CT image-based machine learning models for predicting blood eosinophil levels in acute exacerbation of chronic obstructive pulmonary disease - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice BMC Pulm Med . 2026 Mar 10;26:179. doi: 10.1186/s12890-026-04219-w Search in PMC Search in PubMed View in NLM Catalog Add to search CT image-based machine learning models for predicting blood eosinophil levels in acute exacerbation of chronic obstructive pulmonary disease Shuiqing Zhao Shuiqing Zhao 1 College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, China 2 Key Laboratory of Intelligent Computing in Medical Image, Ministry of Education, Northeastern University, Shenyang, China Find articles by Shuiqing Zhao 1, 2 , Yanan Wu Yanan Wu 3 School of Health Management, China Medical University, Shenyang, China Find articles by Yanan Wu 3 , Lirong Du Lirong Du 4 Department of Respiratory and Critical Care Medicine, Shanxi Bethune Hospital, Shanxi Academy of Medical Sciences, Tongji Shanxi Hospital, Third Hospital of Shanxi Medical University, Taiyuan, China Find articles by Lirong Du 4 , Hui Jia Hui Jia 5 Department of Respiratory and Critical Care Medicine, The Affiliated Center Hospital of Shenyang Medical College, Shenyang, China Find articles by Hui Jia 5 , Patrice Monkam Patrice Monkam 1 College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, China Find articles by Patrice Monkam 1 , Wei Qian Wei Qian 1 College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, China Find articles by Wei Qian 1 , Ruiying Wang Ruiying Wang 4 Department of Respiratory and Critical Care Medicine, Shanxi Bethune Hospital, Shanxi Academy of Medical Sciences, Tongji Shanxi Hospital, Third Hospital of Shanxi Medical University, Taiyuan, China Find articles by Ruiying Wang 4, ✉ , Shuyue Xia Shuyue Xia 5 Department of Respiratory and Critical Care Medicine, The Affiliated Center Hospital of Shenyang Medical College, Shenyang, China Find articles by Shuyue Xia 5, ✉ , Shouliang Qi Shouliang Qi 1 College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, China 2 Key Laboratory of Intelligent Computing in Medical Image, Ministry of Education, Northeastern University, Shenyang, China Find articles by Shouliang Qi 1, 2, ✉ Author information Article notes Copyright and License information 1 College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, China 2 Key Laboratory of Intelligent Computing in Medical Image, Ministry of Education, Northeastern University, Shenyang, China 3 School of Health Management, China Medical University, Shenyang, China 4 Department of Respiratory and Critical Care Medicine, Shanxi Bethune Hospital, Shanxi Academy of Medical Sciences, Tongji Shanxi Hospital, Third Hospital of Shanxi Medical University, Taiyuan, China 5 Department of Respiratory and Critical Care Medicine, The Affiliated Center Hospital of Shenyang Medical College, Shenyang, China ✉ Corresponding author. Received 2025 Mar 26; Accepted 2026 Feb 27; Collection date 2026. © The Author(s) 2026 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 http://creativecommons.org/licenses/by/4.0/ . PMC Copyright notice PMCID: PMC13085593 PMID: 41803845 Abstract Background Acute exacerbation of chronic obstructive pulmonary disease (AECOPD) is characterized by a significant worsening of respiratory symptoms. Blood eosinophil levels are a key predictor of glucocorticoid efficacy in AECOPD patients; however, their stability can present challenges. Predicting stable eosinophil levels from CT images is essential for optimal patient management. Methods This study utilized CT images from 482 AECOPD patients across two hospitals. Dataset 1 comprised 193 patients for model development, while Dataset 2 included 289 patients for external validation. A threshold of 2% eosinophil was used to differentiate between high and low eosinophil levels. A machine learning model was developed to predict eosinophil levels using CT radiomics and quantitative computed tomography (QCT) features. Radiomics features were extracted, and feature selection was performed using random forest (RF) algorithms. Segmentation of pulmonary lobes, airways, and blood vessels yielded 20 QCT features. A Gradient Boosting (GB) classifier was then trained on the fused features. Results The GB classifier with radiomics features demonstrated strong performance, achieving an accuracy (ACC) of 0.734 and an area under the curve (AUC) of 0.838 on the test set of Dataset 1. In external validation, the ACC and AUC were 0.624 and 0.671, respectively. After fusing QCT features, the ACC and AUC improved to 0.786 and 0.843, respectively, with external validation results of 0.673 and 0.697. Conclusion The CT image-based machine learning model can predict blood eosinophil levels in AECOPD patients, providing a noninvasive and stable assessment. It has potential for future clinical application following further validation and external testing. Keywords: Machine learning, Radiomics, Computed tomography, Acute exacerbation of chronic obstructive pulmonary disease, Blood eosinophils Background Chronic obstructive pulmonary disease (COPD) is a prevalent respiratory condition that poses a significant global public health challenge due to its morbidity and high mortality rates [ 1 , 2 ]. Acute exacerbations and associated comorbidities are significant contributors to the mortality of many COPD patients [ 3 , 4 ]. Blood eosinophil is a biomarker to monitor COPD progression among these factors. Numerous studies have demonstrated that eosinophilia can influence the clinical outcomes of acute exacerbations of COPD (AECOPD) [ 5 , 6 ]. Blood eosinophils indicate eosinophilic inflammation in the airways and predict the response to inhaled corticosteroid (ICS) therapy [ 7 , 8 ]. Martin et al. [ 9 ] found that the low blood eosinophil counts in AECOPD patients hospitalized are associated with prolonged hospitalization. Most AECOPD studies use 200 cells/mL and/or 2% as thresholds to identify eosinophil levels [ 10 , 11 ]. Pasco et al. [ 12 ] found that the risk of moderate-to-severe AECOPD increased as eosinophil levels rose, mainly when peripheral blood eosinophil levels were ≥ 1.7%. Conversely, Hong et al. [ 13 ] noted that higher blood eosinophil counts were associated with a more rapid decline in lung function. According to the 2024 GOLD guidelines, blood eosinophil counts can predict the effectiveness of ICS in preventing future exacerbations [ 14 ]. In clinical practice, blood eosinophil counts are typically obtained through routine blood tests. Relevant studies have demonstrated that blood eosinophil counts can be influenced by various factors, including age, gender, atopy, and environmental exposure [ 15 – 18 ]. Significant advances in computed tomography (CT) technology, particularly high-resolution computed tomography (HRCT), which provides more precise and detailed images than conventional CT, have become increasingly crucial for assessing COPD. Developing a fast, accurate, and reliable method for automatically predicting eosinophil levels based on CT images is essential for enhancing the intervention and treatment of COPD patients. This approach provides physicians with targeted intervention recommendations, ensuring timely and appropriate patient care and, to some extent, alleviating the workload of healthcare providers. Radiomics features in lung disease imaging are regarded as one of the cutting-edge tools in medicine [ 19 ]. Previous studies have shown that imaging histology can be applied to diagnosing and predicting COPD and related diseases. Lin et al. [ 20 ] assessed cardiovascular disease (CVD) risk in COPD patients using CT-based whole lung radiomics. Zhao et al. [ 21 ] proposed a method for staging COPD severity by combining lobe radiomics features. Zhou et al. [ 22 ] utilized whole lung radiomics to distinguish between healthy individuals and COPD patients. Amudala et al. [ 23 ] evaluated the performance of standard-dose and low-dose CT radiomics features in diagnosing COPD. The studies above indicate that radiomics features are crucial for the diagnosis and treatment prediction of COPD. In COPD, quantitative computed tomography (QCT) has shown a significant correlation with disease progression and mortality [ 24 – 26 ]. Zhang et al. [ 27 ] proposed a method for the early diagnosis of high-risk COPD patients using QCT. Konietzk et al. [ 28 ] employed QCT to monitor the progression of emphysema in COPD patients at three-month intervals. Moslemi et al. [ 29 ] utilized QCT and CT radiomics to predict the emergency department (ED) visits rate and the length of hospital stays. In summary, while QCT and CT radiomics have shown the potential for predictive diagnosis in COPD, few studies have concentrated on predicting eosinophil levels in AECOPD patients using CT images. In this study, we propose a method for predicting eosinophil levels in AECOPD patients based on CT images. Specifically, the study employed CT radiomics and QCT features to predict eosinophil count levels. The combination of CT radiomics and QCT results in a more comprehensive and enriched characterization, which enhances the predictive model’s performance. The contributions of this paper are as follows: A CT image-based machine learning model is proposed to predict eosinophil levels in AECOPD, addressing the instability of measuring the eosinophil index by routine blood tests. CT radiomics and QCT features are integrated to improve the performance of predicting eosinophil levels in AECOPD patients. Two central datasets are utilized to demonstrate the feasibility and generalization ability of predicting eosinophil levels in AECOPD using CT images. Methods Datasets Dataset 1 : The dataset comprised 193 AECOPD patients screened at the Affiliated Central Hospital of Shenyang Medical College, with slice thicknesses ranging from approximately 0.5 to 1.2 mm. The inclusion criteria were as follows: (1) Complete CT images are available, allowing for all QCT measurements to be analyzed; (2) No QCT measurements have any missing values; (3) Complete demographic information and clinical indicators are provided; (4) All the factors which might affect the measurement of eosinophil level through routine blood tests have been excluded, including diurnal variation, recent use of corticosteroids or antibiotics, allergic reactions, parasitic infections. Based on the eosinophil count at the time of admission, the enrolled AECOPD patients were categorized into two groups: 87 cases in the high eosinophil level group (eosinophil% ≥2%) and 106 cases in the low eosinophil level group (eosinophil% <2%). Table 1 presents the demographic information and clinical characteristics of the 193 AECOPD patients. Dataset 2 : The dataset was obtained from Shanxi Bethune Hospital and comprises 289 CT scans of AECOPD patients, including 115 patients in the high eosinophil group and 174 patients in the low eosinophil group. This dataset was used as an external validation set to further assess the prediction model’s predictive capability for eosinophil level in AECOPD patients. Table 2 presents the patient’s demographic information and clinical characteristics in the external validation set. Table 1. Demographic information and clinical characteristics of patients in Dataset 1 (EOS demonstrates eosinophil) Clinical information High EOS group Low EOS group P value* Case 87 106 - Sex (male/female) 56/31 64/42 0.57 Age (years) 71.34 ± 11.08 72.40 ± 10.05 0.49 Body mass index (BMI-kg/m2) 23.50 ± 3.64 22.49 ± 1.14 0.08 Smoking years 22.95 ± 21.46 25.92 ± 22.29 0.35 Length of hospitalization (days) 9.72 ± 3.87 10.90 ± 5.45 0.09 Open in a new tab * Two-sample t-test Table 2. Demographic information and clinical characteristics of patients in Dataset 2 (EOS demonstrates eosinophil) Clinical information High EOS group Low EOS group P value* Case 115 174 - Sex (male/female) 95/20 141/33 0.84 Age (years) 67.20 ± 8.77 68.56 ± 8.46 0.15 Smoking years 23.2 ± 16.31 29.4 ± 15.69 0.01 Length of hospitalization (days) 6.93 ± 3.23 8.72 ± 4.04 0.01 Open in a new tab * Two-sample t-test Table 3 compares the technical parameters of CT scans from two medical imaging datasets (Dataset1 and Dataset2) with the following information: Table 3. CT acquisition protocols of two cohorts Cohorts Scanning equipment Tube voltage (kV) Pitch Thickness (mm) Matrix Type Dataset1 NeuViz 128 1.0 120 1 0.5–1.2 512 × 512 Standard-dose Dataset2 SIEMENS syngo CT 2012B 120 1.35 1 512 × 512 Standard-dose Open in a new tab Overall framework Figure 1 illustrates the workflow of this study’s proposed CT image-based prediction method for blood eosinophil levels. First, the lung region is segmented from CT images by the lungmask [ 30 ], and the lobes, airway, and blood vessels are segmented by the proposed method [ 31 – 33 ]. All segmenation will be visually inspected by two experienced radiologists to ensure anatomical accuracy and consistency of segmentation results. Second, the same as other studies [ 22 , 34 , 35 ], for small ROI, the features are sensitive to the segmentation methods. However, lung filed is very large, the influence is very slight. CT image radiomics features are extracted from the lung region, while in-house methods extract QCT features. Next, the radiomics features are selected using the random forest (RF) algorithm, and feature fusion is conducted with the QCT features. Finally, a Gradient Boosting (GB) classifier is developed to predict blood eosinophil levels. Fig. 1. Open in a new tab The framework for blood eosinophil levels prediction using CT images Procedure for radiomics model Preprocessing According to the proposed method [ 30 ], automatic lung segmentation from CT images is performed. Before CT radiomics feature extraction, the CT images are resampled to to reduce the effect of different layer thicknesses and the interference of noise. The bin width is set 20. Radiomics feature extraction Pyradiomics [ 36 ] is a tool to extract features from raw or image-processed CT images. With Pyradiomics, feature extraction and analysis can be easily performed to support further research and applications in radionics. Radiomics features can be categorized explicitly into the following three types: first-order statistics, texture features, and higher-order filter transform features. Definitions and detailed explanations of radiomics features can be found in the Pyradiomics documentation. Feature selection Feature selection helps to improve the performance of predictive models by identifying and retaining the most relevant and informative features, reducing the risk of overfitting, and reducing dimensionality by eliminating irrelevant or redundant variables. The RF [ 37 ] is used for feature selection. Specifically, the absolute value of the importance coefficient of each feature is first ranked to reflect the degree of influence of the feature on the classification effect of the model and to analyze the importance of each feature. According the recommendation by Gillies et al. [ 38 ], about 10 samples or patients are required for each feature in a binary classifier. We have 193 patients and 20 features, which approximately according to the recommended requirements. Thus, the top 20 ranked features are taken as the result of feature selection. By randomly selecting features for each tree split, the model reduces the dominance of highly correlated features and helps to better identify informative variables. Although Random Forest does not eliminate multicollinearity, it is less sensitive to it compared to linear models. And the model we use in the training is not linear. Machine learning model prediction The GB is the prediction model [ 39 ]. The GB is a supervised machine learning algorithm known for its effectiveness in various classification tasks. It can handle data with nonlinear and high-dimensional features and has been widely applied across multiple fields. To evaluate the proposed method’s reliability and generalizability, the Dataset1 were divided into a training set (135 samples: 74 patients in the low EOS group and 61 in the high EOS group) and an internal test set (58 samples: 32 patients in the low EOS group and 26 in the high EOS group). The training set was employed for feature selection and training the model, while performance testing was carried out separately on the internal validation set to ensure experimental accuracy. Dataset 2 was used as an external validation set. Procedure for fusion model Preprocessing and QCT feature calculation According to the proposed method [ 31 – 33 ], the lobes, airways, and blood vessels were automatically segmented. The in-house scientific software (Key Laboratory of Intelligent Computing for Medical Imaging, Northeastern University, Shenyang, Liaoning, China) was used to quantify emphysema, airways and blood vessels. Table 4 presents a comprehensive list of QCT measurements utilized in this study and their definitions. (1) Emphysema measurements: The emphysema index [ 40 ] is crucial for quantifying the degree of emphysema. The percentage of low attenuation area with attenuation less than − 950 HU (%LAA-950) for the entire lung as well as for each lobe, including the right upper lobe (RUL), right middle lobe (RML), and right lower lobe (RLL), the right lung (RL), as well as the left upper lobe (LUL), left lower lobe (LLL) and left lung (LL). (2) Airway measurements: The airway measurements focus on the structural characteristics of the airways, offering valuable insights into their morphology and potential pathological changes. Wall thickness (WT) [ 41 ]: This measure assesses the thickness of the airway walls, which may indicate inflammatory processes or structural remodeling associated with various lung diseases. Wall area percentage (WA%) [ 42 ]: This percentage reflects the proportion of the airway cross-section occupied by the wall, providing insights into airway calibre and potential obstruction. Pi10 [ 43 ]: This index estimates the wall thickness of an airway with an ideal internal circumference of 10 mm, which is particularly important for assessing small airway disease. (3) Blood vessel measurements: The vessel measurements encompass a range of quantitative measures computed using the VesselVio [ 44 ], which provides a detailed analysis of the pulmonary vasculature. The following indices include the aggregate vessel volume for vessels less than 5 mm 2 (BV5), network length, surface area, branch points, endpoints, number of segments, mean segment radius, mean segment length, and mean segment tortuosity. Table 4. The definitions of 20 QCT features No. QCT feature Description Category 1 RUL emphysema index %LAA-950 in the right upper lobe Emphysema 2 RML emphysema index %LAA-950 in the right middle lobe 3 RLL emphysema index %LAA-950 in the right lower lobe 4 RL emphysema index %LAA-950 in the right lung 5 LUL emphysema index %LAA-950 in the left upper lobe 6 LLL emphysema index %LAA-950 in the left lower lobe 7 LL emphysema index %LAA-950 in the left lung 8 All lung emphysema index %LAA-950 in the whole lung 9 Pi10 An estimate of the wall thickness of an airway with an ideal internal circumference of 10 mm. Airway 10 WA% Percentage of airway wall area 11 WWT Mean value of airway wall thickness 12 BV5 The aggregate vessel volume for vessels less than 5 mm 2 Blood vessel 13 Vessel network length Total vessel tree length 14 Vessel surface area Total vessel surface area 15 Vessel branch points Total number of vessel branch 16 Vessel endpoints Total number of end branch 17 Number of segments Total number of identifiable vessel segments in the vascular network 18 Mean segment radius Mean value of vessel segment radius 19 Mean segment length Mean value of vessel segment length 20 Mean segment tortuosity Mean value of vessel segment tortuosity Open in a new tab Feature fusion Feature fusion combines multiple features to obtain more prosperous and effective input features and improves the performance of machine learning models. For splicing, 20 CT radionics and 20 QCT features are used. Machine learning model prediction The methods are as same as those in the radiomics model. Implementation details and evaluation metrics Four evaluation metrics are used to assess eosinophil level prediction performance, including accuracy (ACC), sensitivity (SEN), positive predictive value (PPV), area under curve (AUC), receiver operating characteristic (ROC), The formulas are as follows: 1 2 3 where true positive ( ) denotes the number of patients correctly predicted high eosinophil, true negative ( ) denotes the number of patients correctly predicted low eosinophil, false positive ( ) denotes the number of patients which is low eosinophil mispredicted high eosinophil, false negative ( ) denotes the number of patients which is high eosinophil mispredicted low eosinophil. The ROC curve is represented by the false positive rate on the x-axis and the true positive rate on the y-axis [ 45 ]. The closer the ROC curve is to the upper left corner, the larger the AUC and the better the classification model performance [ 46 ]. In this study, other machine learning models are used, including K-nearest neighbour (KNN) [ 47 ], support vector machine (SVM) [ 48 ], decision tree [ 49 ], AdaBoost [ 50 ], and XGBoost [ 51 ]. Classifiers use their default parameter settings to facilitate more direct comparisons between models and highlight each algorithm’s out-of-the-box functionality. Each classifier is independent and has no relationship with other classifiers and no information leakage. There is no need for multiple comparisons or corrections. In machine learning or deep tilting research [ 21 , 52 ], we almost always perform comparisons between different models in this way. Results Prediction performance of radiomics model with different numbers of features and feature selection methods In Table 5 , we analyze the impact of varying the number of features selected on the performance of the RF model for predicting eosinophil levels. With 15 selected features, the RF model achieves an ACC of 0.721 and SEN, PPV and AUC scores of 0.713, 0.724, and 0.800, respectively. However, when the number of selected features is increased to 25, there is a slight decline in performance, resulting in an ACC of 0.716, with SEN, PPV, and AUC scores of 0.708, 0.725, and 0.821. Based on these findings, the best overall performance is achieved when the feature selected number is set to 20, yielding an ACC of 0.734 and SEN, PPV, and AUC scores of 0.732, 0.752, and 0.838. Table 5. Performance comparison using different numbers of radiomics features on Dataset 1 Feature selections number ACC SEN PPV AUC 15 0.721 0.713 0.724 0.800 20 0.734 0.732 0.752 0.838 25 0.716 0.708 0.725 0.821 Open in a new tab Note: Bold values highlight the maximum values Figure 2 illustrates the top 20 essential features of the screening, with wavelet transform features and the Gray-Level Co-Occurrence Matrix showing high importance. The features quantify the spatial relationship between pixel grey values and texture properties and are more suitable for eosinophil prediction. Fig. 2. Open in a new tab 20 Radiomics features and corresponding coefficients after dimensionality reduction by RF Table 6 compares the LASSO and RF methods for feature selection. The LASSO method achieves an ACC of 0.676, with corresponding SEN, PPV, and AUC scores of 0.681, 0.684, and 0.725. In contrast, the RF method exhibits significantly better performance. The results indicate that the RF approach is more effective for predicting eosinophil levels. Table 6. Performance comparison using different feature selection methods on Dataset 1 Method ACC SEN PPV AUC Lasso 0.676 0.684 0.681 0.725 RF 0.734 0.752 0.732 0.838 Open in a new tab Note: Bold values highlight the maximum values Overall, these analyses highlight the importance of feature selection in enhancing predictive ACC, with RF and a feature selection number of 20 emerging as the more robust method. Prediction performance of models with different features Table 7 presents the results of predicting blood eosinophil levels using different features with the GB method on Dataset 1. The model was trained using three distinct feature combinations: CT radionics, QCT features, and a fusion of both. Specifically, the CT radiomics-based model achieves an ACC of 0.734, with SEN, PPV, and AUC around 0.732, 0.752, and 0.838. In contrast, for the QCT-based model, ACC, SEN, PPV, and AUC are 0.602, 0.594, 0.603, and 0.590. Table 7. Performance comparison using different features on Dataset 1 CT radiomics QCT ACC SEN PPV AUC √ 0.734 0.732 0.752 0.838 √ 0.602 0.594 0.603 0.590 √ √ 0.786 0.778 0.786 0.843 Open in a new tab Note: Bold values highlight the maximum values However, combining both features significantly improved performance, resulting in an ACC of 0.786, with corresponding SEN, PPV, and AUC scores of 0.778, 0.786, and 0.843. Overall, these results indicate that eosinophil levels can be effectively predicted using CT radiomics features, and further enhancement in predictive ACC can be achieved by fusing QCT features. Prediction performance of models with different machine learning methods Table 8 comprehensively evaluates machine learning methods for predicting blood eosinophil levels on Dataset 1 (training dataset). The task is to predict high and low blood eosinophil levels. The machine learning methods include KNN, SVM, Decision Tree, AdaBoost, XGBoost, and Gradient Boosting. Table 8. Performance comparison of models with six machine learning methods on Dataset 1 Features Machine learning method ACC SEN PPV AUC CT radiomics KNN 0.612 0.613 0.615 0.624 SVM 0.653 0.645 0.654 0.671 Decision Tree 0.706 0.703 0.705 0.699 AdaBoost 0.584 0.583 0.586 0.671 XGBoost 0.703 0.704 0.704 0.814 GB 0.734 0.732 0.752 0.838 CT radiomics and QCT KNN 0.585 0.576 0.578 0.623 SVM 0.652 0.644 0.648 0.629 Decision Tree 0.724 0.723 0.725 0.719 AdaBoost 0.756 0.744 0.757 0.764 XGBoost 0.733 0.732 0.734 0.801 GB 0.786 0.778 0.786 0.843 Open in a new tab Note: Bold values highlight the maximum values For the CT radiomics model, the performance metrics of each method are as follows: KNN achieves an ACC of 0.612, SEN of 0.613, PPV of 0.615, and AUC of 0.624. The SVM method shows slightly better results with an ACC of 0.653, while the Decision Tree reaches an ACC of 0.706. AdaBoost performs poorly, with an ACC of 0.584. The XGBoost method performs better with an ACC of 0.703. The GB method achieved the highest metrics: an ACC of 0.734, SEN of 0.732, PPV of 0.752, and AUC of 0.838. This highlights the effectiveness of the GB method for CT radiomic features. Most methods show improved performance for CT radiomics and QCT model. KNN’s performance drops slightly, with an ACC of 0.585, while SVM maintains an ACC of 0.652. The Decision Tree method also performs well, achieving an ACC of 0.724. Notably, the AdaBoost method improves significantly, achieving an ACC of 0.756. The XGBoost method improves slightly with an ACC of 0.733. However, the GB method again demonstrates the best performance, reaching an ACC of 0.786, SEN of 0.778, PPV of 0.786, and AUC of 0.843. This improvement highlights the fusion of QCT features, likely providing additional relevant information that enhances the method’s predictive capabilities. Additionally, as shown in Figs. 3 (a) and (b), the evaluation of blood eosinophil level predictions for each method is illustrated through ROC curve analysis, allowing for a visual representation of each method’s performance. The GB method achieves the best results across two feature extraction methods, demonstrating its robustness and effectiveness. Fig. 3. Open in a new tab ROC curves for blood eosinophil levels prediction for different models on Dataset 1. ( a ) With CT radiomics features; ( b ) With CT radiomics and QCT features These findings indicate that the Gradient Boosting method effectively predicts blood eosinophil levels using CT radiomics and significantly enhances performance when combined with QCT features. The results suggest that further research into hybrid features and method optimization could yield better outcomes for predicting blood eosinophil levels. Prediction performance on external validation Dataset 2 As shown in Table 9 ; Fig. 4 (a), and (b), to further validate the method’s reliability, we extend the proposed method to an external validation set (Dataset 2). Dataset 2 is tested on a previously trained model. Utilizing multiple machine learning methods, blood eosinophil levels prediction is performed on Dataset 2 with the same features. Table 9. Performance comparison of models with six machine learning methods on Dataset 2 Features Machine learning method ACC SEN PPV AUC CT radiomics KNN 0.602 0.534 0.556 0.573 SVM 0.613 0.583 0.603 0.661 Decision Tree 0.562 0.560 0.558 0.557 AdaBoost 0.583 0.584 0.586 0.588 XGBoost 0.592 0.603 0.594 0.639 GB 0.624 0.617 0.619 0.671 CT radiomics and QCT KNN 0.614 0.572 0.583 0.602 SVM 0.634 0.603 0.615 0.611 Decision Tree 0.586 0.597 0.582 0.587 AdaBoost 0.634 0.623 0.627 0.645 XGBoost 0.643 0.616 0.627 0.659 GB 0.673 0.667 0.669 0.697 Open in a new tab Note: Bold values highlight the maximum values Fig. 4. Open in a new tab ROC curves for blood eosinophil levels prediction for different models on Dataset 2. ( a ) With CT radiomics features; ( b ) With CT radiomics and QCT features For the CT radiomics method, KNN achieves an ACC of 0.602, SEN of 0.534, PPV of 0.556, and AUC of 0.573, while SVM reaches an ACC of 0.613. The Decision Tree shows lower performance with an ACC of 0.562. AdaBoost and XGBoost also demonstrate limited effectiveness, with ACCs of 0.583 and 0.592, respectively. In contrast, the GB method performs better than others in this method, achieving an ACC of 0.624, SEN of 0.617, PPV of 0.619, and AUC of 0.671. For the CT radiomics with the QCT features method, the KNN method maintains an ACC of 0.614 and SEN of 0.572. The SVM method shows slightly improved results with an ACC of 0.634. The Decision Tree again performs modestly with an ACC of 0.586, while AdaBoost and XGBoost reach an ACC of 0.634 and 0.643. Notably, the GB method excels in this combined feature set, achieving the highest ACC of 0.673, SEN of 0.667, PPV of 0.669, and AUC of 0.697. It is important to note that while training is conducted on Dataset 1, external validation occurs on Dataset 2, which is collected using a different device and from a different geographic region. This variation in the data acquisition process may explain the observed decrease in performance when applying the model to Dataset 2. Overall, these findings reinforce the efficacy of the GB method for predicting blood eosinophil levels, even when validated on an external dataset. The results also highlight the importance of feature selection and model robustness, suggesting that further investigation into improving generalizability across different devices and regions could enhance predictive performance. Discussion This study proposes a new way to predict blood eosinophil levels based explicitly on CT radiomics and QCT. The combination of QCT provided a better overview when predicting blood eosinophil levels compare to the use of CT radiomics alone. The proposed method introduces a new direction for future blood eosinophil detection. Previous studies have demonstrated the use of CT imaging to predict similar physiological indices. Geraghty et al. proposed multiple linear regression to predict an individual’s height, weight, body mass index, and body surface area with a single abdominal CT [ 53 ]. Park et al. developed a deep learning-based method for predicting lung function from low-dose CT images [ 54 ]. With the advancement of computer vision and machine learning techniques, the analysis of CT imaging has been significantly improved. Researchers have used algorithms to extract complex features from CT images to achieve more accurate predictions of physiological indicators. The combination of these techniques not only improves prediction ACC but also provides more substantial support for clinical decision-making. This also provides technical support and new ideas for this study. Some studies have used blood eosinophil as a biomarker for ICS, suggesting that > 2% may be an appropriate threshold for the relative efficacy of ICS [ 55 – 57 ]. Meanwhile, blood eosinophils can potentially be a prognostic indicator in AECOPD [ 58 ], and there is an association between emphysema and blood eosinophil levels [ 59 ]. In conclusion, eosinophils not only have potential as a biomarker for assessing the efficacy of ICS but also show significant value in the prognostic assessment of AECOPD and studies related to emphysema. Routine blood counts make routine blood eosinophil measurements. Thus, we used 2% blood eosinophil (EOS) percentage as the cut-off threshold to construct a model predicting the classification of “high EOS (≥ 2%)” vs. “low EOS (< 2%)”. Thus, indirectly predicting the predicted eosinophil level. The CT scans are typically performed when COPD patients require imaging for other clinical indications. Our model offers an “opportunistic” assessment of eosinophil levels using existing CT images, eliminating the need for additional tests or radiation exposure. Blood eosinophil levels can fluctuate due to various factors, leading to measurement instability. In contrast, CT imaging captures long-term pathological changes in the lungs, providing a more stable and representative assessment of eosinophilic inflammation in COPD. For patients needing frequent monitoring, noninvasive CT assessment may reduce the necessity for repeat blood draws, thereby alleviating patient discomfort and potential complications. This is the first study to use CT images to predict blood eosinophil levels and the first to investigate using hybrid features to predict blood eosinophil levels. Blood eosinophil levels fluctuate throughout the day, leading to inconsistent readings based on the timing of blood draws. Factors such as stress, infection, recent corticosteroid use, or allergic reactions can temporarily alter eosinophil levels, affecting their reliability as stable biomarkers in certain clinical contexts. Our CT-based approach does not eliminate these external influences directly but aims to reduce measurement variability by offering a stable assessment of eosinophilic inflammation. Specifically, CT-based radiomics and quantitative CT features capture structural and pathological changes in the lungs, such as emphysema, airway remodeling, and vascular alterations. These changes are less affected by short-term physiological factors and instead reflect the cumulative effects of chronic eosinophilic inflammation, providing a more stable assessment that complements blood eosinophil measurements. Additionally, our approach posits that CT imaging features may capture chronic disease phenotypes linked to eosinophilic inflammation, potentially offering predictive insights into future eosinophil levels or treatment responses. This represents an advantage over isolated blood test results, which can vary significantly due to short-term factors. In conclusion, while we acknowledge that blood tests are the current gold standard for measuring eosinophil levels and that our methods were evaluated against these measurements, our study’s intent is not to replace blood tests but to provide a complementary tool for situations where blood eosinophil measurements are unreliable or unavailable. In contrast, the model employing gradient boosting as a predictor has an ACC of 0.734 when relying exclusively on CT imaging features. However, in the external validation set, the ACC exceeds 0.624, which is the best performance among the selected classification models. Notably, when quantitative metrics are integrated into the model, the ACC improves significantly, reaching 0.786 on the training set and 0.673 on the external validation set. This improvement validates the model’s generalization ability and highlights the advantages of combining imaging features with CT quantitative metrics. Fusing these data types led to a partial improvement in predictive results, highlighting the importance of a multifaceted approach in predictive modeling. Despite these encouraging findings, our study has some limitations. Firstly, sample size limitations are one of the main limitations of this study, and future work must include larger, multi-center datasets for external validation. Secondly, although the experiment was conducted on data from two centralized datasets, the effect was reduced in the external validation set (Dataset 2) due to equipment, region, and other factors. While the model showed good predictive performance on the training dataset (Dataset 1), further validation on more centralized datasets is needed to confirm its clinical utility. Thirdly, the performance of the radiomics approach across different data domains has not been thoroughly evaluated. Therefore, methods such as domain adaptation or other generative artificial intelligence techniques should to be considered in future studies [ 60 – 62 ]. Thirdly, the study focused on predicting blood eosinophils using CT imaging. Significant differences in smoking history and length of hospitalization were observed between the comparison groups in Data Set 2. However, confounding could not be fully corrected due to the lack of quantitative data on smoking pack-years and limitations in sample size. This residual confounding may affect the interpretation of outcomes, particularly for smoking-related endpoint indicators. Future prospective studies should prioritize obtaining refined exposure data to isolate true effects. And, this study did not include other clinical information or genetic data such as respiratory parameters, BMI, family history, and genetic factors. Finally, deep learning methods could be explored in future studies to predict blood eosinophil levels, as they have been successfully used in COPD identification, staging, and quantitation [ 63 ]. Conclusion This study demonstrates that CT image-based can provide a promising method for blood eosinophil levels prediction that is non-invasive and stable and, when further validated, can guide subsequent treatment regimens for COPD patients. In addition, CT radiomics features can provide global information. In contrast, quantitative CT can provide an intrinsic feature, and the fusion of CT radiomics and quantitative CT features can better enhance the prediction performance. This also demonstrates that there may be a link between macroscopic CT imaging phenotypes and inflammatory microphenotypes. Machine learning methods can identify this link and thus establish predictive relationships. Acknowledgements Not applicable. Abbreviation AECOPD Acute exacerbation of chronic obstructive pulmonary disease QCT Quantitative computed tomography RF Random forest GB Gradient boosting ACC Accuracy AUC Area under the curve COPD Chronic obstructive pulmonary disease ICS Inhaled corticosteroid CT Computed tomography HRCT High-resolution computed tomography CVD Cardiovascular disease EOS Eosinophil BMI Body mass index %LAA-950 The percentage of low attenuation area with attenuation less than − 950 HU RUL Right upper lobe RML Right middle lobe RLL Right lower lobe RL Right lung LUL Left upper lobe LLL Left lower lobe LL Left lung WT Wall thickness WA% Wall area percentage BV5 The aggregate vessel volume for vessels less than 5 mm2 SEN Sensitivity PPV Positive predictive value ROC Receiver operating characteristic TP True positive TN True negative FP False positive FN False negative KNN K-nearest neighbour SVM Support vector machine Authors' contributions SZ conducted research, analysis and writing original draft preparation. LD and HJ conducted the data collection. YW conducted the analysis, PM, WQ, RW, SX and SQ contributed to the manuscript review and supervision. All authors contributed to the manuscript and read and approved the final manuscript. Funding This work was partly supported by the National Natural Science Foundation of China under Grant (Nos. 82472076 and 62271131) and the Fundamental Research Funds for the Central Universities (N25BJD013). Data availability No datasets were generated or analysed during the current study. Declarations Ethics approval and consent to participate All procedures performed in studies involving human participants were in accordance with the ethical standards of the ethics committee of Shanxi Bethune Hospital and the Affiliated Center Hospital of Shenyang Medical College and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. Consent to participate Informed consent was obtained from all individual participants included in the study. Consent for publication Not applicable. Competing interests The authors declare no competing interests Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Ruiying Wang, Email: [email protected]. Shuyue Xia, Email: [email protected]. Shouliang Qi, Email: [email protected]. References 1. Roth GA, Abate D, Abate KH, Abay SM, Abbafati C, Abbasi N, Abbastabar H, Abd-Allah F, Abdela J, Abdelalim A. Global, regional, and national age-sex-specific mortality for 282 causes of death in 195 countries and territories, 1980–2017: a systematic analysis for the Global Burden of Disease Study 2017. lancet. 2018;392(10159):1736–88. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Halpin DM, Criner GJ, Papi A, Singh D, Anzueto A, Martinez FJ, Agusti AA, Vogelmeier CF. Global initiative for the diagnosis, management, and prevention of chronic obstructive lung disease. The 2020 GOLD science committee report on COVID-19 and chronic obstructive pulmonary disease. Am J Respir Crit Care Med. 2021;203(1):24–36. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Roman-Rodriguez M, Kaplan A. GOLD 2021 strategy report: implications for asthma–COPD overlap. Int J Chronic Obstr Pulm Dis. 2021;14:1709–15. [ DOI ] [ PMC free article ] [ PubMed ] 4. Halpin DM, Healey H, Skinner D, Carter V, Pullen R, Price D. Exacerbation history and blood eosinophil count prior to diagnosis of COPD and risk of subsequent exacerbations. Eur Respir J. 2024;64(4):230224. [ DOI ] [ PMC free article ] [ PubMed ] 5. Mathioudakis AG, Bikov A, Foden P, Lahousse L, Brusselle G, Singh D, Vestbo J. Change in blood eosinophils following treatment with inhaled corticosteroids may predict long-term clinical response in COPD. Eur Respir J. 2020;55(5):1902119. [ DOI ] [ PubMed ] 6. Mathioudakis AG, Bate S, Sivapalan P, Jensen JS, Singh D, Vestbo J. Rethinking blood eosinophils for assessing inhaled corticosteroids response in COPD: a post hoc analysis from the FLAME trial. Chest. 2024;166(5):987–97. [ DOI ] [ PMC free article ] [ PubMed ] 7. Zysman M, Deslee G, Caillaud D, Chanez P, Escamilla R, Court-Fortune I, et al. Relationship between blood eosinophils, clinical characteristics, and mortality in patients with COPD. Int J Chron Obstruct Pulmon Dis. 2017;12:1819–24. [ DOI ] [ PMC free article ] [ PubMed ] 8. Wu X, Deng Z, Wu F, Zheng Y, Huang P, Yang H, et al. Clinical characteristics and 2-year outcomes of chronic obstructive pulmonary disease patients with high blood eosinophil counts: a population-based prospective cohort study in China. Archivos de Bronconeumología. 2024;60(7):402–9. [ DOI ] [ PubMed ] 9. MacDonald MI, Osadnik CR, Bulfin L, Hamza K, Leong P, Wong A, King PT, Bardin PG. Low and high blood eosinophil counts as biomarkers in hospitalized acute exacerbations of COPD. Chest. 2019;156(1):92–100. [ DOI ] [ PubMed ] [ Google Scholar ] 10. Petite SE, Murphy JA. Systemic corticosteroid and antibiotic use in hospitalized patients with chronic obstructive pulmonary disease exacerbation. Ann Pharmacother. 2019;53(2):144–50. [ DOI ] [ PubMed ] [ Google Scholar ] 11. David B, Bafadhel M, Koenderman L, De Soyza A. Eosinophilic inflammation in COPD: from an inflammatory marker to a treatable trait. Thorax. 2021;76(2):188–95. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Park YM, Bochner BS. Eosinophil survival and apoptosis in health and disease. Allergy Asthma Immunol Res. 2010;2(2):87–101. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 13. Hong YS, Park HY, Ryu S, Shin SH, Zhao D, Singh D, et al. The association of blood eosinophil counts and FEV1 decline: a cohort study. Eur Respir J. 2024;63(5):2301037. [ DOI ] [ PubMed ] 14. Venkatesan P. GOLD COPD report: 2024 update. Lancet Respiratory Med. 2024;12(1):15–6. [ DOI ] [ PubMed ] [ Google Scholar ] 15. Haldar P, Pavord ID, Shaw DE, Berry MA, Thomas M, Brightling CE, Wardlaw AJ, Green RH. Cluster analysis and clinical asthma phenotypes. Am J Respir Crit Care Med. 2008;178(3):218–24. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Patelis A, Janson C, Borres MP, Nordvall L, Alving K, Malinovschi A. Aeroallergen and food I g E sensitization and local and systemic inflammation in asthma. Allergy. 2014;69(3):380–7. [ DOI ] [ PubMed ] [ Google Scholar ] 17. Higuchi T, Omata F, Tsuchihashi K, Higashioka K, Koyamada R, Okada S. Current cigarette smoking is a reversible cause of elevated white blood cell count: cross-sectional and longitudinal studies. Prev Med Rep. 2016;4:417–22. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Hartl S, Breyer M-K, Burghuber OC, Ofenheimer A, Schrott A, Urban MH, et al. Blood eosinophil count in the general population: typical values and potential confounders. Eur Respir J. 2020; 55(5):1901874. [ DOI ] [ PubMed ] 19. Frix A-N, Cousin F, Refaee T, Bottari F, Vaidyanathan A, Desir C, Vos W, Walsh S, Occhipinti M, Lovinfosse P. Radiomics in lung diseases imaging: state-of-the-art for clinicians. J Personalized Med. 2021;11(7):602. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 20. Lin X, Zhou T, Ni J, Li J, Guan Y, Jiang Xa, et al. CT-based whole lung radiomics nomogram: a tool for identifying the risk of cardiovascular disease in patients with chronic obstructive pulmonary disease. Eur Radiol. 2024;34(8):4852–63. [ DOI ] [ PubMed ] 21. Zhao M, Wu Y, Li Y, Zhang X, Xia S, Xu J, Chen R, Liang Z, Qi S. Learning and depicting lobe-based radiomics feature for COPD severity staging in low-dose CT images. BMC Pulm Med. 2024;24(1):294. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Zhou T-H, Zhou X-X, Ni J, Ma Y-Q, Xu F-Y, Fan B, Guan Y, Jiang X-A, Lin X-Q, Li J. CT whole lung radiomic nomogram: a potential biomarker for lung function evaluation and identification of COPD. Military Med Res. 2024;11(1):14. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Amudala Puchakayala PR, Sthanam VL, Nakhmani A, Chaudhary MF, 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(5):e222998. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. 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):1139–49. [ DOI ] [ PubMed ] [ Google Scholar ] 25. Castillo-Saldana D, Hague CJ, Coxson HO, Ryerson CJ. Using quantitative computed tomographic imaging to understand chronic obstructive pulmonary disease and fibrotic interstitial lung disease: state of the art and future directions. J Thorac Imaging. 2020;35(4):246–54. [ DOI ] [ PubMed ] [ Google Scholar ] 26. Park J, Hobbs BD, Crapo JD, Make BJ, Regan EA, Humphries S, Carey VJ, Lynch DA, Silverman EK, Investigators C. Subtyping COPD by using visual and quantitative CT imaging features. Chest. 2020;157(1):47–60. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 27. Zhang W, Zhao Y, Tian Y, Liang X, Piao C. Early diagnosis of high-risk chronic obstructive pulmonary disease based on quantitative high-resolution computed tomography measurements. Int J Chronic Obstr Pulm Dis. 2023;18:3099–114. [ DOI ] [ PMC free article ] [ PubMed ] 28. Konietzke P, Wielpütz MO, Wagner WL, Wuennemann F, Kauczor H-U, Heussel CP, Eichinger M, Eberhardt R, Gompelmann D, Weinheimer O. Quantitative CT detects progression in COPD patients with severe emphysema in a 3-month interval. Eur Radiol. 2020;30:2502–12. [ DOI ] [ PubMed ] [ Google Scholar ] 29. Moslemi A, Makimoto K, Tan WC, Bourbeau J, Hogg JC, Coxson HO, Kirby M. Quantitative CT lung imaging and machine learning improves prediction of emergency room visits and hospitalizations in COPD. Acad Radiol. 2023;30(4):707–16. [ DOI ] [ PubMed ] [ Google Scholar ] 30. 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;4:1–13. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Wu Y, Zhao S, Qi S, Feng J, Pang H, Chang R, Bai L, Li M, Xia S, Qian W. Two-stage contextual transformer-based convolutional neural network for airway extraction from CT images. Artif Intell Med. 2023;143:102637. [ DOI ] [ PubMed ] [ Google Scholar ] 32. Pang H, Wu Y, Qi S, Li C, Shen J, Yue Y, Qian W, Wu J. A fully automatic segmentation pipeline of pulmonary lobes before and after lobectomy from computed tomography images. Comput Biol Med. 2022;147:105792. [ DOI ] [ PubMed ] [ Google Scholar ] 33. Wang M, Qi S, Wu Y, Sun Y, Chang R, Pang H, Qian W. CE-NC-VesselSegNet: supervised by contrast-enhanced CT images but utilized to segment pulmonary vessels from non-contrast-enhanced CT images. Biomed Signal Process Control. 2023;82:104565. [ Google Scholar ] 34. Qi X, Jiang Z, Yu Q, Shao C, Zhang H, Yue H, et al. Machine learning-based CT radiomics model for predicting hospital stay in patients with pneumonia associated with SARS-CoV-2 infection: a multicenter study. Ann Transl Med. 2020;8(14):859. [ DOI ] [ PMC free article ] [ PubMed ] 35. Liao C-Y, Chen Y-M, Wu Y-T, Chao H-S, Chiu H-Y, Wang T-W, Chen J-R, Shiao T-H, Lu C-F. Personalized prediction of immunotherapy response in lung cancer patients using advanced radiomics and deep learning. Cancer Imaging. 2024;24(1):129. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Van Griethuysen JJ, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, Beets-Tan RG, Fillion-Robin J-C, Pieper S, Aerts HJ. Computational radiomics system to decode the radiographic phenotype. Cancer Res. 2017;77(21):e104–7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 37. Rigatti SJ. Random forest. J Insur Med. 2017;47(1):31–9. [ DOI ] [ PubMed ] [ Google Scholar ] 38. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. 2016;278(2):563–77. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 39. Friedman JH. Greedy function approximation: a gradient boosting machine. Annals of statistics Ann Stat. 2001;29(5):1189–232. 40. Grydeland TB, Dirksen A, Coxson HO, Eagan TM, Thorsen E, Pillai SG, Sharma S, Eide GE, Gulsvik A, Bakke PS. Quantitative computed tomography measures of emphysema and airway wall thickness are related to respiratory symptoms. Am J Respir Crit Care Med. 2010;181(4):353–9. [ DOI ] [ PubMed ] [ Google Scholar ] 41. Schmidt M, Kuhnigk J-M, Krass S, Owsijewitsch M, de Hoop B, Peitgen H-O. Reproducibility of airway wall thickness measurements. In Medical Imaging 2010: Computer-Aided Diagnosis. 2010;7624:487–96. SPIE. 42. Bhatt SP, Bodduluri S, Nakhmani A, Kim Y-i, Reinhardt JM, Hoffman EA, Motahari A, Wilson CG, Humphries SM, Regan EA. Sex differences in airways at chest CT: results from the COPDGene cohort. Radiology. 2022;305(3):699–708. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 43. Kahnert K, Jörres RA, Kauczor H-U, Alter P, Trudzinski FC, Herth F, Jobst B, Weinheimer O, Nauck S, Mertsch P. Standardized airway wall thickness Pi10 from routine CT scans of COPD patients as imaging biomarker for disease severity, lung function decline, and mortality. Ther Adv Respir Dis. 2023;17:17534666221148663. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. Bumgarner JR, Nelson RJ. Open-source analysis and visualization of segmented vasculature datasets with VesselVio. Cell Rep methods 2022;2(4):100189. [ DOI ] [ PMC free article ] [ PubMed ] 45. Fawcett T. An introduction to ROC analysis. Pattern Recognit Lett. 2006;27(8):861–74. [ Google Scholar ] 46. Sun X, Xu W. Fast implementation of DeLong’s algorithm for comparing the areas under correlated receiver operating characteristic curves. IEEE Signal Process Lett. 2014;21(11):1389–93. [ Google Scholar ] 47. Zhang M-L, Zhou Z-H. ML-KNN: a lazy learning approach to multi-label learning. Pattern Recogn. 2007;40(7):2038–48. [ Google Scholar ] 48. Joachims T. Making large-scale SVM learning practical. In: technical report. 1998. 49. Song Y-Y, Ying L. Decision tree methods: applications for classification and prediction. Shanghai archives psychiatry. 2015;27(2):130. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 50. Hastie T, Rosset S, Zhu J, Zou H. Multi-class adaboost. Stat its Interface. 2009;2(3):349–60. [ Google Scholar ] 51. Chen T, Guestrin C. Xgboost: a scalable tree boosting system. In: Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining . 2016. 9. 785-794. 52. Chang R, Qi S, Wu Y, Yue Y, Zhang X, Qian W. Nomograms integrating CT radiomic and deep learning signatures to predict overall survival and progression-free survival in NSCLC patients treated with chemotherapy. Cancer Imaging. 2023;23(1):101. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 53. Geraghty EM, Boone JM. Determination of height, weight, body mass index, and body surface area with a single abdominal CT image. Radiology. 2003;228(3):857–63. [ DOI ] [ PubMed ] [ Google Scholar ] 54. Park H, Yun J, Lee SM, Hwang HJ, Seo JB, Jung YJ, Hwang J, Lee SH, Lee SW, Kim N. Deep learning–based approach to predict pulmonary function at chest CT. Radiology. 2023;307(2):e221488. [ DOI ] [ PubMed ] [ Google Scholar ] 55. Pascoe S, Locantore N, Dransfield MT, Barnes NC, Pavord ID. Blood eosinophil counts, exacerbations, and response to the addition of inhaled fluticasone furoate to vilanterol in patients with chronic obstructive pulmonary disease: a secondary analysis of data from two parallel randomised controlled trials. lancet Respiratory Med. 2015;3(6):435–42. [ DOI ] [ PubMed ] [ Google Scholar ] 56. Hinds DR, DiSantostefano RL, Le HV, Pascoe S. Identification of responders to inhaled corticosteroids in a chronic obstructive pulmonary disease population using cluster analysis. BMJ open. 2016;6(6):e010099. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 57. Pavord ID, Lettis S, Locantore N, Pascoe S, Jones PW, Wedzicha JA, Barnes NC. Blood eosinophils and inhaled corticosteroid/long-acting β-2 agonist efficacy in COPD. Thorax. 2016;71(2):118–25. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 58. Ruiying W, Jianying X. Clinical features and three-year prognosis of AECOPD patients with different levels of blood eosinophils. Heart Lung. 2022;56:29–39. [ DOI ] [ PubMed ] [ Google Scholar ] 59. Hastie AT, Martinez FJ, Curtis JL, Doerschuk CM, Hansel NN, Christenson S, Putcha N, Ortega VE, Li X, Barr RG. Association of sputum and blood eosinophil concentrations with clinical measures of COPD severity: an analysis of the SPIROMICS cohort. Lancet Respiratory Med. 2017;5(12):956–67. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 60. Huang H, Wu R, Li Y, Peng C. Self-supervised transfer learning based on domain adaptation for benign-malignant lung nodule classification on thoracic CT. IEEE J Biomedical Health Inf. 2022;26(8):3860–71. [ DOI ] [ PubMed ] [ Google Scholar ] 61. Yu P, Zhang H, Wang D, Zhang R, Deng M, Yang H, Wu L, Liu X, Oh AS, Abtin FG. Spatial resolution enhancement using deep learning improves chest disease diagnosis based on thick slice CT. npj Digit Med. 2024;7(1):335. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 62. Angelini ED, Yang J, Balte PP, Hoffman EA, Manichaikul AW, Sun Y, Shen W, Austin JH, Allen NB, Bleecker ER. Pulmonary emphysema subtypes defined by unsupervised machine learning on CT scans. Thorax. 2023;78(11):1067–79. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 63. Wu Y, Xia S, Liang Z, Chen R, Qi S. Artificial intelligence in COPD CT images: identification, staging, and quantitation. Respir Res. 2024;25(1):319. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement No datasets were generated or analysed during the current study. Articles from BMC Pulmonary Medicine are provided here courtesy of BMC ACTIONS View on publisher site PDF (1.6 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top