Detection and airflow obstruction severity grading of chronic obstructive pulmonary disease based on capnography

Chronic obstructive pulmonary disease (COPD) is a respiratory disease characterized by irreversible airflow limitation,¹ and as the third leading cause of death worldwide, COPD is increasingly being recognized as a public health concern with high morbidity and mortality.² To slow the progression and reduce flareups of this disease, early diagnosis including screening and grading is of great importance.

The current gold standard for detecting COPD is spirometry. According to the Global Initiative for Chronic Obstructive Lungs Disease (GOLD) criteria,³ COPD is considered when the post-bronchodilator FEV1/FVC (the ratio of forced expiratory volume to forced vital capacity) is less than 70%, and airflow obstruction severity is defined according to GOLD grades based on FEV1% predicted (FEV1%pred). However, spirometry can only be used under strict criteria and with patient cooperation,⁴ so it is difficult to use this diagnosis method with unconscious patients, children, and those who are suffering from severe respiratory conditions. The ensuing low-quality results are unacceptable and increase the risk of inappropriate treatment, so there is an urgent need for effective, noninvasive, and effort-independent tools for evaluating pulmonary functions.

In recent years, capnography (Capno) has received widespread attention because of its monitoring ability during natural breathing and it being easier to perform than conventional pulmonary function tests (PFTs). Capno can be used to describe the CO₂ concentration or partial pressure in the expiratory phase as a function of time or exhaled air volume, giving additional information about the ventilator perfusion dynamics of the lungs.

Quantitative features derived from Capno are recognized as effective indicators for assessing respiratory functions,⁵ and spirometry parameters such as FEV1/FVC and FEV1%pred are crucial for diagnosing COPD and evaluating the ventilatory function of patients. Related studies to date have mostly been focused on dataset construction based on volumetric capnography (VCap) or time-based capnography (TCap) individually to investigate their relationships with conventional pulmonary function parameters. Via path analyses, Kellerer et al. identified four key parameters of VCap (slope in phase 3, ratio of slopes in phases 3 and 2, volume in phase 2, and area/volume ratio in phase 3) that reflect various aspects of the lungs, including malignant inflation, overall ventilatory impairment, bronchoconstriction, and ventilated lungs volume.⁶ Additionally, Pertzov et al. leveraged the waveform characteristics of TCap to establish an FEV1 regression model that can be used to assess the severity of airway obstruction in patients with lungs diseases.⁷ Those studies collectively demonstrate the value of Capno parameters in discriminating pulmonary functions comparable to spirometry. Moreover, Capno provides a distinct perspective for assessing pulmonary functions, allowing for standards independent of such conventional indicators as FEV1/FVC and FEV1%pred.

As an effective computational analysis tool, machine learning (ML) is now used widely in respiratory medicine.⁸ Based on continuous Capno monitoring, some scholars have used ML models such as convolutional neural networks⁹ and quadratic discriminant analysis¹⁰ and have input signals directly into these models. This has demonstrated the efficacy of Capno in tasks such as discriminating between COPD-Normal and COPD-CHF patients,^11–13 and the results indicate that as an alternative or complementary tool for personalized detection and treatment of respiratory diseases, Capno has significant potential.

However, despite advances, Capno has not yet reached its full potential in the appropriate grading of COPD severity. Previous related studies were mostly focused on discriminating between COPD and other conditions, which can only indicate the presence of COPD but cannot provide information for airflow obstruction assessment.¹⁴ Also, there is still a lack of a consensus on the optimal approach to taking full advantage of Capno. Only a few parameters of TCap or VCap are considered, and most investigators analyze the correlation of either one of them with the spirometry parameter,^15,16 which ignores the value of Capno as an independent factor in the diagnosis of COPD.

Our aim is to investigate the pivotal role of quantitative features derived from Capno in the assessment of respiratory functions. In this work as shown in Fig. 1, five different classification models for assessing airflow obstruction are evaluated. TCap and VCap are analyzed separately, and a subset including abundant characteristics is also constructed. Furthermore, a hybrid feature-selection scoring strategy is formulated to identify the significant features for improved classifier performance. Unlike conventional PFTs, Capno may be used for the diagnosis of COPD and the grading of airflow obstruction in cases where the test criteria cannot be met. These advancements are expected to be applied in clinical practice, thereby enhancing the outcomes of patients in the field of respiratory healthcare.

Capno consists of TCap and VCap, which imply the process of pulmonary ventilation and gas exchange from two aspects. Figure 2 shows the curves of TCap and VCap in one respiratory cycle. TCap can be divided into four phases: phase I represents the CO₂-free gas from the airways (anatomical and apparatus dead space); phase II is rapidly S-shaped because of the mixing of dead-space gas with alveolar gas; phase III is an alveolar plateau representing CO₂-rich gas from the alveolar; phase IV represents the concentration of CO₂ dropping sharply to form the descending limb at the end of expiration. VCap can be divided into three phases: phase I is the baseline, representing the CO₂-free gas from the dead space in the early exhalation; phase II is the ascending branch, where dead-space and alveolar gas are mixed and exhaled, and CO₂ concentration increases; phase III is the plateau, which is almost horizontal in normal conditions with a high concentration of CO₂. TCap and VCap yield similar morphology and physiological mechanisms. Capno approximates an asymmetrical sigmoid curve in the expiratory phase with CO₂ elimination.

Deriving meaningful features from original Capno is crucial for airflow obstruction assessment in patients with COPD. However, the lack of consensus on quantitative analysis methods for Capno (TCap or VCap) has limited the understanding of curves. In this study, the specific features that can characterize the complete Capno were extracted from the chaotic information to perform quantitative analyses, including demographic information, morphological features (such as slope, angle, and the area under the curve), and physiological features (such as dead-space volume, expiratory rate, and tidal volume) in TCap and VCap within an effective respiratory cycle. Also, time–frequency domain analysis was performed on the signals (such as segmented wavelet decomposition, and short-time Fourier transform). In practice, the continuous features were further processed by taking logarithms and variance distributions. In total, 110 features were extracted from the raw data.

The features were ranked in descending order by score, with a higher score indicating greater suitability. We set a threshold of 4 based on model performance, and we selected the features above that threshold as the final subset, obtaining 26 features to assess ventilation dysfunction in COPD patients. Table I summarizes the selected features.

The selected features can be categorized into three main groups: demographic features, VCap features, and TCap features. Demographic features including weight, age, BMI, height, and gender provide essential background information about the patients, being closely related to lungs function and COPD risk. VCap features such as the ratio of exhaled CO₂ concentrations (Ratio_VM5075_VT, Ratio_VM2550_VT), alveolar tidal volume (Valv), standard deviation and minimum of expiratory volume (V_std, V_min), tidal volume (VT_R), volume of CO₂ in phase II (VCO2_II), and various dead-space volumes (VDaw, VDW, VDT) reflect changes in CO₂ concentration, ventilation efficiency, and lungs volume; these features are crucial for assessing ventilation-perfusion efficiency and identifying airway obstruction. TCap features including the standard deviation and variance of flow (Flow_std, Flow_var), the angle between phases II and III (Angle), and the activity levels during phases II and III (HP1_Activity, HP2_Activity) capture the dynamic changes in airflow and CO₂ concentration during the respiratory cycle. Higher variability and instability in airflow indicate fluctuating airway resistance and lungs compliance, which are characteristic of COPD. By combining these features, we can comprehensively assess COPD severity, enhance the model’s predictive performance, and improve clinical interpretability.

This study involved a comprehensive selection of classic ML algorithms as benchmark models, i.e., logistic regression (LR), support vector machine (SVM), k-nearest neighbors (KNN), random forest (RF), and extreme gradient boosting (XGBoost).^19–23 These algorithms represent diverse learning paradigms and modeling approaches, encompassing aspects of supervised learning, ensemble learning, and probabilistic modeling. The primary objective was to rigorously evaluate their efficacy in the diagnosis and classification of COPD. LR is a widely used linear model for classification, SVM excels in classification by finding the optimal hyperplane in high-dimensional space, and KNN can perform classification based on the distance between instances. RF can enhance performance by ensemble learning with multiple decision trees, while XGBoost can construct a powerful ensemble model iteratively. Their selection was based on their distinct advantages: interpretability (LR), robustness in high-dimensional spaces (SVM), simplicity and effectiveness in capturing local patterns (KNN), ensemble robustness (RF), and advanced regularization with computational efficiency (XGBoost). These algorithms were collectively applied to the baseline models in this study, providing a performance benchmark for the proposed methods. In the results analysis, their performance on different tasks and datasets was assessed to comprehensively evaluate their strengths and weaknesses.

In this study, patients with respiratory diseases who received treatment in the pneumology department with China–Japan Friendship Hospital approval over four months were selected as research subjects. Participants included but were not limited to those with respiratory diseases such as COPD, allergic asthma, interstitial lungs disease, and pulmonary fibrosis and those for preoperative screening. All subjects underwent Capno tracing and then accepted PFTs by a Masterscreen multifunctional pulmonary function instrument (JEGER, Germany); the standard results from JEGER were recorded as labels.

In total, 1007 participants comprising 535 males and 472 females completed the experiment, as listed in Table II. All participants provided written informed consent prior to their inclusion in the study. Approximately 25% of these participants presented normal pulmonary functions, while the remaining individuals exhibited pulmonary dysfunction with varying degrees, such as small-airway obstruction, restrictive ventilation impairment, and reduced spirometry. Subjects with abnormal pulmonary functions underwent bronchodilator inhalation before subsequent PFTs, strictly adhering to clinical norms. Records not meeting the specified criteria were excluded.

Capno was repeated on multiple occasions to ensure both validity and consistency. In total, 2983 breathing cycles were recorded, and patients were categorized based on established clinical practice. Following the GOLD guidelines, patients with COPD were classified into four levels (mild, moderate, severe, and very severe) based on the severity of airflow obstruction. The comprehensive exploration of the diagnostic value of Capno in COPD focused on the performance metrics for COPD detection and the grading of airflow obstruction in COPD patients. Ultimately, 1207 recordings were collected from both COPD (148) and normal subjects (279) and incorporated into the analysis. Ethical approval for data collection was obtained from committees in strict adherence to local regulations, and all participants provided written informed consent.

As shown in Fig. 3, the Capno of normal subjects and COPD patients shows different patterns, such as the slope of the ascending phase and its angle to the plateau phase. The same feature extraction was used for each recording, and the hybrid feature-selection scoring strategy was performed before input to the classifiers.

Here, we compare the performances of the five ML models (LR, SVM, KNN, RF, and XGBoost) in assessing ventilatory function. For each model, the optimal combination of hyperparameters was found using grid search. The performances of the models on the test set are given in Table III, and as can be seen, XGBoost significantly outperforms the other ML models in discriminating airflow obstruction, achieving accuracy, precision, recall, and F1 score of 90.08%, 90.33%, 90.08%, and 89.72%, respectively. Figure 4 shows the XGBoost confusion matrix, demonstrating that its ability to discriminate between moderate and severe airflow obstruction in COPD patients is slightly higher than its ability to detect mild and very severe patients.

We also analyze the performance of each model in COPD diagnosis as given in Table IV, and as can be seen, XGBoost remains significantly superior to the other algorithms. In tenfold cross-validation, XGBoost achieved an AUC of 0.96 on the test set (Fig. 5). XGBoost outperformed the other models because of its robust ensemble learning capabilities and efficient handling of complex data relationships. Its ability to capture nonlinear interactions and perform well with imbalanced datasets likely contributed to its superior performance in diagnosing and assessing COPD. In contrast, models such as SVM and KNN may have struggled with the complexity and high dimensionality of the features extracted from Capno.

Selecting quantitative features in Capno is a long-standing research focus. To validate the enhancement effect of the hybrid feature-selection scoring strategy, we compare the results of models using this strategy with those of the basic models. As shown in Fig. 6, significant performance improvements are achieved in both tasks when the hybrid scoring strategy is used to discriminate between normal-COPD and ventilatory function assessment in COPD patients. We set thresholds for the selection of representative head features. Specifically, when using all 26 selected features, XGBoost achieved the best performance. Compared to the F1 scores of 85.08% in COPD detection and 83.52% in airflow obstruction assessment without feature selection, our proposed feature selection strategy gives an approximate 6% improvement in overall performance.

We investigate the application of Capno in the assessment of COPD, focusing on two curves: VCap and TCap. With the aim of exploring the performance of different data sources in the diagnosis of COPD and the assessment of ventilatory function, we evaluate the use of VCap and TCap individually and in combination. Table V indicates that in airflow obstruction assessment, using VCap alone is better than using TCap alone. In particular, when VCap and TCap are used in combination, the model shows a more significant performance advantage.

In this study, we used five different ML models for assessing lungs ventilation function, and XGBoost excelled in intricate ventilation function assessment and straightforward COPD diagnosis. We attribute XGBoost’s superior performance to its robust capabilities in ensemble learning and efficient fitting of complex data relationships. Physiologically, XGBoost is adept at capturing the nonlinear relationships between physiological features in Capno, and given the influence of multiple physiological factors on lungs ventilation function, this flexibility is crucial. Moreover, discrepancies in model performance were observed in assessing airflow obstruction in COPD patients; these stem from the relatively few samples for very severe patients in the available dataset. Also, it was noted that for patients with mild ventilation disorders, their Capno morphology closely resembled that of normal individuals, leading to a propensity for misclassification. We emphasize the utilization of the complete Capno, encompassing TCAP and VCap. Results from VCap indicate that Capno exhibits significant advantages in multiple aspects when integrating VCap and TCap features compared to conventional single-feature analysis. This integration furnishes volumetric information regarding patient ventilation and perfusion, while TCap contributes temporal information concerning the respiratory cycle and gas exchange. A complementary relationship exists between VCap and TCap, and the synergistic application of these two signals aids in detecting abnormalities such as hypoventilation and dead-space ventilation.

Many researchers have used physiological signals such as lungs sounds, computed tomography (CT), digital oximetry biomarkers, and nighttime features to avoid misdiagnosis due to PFT failures.^24–27 Although those researchers tried to apply ML to the diagnosis or severity grading of COPD, agreement is yet to be reached. Table VI compares our work against related studies; we considered the F1 score as an evaluation indicator and included only those who completed the assessment of ventilatory function. Spina et al.²⁷ reported a multi-sensor detection system that modeled nighttime topics to discriminate healthy subjects from those with COPD; however, even though multiple classification signals were involved in the model’s pre-training, the final model’s performance was still unsatisfactory, achieving an F1 score of 0.78 for classification in the four COPD grades. Levy et al.²⁶ used an RF classifier to diagnose COPD based on digital oximetry biomarkers and polysomnography; the F1 score for the RF classifier on the test set was 0.89, similar to our work, but acquiring oximetry biomarkers for sleep polysomnography still requires patient cooperation and adequate medical conditions.

Our work has shown that Capno appears to be an objective and valid direct description of lungs function. The proposed method may help clinicians to understand the ventilatory function of patients who are unable to perform PFTs. Compared to previous methods, the results of this study improve the understanding of Capno in the diagnosis and classification of COPD, outperforming other physiological information. This work indicates that Capno combined with an ML model such as XGBoost can effectively diagnose and assess COPD. This approach offers a noninvasive and cost-effective alternative to traditional PFTs, making it suitable for use in various clinical settings including primary healthcare and remote monitoring. However, potential barriers such as the need for standardized protocols, training for healthcare providers, and integration with existing healthcare systems must be addressed in order to facilitate widespread adoption.

To the best of our knowledge, this is the first scientific effort to demonstrate the feasibility of extracting features from Capno in order to evaluate airflow obstruction severity in patients with COPD. The results presented herein confirm the reliability of Capno as a cost-effective and portable method for COPD management, and as a single indicator, Capno can play a role in the initial classification of COPD. Compared with other conventional supervised ML algorithms, XGBoost was validated as a helpful tool for airflow obstruction assessment in COPD, and the proposed feature-selection strategy exhibited favorable performance in model learning, improving the classification accuracy. XGBoost has good interpretability and offers physicians convenient understanding of Capno.

In conclusion, the method proposed herein can accurately assess the severity of patients with COPD, thereby promoting the large-scale screening of COPD in primary healthcare institutions. These findings offer a practical guide for classifying COPD based on Capno using ML. Furthermore, Capno as an alternative to traditional lungs function is expected to provide decision support in diagnosing COPD.

This study was supported by the National Natural Science Foundation of China (Grant Nos. 62331025, 62071451, and U21A20447), the National Key Research and Development Project (Grant No. 2021YFC3002204), the Special Equipment Scientific Research Key Project (Grant No. LB2020LA060003), and the CAMS Innovation Fund for Medical Sciences (Grant No. 2019-I2M-5-019).

The author has no conflicts to disclose.

The data that support the findings of this study are available from the corresponding authors upon reasonable request.

Xiuying Mou received a bachelor’s degree from Yunnan University in 2018 and is currently a Ph.D. candidate at the Aerospace Information Research Institute of the Chinese Academy of Sciences. Her research is focused on bio-health sensing, with a particular emphasis on the development of algorithms for COPD diagnosis and exercise health monitoring.

Peng Wang received a B.E. degree from the School of Information Science and Engineering, Shandong University, China, in 2017, and an M.E. degree from the University of Chinese Academy of Sciences, China, in 2020. From 2020 to the present, he was a Research Assistant with the Aerospace Information Research Institute, Chinese Academy of Sciences. His current research interests include emergency medical information processing, biomedical signal processing, and Research on multi-channel bioelectrical equipment.

Xianya Yu received a Bachelor degree in measurement and control technology and instrumentation from Zhengzhou University, Zhengzhou, Henan, China. She is currently working toward her Ph.D. degree at the Aerospace Information Research Institute, Chinese Academy of Sciences. Her research interests were biomedical signal processing artificial intelligence algorithm.

Sun Jie, a master in the field of electronic information from the University of Chinese Academy of Sciences, has been the CTO of Zhongke Guangrun (Zhongshan) Technology Co., Ltd. since 2020. His main research fields are intelligent perception and computing, multi-source information fusion and intelligent analysis, biometrics and identity recognition.

Xianxiang Chen received a Ph.D. degree from the Institute of Electrics, Chinese Academy of Sciences. He is an associate researcher at the Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China. His study focuses on wearable devices for physiological parameter measurement and biomedical signal processing based on machine learning.

Lidong Du received a Ph.D. degree from the Institute of Electrics, Chinese Academy of Sciences. He is an associate researcher at the Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China. He is a member of the China Society of Micron and Nano Technology. His research interests include flexible sensing technology, and medical and environmental signal acquisition and processing technology.

Zhenfeng Li received the Ph.D. degree in biomedical engineering from the Beijing Institute of Technology, Beijing, China, in 2022. He is currently an assistant research fellow at the Aerospace Information Research Institute, Beijing, China. His research interests include wearable devices development, biomedical signal processing, and artificial intelligence in medicine.

Xia Jingen received a M.D. from Beihang University. He is a Chief respiratory therapist at China-Japan Friendship Hospital and specializes in respiratory physiology, mechanical ventilation, airway management and pulmonary rehabilitation.

Xiaoran Li received the M.D. and Ph.D. degree from the Southern Medical University, Guangzhou, China. She is currently a doctor in the department of cardiology, Beijing Friendship Hospital, Capital Medical University, focusing on the management of arrhythmia, and cardiovascular related chronic diseases such as hypertension, hyperlipidemia. She has published several academic papers. Her research fields are mainly about myocardial reperfusion and imaging analysis, endocrine metabolism, and other fields.

Zhen Fang received a Ph.D. degree from the Institute of Electrics, Chinese Academy of Sciences. He is a professor at the Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China. His research interests include wearable device development, medical artificial intelligence technology, and medical big data platform management technology.