Nomogram models integrating multiregion radiomics, quantitative computed tomography, and clinical indicators for distinguishing connective tissue disease-associated interstitial lung disease from idiopathic pulmonary fibrosis and improving the prediction of gender-age-physiology stage Authors: Xinyao Jiao, Lihua Gao, Yuanbo Huang, Ningxia Jia, Jiapeng Li, Han Song, Sa Huang Published: 2026 ABSTRACT BackgroundAccurate diagnosis of idiopathic pulmonary fibrosis (IPF) and connective tissue disease-associated interstitial lung disease (CTD-ILD), along with assessment of the interstitial lung disease gender-age-physiology (ILD-GAP) index, is crucial for formulating effective therapeutic interventions. The aim of this study was to examine whether multiregion radiomics combined with clinical indicators and quantitative computed tomography (QCT) can enhance the diagnostic accuracy for IPF and CTD-ILD. We also assessed the potential of radiomics as a noninvasive alternative to pulmonary function testing (PFT) for evaluating ILD-GAP staging, a prognostic marker.MethodsThis retrospective study included patients diagnosed with IPF and CTD-ILD at two centers. Honeycombing (H), reticulation (R), and ground-glass opacity (GGO) were outlined in representative regions, and radiomics features were then extracted. Multidimensional data included multiregional radiomics, QCT (lesion volume), and clinical indicators. The most appropriate machine learning algorithms were selected to develop models. Ultimately, the logistic regression algorithm was employed to establish a nomogram model for distinguishing IPF from CTD-ILD and a nomogram gender-age-physiology (nomogram-GAP) model for predicting ILD-GAP staging. Model performance was evaluated via the area under the receiver operating characteristic curve (AUC), calibration, and decision curve analysis (DCA).ResultsTo differentiate IPF from CTD-ILD, we enrolled patients from Center A into training (n=153), test (n=64), and internal validation (n=42) sets, as well as 40 patients from Center B into an external validation set. The nomogram model yielded an AUC of 0.854 in the internal validation set and 0.823 in the external validation set. For ILD-GAP staging prediction, Center A patients were split into training (n=155) and internal validation sets (n=31). The nomogram-GAP model had an AUC of 0.881 in the internal validation set. Calibration and DCA demonstrated that the nomogram model and nomogram-GAP model exhibit strong consistency and clinical utility.ConclusionsTwo types of nomograms integrating multiregion radiomics, QCT, and clinical indicators demonstrated promising performance in distinguishing IPF from CTD-ILD and predicting ILD-GAP staging. When PFTs are unreliable or impracticable, the nomogram-GAP model may serve as a reliable marker for assessing ILD-GAP staging. Introduction Interstitial lung disease (ILD) is a general term for a group of heterogeneous diseases featuring inflammatory or fibrotic lung tissue damage. ILD includes multiple subtypes, with idiopathic pulmonary fibrosis (IPF) and connective tissue disease-associated interstitial lung disease (CTD-ILD) being the most prevalent (1-3). The diagnosis and prognostic assessment of CTD-ILD and IPF remain challenging due to high misdiagnosis rates and the limited applicability of existing tools. Currently, multidisciplinary discussion (MDD) is used for diagnosing CTD-ILD and IPF (2). This process involves the integration of comprehensive patient data, including medical history, clinical signs, specific serum antibody levels, imaging phenotypes from high-resolution computed tomography (HRCT), and pathological findings. However, 18.7% of patients with CTD-ILD consistently test negative for autoantibodies (4), indicating serum antibody testing alone cannot differentiate CTD-ILD and IPF. Additionally, CTD-ILD and IPF exhibit similar imaging characteristics (5). A previous study reported limited agreement among chest radiologists in identifying ILD patterns on HRCT images (6). The inconsistency in imaging-based phenotype recognition and negative antibody results hinder definitive diagnosis, thereby affecting the application of antifibrotic and immunomodulatory therapies. For prognostic assessment, the ILD gender-age-physiology (ILD-GAP) index proposed by Ryerson et al. predicts 1- to 3-year mortality rates for patients with chronic ILD (7). This system incorporates ILD subtype, gender, age, and two pulmonary function indicators. However, its efficacy is limited in patients with severe disease, as they are unable to complete pulmonary function tests (PFTs). These challenges highlight the need to develop more accurate methods to distinguish CTD-ILD from IPF and to identify surrogate markers of PFT to assess ILD-GAP staging. HRCT has the ability to characterize the spatial heterogeneity of ILD-affected lungs, specifically, the diversity of abnormal radiographic features and their distribution patterns, which can offer diagnostic insights for disease identification. For instance, IPF often manifests as a honeycomb patterns in the lungs, predominantly distributed in the lower pulmonary lobes (8). Meanwhile, the extent of abnormal imaging findings may indicate disease progression in patients. For instance, a greater extent of honeycombing (H) and reticulation (R) suggests a poorer prognosis (9,10). However, reliance solely on qualitative or quantitative analysis (e.g., lesion volume) fails to provide a comprehensive depiction of the spatial heterogeneity within ILD-affected lungs. Consequently, achieving precise diagnosis and prognostic assessment remains challenging. Research has indicated that the consistency of qualitative analysis in disease diagnosis is only 0.56 (6). Additionally, Oh et al. reported that the area under the receiver operating characteristic curve (AUC) values for quantitative lesion extent parameters in predicting the 5-year mortality for patients with ILD ranged from 0.648 to 0.744 (11). Consequently, novel techniques are required for the quantitative analysis of intrapulmonary spatial heterogeneity, as this can enhance both disease diagnosis and the prediction of ILD-GAP stage. Radiomics-based analysis of HRCT images has opened new avenues for the objective analysis of ILD (12-15). Radiomics, beyond providing simple features such as volume, allows for the extraction of high-throughput features from radiological images, including microscopic textures and morphology (16,17). Compared to traditional quantitative methods for assessing lesion extent, multiregion radiomics enables detailed characterization of spatial heterogeneity in ILD by extracting high-dimensional computed tomography (CT) features from various lesions. Consequently, multiregion radiomics has the potential to more accurately distinguish CTD-ILD from IPF and improve the prediction of ILD-GAP stage. To our knowledge, no study has comprehensively documented the value of extracting radiomics features from distinct regions and lesions for distinguishing CTD-ILD from IPF and predicting ILD-GAP stage. In our study, we realized multiregion analysis by delineating three typical lesions on representative slices of HRCT: H, ground-glass opacity (GGO), and R. Additionally, quantitative computed tomography (QCT) metrics, comprehensive metabolic panel (CMP), and complete blood count (CBC) data were collected to enhance the diagnostic accuracy for CTD-ILD and IPF and improve the efficacy of prognostic evaluation (12,18-20). These metrics provide supplementary information distinct from multiregion radiomics. Consequently, we hypothesized that multiregion radiomics combined with QCT and clinical indicators can differentiate CTD-ILD from IPF and predict ILD-GAP stage. To address the above-mentioned issues, we developed predictive models using machine learning approaches and sought to achieve the following objectives: (I) to differentiate between IPF and CTD-ILD using multimodal data (radiomics features, clinical indicators, and QCT); (II) to predict ILD-GAP staging for IPF and CTD-ILD using multimodal data (radiomics features, clinical indicators, QCT, and qualitative index); (III) to visualize predictive models using nomograms, thereby equipping clinicians with accessible tools for diagnosis and ILD-GAP stage prediction. We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0903/rc). Methods Participants To differentiate between CTD-ILD and IPF, this retrospective study enrolled 299 patients with IPF or CTD-ILD as confirmed through MDD from The Second Hospital of Jilin University (Center A) and the China-Japan Friendship Hospital (Center B). Our study was conducted at these two medical centers because they are tertiary referral centers for ILDs in China. The core members of the multidisciplinary team (MDT) comprised senior respiratory physicians with expertise in ILD, rheumatologists specializing in connective tissue diseases, thoracic radiology experts, and pathologists specializing in ILD. The MDT established a diagnosis by synthesizing data from clinical assessments, serological tests, HRCT, and histopathological biopsies. Among the cases, 259 were diagnosed at Center A between January 2019 and November 2025, while 40 cases were randomly selected from patients diagnosed at Center B. The patients with IPF and CTD-ILD included in this study received definitive diagnoses from an ILD-specialized MDT team who followed the consensus guidelines (8,21-25). All enrolled patients attained a diagnostic confidence level of at least 70% via a multidisciplinary diagnostic approach (26). The comprehensive diagnostic procedures for IPF and CTD-ILD are detailed in Appendix 1. The inclusion criteria were (I) a clinical diagnosis of CTD-ILD or IPF and (II) completion of HRCT examination in the supine position. Meanwhile, the exclusion criteria were the following: (I) pulmonary edema, infection, tumors, drug poisoning, heart disease, or occupational ILD; (II) pregnancy; and (III) substandard quality of HRCT images. Patients from Center A treated between January 2019 and March 2025 were retrospectively collected and then divided into a training set (n=153) and a test set (n=64) at a 7:3 ratio via a random sampling method. Patients from Center A treated between April 2025 and November 2025 were retrospectively collected as the internal validation set (n=42). The external validation set (n=40) consisted of patients from Center B treated between November 2016 and May 2023 (Figure 1A). Flowchart of the study datasets. (A) Training, test, and validation sets for distinguishing between IPF and CTD-ILD. (B) Training and internal validation sets for predicting ILD-GAP stage. Center A: The Second Hospital of Jilin University; Center B, China-Japan Friendship Hospital. CTD-ILD, connective tissue disease-associated interstitial lung disease; GAP, gender-age-physiology; HRCT, high-resolution computed tomography; ILD, interstitial lung disease; IPF, idiopathic pulmonary fibrosis; PFT, pulmonary function testing. In the ILD‑GAP staging prediction task, patients from Center A treated between January 2019 and March 2025 were retrospectively collected as the training set (n=155), while patients from Center A treated between April 2025 and November 2025 were retrospectively collected as the internal validation set (n=31) (Figure 1B). Exclusion criteria were incomplete PFT data or a time interval between PFTs and HRCT examination exceeding 2 months. ILD-GAP stage was assessed according ILD subtype, gender, age, and two PFT outcomes, namely forced vital capacity (FVC) and diffusing capacity of the lung for carbon monoxide (DLCO) (7). Patients were classified into four stages via the ILD-GAP staging system, with each stage corresponding to distinct mortality rates over 1 to 3 years. In total, 73 patients were classified as ILD-GAP stage I, 55 as ILD-GAP stage II, 41 as ILD-GAP stage III, and 17 as ILD-GAP stage IV. Due to the limited number of participants with ILD-GAP stage III and IV, these stages were combined into the ILD-GAP stage III/IV composite stage. Similarly, stages I and II were combined into the ILD-GAP stage I/II composite stage. Patients with ILD-GAP stage I/II generally have a low 1-year mortality rate, which may delay referral for lung transplant assessment and necessitate careful clinical monitoring alongside active treatment. Conversely, patients with ILD-GAP stage III/IV have a considerably elevated risk of mortality, and lung transplantation should be considered when appropriate (7). Figure 2 depicts the study workflow. Schematic workflow of the study. Center A: The Second Hospital of Jilin University; Center B, China-Japan Friendship Hospital. CBC, complete blood count; CMP, comprehensive metabolic panel; CTD-ILD, connective tissue disease-associated interstitial lung disease; DCA, decision curve analysis; GAP, gender-age-physiology; GGO, ground-glass opacity; HRCT, high-resolution computed tomography; ILD-GAP, interstitial lung disease gender-age-physiology; IPF, idiopathic pulmonary fibrosis; PFT, pulmonary function test; RF, random forest; ROC, receiver operating characteristic; ROI, region of interest; SHAP, Shapley additive explanations; SVM, support vector machine; XGBoost, extreme gradient boosting. This two-center retrospective study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments and was approved by the Ethics Committees of The Second Hospital of Jilin University [Institutional Review Board (IRB) No. 2024-385] and China-Japan Friendship Hospital (IRB No. 2022-KY-031). The requirement for informed consent from study participants was waived, as this study was based on a retrospective review of information that had been previously collected. Clinical data acquisition The clinical indicators were extracted from medical records and included variables such as age, gender, smoking history, PFT, CBC, and CMP. The reason these indicators were selected was because previous research has demonstrated that these markers enhance the accuracy of both diagnosis and prognostic assessment (12,18-20). For data with under 35% missing values, an imputation method based on random forest (missForest) was employed (27). HRCT image acquisition and QCT measurements The HRCT images were stored in Digital Imaging and Communications in Medicine format. All patients underwent HRCT scans in the supine position via multidetector CT systems, specifically the Revolution CT (GE HealthCare, Chicago, IL, USA), NeuViz Prime (Neusoft Corp, Shenyang, China), Brilliance iCT 256 (Philips, Amsterdam, the Netherlands), and SOMATOM Force (Siemens Healthineers, Erlangen, Germany) devices. For each scanner, the patients were positioned supine and instructed to hold their breath at the end of inspiration. The scanning protocol extended from the lung apex to the lung base to ensure comprehensive imaging of all lung tissue. The scanning parameters were as follows: a tube voltage range of 120–140 kV; a reference tube current of 200–280 mA, real-time dynamic exposure dose adjustment (Care-Dose 4D), a slice thickness from 0.625 to 1 mm, and a pitch factor of 0.575–1. The convolution kernels for different devices were as follows: Revolution CT, D kernel; NeuViz Prime, Lung30; Brilliance iCT 256, HDLung; and SOMATOM Force, B157. Scanning procedures and protocols are detailed in Appendix 2 and Table S1. Next, all patient imaging data were imported into ITK-SNAP version 3.8.0 software. The regions of interest (ROIs) were delineated independently by two radiologists who were blinded to the patients’ final diagnoses. It was feasible to divide the entire lung into upper, middle, and lower regions and assess the extent of disease via spaced CT slices (28,29). The labeling slicers included the apex pulmonis, aortic arch, trachea carina, right pulmonary artery, right inferior pulmonary vein, and basis pulmonis slices. Six labeling slices extended longitudinally from the apex to the base of the lung, facilitating the acquisition of lesion data from the upper, middle, and lower lung regions. The cases examined in this study included various imaging patterns, usual interstitial pneumonia (UIP), nonspecific interstitial pneumonia (NSIP), and organizing pneumonia (OP), each characterized by distinct interstitial lesion distribution. The selection of multiple slices enabled the capture of spatial heterogeneity. Furthermore, in the evaluation of ILD, radiologists generally conduct systematic examinations of critical anatomical regions, including the lung apex, hilum, and base. Consequently, the selection of six representative CT slices was more consistent with the diagnostic workflow employed by radiologists. Therefore, this study selected CT slices from six typical anatomical locations to effectively represent the spatial heterogeneity of the lesion distribution across the lungs. GGO, H, and R, typical radiographic signs of ILD, exhibited distinct patterns of pulmonary distribution in IPF and CTD-ILD. Consequently, these features served as valuable tools for differential diagnosis and might also assist in ILD-GAP prediction (1,9). The labeling ROIs were GGO, H, and R. All imaging sign identifications were based on the CT terminology definitions established by the Fleischner Society (30). The details of lesion annotations for GGO, H, and R are detailed in Figure S1. Thirty patients were randomly selected from the training cohort to assess the intraclass correlation coefficients (ICCs) for these two radiologists. The ICCs for each lesion exceeded 0.85 (Table S2). The calculation of QCT metrics and qualitative indicator are detailed in Appendix 3. Radiomics feature extraction and selection We extracted features from ROIs in each region using PyRadiomics (31). To homogenize discrepancies across the scanners, we conducted image preprocessing prior to feature extraction. The procedures were as follows: (I) resampling the images to isotropic 1 mm × 1 mm × 1 mm voxels via B-spline interpolation as implemented with SimpleITK; (II) converting raw CT pixel values to Hounsfield units (HU) by applying a voxel array shift with an intercept offset of −1,024; (III) performing intensity normalization by initially applying Z-score standardization, followed by scaling of the gray values to a standard deviation of 100 to ensure a consistent gray-level dynamic range across samples; and (IV) discretizing gray levels with a fixed bin width of 25. All available first-order, geometric, and texture features were extracted. Eventually, 1,427 radiomics features were extracted in every subregion. Prior to data analysis, all extracted features underwent z-score normalization to ensure consistent parameter magnitudes. For the tasks of distinguishing CTD-ILD from IPF and predicting ILD-GAP stage, the feature selection process was carried out with the following in consecutive fashion: the Welch t-test, Pearson correlation analysis, the minimum redundancy maximum relevance algorithm, and least absolute shrinkage and selection operator (LASSO) regression. The process of selecting radiomics features and clinical and QCT predictors is detailed in Appendix 4. Model construction and evaluation The radiomics, QCT, and clinical features derived from the aforementioned feature selection methods were employed to establish machine learning models. To select the optimal predictive model, we employed six machine learning algorithms: support vector machine, decision tree, random forest, extreme gradient boosting, logistic regression, and light gradient boosting machine. To differentiate CTD-ILD from IPF, we developed three models: (I) a radiomics models consisting of radiomics features; (II) a combined model integrating radiomics with QCT and clinical indicators; and (III) a nomogram model combining optimal radiomics scores with key clinical indicators and QCT. Considering the diverse imaging patterns associated with CTD-ILD, we performed subgroup analyses to assess the model’s ability to discriminate among these patterns. Specifically, we compared (I) CTD-ILD patients with a radiological UIP pattern to IPF patients with UIP and (II) CTD-ILD patients with other radiological patterns (NSIP, OP, NSIP with OP overlap, and unclassified ILD) to IPF patients with other radiological patterns (probable UIP and indeterminate UIP). Due to the limited number of individual imaging patterns and the similarity imaging characteristics, we integrated the remaining imaging patterns, excluding UIP, for analysis. To predict the ILD-GAP stage for patients with CTD-ILD and IPF, we developed three additional models: (I) a radiomics-GAP model consisting of radiomics features; (II) a combined-GAP model integrating radiomics features with QCT parameters, qualitative indicator and clinical indicators, and (III) a nomogram-GAP model combining optimal radiomics features with key clinical indicators, QCT and qualitative indicator. The efficacy of each model was assessed via receiver operating characteristic (ROC) curve analysis, and the performance of each model was assessed according to the AUC and other metrics. Decision curve analysis (DCA) was also applied to provide more robust evidence of clinical benefit, while calibration curves were used to evaluate the fit of the model. Statistical analysis Statistical analysis was performed using with R software version 4.4.1 (The R Foundation for Statistical Computing, Vienna, Austria). Categorical data are expressed as numbers and percentages, while continuous data are expressed as the median and interquartile range or as the mean ± standard deviation. To compare variables across the training, test, and validation sets, the Kruskal-Wallis test, χ2 test, and Fisher exact test were employed. The DeLong test was used to compare the AUCs of the various models. Shapley additive explanations (SHAP) analysis was performed to quantify the influence of features on model predictions, with higher SHAP values indicating a greater contribution. All statistical tests were two-sided, and P values below 0.05 were considered statistically significant. Results Patient characteristics Initially, data from 540 patients with CTD-ILD and IPF were obtained from two centers. Following the exclusion of 241 patients, 299 patients were included in the study (Figure 1). To differentiate between CTD-ILD and IPF, the 299 enrolled patients from Center A and Center B were divided into a training set (84 cases with CTD-ILD and 69 cases with IPF), a test set (35 cases with CTD-ILD and 29 cases with IPF), an internal validation set (24 cases with CTD-ILD and 18 cases with IPF), and an external validation set (20 cases with CTD-ILD and 20 cases with IPF) (Table 1). Multivariable logistic regression analysis identified age, gender, smoking history, red blood cell count (RBC), lung volume (%), the proportion of GGO volume in the right inferior pulmonary vein slice, the proportion of R volume in the pulmonary artery slice, and the proportion of H volume in the aortic arch slice as independent predictors for differentiating between CTD-ILD and IPF (Tables S3,S4). These predictors were subsequently integrated into the combined model. Table 1Clinical characteristics of patients in the four datasets used for distinguishing between IPF and CTD-ILDCharacteristicTraining setTest setInternal validation setExternal validation setP valueCTD-ILD (n=84)IPF (n=69)CTD-ILD (n=35)IPF (n=29)CTD-ILD (n=24)IPF (n=18)CTD-ILD (n=20)IPF (n=20)Age (years)61.56±10.9366.52±7.4660.60±9.4867.45±6.6660.71±11.2669.67±6.9556.95±13.5462.00±4.930.075Gender0.840 Male26 (31.0)55 (79.7)10 (28.6)21 (72.4)7 (29.2)16 (88.9)10 (50.0)16 (80.0) Female58 (69.0)14 (20.3)25 (71.4)8 (27.6)17 (70.8)2 (11.1)10 (50.0)4 (20.0)BMI (kg/m2)24.11±3.6224.72±2.5523.60±3.5224.81±3.6823.19 (21.22, 26.17)23.13 (21.83, 25.34)25.79±3.8925.41±3.180.075Smoking history1 Former or current17 (20.2)40 (58.0)9 (25.7)15 (51.7)4 (16.7)9 (50.0)5 (25.0)13 (65.0) Never67 (79.8)29 (42.0)26 (74.3)14 (48.3)20 (83.3)9 (50.0)15 (75.0)7 (35.0)Pulmonary function DLCO (% predicted)46.96±18.6549.04±22.6439.80 (31.95, 48.08)51.00 (38.70, 64.05)49.55±14.0941.27±17.0473.00 (59.88, 79.88)54.80 (45.05, 73.67)0.003† VC (% predicted)74.81±18.6678.71±19.0781.13±17.6378.90±16.0378.60±11.8081.43±15.7491.40 (80.83, 99.90)82.60 (68.20, 84.40)0.253 TLC (% predicted)69.20 (57.80, 80.00)66.35 (56.38, 80.38)66.90 (60.85, 79.65)73.90 (62.32, 88.45)71.54±11.4565.42±11.4380.15±19.2270.03±16.470.253 FEV1 (% predicted)77.80 (62.45, 86.00)85.10 (71.05, 93.70)78.20 (69.40, 86.45)82.45 (73.12, 96.00)80.00 (72.30, 83.60)84.50 (75.65, 92.45)90.75 (86.50, 99.70)89.80 (75.10, 91.40)0.116 FVC (% predicted)77.26±20.0980.25±19.3283.58±19.0980.79±16.2781.20±12.3584.16±16.6595.65 (79.90, 107.30)85.40 (73.18, 88.50)0.202CBC WBC count (109 cells/L)8.20 (6.00, 11.10)7.50 (6.20, 9.75)8.31±2.587.09±1.707.40 (5.00, 11.40)9.30 (8.00, 10.80)6.20 (5.08, 7.19)5.09 (4.98, 6.13)<0.001† Neutrophil (%)68.30 (61.50, 77.30)60.90 (54.85, 69.70)71.41±10.8360.55±10.3170.00±12.6672.02±9.7663.65±9.3656.63±7.420.003† Lymphocyte (%)22.02±10.5427.12±10.6719.81±9.9729.56±6.8422.59±11.2318.76±7.9226.27±8.4232.67±7.040.004† Monocyte (%)6.48±2.847.10±2.086.47±2.687.32±2.626.10±2.387.15±2.406.67±1.567.10±1.750.806 Eosinophil (%)1.20 (0.40, 2.60)1.70 (1.25, 2.80)1.45 (0.62, 3.70)2.40 (1.80, 3.70)0.70 (0.15, 1.65)1.60 (0.20, 2.35)2.20 (1.40, 3.35)3.00 (2.08, 3.70)<0.001† Basophil (%)0.30 (0.20, 0.60)0.40 0.30, 0.60)0.39±0.250.47±0.280.28±0.150.35±0.240.40 (0.20, 0.60)0.60 (0.47, 0.70)0.016† RBC count (1012 cells/L)4.37±0.584.81±0.504.46 (4.21, 4.60)4.26 (4.12, 4.91)4.32±0.574.53±0.534.20 (3.95, 4.94)4.64 (3.80, 4.99)0.504 Hemoglobin (g/dL)134.48±16.64149.23±19.53134.12±15.57140.08±14.17130.00 (112.00, 144.00)139.00 (122.00, 150.00)143.00 (122.25, 152.50)146.50 (129.00, 163.00)0.174 Platelet (109/L)239.24±79.41223.67±61.34255.88±60.75224.38±86.32235.00 (178.00, 277.00)234.00 (184.00, 274.50)159.00 (148.00, 253.25)150.00 (147.00, 242.00)0.016†CMP ALT (U/L)20.00 (12.75, 31.00)19.50 (14.00, 25.75)17.00 (14.00, 25.00)10.50 (10.00, 14.00)17.00 (13.50, 25.50)21.50 (16.75, 41.25)21.00 (14.75, 26.00)18.50 (13.75, 23.75)0.253 AST (U/L)23.00 (18.00, 31.00)21.00 (17.00, 27.50)21.50 (18.00, 25.50)17.50 (14.75, 20.50)19.00 (14.50, 29.50)22.00 (18.00, 35.50)20.00 (17.75, 24.50)18.50 (16.50, 22.00)0.232 Albumin (g/L)39.80 (36.18, 42.82)40.60 (38.05, 42.82)39.05 (34.85, 41.22)41.35 (38.70, 43.08)40.69±4.8842.59±7.0938.10 (36.67, 40.30)40.10 (37.80, 40.47)0.164Radiographic pattern0.608 UIP14 (16.7)50 (72.5)8 (22.9)22 (75.9)5 (20.8)14 (77.8)3 (15.0)12 (60.0) Probable UIP0 (0.0)17 (24.6)0 (0.0)6 (20.7)0 (0.0)4 (22.2)0 (0.0)8 (40.0) Indeterminate UIP0 (0.0)2 (2.9)0 (0.0)1 (3.4)0 (0.0)0 (0.0)0 (0.0)0 (0.0) NSIP45 (53.6)0 (0.0)20 (57.1)0 (0.0)11 (45.8)0 (0.0)8 (40.0)0 (0.0) OP4 (4.8)0 (0.0)4 (11.4)0 (0.0)2 (8.3)0 (0.0)2 (10.0)0 (0.0) NSIP with OP overlap5 (5.9)0 (0.0)2 (5.7)0 (0.0)2 (8.3)0 (0.0)3 (15.0)0 (0.0) Unclassifiable or other16 (19.0)0 (0.0)1 (2.9)0 (0.0)4 (16.7)0 (0.0)4 (20.0)0 (0.0)Categorical variables are presented as the number of patients, with percentages in parentheses. Continuous variables are presented as the median (interquartile range), or as mean ± standard deviation. †, significant difference (P<0.05). ALT, alanine aminotransferase; AST, aspartate transaminase; BMI, body mass index; CBC, complete blood count; CMP, comprehensive metabolic panel; CTD-ILD, connective tissue disease-associated interstitial lung disease; DLCO, diffusion capacity for carbon monoxide; FEV1, forced expiratory volume in the first second; FVC, forced vital capacity; IPF, idiopathic pulmonary fibrosis; NSIP, nonspecific interstitial pneumonia; OP, organizing pneumonia; RBC, red blood cell; TLC, total lung capacity; UIP, usual interstitial pneumonia; VC, vital capacity; WBC, white blood cell. Clinical characteristics of patients in the four datasets used for distinguishing between IPF and CTD-ILD Categorical variables are presented as the number of patients, with percentages in parentheses. Continuous variables are presented as the median (interquartile range), or as mean ± standard deviation. †, significant difference (P<0.05). ALT, alanine aminotransferase; AST, aspartate transaminase; BMI, body mass index; CBC, complete blood count; CMP, comprehensive metabolic panel; CTD-ILD, connective tissue disease-associated interstitial lung disease; DLCO, diffusion capacity for carbon monoxide; FEV1, forced expiratory volume in the first second; FVC, forced vital capacity; IPF, idiopathic pulmonary fibrosis; NSIP, nonspecific interstitial pneumonia; OP, organizing pneumonia; RBC, red blood cell; TLC, total lung capacity; UIP, usual interstitial pneumonia; VC, vital capacity; WBC, white blood cell. To predict ILD-GAP stage, the 186 patients enrolled in Center A were split into a training set and an internal validation set. Finally, the training set comprised 155 patients (107 cases with ILD-GAP I/II and 48 cases with ILD-GAP III/IV), while the internal validation set included 31 patients (21 cases with ILD-GAP I/II and 10 cases with ILD-GAP III/IV) (Table S5). Multivariable logistic regression analysis identified age, gender, traction bronchiectasis, and the proportion of R in the pulmonary artery slice as independent predictors of ILD-GAP, which were subsequently incorporated into the combined-GAP model (Tables S6,S7). Performance of the radiomics-based model and multimodal-based model in differentiating between CTD-ILD and IPF Following the selection of features based on radiomics, clinical, and QCT parameters, a total of six radiomics features, four clinical indicators, and four QCT parameters were included in model construction. The radiomics and combined models were constructed with the most efficacious algorithm, support vector machine (Figure S2). The nomogram model was constructed through logistic regression, with clinical indicators and QCT parameters being combined with the radiomics model prediction score (Radscore). Table 2 presents the performance indicators for each model. Table 2Performance of three models in distinguishing between IPF and CTD-ILD in the training, test, and validation setsModelCohortAUC95% CIAccuracySensitivitySpecificityPPVNPVF1 scoreRadiomicsTraining0.8780.826–0.9310.7840.8690.6810.7680.8100.816Test0.8000.684–0.9160.7660.9140.5860.7270.8500.810Internal0.7410.580–0.9020.7140.6670.7780.8000.6360.727External0.8230.686–0.9590.6500.9000.4000.6000.8000.720CombinedTraining0.9540.923–0.9850.8950.9170.8700.8950.8960.906Test0.8770.792–0.9620.7810.8000.7590.8000.7590.800Internal0.8900.790–0.9880.7860.8750.6670.7780.8000.824External0.8180.681–0.9540.7250.8500.6000.6800.8000.756NomogramTraining0.9540.923–0.9850.8890.8810.8990.9140.8610.897Test0.9060.833–0.9800.8130.8570.7590.8110.8150.833Internal0.8540.736–0.9720.8330.8750.7780.8400.8240.857External0.8230.687–0.9590.7500.9000.6000.6920.8570.783AUC, area under the receiver operating characteristic curve; CI, confidence interval; CTD-ILD, connective tissue disease-associated interstitial lung disease; IPF, idiopathic pulmonary fibrosis; NPV, negative predictive value; PPV, positive predictive value. Performance of three models in distinguishing between IPF and CTD-ILD in the training, test, and validation sets AUC, area under the receiver operating characteristic curve; CI, confidence interval; CTD-ILD, connective tissue disease-associated interstitial lung disease; IPF, idiopathic pulmonary fibrosis; NPV, negative predictive value; PPV, positive predictive value. In the test set, the AUC value for the nomogram model was 0.906 [95% confidence interval (CI): 0.833–0.980], outperforming both the combined model (AUC =0.877; 95% CI: 0.792–0.962) and the radiomics model (AUC =0.800, 95% CI: 0.684–0.916). In the internal validation cohort, the AUC value for the combined model was 0.890 (95% CI: 0.790–0.988), outperforming both the nomogram model (AUC =0.854; 95% CI: 0.736–0.972) and the radiomics model (AUC =0.741; 95% CI: 0.580–0.902). In the external validation cohort, the nomogram model achieved an AUC value of 0.823 (95% CI: 0.687–0.959), outperforming the combined model (AUC =0.818; 95% CI: 0.681–0.954) and equaling the radiomics model (AUC =0.823; 95% CI: 0.686–0.959) (Figure 3A). The DeLong test revealed that the combined model and nomogram model outperformed the radiomics model in the internal validation set (P<0.05). In both the internal and external validation cohorts, the DeLong test revealed no significant difference in AUC between the combined model and the nomogram model (P>0.05) (Figure S3). The calibration curves confirmed that the nomogram model exhibited superior agreement between the predicted and actual probabilities (Figure S4A). Moreover, DCA indicated that the combined model and nomogram model outperform the radiomics model in terms of net gain for most threshold probabilities (Figure S4B). In the external validation set, the DCA results indicated that within the low threshold range (0.2–0.4), the nomogram model demonstrated a superior clinical net benefit compared to the radiomics and combined models. Meanwhile, in the medium-to-high threshold range (0.4–0.7), the combined model exhibited a superior clinical net benefit relative to the nomogram model (Figure S4B). Compared with the other models, the nomogram model demonstrated moderate improvements and robust performance across various metrics in the internal and external validation sets including accuracies of 0.833 and 0.750 and F1 scores of 0.857 and 0.783, respectively (Table 2 and Figure 3B). Given its superior predictive performance and statistical robustness, the nomogram model was selected as the definitive predictive framework for clinical application (Table 2 and Figure 3). Performance of the models in distinguishing between IPF and CTD-ILD. (A) Graphs showing the ROC for the training, test, internal validation, and external validation sets. (B) Confusion matrix results for the nomogram model in the training, test, internal validation, and external validation sets. (C) Nomogram model for distinguishing between IPF and CTD-ILD. (D) SHAP analysis for the interpretation of the combined model in the differential diagnosis between IPF and CTD-ILD. Importance ranking of feature variables for distinguishing between IPF and CTD-ILD in the SVM machine learning model. In these figures, yellow indicates variables that act as risk factors for predicting CTD-ILD, while purple indicates variables acting as protective factors. AUC, area under the receiver operating characteristic curve; BpGGO, ground-glass opacity at the basis pulmonis slice; BpR, reticulation at the basis pulmonis slice; CTD-ILD, connective tissue disease-associated interstitial lung disease; H, honeycombing; IPF, idiopathic pulmonary fibrosis; PaR, reticulation at the pulmonary artery slice; R, reticulation; RBC, red blood cell; ROC, receiver operating characteristic curve; RvGGO, ground-glass opacity at the right inferior pulmonary vein slice; SVM, support vector machine; TR, reticulation at the trachea carina slice. In the subgroup analysis based on imaging patterns, the nomogram model exhibited promising performance in differentiating the UIP pattern in IPF and CTD-ILD. Within the internal validation set, the nomogram model achieved an AUC of 0.743 and an accuracy of 0.842. In the external validation set, the AUC and accuracy were 0.639 and 0.667, respectively (Table S8 and Figure S5A). The nomogram model exhibited enhanced performance in differentiating other imaging patterns of IPF and CTD-ILD compared to its ability to distinguish UIP patterns among these diseases. In the internal validation set, the nomogram model achieved an AUC of 0.790 and an accuracy of 0.826, while in the external validation set, it attained an AUC of 0.868 and an accuracy of 0.800 (Table S8 and Figure S5B). Performance of the radiomics-based and multimodal-based models in predicting ILD-GAP Following the selection of features, a total of six radiomics features, two clinical indicators, one QCT parameter, and one qualitative indicator were included in the models. The radiomic-GAP model and combined-GAP model were developed with the logistic regression algorithm (Figure S6). SHAP analysis was performed on the combined-GAP model, with the top six features selected for the construction of the nomogram-GAP model (Figure 4A). Table 3 presents the evaluation metrics for each model. Performance of the models in predicting ILD-GAP. (A) SHAP analysis for interpretation of the combined-GAP model in predicting ILD-GAP. (A) Importance ranking of feature variables for predicting ILD-GAP in the logistic regression machine learning model. In these figures, yellow indicates variables that act as risk factors for predicting ILD-GAP I/II, while purple indicates variables acting as protective factors. (B) Graphs showing ROC for the training set and internal validation set. (C) Confusion matrix results for the nomogram-GAP model in the training (left) and validation (right) sets. (D) Nomogram-GAP model for predicting ILD-GAP. AUC, area under the receiver operating characteristic curve; BpGGO, ground-glass opacity at the basis pulmonis slice; BpR, reticulation at the basis pulmonis slice; CTD-ILD, connective tissue disease-associated interstitial lung disease; GAP, gender-age-physiology; H, honeycombing; ILD-GAP, interstitial lung disease gender-age-physiology; IPF, idiopathic pulmonary fibrosis; PaR, reticulation at the pulmonary artery slice; R, reticulation; ROC, receiver operating characteristic curve; RvR, reticulation at the right inferior pulmonary vein slice; TR, reticulation at the trachea carina slice. Table 3Performance of three models in predicting ILD-GAP in the training and internal validation setsModelCohortAUC95% CIAccuracySensitivitySpecificityPPVNPVF1 scoreRadiomicTraining0.8540.791–0.9160.7870.8970.5420.8140.7030.853Internal0.6950.451–0.9400.7740.8100.7000.8500.6360.829CombinedTraining0.9300.891–0.9680.8260.8880.6880.8640.7330.876Internal0.8950.782–1.0000.8070.9050.6000.8260.7500.864NomogramTraining0.9080.862–0.9530.8190.8970.6460.8500.7380.873Internal0.8810.763–0.9990.8070.8570.7000.8570.7000.857AUC, area under the receiver operating characteristic curve; CI, confidence interval; ILD-GAP, interstitial lung disease gender-age-physiology; NPV, negative predictive value; PPV, positive predictive value. Performance of three models in predicting ILD-GAP in the training and internal validation sets AUC, area under the receiver operating characteristic curve; CI, confidence interval; ILD-GAP, interstitial lung disease gender-age-physiology; NPV, negative predictive value; PPV, positive predictive value. In the internal validation cohort, the AUC value for the combined-GAP model was 0.895 (95% CI: 0.782–1.000), outperforming both the nomogram-GAP model (AUC =0.881; 95% CI: 0.763–0.999) and the radiomics-GAP model (AUC =0.695; 95% CI: 0.451–0.940) (Figure 4B and Table 3). The DeLong test revealed no statistically significant difference in AUC between the combined-GAP model and the nomogram-GAP model (P>0.05) (Figure S3). For the internal validation cohort, DCA demonstrated that the combined-GAP model and nomogram-GAP model surpassed the radiomics-GAP model in terms of net gain across a majority of threshold probabilities while also exhibiting strong calibration capabilities (Figure S7A,S7B). For the internal validation set, DCA indicated that within the low-to-moderate threshold range (0.3–0.6), the nomogram-GAP model demonstrated a superior clinical net benefit compared to both the radiomics-GAP and combined-GAP models. Meanwhile, in the high threshold range (0.6–0.8), the clinical net benefit of the combined-GAP model was marginally superior to that of the nomogram-GAP model (Figure S7B). Moreover, compared with other models, nomogram-GAP model demonstrated robust performance across additional metrics in the internal validation set, yielding an accuracy of 0.807, a sensitivity of 0.857, a specificity of 0.700, and an F1 score of 0.857 (Table 3 and Figure 4C). The nomogram-GAP model was selected as the final visual prediction model for predicting ILD-GAP stage due to its statistical robustness and greater suitability for clinical practice (Figure 4D). Interpretability of the models To facilitate interpretation of the model, SHAP analysis was employed to examine all of its features, and the findings were visualized as importance plots and beeswarm plots. For the combined model, both importance and beeswarm plots consistently identified gender, lung volume, original_shape_SurfaceArea of GGO at the basis pulmonis slice, aortic arch H (honeycombing volume percentage at the aortic arch slice level), and lbp.2D_gldm_LargeDependenceHighGrayLevelEmphasis of GGO at the right inferior pulmonary vein slice as the top five most significant features for differentiating CTD-ILD from IPF (Figure 3D). For the combined-GAP model, both importance and beeswarm plots consistently identified gender, age, pulmonary artery R (reticulation volume percentage at the pulmonary artery slice level), traction bronchiectasis, and squareroot_firstorder_Maximum of GGO at the basis pulmonis slice as the top five most significant features for predicting ILD-GAP stage (Figure 4A). In Figure 5, each of the two patient groups is represented by a single case, with all features displayed as z-score-normalized results. Tables S9,S10 present the mean and standard deviation for each feature. Schematic illustrating the application of the nomogram model and nomogram-GAP model in two patients. The HRCT scans shown are axial sections. Patient A and patient B differed in clinical data and QCT features. Patient A had a Radscore of 0.61, whereas patient B had a Radscore of 0.19. When age, gender, RBC, the ratio of lung volume, the ratio of aortic arch H, the ratio of pulmonary artery R, the ratio of right inferior pulmonary vein GGO, and Radscore were entered into the nomogram model, patient A had 264 points, with a corresponding probability of CTD-ILD of 86.36%, while patient B had 159 points, with a corresponding probability of CTD-ILD of 0.99%. When age, gender, traction bronchiectasis, the ratio of pulmonary artery R, and two radiomics features were entered into the nomogram-GAP model, patient A had 192 points, with a corresponding probability of ILD-GAP I/II of 75.87%, while patient B had 140 points, with a corresponding probability of ILD-GAP I/II of 12.65%. Patient A was diagnosed with ILD-GAP I/II CTD-ILD, and patient B was diagnosed with ILD-GAP III/IV IPF. AaH, honeycombing at the apex pulmonis slice; BpGGO, ground-glass opacity at the basis pulmonis slice; CTD-ILD, connective tissue disease-associated interstitial lung disease; GAP, gender-age-physiology; IPF, idiopathic pulmonary fibrosis; PaR, reticulation at the pulmonary artery slice; QCT, quantitative computed tomography; RBC, red blood cell; RvGGO, ground-glass opacity at the right inferior pulmonary vein slice. Discussion The diagnosis of CTD-ILD and IPF and the prediction of ILD-GAP stage remain a clinical challenge. This study supports the value of integrating multiregion radiomics with QCT and clinical data for distinguishing CTD-ILD from IPF and for predicting ILD-GAP stage. The nomogram model, combining gender, age, RBC, QCT, and radiomics features, demonstrated enhanced predictive performance in differentiating between IPF and CTD-ILD. Additionally, the nomogram-GAP model, consisting of gender, age, traction bronchiectasis, QCT, and radiomics features, exhibited superior predictive performance for predicting ILD-GAP stage. A notable strength of this study was its facilitation of clinical diagnoses of CTD-ILD and IPF through the nomogram model, as this can circumvent the diagnostic delays associated with the complex process of conducting CT imaging-based phenotyping in combination with clinical history and serum antibody screening. For critically ill patients unable to undergo PFTs, we offer a noninvasive alternative method—the nomogram-GAP model—for assessing ILD-GAP stage. These nomogram models can serve clinicians as visual diagnostic and ILD-GAP prediction tools, aiding in the formulation of personalized treatment regimens. Radiomics has shown significant potential in distinguishing CTD-ILD from IPF and in predicting ILD-GAP stage. Unlike approaches that merely quantify volumetric metrics of lesions, radiomics facilitates the extraction of high-dimensional lesion characteristics, through which a more detailed characterization of spatial heterogeneity within the lung can be achieved. In our study, multiregion radiomics analysis effectively distinguished between CTD-ILD and IPF and predicted ILD-GAP stage. This suggests it is a practical tool for the accurate diagnosis and prediction of ILD-GAP stage among patients with CTD-ILD and IPF. Multiregion radiomics, QCT, and clinical indicators for differentiating CTD-ILD and IPF Multiregional radiomics identified key lesion features among imaging signs, aiding differentiation between CTD-ILD and IPF. The radiomics features of GGO and R in the lower lung regions (basis pulmonis and right inferior pulmonary vein slices) and R in the middle lung regions (pulmonary artery and trachea carina slices) could effectively distinguish CTD-ILD from IPF. This result may be attributable to IPF pathology, which involves damage to alveolar epithelial cells and repair processes. Activation of fibroblasts within the pulmonary interstitium can lead to the secretion of extracellular matrix components, including collagen and elastic fibers. This process leads to the thickening of the interlobular septa and intralobular interstitium, which manifests as R on HRCT (32). GGO observed on CT scans of patients of IPF might consist of fine R rather than inflammatory exudate (33). Meanwhile, CTD-ILD is driven by autoimmune responses, with R formation arising due to inflammatory cell infiltration into the alveolar septa. Additionally, GGO formation has been associated with exudate and cellular debris within the pulmonary alveolus (34). Multiregion radiomics, through the high-throughput extraction of lung lesion features, may reflect the aforementioned potential differences in the occurrence of R and GGOs between CTD-ILD and IPF. This underscores the advantages of multiregion radiomics in distinguishing CTD-ILD from IPF. The nomogram model incorporating multiregion radiomics, QCT, and clinical indicators demonstrated robust performance, surpassing the radiomics-based classifier in accuracy. This further supports the potential of integrating multidimensional data for distinguishing CTD-ILD from IPF. We found that the presence of increased H in the aortic arch slice, increased R in the pulmonary artery slice, and diminished GGO in the right inferior pulmonary vein slice substantiated the diagnosis of IPF. This can be attributed to the prevalence of H and R in UIP. Although patients with CTD-ILD might exhibit a UIP pattern, this imaging phenotype is more frequent in those with IPF (35). The presence of GGO as the primary CT finding can serve as an exclusion criterion for IPF and indicates CTD-ILD (26,32). In our study, the lung volume percentage in patients with IPF was marginally higher than in patients with CTD-ILD. This might be explained by the reduction in fibrotic lung tissue volume in IPF, with remaining healthy lung tissue compensating through hyperinflation, which results in an increased lung volume in patients with IPF (9). Additionally, we found that a higher RBC count was associated with a higher probability of IPF diagnosis. This is likely due to the chronic hypoxia present in patients with IPF, which stimulates elevated erythropoietin production, promoting the proliferation and differentiation of hematopoietic stem cells into RBCs (36). In contrast, CTD-ILD results from an excessive inflammatory response and autoantibodies suppressing erythropoiesis in the bone marrow, leading to RBC destruction (37,38). The nomogram model integrates clinical baseline information with HRCT imaging features, eliminating the need for antibody testing and invasive tissue biopsies. Imaging features derived from high-throughput radiomics address the limitations of traditional subjective interpretation, thereby aiding in the differentiation between IPF and CTD-ILD. The nomogram model can thus be used to examine patients with early-stage seronegative autoimmune antibodies and provides physician with a noninvasive, visual diagnostic tool. The similarity in imaging patterns between IPF and CTD-ILD is a major reason for the difficulty in diagnosing these diseases (5). To ascertain whether the nomogram model developed in this study effectively captured information beyond visual pattern interpretation, we performed a differential analysis comparing patients with CTD-ILD exhibiting UIP patterns on HRCT to those with IPF exhibiting UIP patterns. The findings indicated that the nomogram model consistently demonstrated robust differential diagnostic performance in distinguishing between IPF and CTD-ILD with an overlapping UIP imaging pattern. Additionally, the nomogram model exhibited good performance in differentiating between various CTD-ILD imaging patterns, including NSIP, OP, NSIP with OP overlap, and unclassified ILD, as well as IPF categories, such as probable UIP and indeterminate UIP. This finding suggests that the nomogram model that includes CT radiomic features, quantitative CT features, and critical clinical variables (e.g., sex, age, and RBC count) can be effective in distinguishing potential differences between IPF and CTD-ILD, even when imaging patterns are similar. Therefore, radiomics offers supplementary value beyond subjective radiological pattern interpretation. The nomogram model is applicable to a diversity of imaging patterns of IPF and CTD-ILD in clinical practice. However, given the limited sample size of each imaging pattern among the patients with CTD-ILD in our study, further validation in larger prospective cohorts is warranted. A number of models for classifying ILD subtypes have been developed. Refaee et al. constructed a radiomics model from distinguish IPF from other ILDs, achieving 65.0% accuracy in the validation set (16). Walsh et al. used deep learning algorithms to classify HRCT images of fibrotic lung diseases, reporting a 73.3% accuracy; however, the output represented only HRCT phenotype classification, not ILD diagnosis (39). Furukawa et al. combined CT features with clinical indicators to distinguish IPF from other ILDs, achieving 83.6% accuracy, but they did not conduct external validation (12). In contrast, our nomogram model achieved 83.3% accuracy in the internal validation set and 75.0% in the external validation set, outperforming previous models (12,16,40). Consequently, our findings support the benefits of multiregional strategies and multidimensional data fusion in improving diagnostic accuracy. Multiregion radiomics, QCT, and clinical indicators for predicting ILD-GAP stage Accurate pretreatment prediction of ILD-GAP stage is crucial for the treatment of patients with CTD-ILD and IPF, as different stages require distinct treatment strategies, such as close monitoring or lung transplantation. The ILD-GAP index, proposed by Ryerson et al., demonstrates satisfactory performance in predicting mortality rates in chronic ILD disease over 1–3 years (7); however, patients with severe disease may struggle to perform demanding respiratory maneuvers (41,42). To address this challenge, we developed an integrated multimodal model combining multiregional radiomics, demographic, traction bronchiectasis, and quantitative CT metrics. Compared to PFTs, which require a high degree of respiratory cooperation, HRCT examinations are easier to perform and suited for patients with severe disease. The nomogram-GAP model’s superior performance demonstrates the complementary value of CT imaging and clinical information in predicting ILD-GAP stage and thus can facilitate ILD-GAP stage assessment without the need for PFTs. Our study confirmed several risk factors capable of predicting ILD-GAP stage, including radiomic features, volumetric metrics of R, and traction bronchiectasis in HRCT. Traction bronchiectasis has been reported to be a significant predictor of ILD progression and mortality (32,43,44). In line with this, we found that patients with traction bronchiectasis were more likely to be classified as ILD-GAP stage III/IV. This might be because, as interstitial fibrosis progresses, the release of profibrotic mediators leads to the activation of mesenchymal cells, the increase in the extracellular matrix, the differentiation of fibroblasts, and the apoptosis of epithelial cells. Further damage to epithelial cells activates airway basal cells, leading to abnormal re-epithelialization or bronchiolization, which manifests as traction bronchiectasis on HRCT (43). Therefore, traction bronchiectasis could be used to anticipate the prognosis of patients with ILD. Although previous studies have suggested that H is a significant predictor of mortality among patients with ILD (45,46), we found no significant correlation between H and ILD-GAP stage. Rather, we found radiomics features and volumetric metrics of R and GGO in the middle and lower lung regions were predictive of ILD-GAP stage. This inconsistency might be due to the processes relevant to disease progression and severity. H emerges following GGO and R. Over time, areas of R increase while those of GGO diminish, resulting in H (47). However, not all patients with CTD-ILD or IPF showed H signs in our study. Most patients were classified as ILD-GAP stage I/II and either exhibited H only in costophrenic angle regions or lacked it entirely. Furthermore, these patients predominantly exhibited R and/or GGO in the lower and middle lung regions. The nomogram-GAP model, combining R, GGO, and traction bronchiectasis, demonstrated superior performance in predicting ILD-GAP stage. These findings highlight the value of combining complementary features to enhance the accuracy of predicting ILD-GAP stage. However, we incorporated the overlapping variables of sex and age from the ILD-GAP stage into both the combined-GAP model and the nomogram-GAP model, which may entail a risk of information leakage. Nonetheless, we aimed to substitute PFT with CT radiomics and quantitative CT metrics, thereby developing an ILD-GAP stage prediction model for patients with severe disease that does not depend on PFT results. By employing imaging radiomics to address the absence of pulmonary function data while maintaining accessible variables such as age and sex, we were able to enhance the clinical utility of the model. In contrast to previous work that has examined imaging features for predicting the mortality of patients with ILD (11,48,49), our study additionally aimed to devise a system for predicting ILD-GAP stage for patients with early-stage disease. This can allow for the early intervention of cases with poorer prognosis and prevent fibrotic progression. Furthermore, this study combined four stages into composite stages I/II and III/IV. Although this approach may facilitate disease management for clinicians, it provides limited refinement in mortality risk stratification. Future studies should seek to achieve greater precision in the prediction of each of the stage through use of larger samples. Furthermore, the nomogram-GAP model forecasted the ILD-GAP stage, which is a validated surrogate prognostic indicator that indirectly reflects disease prognosis. However, the nomogram-GAP model did not directly predict the actual survival endpoints. Survival analysis on a larger sample is needed to validate the model’s predictive accuracy for actual transplant-free survival. We employed nomograms to create two visual decision support tools. These nomograms facilitate the intuitive presentation of both predictive models, enabling their sharing with a broader audience of clinical practitioners. We recommend using these nomogram models as clinical decision tools to facilitate personalized diagnostic and therapeutic decisions. This study involved several limitations that should be addressed. First, the data for this study were gathered retrospectively, and the external validation set was constrained by a limited sample size. Further prospective studies conducted across larger, multicenter populations are needed to validate the model’s performance. Second, although the manually delineated ROIs in this study yielded particularly high ICCs, subjective bias could not be excluded, and thus future work in this direction should incorporate automated segmentation methods. Third, ILD, a diffuse lung disease, demonstrates varied patterns of lesion distribution on CT imaging. Representing the distribution of lesions throughout the lung through use of only six typical slices inherently introduced bias, as it focal lesions or early, subtle changes might have been overlooked. In clinical practice, CT data are typically acquired through continuous volumetric scans. Consequently, the studies relying on six typical slices lacked the continuity necessary to capture changes in lesions between slices. Although the slices selected in this study adequately captured the heterogeneity of the entire pulmonary lesion, a more detailed comparison of the performance differences between these representative slices and whole-lung analyses is needed. This would enhance the clinical applicability of the models developed in our study. Fourth, the nomogram-GAP model was established with qualitative indicators of traction bronchiectasis. Future research should assess the precise severity of traction bronchiectasis from multiple perspectives, including quantitative analysis. Finally, we used different scanners and imaging protocols for HRCT scans, and thus the efficacy of nomogram-GAP model should be subjected to external validation. Conclusions Our study provided robust, noninvasive, and interpretable tools for distinguishing CTD-ILD from IPF and predicting ILD-GAP stage through the integration of multiregion radiomics features with QCT and clinical indicators. These nomogram models demonstrated robust generalizability across cohorts and hold promise for guiding precision diagnosis, ILD-GAP stage prediction, and treatment planning in clinical practice. Supplementary The article’s supplementary files as