients from Institution 1 were retrospectively enrolled and were randomly divided into the training dataset (n = 119) and internal validation dataset (n = 41), while 29 patients from Institution 2 were retrospectively enrolled as external validation dataset. The flowchart of patient enrollment is shown in Fig. 1. The flowchart of research is shown in Fig. S1.

Patient enrollment flowchart. Note: HRCT, high-resolution computed tomography; RP-ILD, rapidly progressive interstitial lung disease; NRP-ILD, non-rapidly progressive interstitial lung disease

HRCT scans were performed in the axial plane with 1-mm-thick sections by multidetector CT scanner including Siemens Somatom Definition (Siemens Healthcare, Erlangen, Germany). Image reconstruction convolution kernels included I70f, B10f, B30f, B80f, and B31f. In all patients, HRCT images were acquired in the supine position and at full inspiration.

All clinical data as well as laboratory samples were collected on the first day of admission and stored in the electronic medical records. Basic demographics, including age, gender, smoking history, and medical history, were assessed. The “infection” recorded in this study refers to patients with a clear diagnosis of “infection” at the time of discharge. The patients may have etiological evidence, or indirect evidence of a diagnosis of “infection”, such as symptoms, signs, and image findings suggestive of infection. Patients diagnosed with cancer on or before March 26, 2023, were documented. The clinical presentation at diagnosis, includes fever, skin changes, arthritis/arthralgia, myalgia, dyspnea, infection, oral pain/ulcers, and acataposis. Fever was defined as an armpit temperature exceeding 37.4°. Laboratory findings at diagnosis, including C-reactive protein, erythrocyte sedimentation rate, rheumatoid factor, anti-CCP antibody, creatine kinase, and blood cell count, were obtained. The neutrophil-to-lymphocyte ratio was calculated by dividing the absolute neutrophil count by the absolute lymphocyte count [11]. Myositis-specific autoantibodies and Myositis-associated autoantibodies assessments were conducted by utilizing immunoblotting technology (YHLO Biotech Co.). Positive findings in patients were validated in duplicate. Time to death was recorded as the period between the time of the HRCT examination to the time of death.

Pneumomediastinum was diagnosed based on CT images. Four HRCT-based scoring systems, namely the Idiopathic pulmonary fibrosis-score, Ground-glass opacity (GGO)-score, Consolidation-score, and Fibrosis-score, were assessed [23, 24]. The overall Idiopathic pulmonary fibrosis-score was calculated by summing the score of six zones (upper, middle, and lower zones on both sides). HRCT findings in each zone were graded 1 (normal attenuation), 2 (GGO without traction bronchiectasis), 3 (consolidation without traction bronchiectasis), 4 (GGO associated with traction bronchiectasis), 5 (consolidation associated with traction bronchiectasis), and 6 (honeycombing). GGO, consolidation, and fibrosis were separately assessed and recorded according to the pulmonary involvement area of the five pulmonary lobes. 0 (no involvement), 1 (≤ 5% involvement), 2 (5–24% involvement), 3 (25–49% involvement), 4 (50–75% involvement), and 5 (> 75% involvement) were recorded for GGO or consolidation at each lobe. And the fibrotic changes in each lobe were classified into 0 (no fibrosis), 1 (interlobular septal thickening without honeycombing), 2 (honeycombing < 25%), 3 (25–49%), 4 (50–75%), and 5 (> 75%) as fibrosis score. The respective total score of each component (GGO-score, Consolidation-score, and fibrosis-score) was the sum of each lobe’s score. Four HRCT scores were assessed independently by chest radiologists 1 and 2 (with 4 and 18 years of experience in chest imaging diagnosis, respectively) in training, internal validation, and external validation datasets. The final scores were averaged by the scores of two radiologists. One month later, 30 patients were randomly selected to be assessed by radiologist 1 to calculate the intra-observer correlation coefficient. The inter-observer correlation coefficient was calculated from the results of the first assessment by the two radiologists.

The segmentation of the three-dimensional region of interest (ROI) was performed using the open-source software 3D Slicer (Version 5.0.2). The bilateral lung regions, including 5 lobes as well as corresponding bronchial and vascular bundles, were first labeled as ROI 2, while the areas only with Hu values from − 950 to -150 were labeled as ROI 1. Furthermore, the subpleural 1 cm areas were annotated as ROI 4, while the rest areas of the lung were marked as ROI 3 (Fig. 2). Three months later, 60 patients were randomly selected to be segmented by the same radiologist to calculate the intra-observer correlation coefficient. The radiologist was aware of the diagnosis of ILD but was blinded to clinical information.

Examples of the 4 types of ROIs. Note: ROI, region of interest; 3D, three dimension; ROI 1, with Hu values from − 950 to -150 regions; ROI 2, the bilateral lung regions; ROI 3, without subpleural 1 cm area; ROI 4, the subpleural 1 cm area

Note: ROC, receiver operating characteristic curve; AUC, area under the curve

The pixel resampling was applied before feature extraction and the CT images were reconstructed to a target voxel of 1 mm×1 mm×1 mm. The pixel values were also converted to HU using the following formula: HU = pixel_value × slope + intercept, where slope = 1, intercept = -1024.

The radiomics features were extracted from the manually labeled ROIs in HRCT images by using the IBSI-compliant Python package named PyRadiomics (version 3.0) with the bin size fixed to 32. Multiple filters including Exponential, Gradient, Logarithm, Log-sigma (1.0 mm, 2.0 mm, 3.0 mm, 4.0 mm, and 5.0 mm), Square, Squareroot, and Wavelet (HHH, HHL, HLH, HLL, LHH, LHL, LLH, and LLL) were applied; and finally, a total of 1729 radiomics features were extracted from each ROI.

Shape features were excluded because the relationship between shape features and diffuse lung diseases is not clear. Histogram and texture features that were robust to variation in contour delineation (Intraclass correlation coefficient > 0.80) and also not highly correlated with each other (Pearson correlation coefficient > 0.95) were retained for subsequent analysis. The Least Absolute Shrinkage and Selection Operator regression analysis was applied to select radiomics features, and 10-fold cross-validation was used to select the appropriate value of the penalty parameter and avoid overfitting.

The z-score normalization was used to standardize the values of the selected histogram and texture features before model development. The formula for z-score standardization is: z-score = (X-mean)/SD, where: X represents the value of the original sample, mean represents the mean value of the original sample values, and SD represents the standard deviation of the original sample values. Four radiomics models were developed based on the selected radiomics features from ROI 1, ROI 2, ROI 3, and ROI 4, respectively. For each radiomics model, the risk-score was calculated based on the selected radiomic features with the support vector machine classifier, and the parameters were as follows: kernel = RBF, tolerance = 0.001, class _weight = balanced. In addition, the optimal c and gamma were determined by cross-validation and grid-search. By using multivariate regression, a nomogram was established by integrating the selected clinico-radiologic variables and the Risk-score of the most discriminative radiomics model in the training dataset.

The prognostic value of the nomogram was evaluated by Kaplan-Meier curves, the Mantel-Haenszel test, and Cox regression. In Institutions 1 and 2, patients were classified into high-risk and low-risk groups at a 50:50 ratio according to the score calculated by the nomogram, respectively.

The discriminative capabilities were evaluated by the receiver operating characteristic analysis with respect to the area under the curve (AUC). Uni- and multivariable logistic regression analyses were used to select clinico-radiologic variables according to the onset of RP-ILD. The goodness-of-fit of each model was calculated via the Hosmer-Lemeshow test and the calibration curve was generated by applying the 1,000 times bootstrapping resampling method. Decision curve analysis was plotted to compare the clinical usefulness of different models.

Statistical analyses were performed on the SPSS (SPSS Institute, Inc., Chicago, IL, USA, version 26.0) and MedCalc (version 20.0) software. The Chi-square test and the analysis of variance (ANOVA) were used to compare qualitative and category characteristics. The AUCs between different models were compared by Delong’s test. The heatmap of the selected radiomics features was generated by using HemI v1.0 software. The calibration analysis and decision curve analysis were performed with R language (version 3.4.4) by using the “RMS” package and the “rmda” package, respectively. Pearson’s correlation coefficient was used to determine correlations between radiomics features, Risk-score, and four HRCT scores. A 2-tailed p-value < 0.05 was considered to be statistically significant.