Skip to main content

Development and application of a deep learning-based comprehensive early diagnostic model for chronic obstructive pulmonary disease



Chronic obstructive pulmonary disease (COPD) is a frequently diagnosed yet treatable condition, provided it is identified early and managed effectively. This study aims to develop an advanced COPD diagnostic model by integrating deep learning and radiomics features.


We utilized a dataset comprising CT images from 2,983 participants, of which 2,317 participants also provided epidemiological data through questionnaires. Deep learning features were extracted using a Variational Autoencoder, and radiomics features were obtained using the PyRadiomics package. Multi-Layer Perceptrons were used to construct models based on deep learning and radiomics features independently, as well as a fusion model integrating both. Subsequently, epidemiological questionnaire data were incorporated to establish a more comprehensive model. The diagnostic performance of standalone models, the fusion model and the comprehensive model was evaluated and compared using metrics including accuracy, precision, recall, F1-score, Brier score, receiver operating characteristic curves, and area under the curve (AUC).


The fusion model exhibited outstanding performance with an AUC of 0.952, surpassing the standalone models based solely on deep learning features (AUC = 0.844) or radiomics features (AUC = 0.944). Notably, the comprehensive model, incorporating deep learning features, radiomics features, and questionnaire variables demonstrated the highest diagnostic performance among all models, yielding an AUC of 0.971.


We developed and implemented a data fusion strategy to construct a state-of-the-art COPD diagnostic model integrating deep learning features, radiomics features, and questionnaire variables. Our data fusion strategy proved effective, and the model can be easily deployed in clinical settings.

Trial registration

Not applicable. This study is NOT a clinical trial, it does not report the results of a health care intervention on human participants.


Chronic obstructive pulmonary disease (COPD), a prevalent and preventable chronic lung ailment, is a well-established risk factor for lung cancer [1,2,3]. In 2019, COPD resulted in approximately 3.23 million global deaths [4], with a reported prevalence of 13.7% in individuals aged 40 and above in China [5]. The insidious onset, atypical symptoms, and limited public awareness contribute to a low success rate in early diagnosis and a high case-fatality rate [6, 7]. Strategies to enhance early COPD diagnosis are therefore in urgent demand [8, 9].

Conventional diagnostic methods, primarily relying on pulmonary function tests, exhibit limitations in accuracy, demonstrating a sensitivity of 79.9% (74.2–84.7%) and specificity of 84.4% (68.9–93.0%) [10] due to the subtle clinical symptoms and modest pulmonary function changes in early-stage COPD [10, 11]. In contrast, Computed Tomography (CT) imaging emerges as a more objective, precise, and efficient tool for the early diagnosis and assessment of COPD [12, 13], and improves sensitivity to 83.95% (73.0–89.0%) and specificity to 87.95% (70.0-95.6%) [14]. However, the disease’s heterogeneity poses practical challenges for conventional manual reading methods, including subjective interpretation variability among medical personnel and time-consuming processes [12,13,14,15].

To address these challenges, artificial intelligence (AI) techniques, particularly machine learning, have been applied in the analysis of CT-based radiomics features and demonstrated favorable performance [16]. In recent years, deep learning, a subset of machine learning known for its ability to handle complex problems, has further expanded the application of AI across various fields, including medicine [17, 18]. Within the context of COPD, deep learning algorithms such as convolutional neural networks (CNNs) have significantly improved diagnostic performance [19,20,21,22,23,24,25]. Notably, an AUC of 0.82 was achieved for COPD detection using DenseNet [26], a robust deep learning network. Additionally, leveraging radiomics features extracted from deep learning models to predict survival for COPD patients resulted in a concordance index of 0.73 (95% CI, 0.70–0.73) [27]. Furthermore, some recent studies have utilized deep learning approaches to model respiratory sound for COPD diagnosis, achieving an accuracy of 0.953 [28], an AUC of 0.966 [29] and an accuracy of 0.958 [30], respectively. These deep learning models were able to autonomously extract intricate features from CT images, and provide exceptional precision in image analysis and alleviate the workload of clinical practitioners [18,19,20, 26].

Fusion strategies were recently introduced where radiomics features and abstract deep learning features were simultaneously extracted from CT images, and computational algorithms such as Multi-Layer Perceptron (MLP) [31] were then employed to seamlessly integrate these cross-modal features, leading to the construction of fusion models [32]. Leveraging the complementary nature of radiomics and deep learning features, this fusion approach holds promise in enhancing precision and reliability in diagnosis and risk prediction [32]. Moreover, as efforts continue in constructing large population cohorts, questionnaire data are becoming more widely available. Existing research has demonstrated that adding radiomic features to epidemiological information can improve the predictive capacities of machine learning models for COPD progression and mortality [33, 34].However, whether these questionnaire data can enhance the diagnostic performance of deep-learning based models of COPD remains unclear.

In this study, we combined deep learning with radiomics profiles from CT images using MLP to build a COPD diagnostic fusion model. The epidemiological questionnaire data were then incorporated to develop a comprehensive model. A comparative analysis was conducted to assess the performance of standalone, fusion, and comprehensive models.

Materials and methods

Study design and COPD ascertainment

We obtained a CT dataset from Lanxi People’s Hospital, comprising 23,552 chest CT scans from 12,328 adults, with each adult contributing a varying number of scans. A team of respiratory specialists at the hospital reviewed CT diagnostic reports and related symptoms, and confirmed the presence of normal lung radiographic findings or imaging manifestations indicative of COPD in 2,983 adults. Specifically, normal lung radiographic findings were defined as the absence of obvious pathological lesions in both lungs, as indicated in the imaging reports. Imaging manifestations of COPD were characterized by the presence of chronic bronchitis combined with emphysema or bullae, or a definitive diagnosis of COPD. These manifestations could also be accompanied by pulmonary nodules, calcified lesions, fibrotic changes, minor inflammation (all less than 1 cm), or minor pleural effusion (less than 3 cm), as well as localized bronchial dilation. CT imaging reports showing other pathological abnormalities, as well as asthma patients, were excluded from the study. Furthermore, 2,317 of these 2,983 individuals had responded to a comprehensive epidemiological questionnaire as part of the Healthy Zhejiang One Million People Cohort, a newly established prospective cohort in Zhejiang province, China. No patients were excluded based on gender or ethnicity.Although pulmonary function tests (PFTs) are commonly recommended for the diagnosis of COPD, they possess certain limitations in early detection of COPD, and some patients with severe symptoms may be unable to tolerate or fully complete these tests, resulting in incomplete PFT results in our daily practice. Using clinicians’ actual diagnoses as the criteria aligns with our goal of creating a reliable, non-invasive, and practical diagnostic tool.

Questionnaire data acquisition

The epidemiological questionnaire gathered data on participants’ demographic variables, lifestyles, and health status. All questions were asked and answered in Chinese.

Demographic variables included age, gender, residence (urban or rural), marital status (married, divorced/separated, widowed, not married, or others), years of education (0, ≥ 1 and ≤ 6, ≥7 and ≤ 12, or ≥ 13 years), annual income rounded in thousands (< 50, 50–100, 101–200, 201–300, or > 300, thousand Chinese Yuan), and number of cohabitants in the household (< 2, 2, 3, 4, > 4).

Body Mass Index (BMI) was calculated by dividing weight (in kilograms) by the square of height (in meters). Subsequently, individuals were categorized into different BMI groups, including underweight (< 18.5 kg/m2), healthy weight (18.5 ≤ BMI < 24 kg/m2), overweight (24 ≤ BMI < 28 kg/m2), and obesity (≥ 28 kg/m2) categories.

Smoking status was classified based on participants’ self-reported smoking history in the questionnaire. Specifically, participants who reported smoking fewer than 100 cigarettes in their lifetime were classified as “never smokers”. Those who reported smoking more than 100 cigarettes in their lifetime and had quit smoking at the time of the questionnaire survey were classified as “former smokers”. Lastly, participants who reported smoking more than 100 cigarettes in their lifetime and were still smoking at the time of the survey were classified as “current smokers”. Smoking addictiveness was categorized as none, moderately addicted or highly addicted based on the Fagerstrom Test for Nicotine Dependence (FTND). Participants also provided information on the weekly frequency of secondhand smoke exposure, categorized as never, sometimes, 1–2 days/week, 3–5 days/week, or almost daily. Drinking status was categorized as never, former drinker or less than once per week, more than once per week for less than 12 years, or more than once per week for 12 or more years.

Diet preferences, habits, and frequency of behaviors were also collected. Preferences included food temperature, dryness, texture, saltiness, spiciness. Habits included speed of eating, meal regularity, and balance of diet (rich in meat, rich in vegetables, or balanced). Frequencies of dietary behaviors per week included intake of red meat, vegetables, whole grains, oil-rich food, sugar-rich food, sugar-sweetened beverages, tea, and leftovers, along with the frequency of dining out, and breakfast. Participants also disclosed whether they were vegan.

Weekly work intensity was classified based on the work duration (< 20 h, 20–40 h, currently working and ≥ 40 h, retired and ≥ 40 h).

Participants were asked to report the intensity, frequency, and duration of physical activity that they did during the past weeks, and then divided as inactive (0 metabolic equivalent (MET) hours /week), > 0 and < 7.5 MET hours/week, ≥ 7.5 and < 15.0 MET hours/week, and ≥ 15.0 MET hours/week.

Sleep quality was assessed based on items from the Pittsburgh sleep quality index (PSQI); depressive symptoms was evaluated based on the Center for Epidemiological Studies Depression-10 (CES-D-10) scale; and cognitive function was assessed based on the AD8 scale.

Health status data including prevalence of tumors, coronary heart disease, stroke, diabetes mellitus, osteoarthritis, Helicobacter pylori infection, gastritis, uremia, rheumatoid arthritis, and alcoholic fatty liver were also collected.

In modelling phase, Z-score normalization was applied to numerical questionnaire variables, while categorical questionnaire variables were encoded manually as binary (0 and 1) or ordinal values.

Chest CT Image Acquisition and Preprocessing

This study utilized two CT scanners, the GE Optima CT680Q and the GE Optima CT540, for CT scanning. The scans were performed in a supine position after complete inspiration, from the lung apex to the lung base. Scanning parameters: tube voltage 120KV, adaptive tube currents according to the body size (ranging from 60 to 350 mA), detector configuration 64 × 0.6 mm, gantry rotation time 0.5s/rotation, pitch 1.0, matrix 512 × 512. The CT scanning had an attenuation coefficient range of -1024 to 3072 HU. Slice thickness was 1.25 mm. The scanned CT data were stored in DICOM format. For image preprocessing, we initially normalized the pixel density values of all CT scans to a range between 0 and 1. Subsequently, we employed the LungMask tool [35], which is specifically designed for automated lung segmentation in chest CT scans, and can be used as a python module, to obtain the lung parenchyma, with pixel values outside this region being set to zero. Its robust performance has been validated across diverse datasets and disease contexts. Finally, the resulting images were resized to a dimension of 128 × 256 × 256.

Model development and evaluation

We propose a two-stage multimodal prediction modelling framework, wherein the first stage focused on feature extraction from various modalities. Specifically, epidemiological features were derived from preprocessed raw questionnaire data, while deep learning features were extracted from preprocessed CT images using self-supervised learning generative model, and radiological features were obtained via lung image analysis tools. In the second stage, multimodal features were fused. Initially, features from different modalities were dimensionally reduced to preset dimensions using different shaped multilayer perceptron (MLP). Subsequently, these features were concatenated to construct multimodal feature utilized for prediction. Finally, the multimodal features were fed into the stacked MLP to output the final prediction.

Radiomics feature extraction

The radiomic features were extracted from the resultant CT images using the PyRadiomics [36] package. Each CT image was resampled to (1,1,1) before feature extraction. The extraction parameters were configured as follows: ‘minimumROIDimensions’ set to 2, ‘minimumROISize’ left as None, ‘normalize’ set to False, ‘normalizeScale’ set to 1, ‘removeOutliers’ left as None, ‘resampledPixelSpacing’ set to (1,1,1), ‘interpolator’ specified as ‘sitkBSpline’, ‘preCrop’ set to False, ‘padDistance’ set to 5, ‘distances’ configured as (1), ‘force2D’ set to False, ‘force2Ddimension’ set to 0, ‘resegmentRange’ left as None, ‘label’ assigned as 1, and ‘additionalInfo’ left as True.

In addition to the original image, we applied a range of filters including the Wavelet, Laplacian of Gaussian (LoG), Square, Square Root, Logarithm, Exponential, Gradient, and Local Binary Pattern (LBP) filters to the images. Features were extracted using First order, Gray Level Cooccurrence Matrix (GLCM), Gray Level Size Zone Matrix (GLSZM), Gray Level Run Length Matrix (GLRLM), Gray Level Dependence Matrix (GLDM), Neighboring Gray Tone Difference Matrix (NGTDM), and Shape methods. A total of 2,983 participants had extractable CT imaging radiomics features using the PyRadiomics package, with a consistent yield of 1,409 radiomics features for each subject. In constructing the model, Z-score normalization was applied.

Self-supervised deep learning feature extraction

We initially utilized Variational Autoencoder (VAE) [37], a large-scale unlabeled self-supervised model to extract deep learning features with dimensions of 16 × 32 × 32 from the 23,552 images. Among them, we subsequently employed 3D convolutional kernels to down-sample the deep learning features in 2,983 CT images containing available radiomics features. The down-sampling was achieved using 3D convolutional kernels with a kernel size of (3, 3, 3) and a stride of (1, 1, 1), while maintaining the same number of input and output channels. After passing through this convolutional kernel, the output feature map’s dimensions are halved compared to the input feature map. This down-sampling resulted in deep learning features with dimensions of 8 × 16 × 16. Finally, we flattened these reduced-dimension features, yielding 2,048 deep learning features for each participant.

The optimization objective during the training of the unlabeled 3D VAE aimed to maximize the Evidence Lower Bound (ELBO), which consists of two components: the reconstruction loss and the prior matching loss. Additionally, to ensure the effectiveness of the feature representations, we incorporated patch adversarial loss and perceptual loss to enhance the quality of image reconstructions. These added components were instrumental in preserving the quality of the reconstructed images while optimizing the feature variables.

Feature fusion strategy

We employed the MLP model for feature fusion. The fusion strategy was feature-specific, meaning a uniform fusion approach was applied to participants concurrently exhibiting different features. Therefore, we utilized MLP models in two subsets: one comprising 2,983 participants with both radiomics and deep learning features for the fusion model, and another with 2,317 participants possessing radiomics, deep learning, and epidemiological features for the comprehensive model.

To reduce dimensionality, we employed a down-sampling process that reduced the deep learning features from 2,048 to 256 dimensions and the radiomic features from 1,409 to 139 dimensions. Here the down-sampling process was achieved by setting the output dimension of the MLP, which was a hyperparameter. However, the original dimension of 49 epidemiological features remained unchanged.

These features were subsequently concatenated and processed through a three-layer MLP module. Each module within the MLP architecture was composed of linear transformation layers, batch normalization, rectified linear unit (ReLU) activation, and dropout layers. Figure 1 shows the workflow of our fusion strategy.

Fig. 1
figure 1

Flowchart of standalone, fusion, and comprehensive models. Model 1: standalone model based on radiomics features; Model 2: standalone model based on deep learning features; Model 3: fusion model; Model 4: comprehensive model. Note: VAE: Variational Autoencoder; Feature EF: epidemiological features; Feature DL: deep learning features; Feature RA: radiomics features; MLP: Multi-Layer Perceptron

Hyperparameters selection


Our VAE model was trained with a carefully chosen set of hyperparameters to optimize performance. For model construction, we set the number of groups for group normalization to 32, and allocated 64, 128, and 128 feature channels per layer within the encoder. In the loss function, we assigned coefficients of 1e-3, 1e-2, and 1e-6 to the perceived loss, confrontation loss, and Kullback-Leibler loss, respectively. We utilized the Adaptive Moment Estimation (Adam) optimizer with an initial learning rate of 1e-4 for generator and 5e-4 for discriminator. Finally, training was conducted over 100 epochs to ensure comprehensive model training and convergence.


Our MLP model underwent training with a set of hyperparameters to optimize performance. We employed the Stochastic Gradient Descent (SGD) optimizer with an initial learning rate of 0.01 and momentum of 0.9. The learning rate was decayed every 30 epochs by a factor of 0.3 to facilitate convergence. To prevent overfitting, L2 weight decay with a coefficient of 0.01 was applied. A batch size of 200 and a hidden layer width of 64 were chosen for enhanced computational efficiency and model capacity. Furthermore, a dropout ratio of 0.5 was introduced to mitigate overfitting. Finally, training was conducted over 160 epochs to ensure comprehensive model training and convergence.

Time complexity analysis

We utilized the big O notation to show the time complexity of models above. The time complexity of the convolution module is

$$\begin{aligned}&Time \sim O ({\sum }_{l=1}^{L}\left({G}_{l}\cdot{{C}_{l-1}\cdot F}_{h}\cdot{F}_{w}\cdot{F}_{d}\right)\\&+\left({F}_{h}\cdot{F}_{w}\cdot{F}_{d} \right)\cdot {K}_{l}^{3}\cdot {C}_{l}\cdot{C}_{l-1})\end{aligned}$$

Where \(G\) represents the number of groups in group normalization, \({F}_{h},{F}_{w},{F}_{c},C\) represents the length, width, height and number of channels of the feature map, \(K\) represents the size of the convolutional kernel and \(L\) represents the number of convolutional kernel modules; The time complexity of the down sampling module is

\(Time \sim O({\sum }_{l=1}^{{ \text{L}}_{\text{d}}}\left({F}_{h}\cdot{F}_{w} \cdot {F}_{d} \right) \cdot {K}_{l}^{3} \cdot {C}_{l} \cdot {C}_{l-1}) )\))

The time complexity of the self-attention module is.

\(Time \sim O({N}^{2}+Nd\))

Where N represents the sequence length and D represents the feature dimension.

Other statistical methods

Continuous variables were presented as mean ± SD and categorical variables were presented as count and percentages. Model performance was evaluated using metrics such as accuracy, precision, sensitivity, F1-score, Brier score, receiver operating characteristic curve (ROC), and area under the curve (AUC). All metrics were obtained using weighted methods except for the Brier score. Additionally, bootstrap resampling was used to compute the 95% confidence intervals for accuracy, precision, sensitivity, and F1-score, with 1000 randomized tests. Accuracy indicates overall correctness, sensitivity measures true positive rate, precision assesses positive prediction accuracy, and F1-score balances both precision and sensitivity through harmonic mean [38]. These metrics are prevalent in classification performance evaluation [39,40,41,42], where higher values of AUC denote model’s superior discrimination ability between positive and negative classes. A two-tailed P-value of less than 0.05 indicated statistical significance. Statistical analysis was conducted using Scipy packages in Python 3.9. The CT images were processed with SimpleITK for preprocessing and Lungmask for radiomics feature extraction. Self-supervised 3D- VAE and MLP models were implemented in Python 3.9 using Pytorch 1.13.1.


Patients characteristics

The study included a total of 2,983 participants, which comprised 497 COPD patients and 2,486 participants with normal lung radiographic findings. Questionnaire data were available for 2,317 participants, where 477 were COPD patients.

To improve the comparability of models, the 2,317 participants with available questionnaire variables were randomly divided into training and testing datasets at an 8:2 ratio. The training set encompassed 1,853 participants, including 373 individuals with COPD and 1,480 individuals without COPD; and the testing set comprised 464 individuals, with 104 having COPD and 360 without COPD. Then, for the additional 666 individuals in the total 2,983 individuals, a same division strategy was applied to establish another training and testing sets. As a result, the final training set included 2,385 individuals, comprising 389 with COPD and 1,996 without COPD; and the final testing set comprised 598 participants, with 108 having COPD and 490 without COPD. Table 1 provided a detailed overview of the basic epidemiological characteristics of both the training and testing sets. Importantly, statistical analysis revealed no significant differences between them.

Table 1 The basic characteristics of training and test sets of the 2,317 participants

Diagnostic performance

We conducted a thorough assessment of the diagnostic performance of both standalone and fusion models. For COPD diagnosis, the AUC was 0.844 for the model utilizing deep learning features, and 0.944 for the model employing radiomics features. Notably, the fusion model outperformed the independent models, achieving an AUC of 0.952. In the comprehensive model, which incorporated deep learning, radiomics and epidemiological features, the performance was further enhanced, yielding an AUC of 0.971. Furthermore, the comprehensive model outperformed all other metrics, including accuracy, precision, recall, F1-score, and Brier score, which were separately 0.886, 0.909, 0.886, 0.891 and 0.080.

Figure 2 was a visual representation of the ROC curves and their corresponding AUC values for various models under consideration. Higher AUC value indicated a superior discriminatory capacity of a model.

Fig. 2
figure 2

The ROC curves and their corresponding AUC values of various COPD diagnostic models. (A) standalone model based on deep learning features. (B) standalone model based on radiomics features. (C) fusion model integrating deep learning and radiomics features. (D) comprehensive model incorporating deep learning, radiomics and epidemiological features. RA: Radiomic Feature, DL: Deep Learning Feature, EF: Epidemiological Features

Table 2 provided a comprehensive summary of the diagnostic metrics, including AUC, accuracy, precision, recall, F1 score and Brier score, for the different models assessed in the testing set. Our observations consistently indicated that among the two independent models, the radiomics-based model exhibited superior performance.

Table 2 Performance of various models for COPD diagnosis in test set

Across all metrics, the fusion model consistently outperformed the independent models, reaffirming its effectiveness in enhancing diagnostic accuracy. Notably, the most remarkable performance was achieved by the comprehensive model.

Model interpretability

We employed SHAP values to evaluate the relative importance of the epidemiologic variables in our comprehensive model, and the results were illustrated in Fig. 3. The SHAP values represented the average impact of each variable on the magnitude of the model output, with higher SHAP values indicating a more substantial role played by the respective variable.

Fig. 3
figure 3

The SHAP values of epidemiological variables in the comprehensive model


We established and validated a robust AI-driven data fusion-based COPD diagnostic model. Leveraging deep learning features, radiomics, epidemiological data, and MLP, our fusion model achieved an AUC of 0.971 for diagnosis of COPD. Moreover, our modelling framework can be readily deployed in assisting COPD diagnosis in clinical practices. To the best of our knowledge, we are among the first studies that successfully applied a multimodal data-fusion strategy in COPD diagnosis in a Chinese population.

In recent years, large population cohorts such as COPDGene [43], ECLIPSE [44], and SPIROMICS [45] have been established in western countries, and these initiatives have significantly advanced the research on COPD. However, it is essential to acknowledge certain drawbacks, including the need for robust validation and potential challenges in generalizing findings to diverse populations. Notably, there remains a substantial gap in research related to COPD in Chinese population. The COMPASS (Investigation of the Clinical, Radiological and Biological Factors, Humanistic and Healthcare Utilisation Burden Associated with Disease Progression, Phenotypes and Endotypes of COPD in China) is a prospective multi-center study started in 2019, which has marked a recent effort in understanding the relevance of COPD biomarkers identified in Western cohorts to Chinese patients [46]. Our study bears unique significance as it employed an extensive dataset of CT scans from individuals residing in the eastern region of China, presenting a COPD diagnostic model tailored specifically for the Chinese population.

Owing to COPD’s slow progression and complex pathophysiology [47, 48], traditional lung function indicators like FEV1 [49, 50] have proven inadequate for early detection of COPD. For example, in a follow-up study, it was reported that approximately half of the 332 individuals with COPD at the end of the observation period exhibited normal FEV1 values by the age of 40 (P < 0.001) [49]. Conversely, CT scans offer the advantage of detecting structural lung changes at an earlier stage [51]. Previous studies have attempted to utilize mathematical models [52, 53] to assess emphysema, which largely represents lung parenchymal destruction [54]. In a study using integrated CT metrics, radiomics, and standard spirometry measurements, the investigators were able to estimate the presence and severity of emphysema at AUCs of 0.86 and 0.88, respectively [53]. However, these methods relied on manual annotation, which led to time consumption and issues related to observer variability [32, 48, 55]. The advent of deep learning has revolutionized the data extraction process and alleviated observer variability concerns [56]. Studies that employed deep learning algorithms with CT data for COPD detection had achieved notable success with AUCs of 0.87, 0.90 [57], and 0.927 [58], which were relatively good. However, previous studies have predominantly focused on radiomics features. For instance, one study aimed to predict COPD progression by combining radiomic features and demographics with a machine learning approach, achieving an AUC of 0.73 [59]. Another study extracted radiomics features for COPD survival prediction using a deep learning approach, resulting in a concordance index of 0.716 [27]. While the results were promising, whether adapting a multimodal feature fusion strategy would enhance the model performance remains unclear.

In the field of multimodal data fusion, MLP emerges as a strong deep learning tool, renowned for its robust modeling capabilities and adeptness at autonomous feature extraction [31], particularly well-suited for handling high-dimensional and heterogeneous data [60]. In the current study, we observed that both deep learning and radiomics models emerged as promising independent diagnostic tools, boasting respective AUC values of 0.844 and 0.944, respectively. Furthermore, the fusion model that combined radiomics and deep learning features yield a higher AUC of 0.952, proving the effectiveness of incorporating epidemiological questionnaire data to assist CT-based diagnosis. Our study presented a holistic diagnostic method that integrates deep learning, radiomics and epidemiological data, and empirically demonstrated the capacity of MLP in the context of COPD diagnosis. This exceptional performance could be attributed to MLP’s proficiency in capturing intricate and nonlinear data relationships [61], a critical advantage in the complex domain of COPD diagnosis. Additionally, the high dimensionality of our dataset resulting from the integration of various features might have provided the MLP model with an advantage in discerning subtle interactions that linear models like LASSO [62] may overlook. Early diagnosis plays a crucial role in improving patient prognosis in COPD management, and our strategy proves to be a useful tool in assisting healthcare providers to integrate epidemiological data such as demographics, lifestyles, and health status of the patients into their diagnostic frameworks, especially in scenarios where COPD is suspected but not definitively diagnosed.

Moreover, our model can be easily deployed to clinical settings. Following chest CT scans acquisition, our all-encompassing workflow, including preprocessing, deep learning feature extraction, and radiomic feature extraction, supplies clinicians with a reference AUC for diagnosing COPD, regardless of the availability of patients’ epidemiological data. While multi-modal data have been collected routinely in clinical practice, how to utilize these data has been elusive. Our approach greatly facilitated this process and provided an example as utilizing the full potential of health big data.

Due to technical constraint, we were only able to investigate the SHAP values of the included predictors in each modal separately. Since both deep learning and radiomics features were abstract and hard to understand, we only investigated the variable importance of the epidemiological questionnaire data. Here we identified age, smoking, and tea intake as the top significant contributors to the model’s performance. It should be noted that these SHAP values were conditional on deep learning and radiomics features. To certain extend, our findings on these variables aligned with prior research on COPD risk factors. For instance, the influence of aging on COPD development might be attributed to factors like cellular senescence and heightened basal levels of inflammation and oxidative stress [63]. Smoking is widely recognized as the primary risk factor for COPD [64], with potential mechanisms of its role in airway inflammation [65] and vascular endothelial cell apoptosis [66]. Additionally, existing studies have indicated that the green tea consumption was associated with a decreased likelihood of developing COPD [67], potentially due to the tea extracts like catechin, which can reduce lung tissue inflammation [68]. These insights provided a robust foundation for the improved performance observed in our comprehensive models.

The strengths of our model include its high accuracy and potential for early intervention, which can improve patient outcomes in clinical practice. Additionally, the present model integrates readily accessible variables derived directly from CT images, thereby streamlining the diagnostic procedure, and rendering it user-friendly for healthcare practitioners. This accessibility promotes early intervention and enhances the management of COPD.

Despite its strengths, there are some limitations. The use of data from a single center necessitates external validation for model generalizability and stability. Future multi-center studies are needed to confirm the robustness of our modeling strategy. Additionally, optimizing feature selection and hyperparameter tuning as well as improving model interpretability are ongoing challenges that warrant further investigation.


We successfully developed and demonstrated a comprehensive COPD early detection model by integrating deep learning and radiomics features from CT imaging, along with epidemiological data from questionnaires. Our proposed model provided clinicians with novel AI tools for COPD diagnosis, and can improve the prediction accuracy by incorporating epidemiological data, shedding lights on utilizing multi-source epidemiological and clinical data in diagnosis of COPD. This modelling strategy holds significant promise for practical implementation in clinical settings and serves as a valuable tool for COPD research.

Data availability

No datasets were generated or analysed during the current study.



Chronic obstructive pulmonary disease


Area under the curve


Artificial Intelligence


Computed Tomography


Convolutional neural networks


Multi-Layer Perceptron


Fagerstrom Test for Nicotine Dependence


Metabolic equivalent


Pittsburgh sleep quality index


Center for Epidemiological Studies Depression-10


Laplacian of Gaussian


Local Binary Pattern


Gray Level Cooccurrence Matrix


Gray Level Size Zone Matrix


Gray Level Run Length Matrix


Gray Level Dependence Matrix


Neighboring Gray Tone Difference Matrix


Variational Autoencoder


Evidence Lower Bound


Rectified linear unit


Adaptive Moment Estimation


Stochastic Gradient Descent


Receiver operating characteristic curve


  1. Viegi G, Maio S, Fasola S, Baldacci S. Global Burden of Chronic Respiratory diseases. J Aerosol Med Pulm Drug Deliv. 2020;33:171–7.

    Article  PubMed  Google Scholar 

  2. Labaki WW, Rosenberg SR. Chronic obstructive Pulmonary Disease. Ann Intern Med. 2020;173:Itc17–32.

    Article  PubMed  Google Scholar 

  3. Houghton AM. Mechanistic links between COPD and lung cancer. Nat Rev Cancer. 2013;13:233–45.

    Article  CAS  PubMed  Google Scholar 

  4. Adeloye D, Song P, Zhu Y, Campbell H, Sheikh A, Rudan I. Global, regional, and national prevalence of, and risk factors for, chronic obstructive pulmonary disease (COPD) in 2019: a systematic review and modelling analysis. Lancet Respir Med. 2022;10:447–58.

    Article  PubMed  PubMed Central  Google Scholar 

  5. Liang Y, Sun Y. COPD in China: current Status and challenges. Arch Bronconeumol. 2022;58:790–1.

    Article  PubMed  Google Scholar 

  6. Safiri S, Carson-Chahhoud K, Noori M, Nejadghaderi SA, Sullman MJM, Ahmadian Heris J, Ansarin K, Mansournia MA, Collins GS, Kolahi AA, Kaufman JS. Burden of chronic obstructive pulmonary disease and its attributable risk factors in 204 countries and territories, 1990–2019: results from the global burden of Disease Study 2019. BMJ. 2022;378:e069679.

    Article  PubMed  PubMed Central  Google Scholar 

  7. Fazleen A, Wilkinson T. Early COPD: current evidence for diagnosis and management. Ther Adv Respir Dis. 2020;14:1753466620942128.

    Article  PubMed  PubMed Central  Google Scholar 

  8. Casaburi R, Duvall K. Improving early-stage diagnosis and management of COPD in primary care. Postgrad Med. 2014;126:141–54.

    Article  PubMed  Google Scholar 

  9. Yang T, Cai B, Cao B, Kang J, Wen F, Yao W, Zheng J, Ling X, Shang H, Wang C. REALizing and improving management of stable COPD in China: a multi-center, prospective, observational study to realize the current situation of COPD patients in China (REAL) – rationale, study design, and protocol. BMC Pulm Med. 2020;20:11.

    Article  PubMed  PubMed Central  Google Scholar 

  10. Haroon S, Jordan R, Takwoingi Y, Adab P. Diagnostic accuracy of screening tests for COPD: a systematic review and meta-analysis. BMJ Open. 2015;5:e008133.

    Article  PubMed  PubMed Central  Google Scholar 

  11. Rabe KF, Watz H. Chronic obstructive pulmonary disease. Lancet. 2017;389:1931–40.

    Article  PubMed  Google Scholar 

  12. Washko GR. Diagnostic imaging in COPD. Semin Respir Crit Care Med. 2010;31:276–85.

    Article  PubMed  PubMed Central  Google Scholar 

  13. Washko GR, Coxson HO, O’Donnell DE, Aaron SD. CT imaging of chronic obstructive pulmonary disease: insights, disappointments, and promise. Lancet Respir Med. 2017;5:903–8.

    Article  PubMed  Google Scholar 

  14. Li JS, Zhang HL, Bai YP, Wang YF, Wang HF, Wang MH, Li SY, Yu XQ. Diagnostic value of computed tomography in chronic obstructive pulmonary disease: a systematic review and meta-analysis. Copd. 2012;9:563–70.

    Article  PubMed  Google Scholar 

  15. Cho MH, Hobbs BD, Silverman EK. Genetics of chronic obstructive pulmonary disease: understanding the pathobiology and heterogeneity of a complex disorder. Lancet Respir Med. 2022;10:485–96.

    Article  CAS  PubMed  Google Scholar 

  16. Hosny A, Parmar C, Quackenbush J, Schwartz LH, Aerts HJWL. Artificial intelligence in radiology. Nat Rev Cancer. 2018;18:500–10.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  17. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015;521:436–44.

    Article  CAS  PubMed  Google Scholar 

  18. Esteva A, Robicquet A, Ramsundar B, Kuleshov V, DePristo M, Chou K, Cui C, Corrado G, Thrun S, Dean J. A guide to deep learning in healthcare. Nat Med. 2019;25:24–9.

    Article  CAS  PubMed  Google Scholar 

  19. González G, Ash SY, Vegas-Sánchez-Ferrero G, Onieva Onieva J, Rahaghi FN, Ross JC, Díaz A. San José Estépar R, Washko GR: Disease staging and prognosis in smokers using deep learning in chest computed Tomography. Am J Respir Crit Care Med. 2018;197:193–203.

    Article  PubMed  PubMed Central  Google Scholar 

  20. Du R, Qi S, Feng J, Xia S, Kang Y, Qian W, Yao YD. Identification of COPD from Multi-view snapshots of 3D lung Airway Tree via Deep CNN. IEEE Access. 2020;8:38907–19.

    Article  Google Scholar 

  21. Cosentino J, Behsaz B, Alipanahi B, McCaw ZR, Hill D, Schwantes-An T-H, Lai D, Carroll A, Hobbs BD, Cho MH, et al. Inference of chronic obstructive pulmonary disease with deep learning on raw spirograms identifies new genetic loci and improves risk models. Nat Genet. 2023;55:787–95.

    Article  CAS  PubMed  Google Scholar 

  22. Almeida SD, Norajitra T, Lüth CT, Wald T, Weru V, Nolden M, Jäger PF, von Stackelberg O, Heußel CP, Weinheimer O et al. Prediction of disease severity in COPD: a deep learning approach for anomaly-based quantitative assessment of chest CT. Eur Radiol 2023.

  23. Zhang L, Jiang B, Wisselink HJ, Vliegenthart R, Xie X. COPD identification and grading based on deep learning of lung parenchyma and bronchial wall in chest CT images. Br J Radiol. 2022;95:20210637.

    Article  PubMed  PubMed Central  Google Scholar 

  24. Xue M, Jia S, Chen L, Huang H, Yu L, Zhu W. CT-based COPD identification using multiple instance learning with two-stage attention. Comput Methods Programs Biomed. 2023;230:107356.

    Article  PubMed  Google Scholar 

  25. Polat Ö, Şalk İ, Doğan ÖT. Determination of COPD severity from chest CT images using deep transfer learning network. Multimedia Tools Appl. 2022;81:21903–17.

    Article  Google Scholar 

  26. Yang Y, Li W, Guo Y, Zeng N, Wang S, Chen Z, Liu Y, Chen H, Duan W, Li X, et al. Lung radiomics features for characterizing and classifying COPD stage based on feature combination strategy and multi-layer perceptron classifier. Math Biosci Eng. 2022;19:7826–55.

    Article  PubMed  Google Scholar 

  27. Yun J, Cho YH, Lee SM, Hwang J, Lee JS, Oh YM, Lee SD, Loh LC, Ong CK, Seo JB, Kim N. Deep radiomics-based survival prediction in patients with chronic obstructive pulmonary disease. Sci Rep. 2021;11:15144.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  28. Gökçen A. Computer-aided diagnosis system for Chronic Obstructive Pulmonary Disease using empirical Wavelet transform on Auscultation sounds. Comput J. 2021;64:1775–83.

    Article  Google Scholar 

  29. Altan G, Kutlu Y, Gökçen A. Chronic obstructive pulmonary disease severity analysis using deep learning on multi-channel lung sounds. Turkish Journal of Electrical Engineering and Computer Sciences. 2020, 28:2979–96.

  30. Altan G, Kutlu Y, Pekmezci AÖ, Nural S. Deep learning with 3D-second order difference plot on respiratory sounds. Biomed Signal Process Control. 2018;45:58–69.

    Article  Google Scholar 

  31. Ye S, Zeng P, Li P, Wang W, Xinan W, Zhao Y. MLP-Stereo: Heterogeneous Feature Fusion in MLP for Stereo Matching. In 2022 IEEE International Conference on Image Processing (ICIP); 16–19 Oct. 2022. 2022: 101–105.

  32. Zhang Y, Lobo-Mueller EM, Karanicolas P, Gallinger S, Haider MA, Khalvati F. Improving prognostic performance in resectable pancreatic ductal adenocarcinoma using radiomics and deep learning features fusion in CT images. Sci Rep. 2021;11:1378.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  33. Moll M, Qiao D, Regan EA, Hunninghake GM, Make BJ, Tal-Singer R, McGeachie MJ, Castaldi PJ, San Jose, Estepar R, Washko GR, et al. Machine learning and prediction of all-cause mortality in COPD. Chest. 2020;158:952–64.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  34. Makimoto K, Hogg JC, Bourbeau J, Tan WC, Kirby M. CT Imaging with Machine Learning for Predicting Progression to COPD in individuals at risk. Chest. 2023;164:1139–49.

    Article  PubMed  Google Scholar 

  35. Hofmanninger J, Prayer F, Pan J, Röhrich S, Prosch H, Langs G. Automatic lung segmentation in routine imaging is primarily a data diversity problem, not a methodology problem. Eur Radiol Experimental. 2020;4:50.

    Article  Google Scholar 

  36. van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, Beets-Tan RGH, Fillion-Robin JC, Pieper S, Aerts H. Computational Radiomics System to Decode the Radiographic phenotype. Cancer Res. 2017;77:e104–7.

    Article  PubMed  PubMed Central  Google Scholar 

  37. Kingma, D. P., & Welling, M. An Introduction to Variational Autoencoders. Foundations and Trends in Machine Learning. 2019;12(4):307-392. arXiv:1906.02691 [cs.LG].

  38. Trajković G. Measurement: Accuracy and Precision, Reliability and ValidityMeasurement: accuracy and precision, reliability and validity. In Encyclopedia of Public Health Edited by Kirch W. Dordrecht: Springer Netherlands; 2008: 888–892.

  39. Altan G. Deep learning-based Mammogram classification for breast Cancer. Int J Intell Syst Appl Eng. 2020;8:171–6.

    Article  Google Scholar 

  40. Altan G. DeepOCT: an explainable deep learning architecture to analyze macular edema on OCT images. Eng Sci Technol Int J. 2022;34:101091.

    Google Scholar 

  41. Muntean M, Militaru F-D. Metrics for evaluating classification algorithms. In; Singapore. Springer Nat Singap; 2023: 307–17.

  42. Aggarwal R, Sounderajah V, Martin G, Ting DSW, Karthikesalingam A, King D, Ashrafian H, Darzi A. Diagnostic accuracy of deep learning in medical imaging: a systematic review and meta-analysis. Npj Digit Med. 2021;4:65.

    Article  PubMed  PubMed Central  Google Scholar 

  43. Maselli DJ, Bhatt SP, Anzueto A, Bowler RP, DeMeo DL, Diaz AA, Dransfield MT, Fawzy A, Foreman MG, Hanania NA, et al. Clinical epidemiology of COPD: insights from 10 years of the COPDGene Study. Chest. 2019;156:228–38.

    Article  PubMed  PubMed Central  Google Scholar 

  44. Faner R, Tal-Singer R, Riley JH, Celli B, Vestbo J, MacNee W, Bakke P, Calverley PM, Coxson H, Crim C, et al. Lessons from ECLIPSE: a review of COPD biomarkers. Thorax. 2014;69:666–72.

    Article  PubMed  Google Scholar 

  45. Hastie AT, Martinez FJ, Curtis JL, Doerschuk CM, Hansel NN, Christenson S, Putcha N, Ortega VE, Li X, Barr RG, et al. Association of sputum and blood eosinophil concentrations with clinical measures of COPD severity: an analysis of the SPIROMICS cohort. Lancet Respir Med. 2017;5:956–67.

    Article  PubMed  PubMed Central  Google Scholar 

  46. Liang Z, Zhong N, Chen R, Ma Q, Sun Y, Wen F, Tal-Singer R, Miller BE, Yates J, Song J et al. Investigation of the clinical, radiological and biological factors Associated with Disease Progression, Phenotypes and endotypes of COPD in China (COMPASS): study design, protocol and rationale. ERJ Open Res 2021, 7.

  47. Wang F, Wei L. Multi-scale deep learning for the imbalanced multi-label protein subcellular localization prediction based on immunohistochemistry images. Bioinformatics. 2022;38:2602–11.

    Article  CAS  PubMed  Google Scholar 

  48. Zhu D, Qiao C, Dai H, Hu Y, Xi Q. Diagnostic efficacy of visual subtypes and low attenuation area based on HRCT in the diagnosis of COPD. BMC Pulm Med. 2022;22:81.

    Article  PubMed  PubMed Central  Google Scholar 

  49. Lange P, Celli B, Agustí A, Boje Jensen G, Divo M, Faner R, Guerra S, Marott JL, Martinez FD, Martinez-Camblor P, et al. Lung-function trajectories leading to Chronic Obstructive Pulmonary Disease. N Engl J Med. 2015;373:111–22.

    Article  CAS  PubMed  Google Scholar 

  50. Anees W, Moore VC, Burge PS. FEV1 decline in occupational asthma. Thorax. 2006;61:751–5.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  51. Kakavas S, Kotsiou OS, Perlikos F, Mermiri M, Mavrovounis G, Gourgoulianis K, Pantazopoulos I. Pulmonary function testing in COPD: looking beyond the curtain of FEV1. NPJ Prim Care Respir Med. 2021;31:23.

    Article  PubMed  PubMed Central  Google Scholar 

  52. Occhipinti M, Paoletti M, Bigazzi F, Camiciottoli G, Inchingolo R, Larici AR, Pistolesi M. Emphysematous and Nonemphysematous Gas Trapping in Chronic Obstructive Pulmonary Disease: quantitative CT findings and pulmonary function. Radiology. 2018;287:683–92.

    Article  PubMed  Google Scholar 

  53. Occhipinti M, Paoletti M, Bartholmai BJ, Rajagopalan S, Karwoski RA, Nardi C, Inchingolo R, Larici AR, Camiciottoli G, Lavorini F, et al. Spirometric assessment of emphysema presence and severity as measured by quantitative CT and CT-based radiomics in COPD. Respir Res. 2019;20:101.

    Article  PubMed  PubMed Central  Google Scholar 

  54. Shah PL, Herth FJ, van Geffen WH, Deslee G, Slebos DJ. Lung volume reduction for emphysema. Lancet Respir Med. 2017;5:147–56.

    Article  PubMed  Google Scholar 

  55. Venkatesan P. GOLD COPD report: 2023 update. Lancet Respir Med. 2023;11:18.

    Article  PubMed  Google Scholar 

  56. Rajpurkar P, Chen E, Banerjee O, Topol EJ. AI in health and medicine. Nat Med. 2022;28:31–8.

    Article  CAS  PubMed  Google Scholar 

  57. Amudala Puchakayala PR, Sthanam VL, Nakhmani A, Chaudhary MFA, Kizhakke Puliyakote A, Reinhardt JM, Zhang C, Bhatt SP, Bodduluri S. Radiomics for Improved Detection of Chronic Obstructive Pulmonary Disease in low-dose and standard-dose chest CT scans. Radiology. 2023;307:e222998.

    Article  PubMed  Google Scholar 

  58. Zhou T, Tu W, Dong P, Duan S, Zhou X, Ma Y, Wang Y, Liu T, Zhang H, Feng Y et al. CT-Based Radiomic Nomogram for the Prediction of Chronic Obstructive Pulmonary Disease in patients with lung cancer. Acad Radiol. 2023;30(12):2894-2903.

  59. Makimoto K, Hogg JC, Bourbeau J, Tan WC, Kirby M. CT Imaging With Machine Learning for Predicting Progression to COPD in Individuals at Risk. Chest. 2023;164(5):1139-1149.

  60. Gardner MW, Dorling SR. Artificial neural networks (the multilayer perceptron)—a review of applications in the atmospheric sciences. Atmos Environ. 1998;32:2627–36.

    Article  CAS  Google Scholar 

  61. McCaw ZR, Colthurst T, Yun T, Furlotte NA, Carroll A, Alipanahi B, McLean CY, Hormozdiari F. DeepNull models non-linear covariate effects to improve phenotypic prediction and association power. Nat Commun. 2022;13:241.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  62. Freijeiro-González L, Febrero-Bande M, Gonz’alez-Manteiga WJISR. A critical review of LASSO and its derivatives for variable selection under dependence among covariates. 2020, 90:118–45.

  63. MacNee W. Is chronic obstructive Pulmonary Disease an Accelerated Aging Disease? Ann Am Thorac Soc. 2016;13(Suppl 5):S429–37.

    Article  PubMed  Google Scholar 

  64. Barnes PJ, Burney PGJ, Silverman EK, Celli BR, Vestbo J, Wedzicha JA, Wouters EFM. Chronic obstructive pulmonary disease. Nat Reviews Disease Primers. 2015;1:15076.

    Article  PubMed  Google Scholar 

  65. Hikichi M, Mizumura K, Maruoka S, Gon Y. Pathogenesis of chronic obstructive pulmonary disease (COPD) induced by cigarette smoke. J Thorac Dis. 2019;11:S2129–40.

    Article  PubMed  PubMed Central  Google Scholar 

  66. Song Q, Chen P, Liu X-M. The role of cigarette smoke-induced pulmonary vascular endothelial cell apoptosis in COPD. Respir Res. 2021;22:39.

    Article  PubMed  PubMed Central  Google Scholar 

  67. Oh CM, Oh IH, Choe BK, Yoon TY, Choi JM, Hwang J. Consuming Green Tea at least twice each day is Associated with reduced odds of Chronic Obstructive Lung Disease in Middle-aged and older Korean adults. J Nutr. 2018;148:70–6.

    Article  PubMed  Google Scholar 

  68. Chan KH, Ho SP, Yeung SC, So WH, Cho CH, Koo MW, Lam WK, Ip MS, Man RY, Mak JC. Chinese green tea ameliorates lung injury in cigarette smoke-exposed rats. Respir Med. 2009;103:1746–54.

    Article  PubMed  Google Scholar 

Download references


Not applicable.


This study was supported in part by grants 2020E10004 from the Key Laboratory of Intelligent Preventive Medicine of Zhejiang Province, 2019R01007 from the Leading Innovative and Entrepreneur Team Introduction Program of Zhejiang, 2020C03002 from the Key Research and Development Program of Zhejiang Province, K20230085 from the Healthy Zhejiang One Million People Cohort, 82203984 from the National Natural Science Foundation for Young Scientists of China, and 2023-3-001 from the Major Research Plan of Jinhua. The funding source had no role in study design, data collection, data analysis, data interpretation, writing of the report, or the decision to submit the article for publication.

Author information

Authors and Affiliations



XW and GC conceived and designed the study, interpreted the data. ZZ, SZ, JL, XL, ZL, and WL analyzed and interpreted the data. ZZ, JL, WL, and XW drafted the first version of the manuscript. JL, YW, YJ, ZL, XL, WL, GC, and XW interpreted the data and critically reviewed the manuscript. All authors contributed to the content and critical revision of the manuscript and agreed to submit the manuscript for publication.

Corresponding authors

Correspondence to Gang Chen or Xifeng Wu.

Ethics declarations

Ethics approval and consent to participate

All participants provided written informed consent. The study was approved by the Ethics Committee of the Second Affiliated Hospital in Zhejiang University School of Medicine. The study was conducted in accordance with the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit The Creative Commons Public Domain Dedication waiver ( applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Zhu, Z., Zhao, S., Li, J. et al. Development and application of a deep learning-based comprehensive early diagnostic model for chronic obstructive pulmonary disease. Respir Res 25, 167 (2024).

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: