Predicting the progression risk of chronic obstructive pulmonary disease in high-risk individuals based on whole-lung radiomics and deep learning: a multicenter study Authors: Jianfei Lin, Yan Zhang, Sheng Lu, Hongxing Zhao, Chentao Zhu, Huifang Du, Yifeng Zheng, Hupo Bian, Haifeng Qian Published: 2026 ABSTRACT AbstractBackgroundAs a highly prevalent and seriously debilitating lung disorder, chronic obstructive pulmonary disease (COPD) adds a considerable burden to global health services. However, many early-stage COPD patients do not exhibit abnormal results in lung function tests. Therefore, accurately identifying the individuals with high progression risk who are about to develop clinical COPD has become a key challenge for achieving early, precise intervention and reducing the disease burden. This study aims to develop and validate a comprehensive model for predicting COPD risk in high-risk populations.MethodsA retrospective analysis was conducted involving 806 patients from two hospitals, with the time span ranging from January 1, 2021, to May 30, 2025. After the entire lung parenchyma region is automatically segmented from the lung computed tomography (CT) images, imaging biomarker features and deep learning features were extracted. An integrated nomogram was constructed and verified, which combines imaging biomarker characteristics, deep learning attributes, and independent clinical predictors. The performance of the model was evaluated and compared using receiver operating characteristic (ROC) curves, decision curve analysis (DCA), and DeLong tests.ResultsIn the training set and test set, the area under the ROC curve (AUC) values of the clinical model were 0.615 and 0.572, respectively; for the radiomics (Rad) model, they were 0.834 and 0.824, respectively; for the deep learning radiomics (DLR) model, they were 0.873 and 0.825, respectively; those for the combined models were 0.878 and 0.836, respectively. The performance of the combined models was superior to that of the individual clinical model, Rad model, and DLR model.ConclusionsThis study developed and validated a combined nomogram by integrating the whole-lung Rad features of chest CT with deep learning features, and combining with clinical independent predictors to predict the risk level of high-risk individuals progressing to COPD.Keywords: Chronic obstructive pulmonary disease (COPD), radiomics (Rad), deep learning, prediction model Highlight box Key findings • This investigation established and validated an integrative nomogram that combines chest computed tomography (CT)‑based whole‑lung radiomic signatures, deep learning features, and clinically independent predictors. The goal was to assess the risk level of progression to COPD in high‑risk subjects. What is known and what is new? • The traditional diagnosis and staging assessment of chronic obstructive pulmonary disease (COPD) rely on pulmonary function tests (PFT) as the gold standard. However, in the preclinical or early stages of COPD, significant airway remodeling and macro/micro structural changes such as emphysema can occur in the lungs, while PFT parameters may still be within the normal range at that time point. This “structural-function dissociation” phenomenon indicates that relying solely on PFT may delay early diagnosis. • Based on the radiomics features of the entire lung from chest CT combined with deep learning features, it is possible to effectively assess the risk of COPD. A comprehensive model incorporating independent clinical predictors can be used by clinicians to predict the risk of patients progressing to COPD. What is the implication, and what should change now? • This comprehensive model provides clinicians with a rapid, cost-effective and non-invasive tool for simultaneously assessing the risk of COPD progression during routine chest CT scans. By early identification of high-risk patients, it promotes timely intervention and optimization of treatment strategies, prevents disease progression, and thereby reduces the economic burden on patients. Introduction Chronic obstructive pulmonary disease (COPD) is a chronic inflammatory lung disease characterized by persistent and progressive airflow limitation. It has caused a severe public health burden globally, with a high prevalence rate, high disability rate, and high mortality rate, ranking third among the causes of death worldwide (1). The traditional diagnosis and staging assessment rely on pulmonary function tests (PFT) as the gold standard (2). However, an increasing body of evidence indicates that in the preclinical or early stages of COPD, significant airway remodeling and macro- and micro-structural changes in the lungs can occur, while PFT parameters may still be within the normal range at that time point (3,4). This “structural-functional dissociation” phenomenon reveals that relying solely on PFT may delay early diagnosis. It is noteworthy that studies have shown that intervention during this window period (such as smoking cessation, bronchodilators, pulmonary rehabilitation, etc.) can effectively slow the rate of lung function decline and improve the long-term prognosis of patients (5). Therefore, accurately identifying the individuals with high progression risk who are about to develop clinical COPD from the large group of high-risk individuals with imaging abnormalities but normal pulmonary function is a key challenge for achieving early, precise intervention and reducing the overall burden of the disease. Computed tomography (CT) imaging provides a powerful tool for this challenge. Compared with PFT, CT can non-invasively and intuitively quantify early structural abnormalities in the lungs, such as by measuring the thickness of airway walls and low-attenuation areas of emphysema, and has become one of the most effective means for identifying subclinical lung injury (6). With the widespread application of low-dose chest CT in lung cancer screening (7), systematically identifying and tracking populations at risk of COPD from large-scale screening images has become possible. In terms of extracting the rich information contained in CT images, radiomics (Rad) has shown great potential. It extracts quantitative Rad features through high-throughput methods, thereby revealing the heterogeneity of diseases that the human eye cannot recognize (8). Rad is mainly applied in oncology (9), but the disease course of COPD has significant heterogeneity, and the application of Rad in COPD is rapidly expanding. It has already shown value in COPD prediction of acute exacerbation (10), risk of concurrent complications (11), and staging (12) assessment. With the significant development of computer performance, artificial intelligence plays an important role in medical imaging diagnosis and treatment. Deep learning, as a type of artificial intelligence, can automatically and effectively learn the advanced features in the original images, and can learn key classification features from the given data set. It has been proven to be superior to the intermediate and low-level features extracted manually (13). It has been widely applied in automated image detection and analysis (14), disease prediction and staging (15), and efficacy evaluation (16), greatly improving the accuracy and efficiency of disease diagnosis. Although deep learning features have certain advantages over Rad features, there are still certain limitations in practical applications, such as the problem of gradient disappearance in complex models. In recent years, the hybrid strategy that combines the advantages of image biomarker analysis and deep learning has become a leading direction in the field of medical image analysis. This “dual-track” approach can not only utilize the rich and potentially biologically significant quantitative features provided by image biomarker analysis, but also leverage deep learning to capture deeper and more abstract hidden patterns in the images (17). When combined, they are expected to construct more robust and powerful predictive models, providing new solutions for achieving early screening, precise diagnosis, and prognosis prediction of diseases (18-20). Currently, some studies have used CT image features to predict the risk of COPD occurrence (21). However, most studies have a relatively broad target population definition, or focus on diagnosed patients, or use healthy individuals as controls, it is difficult to accurately identify progressors within the specific high-risk subgroup of “abnormal CT images but normal lung function” which has significant clinical significance and a critical intervention window. Moreover, studies that solely use image biomarker analysis or deep learning technology may not fully utilize multi-dimensional imaging information. Based on this, unlike previous studies that focused on the diagnosis of COPD or the prognosis of confirmed patients, this study centers on high-risk individuals with normal lung function but existing structural abnormalities in imaging, aiming to predict their future risk of progressing to clinical COPD. By integrating the whole lung imaging biomarkers and deep learning features, the study first conducted a preliminary exploration of diagnostic features in CT images, and then merged these features with the identified independent clinical factors to create a joint prediction model. This model was visualized using a nomogram, thereby significantly improving the ability to stratify the risk for progressing to clinically significant COPD in the future. The value of chest CT can also be fully utilized, providing an objective basis for the clinical early identification of “high-risk individuals who are about to progress”, and the implementation of targeted intervention measures. Ultimately, it contributes effective strategies to curb the progression of COPD and reduce the disease burden for individuals and society. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0549/rc). Methods Patients and clinical data This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethical Committee of Huzhou Central Hospital (No. 20201017-01). The First Affiliated Hospital of Huzhou University was also informed of and agreed to the study. Individual consent for this retrospective analysis was waived. From January 1, 2021, to May 30, 2025, clinical data and thin-slice CT images of eligible patients were gathered from two hospitals: Huzhou Central Hospital and The First Affiliated Hospital of Huzhou University. Over this period, a total of 806 patients were enrolled in the study. Inclusion criteria are as follows: (I) patients diagnosed with COPD through PFT; (II) patients with organic changes in the lungs on CT images but with normal PFT results: on the CT images, there are signs of emphysema, thickening of the airway walls, or gas trapping, but the PFT results are normal; (III) PFT and chest CT completed within 2 weeks; (IV) having complete thin-layer (1 mm) chest CT images. The following conditions led to exclusion from the study: (I) incomplete clinical records, together with the presence of other chest diseases (e.g., pneumonia, atelectasis, pulmonary nodules or masses exceeding 6 mm, and pleural effusion); (II) the coexistence of any malignant tumor; (III) spinal hardware or marked imaging artifacts that result in suboptimal image quality and affect diagnostic assessment; (IV) accompanied by primary or secondary mediastinal abnormalities, such as cardiac hypertrophy, massive pericardial effusion or mediastinal masses; (V) received intervention treatment within the past three months; (VI) lack of thin-layer chest CT images. This study will take 556 patients from Center 1 (The First Affiliated Hospital of Huzhou University) as the training set and 250 patients from Center 2 (Huzhou Central Hospital) as the test set. Clinical information includes age, BMI, gender, smoking status, and laboratory test indicators such as C-reactive protein (CRP), erythrocyte sedimentation rate (ESR), procalcitonin (PCT), etc. For some missing clinical indicators, this study ultimately did not include them in the research. CT image acquisition and pulmonary function examination Non-contrast CT scans were performed on the participants using equipment such as the Aquilion ONE TSX-301C and Brilliance CT 16. Whole-chest axial images were acquired during full inspiration (scan parameters are provided in Appendix 1). Lung function was assessed with the Ganshorn PowerCube. The diagnosis of COPD followed the Global Initiative for Chronic Obstructive Lung Disease (GOLD) criteria (22), specifically defined as a post-bronchodilator forced expiratory volume in one second (FEV1)/forced vital capacity (FVC) ratio below 0.7. The grading of pulmonary function was as follows: GOLD I—FEV1/FVC <0.7 and FEV1 ≥80% of predicted; GOLD II—FEV1/FVC <0.7 and 50% of predicted 10 years of experience independently rated the segmentations using ITK-SNAP (version 3.8.0) (23). Subsequently, the Dice index was assessed, which helps to objectively quantify the spatial overlap of the contours and explore whether the fully automatic segmentation results are in line with the manual segmentation results. After that, the remaining samples were automatically segmented. Extraction and selection of imaging biomarker features This study extracted imaging features from the automatically segmented lung regions using Pyradiomics (http://pyradiomics.readthedocs.io). Before feature extraction, the images needed to undergo three steps of preprocessing to become standard images. The images were resampled to 1 mm × 1 mm × 1 mm, and the gray values were adjusted to a 25 Gy level, which helped reduce the interference from noise. Subsequently, feature extraction was performed using Z-score normalization, yielding three categories of features: first-order, shape, and texture features. The statistical significance of these features was assessed via t-tests, and only those with a P value below 0.05 were kept. To reduce the problem of collinearity, this study used the Pearson correlation coefficient to check the correlation between the features, and any feature pairs with a correlation coefficient higher than 0.9 were excluded from the analysis. This study also further optimized the feature set within a 10-fold cross-validation framework using Least Absolute Shrinkage and Selection Operator (LASSO) regression to determine the optimal regularization parameter λ and select the best features with non-zero coefficients. Thus, the feature set was effectively reduced to those with the most predictive and informative features. The coefficients estimated by LASSO are used for linear weighting to calculate the Rad-score (RS). Extraction and selection of deep learning features In this study, the author used the ResNet50 architecture as the convolutional neural network (CNN) for extracting deep learning features. The intensity distribution of the images was standardized through Z-score normalization. These standardized images were used as the input for the deep learning model. During the entire training phase, real-time data augmentation techniques, including random cropping and flipping, were adopted to enhance the robustness of the model. For test images, only normalization was applied for preprocessing to maintain consistency during the evaluation. ResNet50 performed the best in the test set. Its second-to-last layer was utilized for feature extraction. Within this model, the probability output generated by the CNN was termed “deep learning features”. Because the CNN model is complex, principal component analysis (PCA) was employed in the present study to compress these features down to a more tractable dimensionality of 512. Model construction and evaluation After feature selection, four models were constructed. Through logistic regression (LR) analysis of clinical features, statistically significant features were identified to develop clinical models. After feature selection using LASSO, the selected features and their respective coefficients were directly used to construct a linear LR model, which is the Rad model. At the same time, a fusion algorithm was adopted to combine deep learning features with Rad features. Subsequently, following the same feature selection and construction process as in Rad, a deep learning radiomics (DLR) model was constructed. In order to improve clinical utility, all clinical features underwent both univariate and multivariate analyses to detect clinically significant predictors. Then, these were integrated with the DLR model features to construct a linear model based on LR, namely “combined model”, and it was effectively visualized through a nomogram. In the final step, the performance of the models was assessed using several metrics: area under the curve (AUC), accuracy, sensitivity, specificity, recall, and the F1 score. Positive predictive value (PPV) and negative predictive value (NPV) were used to assess model performance. The AUC of different models was compared through DeLong test. The calibration curve (Hosmer-Lemeshow test) was used to analyze the calibration of the model and verify its reliability. Additionally, decision curve analysis (DCA) was employed to evaluate the clinical utility of the prediction model, thereby providing insights into their potential benefits in clinical contexts. Figure 1 shows the flowchart of the entire research design. The flowchart of the entire research design. AUC, area under the curve; CI, confidence interval; DCA, decision curve analysis; DL, deep learning; LASSO, Least Absolute Shrinkage and Selection Operator; MSE, mean squared error; Rad, radiomics; ROC, receiver operating characteristic curve. Statistical analysis The Shapiro-Wilk test was employed to assess whether the clinical characteristics followed a normal distribution. For continuous variables, the Shapiro-Wilk test is first used to assess their normality; if they conform to the normal distribution, the independent samples t-test is employed; otherwise, the Mann-Whitney U test is adopted; for categorical variables, Chi-squared tests (×2) were used for assessment. All data analyses were conducted using Python 3.7.12. Statistical analysis was performed using Statsmodels 0.13.2. Rad features were extracted using PyRadiomics 3.0.1. Machine learning tasks were implemented using Scikit-learn 1.0.2. A deep learning framework was developed using PyTorch version 1.11.0. The performance was optimized using CUDA version 11.3.1 and cuDNN version 8.2.1. Results Baseline characteristics and clinical feature screening of the patients Figure 2 illustrates the flowchart for patient selection. As of May 30, 2025, after screening according to the inclusion and exclusion criteria, a total of 806 patients were finally included. Center 1 included 556 patients, and Center 2 included 250 patients. Among them, there were 119 females and 687 males. There were 448 smoking patients and 358 non-smoking patients. Among them, 493 were diagnosed as GOLD I grade COPD patients, and another 313 were COPD high-risk patients. Table 1 details the baseline clinical characteristics of the study cohort (standardized units for all indicators in Table 1 are available in Appendix 3). The flowchart of the patient’s selection. Center 1: The First Affiliated Hospital of Huzhou University; Center 2: Huzhou Central Hospital. COPD, chronic obstructive pulmonary disease; CT, computed tomography; PFT, pulmonary function tests. Table 1Baseline characteristics of the study populationClinical factorsTraining cohort (n=556)Test cohort (n=250)High-risk patients (n=219)Patients with COPD (n=337)P valueHigh-risk patients (n=94)Patients with COPD (n=156)P valueAge (years)71.35±9.3174.35±8.56<0.00173.59±9.2675.15±9.460.10BMI (kg/m2)22.46±3.2722.07±3.230.1921.76±2.8122.00±3.020.75Albumin (g/L)39.76±4.5439.49±4.280.4639.24±4.1237.99±4.610.04Globulin (g/L)88.71±40.0291.87±44.970.4589.37±35.9482.56±40.650.16Triglyceride (mmol/L)1.50±2.501.17±0.640.022.65±14.371.16±0.790.63Alkaline phosphatase (IU/L)89.31±62.1686.76±38.960.9781.44±34.2886.54±50.910.49White blood cell count (×109/L)6.95±6.188.27±28.920.166.69±2.606.72±3.720.95Neutrophilic granulocyte (%)66.53±10.9668.05±11.600.0867.52±11.3166.54±15.460.91Lymphocytes (%)24.07±9.2722.09±9.150.0122.47±9.2921.28±10.400.22Absolute eosinophil count (×109/L)0.17±0.400.15±0.260.570.13±0.150.16±0.230.23Plateletcrit (%)0.20±0.060.19±0.060.128.94±11.064.48±6.35<0.001Red blood cell distribution width (%)13.52±1.4913.55±1.630.795.53±6.559.13±6.270.002Mean platelet volume (fL)10.20±1.3010.52±5.430.6310.61±1.6410.46±1.350.74Platelet distribution width (%)16.05±0.9515.97±1.220.9813.96±3.0414.46±3.460.50C-reactive protein (mg/L)87.26±81.5690.36±84.170.61152.65±101.26151.13±101.570.79Erythrocyte sedimentation rate9.15±8.649.43±9.320.529.49±8.8212.50±16.230.52Procalcitonin (μg/L)3.81±5.424.09±6.470.795.39±8.726.47±10.040.09PaO2 (mmHg)89.39±15.0492.08±23.930.5288.78±17.4585.62±15.780.13PaCO2 (mmHg)38.52±4.5137.90±3.770.6838.28±2.2838.62±5.750.5Gender0.0520.46 Female39 (17.81)39 (11.57)18 (19.15)23 (14.74) Male180 (82.19)298 (88.43)76 (80.85)133 (85.26)Smoke0.450.01 Non-smokers91 (41.55)128 (37.98)62 (65.96)77 (49.36) Smokers128 (58.45)209 (62.02)32 (34.04)79 (50.64)Data are presented as mean ± standard deviation or n (%). BMI, body mass index; COPD, chronic obstructive pulmonary disease; PaCO2, arterial carbon dioxide partial pressure; PaO2, arterial oxygen partial pressure. Baseline characteristics of the study population Data are presented as mean ± standard deviation or n (%). BMI, body mass index; COPD, chronic obstructive pulmonary disease; PaCO2, arterial carbon dioxide partial pressure; PaO2, arterial oxygen partial pressure. Univariate and multivariate analyses identified age and gender as predictors of the clinical model, and they were also independent predictors of the progression risk in high-risk COPD patients, as shown in Table 2. The AUC values of the clinical model in the training set and the external validation set were 0.615 [95% confidence interval (CI): 0.5677–0.6623] and 0.572 (95% CI: 0.4981–0.6455), respectively, as shown in Table 3. Table 2Univariable and multivariable analysis of clinical featuresVariableUnivariable analysisMultivariable analysisOR (95% CI)P valueOR (95% CI)P valueBMI1.001 (1.001–1.001)<0.0011 (0.999–1.001)0.64C-reactive protein1.003 (1.002–1.004)<0.0011.001 (0.999–1.002)0.69Globulin1.004 (1.003–1.006)<0.0011.003 (1–1.007)0.14Alkaline phosphatase1.004 (1.002–1.005)<0.0010.997 (0.994–1)0.15PaO21.005 (1.003–1.006)<0.0011.007 (1–1.015)0.12Age1.006 (1.004–1.008)<0.0011.041 (1.024–1.06)<0.001Neutrophilic granulocyte percentage1.007 (1.004–1.009)<0.0010.979 (0.953–1.005)0.18PaCO21.011 (1.007–1.014)<0.0010.97 (0.935–1.006)0.17Albumin1.011 (1.007–1.014)<0.0011.011 (0.975–1.049)0.60Percentage of lymphocytes1.013 (1.007–1.019)<0.0010.963 (0.931–0.995)0.059Platelet distribution width1.027 (1.018–1.036)<0.0010.922 (0.807–1.052)0.31Erythrocyte sedimentation rate1.031 (1.017–1.045)<0.0011.002 (0.984–1.019)0.88Red blood cell distribution width1.032 (1.021–1.043)<0.0010.997 (0.908–1.093)0.95Mean platelet volume1.042 (1.028–1.057)<0.0011.079 (0.957–1.215)0.29White blood cell count1.045 (1.025–1.066)<0.0011.001 (0.991–1.011)0.86Procalcitonin1.046 (1.019–1.073)0.0041.003 (0.976–1.03)0.84Triglyceride1.063 (0.982–1.151)0.20––Smoke1.493 (1.276–1.747)<0.0011.169 (0.889–1.537)0.34Absolute eosinophil count1.538 (0.97–2.438)0.12––Gender1.655 (1.418–1.933)<0.0011.863 (1.165–2.977)0.02Plateletcrit6.1 (3.016–12.342)<0.0010.233 (0.021–2.646)0.32BMI, body mass index; CI, confidence interval; OR, odds ratio; PaCO2, arterial carbon dioxide partial pressure; PaO2, arterial oxygen partial pressure. Univariable and multivariable analysis of clinical features BMI, body mass index; CI, confidence interval; OR, odds ratio; PaCO2, arterial carbon dioxide partial pressure; PaO2, arterial oxygen partial pressure. Table 3A comparison of four models’ performance across distinct cohortsSignatureAccuracyAUC95% CISensitivitySpecificityPPVNPVCohortClinic0.560.6150.5677–0.66230.4490.7310.7190.464TrainRad0.7190.8340.8018–0.86660.5860.9220.9210.592TrainDLR0.7780.8730.8447–0.90130.7080.8860.9050.664TrainCombined0.7770.8780.8501–0.90540.7320.8890.9040.688TrainClinic0.5570.5720.4981–0.64550.4740.6960.7230.441TestRad0.7360.8240.7739–0.87370.5840.9890.9890.587TestDLR0.6830.8250.7754–0.87510.494110.541TestCombined0.740.8360.7874–0.88410.7160.8180.8510.654TestThe models in question are: Clinic (based on clinical signatures), Rad (radiomics-based), DLR, and Combined (which merges clinical, deep learning, and radiomics signatures). AUC, area under the curve; CI, confidence interval; DLR, deep learning radiomics; NPV, negative predictive value; PPV, positive predictive value; Rad, radiomics. A comparison of four models’ performance across distinct cohorts The models in question are: Clinic (based on clinical signatures), Rad (radiomics-based), DLR, and Combined (which merges clinical, deep learning, and radiomics signatures). AUC, area under the curve; CI, confidence interval; DLR, deep learning radiomics; NPV, negative predictive value; PPV, positive predictive value; Rad, radiomics. Feature selection and model construction This study was based on the whole-lung images automatically segmented from chest CT scans. Imaging biomarker features and 3D deep learning features were each extracted in their own procedure. The repeatability was evaluated using intraclass correlation coefficient (ICC), and then after t-test, Spearman correlation, and LASSO filtering, six optimal imaging biomarker features with non-zero coefficients were finally selected, and an imaging biomarker model was constructed. Figure 3A,3B display the coefficients along with the average standard errors derived from 10-fold cross-validation, while Figure 3C shows the feature values corresponding to the non-zero coefficients that were selected. The detailed formula used to select imaging biomarker features can be found in Appendix 4. The feature extraction processes of the DL model and the DLR model. (A,D) The LASSO coefficients for radiomics and DLR features; (B,E) the MSE values for these two feature types; (C,F) the corresponding feature weights. DL, deep learning; DLR, deep learning radiomics; LASSO, Least Absolute Shrinkage and Selection Operator; MSE, mean squared error. In this study, the ResNet50 architecture was simultaneously adopted as the CNN for deep learning feature extraction, yielding a total of 2,048 such features. Following compression, 32 deep learning features remained. To enhance the predictive accuracy for COPD severity, Rad features and deep learning features were integrated. Through the combined use of t-test, Spearman’s correlation coefficient, and LASSO regression analysis, 13 features with non-zero coefficients were identified (as shown in Figure 3D-3F), and a DLR model was subsequently developed. This study used the following formula to screen DLR features (Appendix 5). Finally, this study combined independent clinical predictors with the DLR model and constructed a joint model using multiple LR. Finally, a nomogram was used for visualization for clinical application. Figure 4 is a nomogram specifically designed for clinical use. A nomogram intended for clinical application to enable early prediction of COPD. COPD, chronic obstructive pulmonary disease; DLR, deep learning radiomics. The nomogram (displayed in Figure 4) was built upon the regression coefficients derived from the final multivariate LR model. To compute the “Points” for each predictive variable, the following approach was used. For continuous predictors, namely age and DLR-score, the corresponding regression coefficient was multiplied by either the raw or standardized value of that predictor, and the product was then linearly scaled to an integer between 0 and 100. For the categorical predictor, gender, a fixed point value was directly assigned based on its regression coefficient. In more precise terms, the point contribution from any given variable equals (its regression coefficient divided by the sum of the absolute values of all coefficients) multiplied by 100. Summing the point contributions from all variables yields a total score. This total score is then transformed via the LR function, allowing one to read off the corresponding “probability of COPD” from the bottom axis of the nomogram. In practice, users can locate the point values for age, gender, and DLR-score above Figure 4, add them together, and then find the predicted probability at the bottom. The combined prediction model from this study can be implemented either as an automated computational tool or converted into a manual nomogram. Comparison of clinical models, imaging-based models, deep learning models, DLR models, and combined models The performance of the clinical model, the imaging-omics model, the DLR model, and the combined model is shown in Table 3. For the training cohort (Figure 5A), the combined model has the highest AUC value of 0.878, followed by DLR with 0.873 and Rad with 0.834. The AUC value of the clinical cohort is 0.615, which is relatively poor. For the external validation cohort (Figure 5B), the combined model has the highest AUC value of 0.836, followed by DLR with 0.825 and Rad with 0.824. The AUC value of the clinical cohort is 0.572. These results indicate that in different cohorts, the AUC of the clinical model is consistently lower, highlighting its limitations when used independently. The Rad model performs well and stably, suggesting that the information extracted from CT images has good predictive value. Finally, the combined model shows the best performance in terms of AUC, indicating its robustness and generalizability. These findings emphasize the value of multimodal data fusion in predicting the risk of progression to COPD in high-risk populations. Figure 6 presents the results of the DCA, which is used to evaluate the clinical applicability of the model. As shown in the figure, the combined model consistently shows higher net benefits compared to other models, indicating its better benefit-to-risk ratio in clinical decision-making. The calibration curves for the different cohorts are presented in Figure 7, which illustrate how well the model’s predicted probabilities align with the actual observed outcomes. Meanwhile, Figure 8 shows the results of the DeLong test, used to compare the discriminative performance across models. Taken together, these analyses indicate that the combined model has outstanding predictive performance, demonstrating strong capabilities in both discrimination and calibration. ROC curve of different models in the training (A) and test (B) cohorts, respectively. AUC, area under the curve; CI, confidence interval; DLR, deep learning radiomics; ROC, receiver operating characteristic curve. DCA curve of different models in the training (A) and test (B) cohorts, respectively. DCA, decision curve analysis; DLR, deep learning radiomics; Rad, radiomics. Calibration curve of different models in the training (A) and test (B) cohorts, respectively. DLR, deep learning radiomics; Rad, radiomics. The DeLong test results for various models are shown for the training (A) and test (B) cohorts. DLR, deep learning radiomics; Rad, radiomics. Discussion To forecast the progression risk of COPD among high-risk individuals, a composite nomogram was built and cross-validated in the present work. This nomogram merges three elements: whole-lung radiomic signatures from chest CT, deep learning features, and clinically independent predictors. Moreover, Rad features and deep learning features can independently classify high-risk individuals, indicating that chest CT not only assesses lung structure but also evaluates lung function status. Due to the slow progression and complex pathophysiological characteristics of COPD (24,25), traditional pulmonary function indicators such as FEV1 (26,27) have been proven insufficient for early detection of COPD. Interventions should be carried out after abnormal pulmonary function indicators occur, but this has missed the best timing. A large number of studies have found that CT scans can non-invasively and intuitively detect early structural abnormalities in the lungs (28). At the same time, the incidence and disease burden of COPD in China are high, while the overall detection rate of PFT is still at a low level. In contrast, with the promotion of large-scale lung cancer screening programs, the popularity of chest CT is very high. Therefore, how to detect early COPD through CT images has become the main research direction for many researchers. At present, numerous investigations have demonstrated that certain morphological features on CT—including bronchial wall thickening, alterations in tracheal shape, and the proportion of whole-lung emphysema—are associated with patients who have early-stage COPD (29). Makimoto et al. (6). can reflect the heterogeneous structural changes in the lungs of high-risk individuals by quantifying CT imaging features and combining traditional risk factors, thereby improving the ability to predict the progression of COPD. In recent years, Rad features have demonstrated unique value in the biological analysis of diseases. An increasing number of studies have proved the unique value of Rad features in reflecting the “heterogeneity of COPD”. Zhou et al. (30) developed a nomogram based on the Rad of whole lung CT images, which can be used for the early identification of COPD. However, Rad methods usually require precise manual segmentation of the ROI, and the repeatability of the used method is limited. Therefore, we used a deep learning model to perform automatic segmentation of the whole lung. Currently, CT Rad combined with deep learning algorithms has gradually been applied in COPD. Some studies have constructed a combined model by integrating clinical data and lung function measurement images, which has enhanced the prediction accuracy of acute COPD events compared to only clinical methods (31). This artificial intelligence-based framework promotes the early identification of high-risk individuals by detecting physiological abnormalities that traditional indicators do not capture. The present investigation demonstrated that the predictive capacity derived from any single information source—namely, clinical data alone, imaging biomarkers alone, or deep learning features alone—remains inadequate for fulfilling clinical practice requirements. This finding is consistent with observations from prior multimodal integration studies. Specifically, van Dijk et al. (32) established a diagnostic system that combined imaging manifestations with clinical signs to identify markers of acute heart failure in COPD patients, thereby successfully forecasting the potential risk of acute decompensation. Meanwhile, Zhu et al. (33) formulated a data fusion strategy integrating deep learning features, radiomic signatures, and questionnaire-derived variables. Collectively, these studies underscore the advantages of multimodal data synthesis within the COPD domain. However, the study cohorts in those previous investigations were comparatively heterogeneous, encompassing either patients with established diagnoses or healthy control subjects, neither of which enables precise targeting of the clinically pivotal high-risk subgroup characterized by “CT abnormalities with preserved lung function”—a population exhibiting a critical therapeutic window. In response, the present study specifically designed a composite predictive model for high-risk individuals who present early pulmonary structural alterations on baseline CT while maintaining spirometric parameters within the normal range. This model integrates three dimensions: independent clinical predictors, imaging biomarker features, and deep learning features. It attained an AUC of 0.878, surpassing the clinical-only model (AUC =0.615), the imaging-biomarker-only model (AUC =0.834), and the combined imaging-biomarker-plus-deep-learning model (AUC =0.876). The integrated model exhibited robust performance across both the training and test sets, as evidenced by high AUC values and well-calibrated curves. Notably, relative to previous studies, the current work constitutes a multicenter investigation incorporating an independent external validation cohort, thereby conferring a certain degree of generalizability. This study also has some limitations. Firstly, due to the retrospective nature of this study, there may be selection bias and information bias. Secondly, all the data in this study come from the institutions in the author’s city, which may limit the general applicability of this model in other geographical regions or diverse populations, and also limit its promotional value. Thirdly, because COPD has the characteristic of heterogeneous distribution throughout the lungs, our study is limited to a single time point and does not focus on the lung change characteristics of the same sample at different time points. The future research direction will combine traditional CT features into our model, and also include bronchial features and mediastinal features in the study, thereby improving the model’s efficacy, and combine delta imaging biomarkers to study the dynamic characteristics of COPD. Conclusions Whole-lung parenchyma extraction from CT images was carried out in this study using an automatic segmentation technique. A composite predictive model was then built by merging independent clinical risk factors, whole-lung Rad features, and deep learning features, with the aim of estimating the probability of COPD progression among high-risk subjects. For clinical utility, a nomogram was adopted to visualize the model, and its satisfactory applicability was confirmed. Additionally, the added value of chest CT for evaluating both lung morphology and functional status was demonstrated. Supplementary The article’s supplementary files as Acknowledgments None. Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethical Committee of Huzhou Central Hospital (No. 20201017-01). The First Affiliated Hospital of Huzhou University was also informed of and agreed to the study. Individual consent for this retrospective analysis was waived. Footnotes Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0549/rc Funding: This work was supported by the Science and Technology Project of Huzhou City, Zhejiang Province (No. 2020GY09) and Postgraduate Research and Innovation Project of Huzhou University (No. 2025KYCX99). Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0549/coif). The authors have no conflicts of interest to declare. Data Sharing Statement Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0549/dss References Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0549/prf Supplementary Materials The article’s supplementary files as Data Availability Statement Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0549/dss