Background

Chronic obstructive pulmonary disease (COPD) is a common pulmonary disease characterized by persistent respiratory symptoms and airflow limitation that is due to airway and/or alveolar abnormalities mainly caused by cigarette smoking [1]. Despite many decades of research on the pathogenesis and treatment of COPD, the medical community has failed to decrease its morbidity and mortality to the same degree that has been achieved in other major noncommunicable diseases, An important factor contributing to this slow progress may be that he previous COPD definition showed limitations as the lack of identification of the disorder at its early stages in the absence of flow limitation [2]. The proposed solutions are aimed to encourage novel treatments and translational studies: incorporating into the definition objectivable early computed tomography (CT) scan changes [3]. Epidemiological studies have shown that COPD is the third leading cause of death and accounts for 5% of all deaths worldwide each year [4]. The overall prevalence of COPD in people aged 40 years and older was 12.64% (95% CI 10.75%-14.65%) and 7.38% [5]. About a quarter of adults over 40 years have moderate airflow limitations, yet most of them are unaware of their conditions [6]. Early diagnosis of COPD is critical for early self-management and timely therapy to improve the overall prognosis [7, 8]. However, a considerable proportion of COPD patients are undiagnosed. An estimate of over 40% of COPD patients remain undiagnosed, particularly in develo** countries [9, 10], and only 12% of individuals with chronic airflow limitations had a previous spirometry-defined COPD diagnosis during the recent screening of 57,779 participants in China [11].

The conventional diagnosis and staging measure of COPD is pulmonary function test (PFT) according to the Global Initiative for Chronic Obstructive Lung Disease (GOLD) diagnostic criteria [12]. However, spirometry strongly depends on patients’ cooperation and COPD can either be misdiagnosed or missed entirely when using spirometry alone [13]. The 2017 GOLD report recommended that therapy should be based on clinical criteria rather than isolated PFT [14]. Furthermore, due to the shortage of experienced spirometry experts in poor areas of develo** countries, it’s hard to use PFT to screen asymptomatic patients in regular health examination on a large scale. Consequently, the need to develop new tools for early detection of COPD arises, and cost-effective strategies for case-finding are urgently needed.

In the past few years, growing evidences have shown that chest quantitative computer tomography (CT) has potential in COPD diagnosis and stratification [15,16,17,18]. CT-based imaging can help improve COPD detection and evaluation in patients who cannot undergo PFT [19]. In a recent study, deep learning (DL) models that utilize computed tomography (CT) image data were developed for automated detection and staging of spirometry-defined COPD. The result showed chest CT-DL approach could automatically identify spirometry-defined COPD and categorize patients according to the GOLD scale [11]. In the research reported by Lin Zhang, et al., they trained and tested the deep convolutional neural network (CNN) based on CT images of lung parenchyma and bronchial wall to determine the presence of COPD and GOLD staging, using PFT as reference, so as to infer lung function and determine the existence and severity of COPD. The result demonstrated that CNN can identify emphysema and airway wall remodeling on CT images to infer lung function and determine the existence and severity of COPD. As the CNN reached AUCs of 0.853 to determine the presence of COPD in the training and external test cohorts, and the accuracies of CNN to determine COPD GOLD grade in three- and five- classifications were 77.4 and 67.9%, respectively [27,28]. It’s highly accurate for advanced emphysema [29], but only moderately sensitive in patients with mild to moderate emphysema [30,31,32]. Recently, machine learning (ML) technology is being assessed to perform medical tasks in almost every field of practice [33]. It has been successfully used in automated interpretation of PFT for differential diagnosis of obstructive lung diseases and COPD detection based on HRCT images [34,35,36]. However, none of the previous studies have used deep learning (DL) to predict COPD based on CXR images.

In the present study, we used DL algorithms to detect COPD and predict disease severity based on CXR imaging and clinical parameters, with the purpose to screen potential COPD patients while minimizing the need for additional radiographic examination. We hypothesized that applying DL algorithms to clinical and CXR imaging features would improve early diagnosis and prognosis prediction in COPD.

Methods

Study population

This was a multicenter retrospective study performed at the Third Affiliated Hospital of Sun Yat-sen University, the Third Affiliated Hospital of Sun Yat-sen University. Lingnan Hospital, the Six Affiliated Hospital of Guangzhou Medical University, Qingyuan People’s Hospital, and Huazhou Hospital of Traditional Chinese Medical. The data included frontal CXR images and clinical information of 1055 participants (535 patients with COPD and 520 controls) from outpatient, inpatient, and physical examination center settings between January 2019 and December 2021. This study was reviewed and approved by the Ethics Committee of the Third Affiliated Hospital of Sun Yat-sen University and requirements for written informed consent were waived due to the retrospective nature of the research.

Frontal images were identified by searching image databases for CXRs of the patients who also received PFT within one week, whereas lateral radiographs and oblique views were excluded. COPD diagnosis was confirmed by forced expiratory volume in 1 s (FEV1) to forced vital capacity (FVC) ratio less than 0.7 after inhalation of bronchodilators according to GOLD 2018. The severity of COPD is graded as GOLD 1 (FEV1%pred ≥ 80%), GOLD 2 (50%≤FEV1%pred < 80%), GOLD 3 (30%≤FEV1%pred < 50%), and GOLD 4 (FEV1%pred < 30%), as depending on the FEV1%pred value of PFT based on GOLD 2018. Since the diagnosis and severity of COPD depend on GOLD level based on PFT result, this study used the GOLD level as a reference standard to classify patients.

As COPD usually occurs in patients older than 40 years [6], all the subjects included in the present study were > 40 years old. The exclusion criteria were: (1) pregnant women; (2) other pulmonary diseases with abnormal CXR presentations, such as bronchiectasis, pulmonary fibrosis, atelectasis, pulmonary infectious disease, active pulmonary tuberculosis, pleural effusion, lung cancer, and pneumothorax; (3) severe renal insufficiency, severe liver disease, human immunodefciency virus, or other immune-related diseases; (4) previous chest surgery; (5) severe cardiac insufficiency; (6) in the acute phase of COPD. A total of 1224 subjects ranging in age from 41 to 86 years were recruited. Spirometry data, demographic information, smoking history, clinical indices and manifestations information were collected using a standardized data collection form. After excluding cases with incomplete clinical data (57 cases), substandard pulmonary function (42 cases), and poor CXR image quality (70 cases), finally 1055 participants were enrolled in the cohort. The 950 subjects recruited from the Third Affiliated Hospital of Sun Yat-sen University and the Third Affiliated Hospital of Sun Yat-sen University. Lingnan Hospital were randomly split into training set and internal test set, with a ratio of 70%:30% (Fig. 1). For the training set (n = 666), 49.4% were COPD patients (n = 329). Of the 284 subjects in internal test set, 49.3% (n = 140) were COPD patients. The 105 participants (66 COPD patients and 39 control subjects) from the Six Affiliated Hospital of Guangzhou Medical University, Qingyuan People’s Hospital, and Huazhou Hospital of Traditional Chinese Medical were used for external test set.

Demographic and clinical characteristics

A total of 1055 participants were finally included in the study: 535 COPD patients and 520 control subjects. The median age of COPD patients was higher than that of non-COPD participants (67 vs. 63, P < 0.001) and the majority of the COPD cohort was male (87.85%), which was consistent with COPD gender distributions in China [6]. A higher proportion of smokers (78.50% vs. 20.96%, P < 0.001), a reduced FEV1% (55.02% vs. 92.33%, P < 0.001), and a lower BMI (22.10 vs. 24.13, P = 0.001) were evident among COPD patients compared to control group. In addition, the symptoms of cough, sputum, and dyspnea were more common in COPD patients than in control subjects (77.94%, 51.40%, 50.47% vs. 13.46%, 8.46%, 2.31%, respectively). The percentages of stage 1, 2, 3, and 4 spirometry-defined COPD subjects on the GOLD scale were 18.50%, 42.06%, 25.61%, and 13.83%, respectively. Detailed demographic and clinical characteristics for the participants were provided in Table 1.

Table 1 Demographic and clinical characteristics
Full size table

Data preprocessing

The data set consisted of clinical information and CXR images. The clinical information contained nine characteristics in all: gender, age, average body mass index (BMI), history of smoking, cough, expectoration, carbon dioxide retention, (arterial partial pressure of carbon dioxide greater than 45mmHg), dyspnea, respiratory failure (arterial partial oxygen pressure less than 60mmHg, with or without elevated carbon dioxide levels) and so on. Identity information including patient’s name, hospitalization ID and other DICOM information were removed from CXR images. In two classifications, the data were divided into normal control subjects without COPD (n = 520) and patients with COPD (n = 535). In three-classifications, the data were divided into three groups, normal (n = 520), GOLD 1/2 (n = 318), and GOLD 3/4 (n = 217), while in five-classifications, the data were divided into five groups, including normal (n = 520), GOLD 1 (n = 99), GOLD 2 (n = 225), GOLD 3 (n = 137), and GOLD 4 (n = 74).

Development of the COPD detection and COPD staging model

We first studied the comparison between normal control subjects and patients with COPD based on clinical data only. Clinical information were preprocessed through sk-learn’s data preprocessing interface StandardScaler, by filling in the missing values and standardizing the data to (-1, 1) interval. Three classical machine learning algorithms, decision tree, support vector machine and random forest were selected for modeling. The random forest algorithm showing best area under the curve (AUC) was determined to be the best algorithm for feature mining and extraction of clinical inform clinical information. For classification of CXR images, we generally chose transfer learning for modeling. Three classical deep learning convolution neural networks such as EfficientNet-B5, ResNet50 and DenseNet were selected for modeling [37,

Discussion

In this multicenter retrospective study, we used deep learning algorithms to predict COPD based on clinical parameters and CXR images, with the purpose to find a more sensitive and simple option for early detection of potential COPD patients. The diagnostic accuracies of COPD reached an AUC of 0.965 in internal test and 0.934 in external test to detect COPD according clinical characteristics and CXR images, better than the results of previous study, which used deep residual networks for automated detection of COPD based on low dosed CT images only, with an ACU of 0.889 [4] and 0.899 [49]. Through the gradcam interface of the EfficientNet model, we can get the hot focus areas of the CNN model for the diagnosis of COPD patients. The results indicated that the model paid specific attention to these lesions when distinguishing COPD subjects. This could make doctors and patients better understand what they have learned from DL, and whether they can rest assured that ML can assist them in diagnosis.

There are several limitations in the present study. First, this was a retrospective study only performed in four medical institutes. The outer generalization of the deep learning algorithms needed to be tested further by prospective study including more centers. Second, the prediction was performed based on COPD and normal control. Subjects involved in the present study were mainly COPD patients without other complications or disease, as radiographic images with other lesion were eliminated. For better use in clinical practice, subsequent study with larger sample size including complication or comorbidities of COPD is needed. Third, a limited number of subjects with GOLD 1 and 4 were enrolled, which might constrain the model’s stratification capacity and resulted in the discrepancy of staging efficiency between groups. To improve the efficiency of detection and staging, we are currently recruiting more participants and aim to optimize our cohort in the future.

In conclusion, we developed a more simple, sensitive and safer learning approach for detecting and staging of COPD. The proposed model approach achieved the desired performance and could serve as a powerful tool for COPD screening and evaluation, which may help clinicians easily identify possible suspected COPD patients. Nevertheless further studies are necessary to determine the feasibility of these outcomes in a prospective clinical setting.