Abstract
Background
Acute asthma exacerbations (AAEs) are a leading cause of asthma-related morbidity and mortality, especially in resource-limited settings where pulmonary function tests are unavailable or when patients are unable to cooperate with testing. This study aimed to develop and validate a diagnostic model for AAE using routine blood parameters through machine learning techniques.
Methods
We developed a machine learning-based diagnostic model using routine blood test parameters. Data from 23,013 asthma patients treated at the First Affiliated Hospital of Guangzhou Medical University were analyzed. Significant variables were identified through logistic regression, and 12 machine learning algorithms were used to construct diagnostic models, which were evaluated using Receiver Operating Characteristic (ROC) analysis, calibration, and Decision Curve Analysis (DCA).
Results
The Generalized Linear Model Boosting combined with Random Forest (glmBoost + RF) algorithm using 14 variables achieved comparable performance (Area Under the Curve [AUC] = 0.981) to the more complex Least Absolute Shrinkage and Selection Operator combined with Random Forest (Lasso + RF) algorithm using 25 variables (AUC = 0.985). Both models demonstrated excellent calibration and consistent performance across different demographic subgroups. DCA confirmed superior clinical utility compared to conventional strategies.
Conclusions
This machine learning model provides an efficient and practical tool for detecting AAE using routine blood parameters, offering potential value in clinical practice, especially in resource-limited settings.
Clinical trial number
Not applicable.
Keywords: Asthma, Machine learning, Blood chemical analysis, Diagnosis
Introduction
Asthma is one of the most prevalent chronic respiratory diseases, affecting more than 300 million people worldwide.1 Despite advances in treatment, asthma remains the second leading cause of death among chronic respiratory diseases,2 with at least 250,000 deaths annually.3,4 Acute asthma exacerbation (AAE) represents the most dangerous phase in the disease progression, characterized by severe respiratory symptoms that can be life-threatening. The disease significantly impacts quality of life of patients and is frequently associated with various comorbidities, including allergic rhinitis, gastroesophageal reflux disease, obstructive sleep apnea, and anxiety.5,6
Traditional biomarkers for asthma assessment include eosinophils, neutrophils, immunoglobulin E (IgE), periostin, fractional exhaled nitric oxide (FeNO), and leukotrienes.7 Recent studies have shown that inflammatory indicators derived from complete blood count (CBC) are promising biomarkers for evaluating asthma severity and potential therapeutic targets.8, 9, 10 These readily available parameters could be particularly valuable in resource-limited settings.
Pulmonary function testing plays a crucial role in assessing and monitoring asthma. However, the identification and prediction of AAE remain challenging, especially in primary healthcare settings where pulmonary function testing equipment may be limited or unavailable, and in situations where patients are unable to cooperate with pulmonary function testing, such as elderly individuals or those with severe symptoms. While biomarkers play a crucial role in diagnosing and predicting asthma severity, their value in predicting exacerbations is not well defined.11 Traditional statistical approaches to biomarker analysis may be insufficient due to the complex, non-linear nature of biological systems, suggesting the need for incorporating pattern recognition and machine learning techniques.12
Despite the general decline in age-standardized rates of asthma burden over the past 30 years, the overall burden remains substantial, particularly in low- and middle-income countries (LMICs).13,14 The economic burden is especially significant for patients with poorly controlled asthma and those in low-income countries, highlighting the urgent need for more effective diagnostic and management strategies.7
Therefore, we aimed to develop and validate a diagnostic model based on routine blood tests and biochemical parameters to identify AAE. Our study introduces several methodological innovations to the field of asthma exacerbation diagnosis. We systematically evaluated an extensive array of 12 diverse machine learning algorithms, generating 113 unique algorithmic combinations — a comprehensive approach not previously attempted in AAE detection. This exhaustive algorithm exploration allowed us to identify optimal model configurations that balance complexity with performance. Unlike previous studies that relied heavily on specialized clinical measurements or longitudinal data collection, our approach leverages only routine blood parameters that are readily available in most healthcare settings. Our study utilized machine learning algorithms to analyze these readily available laboratory parameters, potentially providing a practical solution for healthcare settings where advanced pulmonary function testing is not available. This approach could contribute to improved identification of acute exacerbations and better patient outcomes, particularly in resource-limited settings.
Materials and methods
Study design
This retrospective study was approved by the Ethics Committee of the First Affiliated Hospital of Guangzhou Medical University (approval number: ES-2023-204). All patient data were anonymized and de-identified before analysis. We retrospectively collected data from patients with asthma who were treated at the First Affiliated Hospital of Guangzhou Medical University between January 2013 and December 2024. We developed a diagnostic model based on CBC and biochemical parameters to address the limitations of pulmonary function testing equipment availability in primary healthcare settings and patient compliance issues, aiming to improve the identification rate of AAE and patient outcomes (Fig. 1). Baseline characteristic analysis was first conducted to identify statistically significant factors, including age, gender, CBC parameters, and biochemical indicators through logistic regression analysis. For model development, the study population was randomly divided into a training set (n = 16,109, 70%) and a validation set (n = 6,904, 30%). Using the significant parameters identified from logistic regression, we constructed diagnostic models through 12 different machine learning algorithms, resulting in 113 algorithmic combinations. Model performance was comprehensively evaluated through Receiver Operating Characteristic (ROC) analysis, calibration assessment, clinical utility analysis, and nomogram development and validation. External validation was subsequently performed to ensure the model's robustness.
Fig. 1.
Study workflow for developing a machine learning-based diagnostic model for acute asthma exacerbation. Patient data (N = 23,013) were analyzed to identify significant laboratory parameters, followed by model development using 12 machine learning algorithms in training (n = 16,109) and validation (n = 6904) sets. Model performance was evaluated through ROC analysis, calibration, and clinical utility assessment
Study population and data collection
This retrospective study analyzed data from 23,013 asthma patients who received treatment at the First Affiliated Hospital of Guangzhou Medical University between January 2013 and December 2024. The study population comprised 11,524 patients with stable asthma and 11,489 patients with AAE. The collected data included demographic characteristics (age and gender) and laboratory test results. Laboratory parameters encompassed CBC indices, including red blood cell (RBC) parameters, white blood cell parameters, and platelet parameters, as well as biochemical parameters (Fig. 2A). Laboratory test results for stable asthma patients were collected during their routine outpatient visits, while samples from AAE patients were collected upon their emergency department visits or hospital admissions. Blood samples from AAE patients were collected at the time of initial emergency care or hospital admission to ensure they reflected the acute phase of the disease, while samples from stable asthma patients were collected during routine outpatient visits. AAE was diagnosed according to the Global Initiative for Asthma guidelines, defined as progressive worsening of symptoms of shortness of breath, cough, wheezing, and/or chest tightness, accompanied by decreases in expiratory airflow. All AAE patients met at least 1 of the following criteria: requirement for systemic corticosteroids, pulmonary function tests showing FEV1 less than 80% of the patient's personal best, or need for urgent medical intervention (such as emergency department visit or hospitalization). All laboratory tests were performed in the clinical laboratory of our hospital, and test results were collected at the time of patient visits. To maintain statistical robustness, laboratory parameters with missing values exceeding one-third of their respective observations were excluded from the analysis. For the remaining variables included in the modeling process, missing values were imputed using the mean value of each parameter. While we recognize that more sophisticated techniques like multiple imputation or k-NN might offer theoretical advantages for non-normally distributed variables, mean imputation was selected for its computational efficiency, which was an important consideration given our large sample size and the extensive algorithmic exploration (113 combinations) conducted in this study.
Fig. 2.
Overview of clinical laboratory indicators and their association with acute asthma exacerbation. (A) Distribution of 28 clinical laboratory indicators categorized into red blood cell parameters, platelet parameters, white blood cell parameters, and biochemical parameters. (B) Forest plot showing odds ratios and 95% confidence intervals for significant factors associated with acute asthma exacerbation. ∗P < 0.05, ∗∗P < 0.01, ∗∗∗P < 0.001
Statistical analysis of group differences
Demographic characteristics and laboratory parameters were compared between patients with stable asthma and AAE. The normality of continuous variables was assessed using the Shapiro-Wilk test. For continuous variables, Student's t-test was applied for normally distributed data, while the Wilcoxon rank-sum test was used for non-normally distributed data. Categorical variables were compared using either chi-square test or Fisher's exact test, as appropriate. Variables showing significant differences between the 2 groups (P < 0.05) were subsequently included in univariable logistic regression analysis to examine their association with AAE. Statistical significance was set at P < 0.05 for all analyses. These analyses provided the foundation for identifying potential factors for the development of our diagnostic model.
Machine learning algorithms for model development
To construct the diagnostic model, we employed 12 diverse machine learning algorithms: Least Absolute Shrinkage and Selection Operator (Lasso), Ridge Regression, Stepwise Generalized Linear Model (Stepglm), Extreme Gradient Boosting (XGBoost), Random Forest (RF), Elastic Net (Enet), Partial Least Squares Regression for Generalized Linear Models (plsRglm), Generalized Boosted Regression Modeling (GBM), Naive Bayes, Linear Discriminant Analysis (LDA), Generalized Linear Model Boosting (glmBoost), and Support Vector Machine (SVM). These algorithms were selected to provide diverse mathematical approaches. First, we included a balanced mix of linear models (Lasso, Ridge Regression, Stepglm), ensemble methods (RF, XGBoost, GBM), and other specialized approaches to capture different mathematical relationships within the data. Second, we selected algorithms with varying capabilities for handling biomedical data characteristics, including those effective with high-dimensional data (Lasso, Ridge), those robust to outliers (RF, SVM), and those adept at capturing non-linear relationships (XGBoost, RF). These 12 base algorithms were systematically combined through a hierarchical approach, generating 113 different algorithmic combinations. This comprehensive combinatorial strategy allowed us to leverage the complementary strengths of different algorithms, with initial algorithms performing feature selection or dimensionality reduction, followed by secondary algorithms focusing on classification. Variable selection and model development were performed within a 10-fold cross-validation framework to ensure robust model performance and minimize overfitting. Each algorithm combination was evaluated for its ability to accurately identify AAE based on blood count and biochemical parameters.
Model performance evaluation
The diagnostic model's performance was comprehensively evaluated through multiple approaches. The discriminative ability was assessed using ROC curves and Area Under the Curve (AUC) values. In our study population, asthma patients were categorized by age into 4 groups: preschool asthma (age <6 years), school-age asthma (≥6 to <13 years), adolescent asthma (≥13 to <19 years), and adult asthma (≥19 years). Model calibration was evaluated using calibration curves, which compared predicted probabilities with observed probabilities. The Mean Absolute Error (MAE) between predicted and observed probabilities was calculated to quantify calibration accuracy. Clinical utility was assessed through Decision Curve Analysis (DCA), which evaluated the model's net benefit at different threshold probabilities. The net benefit was calculated by subtracting the proportion of false-positive cases from true-positive cases, weighted by the relative harm of false-positive and false-negative results. To facilitate clinical implementation, a nomogram was developed based on the final model. This visual scoring system provides clinicians with a practical tool for estimating the probability of AAE based on routine blood test parameters.
Statistical analysis
The Shapiro-Wilk test was used to assess data distribution characteristics. Normally distributed data were presented as mean ± standard deviation, while non-normally distributed data were expressed as median (interquartile range). Categorical variables were presented as frequencies (percentages).
For continuous variables, Student's t-test was applied for normally distributed data, while the Wilcoxon rank-sum test was used for non-normally distributed data. Categorical variables were compared using either chi-square test or Fisher's exact test, depending on data characteristics. All statistical analyses were performed using R software (version 4.3.1), and statistical significance was set at P < 0.05.
Results
Baseline characteristics
A total of 23,013 patients were included in this study, comprising 11,524 patients with stable asthma and 11,489 with AAE (Table 1). In terms of demographic characteristics, patients in the AAE group were significantly older than those in the stable asthma group (60.00 vs 53.00 years, P < 0.001), with a statistically significant difference in gender distribution (P = 0.029).
Table 1.
Clinical characteristics and laboratory findings in patients with Stable Asthma and Acute Asthma Exacerbation
| Variable | Overall N = 23,013 |
Stable Asthma N = 11,524 |
Acute Asthma Exacerbation N = 11,489 |
P-value |
|---|---|---|---|---|
| Age (years) | 57 (44, 68) | 53.00 (37, 65) | 60.00 (49, 71) | 1.37E-226 |
| Gender | ||||
| Male | 11,148 (48.4%) | 5499 (47.7%) | 5649 (49.2%) | 0.029 |
| Female | 11,865 (51.6%) | 6025 (52.3%) | 5840 (50.8%) | 0.029 |
| HB (g/L) | 128.00 (114.00, 140.00) | 131.00 (119.00, 142.00) | 125.00 (110.00, 138.00) | 1.95E-92 |
| HCT (L/L) | 0.39 (0.35, 0.42) | 0.39 (0.36, 0.42) | 0.38 (0.34, 0.42) | 1.17E-47 |
| Eosinophil percentage (%) | 2.20 (0.50, 5.30) | 3.00 (1.10, 6.00) | 1.30 (0.10, 4.20) | 4.97E-207 |
| Basophil percentage (%) | 0.40 (0.20, 0.70) | 0.50 (0.30, 0.70) | 0.40 (0.20, 0.60) | 6.84E-149 |
| MCV (fL) | 89.20 (85.10, 92.30) | 88.80 (84.50, 92.00) | 89.60 (85.80, 92.50) | 5.93E-18 |
| PLT (10ˆ9/L) | 241.00 (193.00, 294.00) | 248.00 (204.00, 301.00) | 232.00 (182.00, 285.00) | 1.48E-57 |
| Neutrophils count (10ˆ9/L) | 4.96 (3.50, 7.70) | 4.29 (3.20, 6.20) | 6.00 (4.00, 9.05) | 5.03E-288 |
| Lymphocytes count (10ˆ9/L) | 1.70 (1.10, 2.30) | 1.85 (1.40, 2.40) | 1.50 (0.90, 2.10) | 1.04E-182 |
| Monocytes count (10ˆ9/L) | 0.51 (0.40, 0.80) | 0.50 (0.40, 0.70) | 0.60 (0.40, 0.80) | 5.53E-61 |
| Eosinophils count (10ˆ9/L) | 0.17 (0.04, 0.40) | 0.20 (0.10, 0.40) | 0.10 (0.00, 0.32) | 5.21E-147 |
| Basophils count (10ˆ9/L) | 0.03 (0.00, 0.05) | 0.03 (0.01, 0.06) | 0.02 (0.00, 0.05) | 1.75E-40 |
| MPV (fL) | 8.60 (7.90, 9.40) | 8.60 (7.90, 9.40) | 8.60 (7.90, 9.40) | 0.205 |
| Monocyte percentage (%) | 6.90 (5.30, 8.70) | 6.90 (5.50, 8.50) | 6.90 (5.00, 8.97) | 0.005 |
| Lymphocyte percentage (%) | 22.70 (12.60, 31.50) | 26.80 (18.00, 34.30) | 17.30 (9.10, 27.50) | <0.001 |
| Neutrophil percentage (%) | 64.50 (54.20, 78.00) | 60.30 (51.60, 70.50) | 70.80 (58.10, 83.20) | <0.001 |
| Plateletcrit (%) | 0.21 (0.17, 0.25) | 0.21 (0.18, 0.26) | 0.20 (0.16, 0.24) | 1.37E-68 |
| Nucleated red blood cells count (10ˆ9/L) | 0.00 (0.00, 0.00) | 0.00 (0.00, 0.00) | 0.00 (0.00, 0.00) | 1.27E-12 |
| Nucleated RBC percentage (%) | 0.00 (0.00, 0.00) | 0.00 (0.00, 0.00) | 0.00 (0.00, 0.00) | 0.579 |
| Potassium (mmol/L) | 3.86 (3.58, 4.14) | 3.94 (3.68, 4.21) | 3.79 (3.51, 4.07) | 1.17E-109 |
| Sodium (mmol/L) | 139.00 (136.90, 140.80) | 139.10 (137.00, 140.90) | 138.70 (136.40, 140.60) | 4.58E-22 |
| Chloride (mmol/L) | 104.90 (102.20, 107.00) | 104.90 (102.80, 106.90) | 104.90 (101.70, 107.20) | 0.064 |
| Creatinine (umol/L) | 70.00 (56.20, 85.60) | 70.60 (57.00, 86.40) | 69.00 (56.00, 84.90) | 0.001 |
| Glucose (mmol/L) | 5.51 (4.77, 7.13) | 5.20 (4.70, 6.05) | 6.10 (4.89, 8.00) | 4.31E-175 |
| Calcium (mmol/L) | 2.22 (2.13, 2.32) | 2.26 (2.15, 2.36) | 2.20 (2.11, 2.29) | 2.20E-107 |
| MCH (pg) | 29.80 (28.20, 31.00) | 29.70 (28.10, 31.00) | 29.80 (28.30, 31.00) | 0.006 |
| RDW-CV (%) | 13.80 (13.10, 15.00) | 13.60 (13.00, 14.70) | 14.00 (13.30, 15.30) | 7.36E-77 |
| MCHC (g/L) | 332.00 (325.00, 339.00) | 333.00 (326.00, 340.00) | 331.00 (324.00, 339.00) | 1.10E-16 |
| PDW (fL) | 16.40 (16.00, 17.00) | 16.20 (16.00, 17.00) | 16.60 (16.00, 17.00) | 4.25E-72 |
Data are presented as mean ± standard deviation, median (25th percentile, 75th percentile), or count (percentage). P-values were calculated using the Student's t-test, Wilcoxon rank-sum test, or chi-square test, as appropriate. HB: Hemoglobin; HCT: Hematocrit; MCV: Mean Corpuscular Volume; PLT: Platelet Count; MPV: Mean Platelet Volume; MCH: Mean Corpuscular Hemoglobin; RDW-CV: Red Cell Distribution Width Coefficient of Variation; MCHC: Mean Corpuscular Hemoglobin Concentration; PDW: Platelet Distribution Width.
The AAE group demonstrated distinct inflammatory characteristics in hematological parameters compared to the stable asthma group. Neutrophils count (6.00 vs 4.29, P < 0.001) and neutrophil percentage (70.80% vs 60.30%, P < 0.001) were significantly elevated, while eosinophil percentage (1.30% vs 3.00%, P < 0.001) and lymphocytes count (1.50 vs 1.85, P < 0.001) were significantly lower. The AAE group also showed notably lower levels of hemoglobin (HB) (125.00 vs 131.00, P < 0.001) and platelet count (PLT) (232.00 vs 248.00, P < 0.001). Mean platelet volume (MPV) (P = 0.205) and nucleated RBC percentage (P = 0.579) showed no significant differences between groups.
Biochemical analysis revealed significantly lower levels of potassium (3.79 vs 3.94 mmol/L, P < 0.001) and calcium (2.20 vs 2.26 mmol/L, P < 0.001) in the AAE group, while blood glucose levels were significantly elevated (6.10 vs 5.20 mmol/L, P < 0.001). No significant difference was observed in chloride levels between the groups (P > 0.05).
Identification of significant variables
Through comparative analysis between stable asthma and AAE groups, we identified 27 statistically significant variables, including age, gender, and 25 laboratory parameters. To further determine the factors influencing AAE, logistic regression analysis was performed on these 27 variables. The analysis revealed that 25 parameters were significant predictive factors for AAE (Fig. 2B).
Model development and performance analysis
To evaluate model performance and generalizability, the study population was randomly divided into a training cohort (n = 16,109, 70%) and a validation cohort (n = 6,904, 30%). The baseline characteristics were comparable between the 2 cohorts, with no statistically significant differences observed across all variables (Table 2, all P values > 0.05), indicating successful randomization and providing a reliable foundation for model development and validation.
Table 2.
Comparison of clinical characteristics and laboratory findings between training and validation cohorts
| Variable | Overall N = 23,013(100%) |
Training Cohort N = 16,109(70%) |
Validation Cohort N = 6904(30%) |
P-value |
|---|---|---|---|---|
| Age (years) | 57 (44, 68) | 57 (44, 68) | 57 (44, 68) | 0.842 |
| Gender | ||||
| Male | 11,148 (48.4%) | 7807 (48.5%) | 3341 (48.4%) | 0.932 |
| Female | 11,865 (51.6%) | 8302 (51.5%) | 3563 (51.6%) | 0.932 |
| HB (g/L) | 128.00 (114.00, 140.00) | 128.00 (114.00, 140.00) | 128.00 (114.00, 139.00) | 0.309 |
| HCT (L/L) | 0.39 (0.35, 0.42) | 0.39 (0.35, 0.42) | 0.39 (0.35, 0.42) | 0.443 |
| Eosinophil percentage (%) | 2.20 (0.50, 5.30) | 2.20 (0.50, 5.30) | 2.20 (0.50, 5.30) | 0.651 |
| Basophil percentage (%) | 0.40 (0.20, 0.70) | 0.40 (0.20, 0.70) | 0.40 (0.20, 0.70) | 0.816 |
| MCV (fL) | 89.20 (85.10, 92.30) | 89.20 (85.10, 92.20) | 89.20 (85.10, 92.40) | 0.491 |
| PLT (10ˆ9/L) | 241.00 (193.00, 294.00) | 241.00 (193.00, 294.00) | 240.00 (193.00, 294.00) | 0.865 |
| Neutrophils count (10ˆ9/L) | 4.96 (3.50, 7.70) | 5.00 (3.50, 7.70) | 4.90 (3.50, 7.60) | 0.608 |
| Lymphocytes count (10ˆ9/L) | 1.70 (1.10, 2.30) | 1.70 (1.10, 2.30) | 1.70 (1.10, 2.30) | 0.252 |
| Monocytes count (10ˆ9/L) | 0.51 (0.40, 0.80) | 0.51 (0.40, 0.80) | 0.50 (0.40, 0.80) | 0.423 |
| Eosinophils count (10ˆ9/L) | 0.17 (0.04, 0.40) | 0.17 (0.03, 0.39) | 0.18 (0.04, 0.40) | 0.605 |
| Basophils count (10ˆ9/L) | 0.03 (0.00, 0.05) | 0.03 (0.00, 0.05) | 0.03 (0.00, 0.05) | 0.678 |
| MPV (fL) | 8.60 (7.90, 9.40) | 8.60 (7.90, 9.40) | 8.60 (7.90, 9.50) | 0.269 |
| Monocyte percentage (%) | 6.90 (5.30, 8.70) | 6.90 (5.30, 8.70) | 6.90 (5.40, 8.70) | 0.329 |
| Lymphocyte percentage (%) | 22.70 (12.60, 31.50) | 22.70 (12.50, 31.50) | 22.90 (12.80, 31.60) | 0.443 |
| Neutrophil percentage (%) | 64.50 (54.20, 78.00) | 64.50 (54.30, 78.10) | 64.40 (54.10, 77.80) | 0.504 |
| Plateletcrit (%) | 0.21 (0.17, 0.25) | 0.21 (0.17, 0.25) | 0.21 (0.17, 0.25) | 0.813 |
| Nucleated red blood cells count (10ˆ9/L) | 0.00 (0.00, 0.00) | 0.00 (0.00, 0.00) | 0.00 (0.00, 0.00) | 0.759 |
| Nucleated RBC percentage (%) | 0.00 (0.00, 0.00) | 0.00 (0.00, 0.00) | 0.00 (0.00, 0.00) | 0.835 |
| Potassium (mmol/L) | 3.86 (3.58, 4.14) | 3.86 (3.58, 4.14) | 3.87 (3.58, 4.14) | 0.943 |
| Sodium (mmol/L) | 139.00 (136.90, 140.80) | 139.00 (136.90, 140.80) | 139.00 (136.80, 140.70) | 0.633 |
| Chloride (mmol/L) | 104.90 (102.20, 107.00) | 104.90 (102.10, 107.00) | 104.90 (102.20, 107.00) | 0.696 |
| Creatinine (umol/L) | 70.00 (56.20, 85.60) | 70.00 (56.20, 85.30) | 70.00 (56.30, 86.00) | 0.838 |
| Glucose (mmol/L) | 5.51 (4.77, 7.13) | 5.50 (4.76, 7.11) | 5.56 (4.79, 7.17) | 0.294 |
| Calcium (mmol/L) | 2.22 (2.13, 2.32) | 2.23 (2.13, 2.32) | 2.22 (2.13, 2.32) | 0.801 |
| MCH (pg) | 29.80 (28.20, 31.00) | 29.80 (28.20, 31.00) | 29.80 (28.30, 31.00) | 0.502 |
| RDW-CV (%) | 13.80 (13.10, 15.00) | 13.80 (13.10, 15.00) | 13.80 (13.10, 15.00) | 0.766 |
| MCHC (g/L) | 332.00 (325.00, 339.00) | 332.00 (325.00, 340.00) | 332.00 (325.00, 339.00) | 0.747 |
| PDW (fL) | 16.40 (16.00, 17.00) | 16.40 (16.00, 17.00) | 16.40 (16.00, 17.00) | 0.829 |
Data are presented as mean ± standard deviation, median (25th percentile, 75th percentile), or count (percentage). P-values were calculated using the Student's t-test, Wilcoxon rank-sum test, or chi-square test, as appropriate. HB: Hemoglobin; HCT: Hematocrit; MCV: Mean Corpuscular Volume; PLT: Platelet Count; MPV: Mean Platelet Volume; MCH: Mean Corpuscular Hemoglobin; RDW-CV: Red Cell Distribution Width Coefficient of Variation; MCHC: Mean Corpuscular Hemoglobin Concentration; PDW: Platelet Distribution Width.
Using 10-fold cross-validation, we combined 12 machine learning algorithms to construct diagnostic models based on 25 significant variables. The analysis was conducted in both training and validation cohorts (Fig. 3A). Results demonstrated that the Least Absolute Shrinkage and Selection Operator combined with Random Forest (Lasso + RF) algorithm using 25 variables showed sensitivity of 93.47%, specificity of 98.78%, and F1-score of 96.05%, while the Generalized Linear Model Boosting combined with Random Forest (glmBoost + RF) algorithm achieved comparable diagnostic efficacy with sensitivity of 92.64%, specificity of 98.23%, and F1-score of 95.35% using only 14 variables (Fig. 3B and C).
Fig. 3.
Performance comparison of machine learning algorithms and stratified analysis. (A) Performance metrics of 113 algorithm combinations in both training and validation sets, ordered by AUC values. The color intensity represents model performance. (B) ROC curves for Lasso + RF model in training and validation sets. (C) ROC curves for glmBoost + RF model in training and validation sets. (D) Stratified analysis showing ROC curves for the Lasso + RF model across gender subgroups (male and female). (E) Stratified analysis showing ROC curves for the glmBoost + RF model across gender subgroups (male and female). (F) Stratified analysis showing ROC curves for the Lasso + RF model across age categories (preschool asthma, school-age asthma, adolescent asthma, and adult asthma). (G) Stratified analysis showing ROC curves for the glmBoost + RF model across age categories (preschool asthma, school-age asthma, adolescent asthma, and adult asthma)
The selection of 14 variables in the glmBoost + RF model occurred through the algorithm's inherent feature selection mechanism. During model training, the glmBoost algorithm iteratively builds the model by sequentially adding predictors that most significantly reduce the prediction error at each step. This process naturally assigns importance to variables with stronger predictive power, while effectively eliminating less informative features by assigning them negligible or zero coefficients. The 14 variables retained in the final model were those that received substantial weights during this iterative process, reflecting their significant contribution to the model's predictive accuracy. This embedded feature selection capability of glmBoost provided a data-driven approach to identifying the most informative subset of laboratory parameters for AAE diagnosis.
Further stratified analysis revealed that both algorithms demonstrated similar diagnostic performance across gender groups and age categories (preschool asthma, school-age asthma, adolescent asthma, and adult asthma) (Fig. 3D–G), indicating consistent diagnostic effectiveness across all population subgroups.
The glmBoost + RF algorithm selected 14 most valuable routine laboratory parameters to construct the model, offering a more streamlined and efficient clinical diagnostic approach while maintaining near-optimal diagnostic performance. By optimizing the number of variables from 25 to 14, this model significantly enhanced clinical utility while maintaining robust diagnostic capability. The detailed composition of feature variables for all 113 model combinations is presented in Supplementary Table 1.
Comprehensive model evaluation and clinical utility assessment
Further comprehensive evaluation of the diagnostic models' performance was conducted. Calibration curves demonstrated excellent agreement between predicted probabilities and actual observations, with apparent and bias-corrected curves closely aligned with the ideal line. Both the glmBoost + RF algorithm (MAE = 0.017) and the Lasso + RF algorithm (MAE = 0.01) exhibited exceptional calibration performance (Fig. 4A and B), further supporting the feasibility of the simplified model (glmBoost + RF) in maintaining model performance.
Fig. 4.
Model calibration, clinical utility, and nomogram development. (A) Calibration curve for Lasso + RF model. (B) Calibration curve for glmBoost + RF model. (C) Decision curve analysis comparing model net benefits. (D,E) Nomograms for individual patient risk prediction
DCA revealed that both Lasso + RF and glmBoost + RF algorithms demonstrated similar net benefits across various risk thresholds. The curves for both algorithms significantly outperformed the baseline strategies of "treat all" and "treat none" (Fig. 4C). This finding indicates that despite using fewer variables, the glmBoost + RF algorithm maintains comparable clinical decision-making utility to the Lasso + RF algorithm.
To facilitate clinical implementation, we developed nomograms incorporating age, gender, and algorithm-predicted values (Lasso + RF/glmBoost + RF). The nomograms for both algorithms exhibited similar predictive structures (Fig. 4D and E), demonstrating comprehensive risk assessment capabilities for AAE. Nomograms are visual scoring tools that allow clinicians to estimate individual risk probabilities through a simple point-based system. In our study, the nomograms integrate patient age, gender, and algorithm-predicted values, providing clinicians with a practical method to assess the probability of AAE. To use the nomogram, one first identifies the position corresponding to each of the patient's variables on their respective axes, projecting upward to determine the points on the 'Points' axis. These points are then summed to obtain a total score, which is mapped to the corresponding probability of AAE on the probability axis. This intuitive risk assessment tool facilitates rapid clinical decision-making, particularly valuable in resource-limited healthcare settings.
Discussion
In this large-scale retrospective study involving 23,013 asthma patients, we developed and validated a machine learning-based diagnostic model for AAE using routine blood parameters. Our study revealed distinct clinical, hematological, and biochemical characteristics associated with AAE and demonstrated the effectiveness of machine learning approaches in predicting exacerbations.
The demographic analysis identified age as a significant risk factor, with AAE patients being notably older. This age-related vulnerability likely reflects multiple factors,15 including reduced lung function, increased comorbidities, and altered immune responses in older populations.16, 17, 18 Our study revealed significant changes in inflammatory patterns during AAE, characterized by elevated neutrophil counts and percentages, alongside decreased eosinophil percentages. The observed neutrophilic predominance in AAE suggests a potential shift in inflammatory patterns during exacerbation, which may have important implications for treatment strategies.19 Despite the observed neutrophilic inflammation pattern in AAE, it is important to clarify that we do not suggest antibiotic therapy as a treatment approach based solely on this finding. Neutrophilic airway inflammation in asthma can be triggered by viral infections, environmental factors, and oxidative stress without bacterial infection. Current evidence does not support routine antibiotic use for asthma exacerbations unless clear signs of bacterial infection are present. The neutrophilic pattern we observed likely represents the acute inflammatory response during exacerbation rather than indicating a specific microbial etiology. Treatment should follow established asthma management guidelines focusing on bronchodilators and anti-inflammatory therapies. While traditional asthma biomarkers have focused on eosinophilic inflammation,20, 21, 22 our results support the growing evidence that neutrophilic inflammation plays a crucial role in acute exacerbations. This finding could potentially influence therapeutic approaches,23 suggesting the need to consider treatments targeting neutrophilic inflammation in AAE management.
Lower HB and PLT levels in AAE patients, combined with decreased potassium and calcium levels and elevated blood glucose, indicate a systemic response to acute exacerbation. These findings align with emerging evidence of metabolic dysfunction in exacerbation-prone asthma24, 25, 26 and suggest potential new avenues for monitoring and intervention. The identification of these systemic changes could lead to the development of novel therapeutic strategies targeting metabolic pathways in AAE management.
In the model development phase, we compared 2 machine learning approaches: the Lasso + RF algorithm using 25 variables and the glmBoost + RF algorithm using 14 variables. Both models demonstrated excellent calibration performance, with low mean absolute errors (MAE = 0.01 and 0.017, respectively). The comparable performance of the streamlined glmBoost + RF model represents an important advancement in clinical applicability while maintaining diagnostic accuracy. Notably, both algorithms demonstrated consistent performance across different demographic subgroups, including age categories and gender groups, suggesting broad applicability despite the known heterogeneous nature of asthma.27,28
DCA validated the clinical utility of both models, showing superior net benefits compared to default clinical strategies. The development of user-friendly nomograms further enhances the practical implementation of these models in clinical settings. The similar predictive structures of both nomograms provide additional validation of our approach, offering clinicians intuitive tools for risk assessment in routine practice.
Compared to existing AAE diagnostic approaches, our model offers distinct advantages. Traditional clinical scoring systems (PASS, Pulmonary Score, PRAM) rely on subjective observations that may introduce variability, while specialized biomarker approaches often require equipment unavailable in primary care settings. Our model demonstrates superior performance to previous biomarker models using only routine laboratory parameters. However, we recognize these different approaches serve complementary purposes - clinical scoring systems excel in rapid emergency assessment, while our model may be particularly valuable in resource-limited settings lacking specialized diagnostic capabilities.
Several limitations of our study should be acknowledged. First, as a single-center retrospective study, our findings may be influenced by selection bias and regional characteristics. Second, while our model shows robust performance across different demographic subgroups, external validation in diverse populations and healthcare settings is needed. To address these limitations, future work should include prospective, multi-center validations across geographically and demographically diverse populations. External validation using independent datasets from multiple healthcare systems would be particularly valuable to assess the model's generalizability and potential implementation challenges. Additionally, validation in settings with different laboratory reference ranges and testing methodologies would help establish the robustness of our approach in varied clinical environments. Third, the temporal relationship between laboratory parameter changes and exacerbation onset requires further investigation through prospective studies. Our cross-sectional design captures laboratory parameters only at specific time points during acute exacerbation or stable state, without the ability to track their evolution over time. This approach limits our understanding of how these parameters might change in the days or hours preceding an exacerbation, which could be valuable for early intervention. A longitudinal design with serial measurements would be optimal for future studies to characterize the dynamic changes that might serve as early warning signals before clinical symptoms become apparent.
Another significant limitation is that our model does not incorporate important clinical variables beyond age and gender. Asthma is a heterogeneous disease with multiple phenotypes and endotypes, and factors such as allergic versus non-allergic presentation, comorbidities, and treatment regimens could significantly influence both laboratory parameters and exacerbation risk. For example, corticosteroid therapy typically suppresses eosinophil counts, while biologics targeting specific inflammatory pathways alter the inflammatory profile reflected in blood tests. The absence of these variables may affect model performance when applied to heterogeneous asthma populations under various treatment protocols. While these laboratory markers provide valuable objective data, comprehensive AAE diagnosis requires integration with clinical symptoms, physical examination findings, and when available, pulmonary function testing. Relying exclusively on blood parameters without considering the broader clinical context could potentially lead to misdiagnosis in certain cases, particularly in patients with comorbidities that might affect hematological parameters independently of asthma status. Therefore, we strongly emphasize that our model should be implemented as a complementary diagnostic tool to support, rather than replace, clinical assessment. Furthermore, laboratory parameter changes during exacerbations might differ substantially between asthma phenotypes. The inflammatory patterns in T2-high versus T2-low asthma could result in distinctly different biomarker profiles during exacerbations. The optimal approach involves interpreting our model's predictions within the context of the patient's clinical presentation, medical history, and response to treatment, leveraging both the objective nature of laboratory data and the nuanced insights from clinical evaluation. Future model refinement should consider stratification by asthma phenotype to maximize predictive accuracy across the spectrum of asthma presentations.
Future research directions should include prospective multi-center validation studies, investigation of the temporal dynamics of biomarker changes during exacerbation progression, and evaluation of the model's impact on clinical outcomes and healthcare resource utilization. Additional stratified analyses examining model performance across patients with different comorbidities, medication regimens, and asthma phenotypes would be particularly valuable to determine if model recalibration is needed for specific patient subgroups. Studies examining the integration of our laboratory parameter-based model with clinical variables including phenotyping data, comorbidity assessment, and treatment information would be valuable for developing more comprehensive predictive tools. Despite these limitations, our current model represents an important step toward developing accessible diagnostic tools for AAE, particularly in settings where advanced diagnostic capabilities may be limited.
To facilitate real-world implementation, our model could be integrated into electronic health record systems as a clinical decision support tool. The limited number of required inputs (14 routine laboratory parameters plus basic demographics) makes it feasible for integration into existing laboratory information systems. Such implementation could automate risk assessment during routine blood work, potentially enabling earlier intervention for patients at high risk of AAE, particularly in primary care settings with limited specialized diagnostic capabilities.
Our study's findings have particular relevance for resource-limited settings, where access to specialized pulmonary function testing may be restricted.29, 30, 31 The use of readily available blood parameters, combined with machine learning techniques, offers a practical and efficient approach to predicting asthma exacerbations. This approach aligns with the growing recognition that sophisticated analytical methods can effectively leverage routine clinical data to improve patient care.12
These results suggest that machine learning-based analysis of standard blood parameters can provide valuable insights into asthma exacerbation risk, potentially transforming how we approach monitoring and early intervention in asthma care. The practical utility and broad applicability of our model represent a significant advancement in asthma management, particularly in resource-limited settings.
Abbreviations
AAE: Acute Asthma Exacerbation, AUC: Area Under the Curve, CBC: Complete Blood Count, DCA: Decision Curve Analysis, Enet: Elastic Net, FeNO: Fractional Exhaled Nitric Oxide, GBM: Generalized Boosted Regression Modeling, glmBoost: Generalized Linear Model Boosting, glmBoost + RF: Generalized Linear Model Boosting combined with Random Forest, HB: Hemoglobin, IgE: Immunoglobulin E, Lasso: Least Absolute Shrinkage and Selection Operator, Lasso + RF: Least Absolute Shrinkage and Selection Operator combined with Random Forest, LDA: Linear Discriminant Analysis, LMICs: Low- and Middle-Income Countries, MAE: Mean Absolute Error, MPV: Mean Platelet Volume, PLT: Platelet Count, plsRglm: Partial Least Squares Regression for Generalized Linear Models, RBC: Red Blood Cell, RF: Random Forest, ROC: Receiver Operating Characteristic, Stepglm: Stepwise Generalized Linear Model, SVM: Support Vector Machine, XGBoost: Extreme Gradient Boosting.
Availability of data and materials
The datasets generated and analyzed during the current study are available from the corresponding author on reasonable request.
Authors' contributions
YC, JS: Conceptualization, Investigation, Visualization, and Writing - Original Draft. YC, EL: Formal Analysis, Writing - Review & Editing. JL, HT: Writing - Original Draft and Formal Analysis. YX: Visualization and Formal Analysis. JZ, LP, HW: Visualization and Formal Analysis. ZJC, BS: Project Administration and Supervision. All authors (YC, JS, YC, EL, JL, HT, YX, JZ, LP, HW, ZJC, BS) contributed to the article and approved the submitted version.
Ethics approval
This study was conducted in accordance with the Declaration of Helsinki. This retrospective study was approved by the Ethics Committee of the First Affiliated Hospital of Guangzhou Medical University (approval number: ES-2023-204). Patient confidentiality was strictly maintained throughout the study by using coding systems and restricting data access to authorized research personnel only. Informed consent to participate was obtained from all human participants involved in this study.
Consent for publication
All authors have reviewed the final version of this manuscript and approve its publication.
Funding
This study was supported by the State Key Laboratory of Respiratory Disease (SKLRD-OP-202402), the Guangzhou Municipal Science and Technology Bureau (SL2024A04J00706), and the Multi-Center Clinical Research Project of Guangzhou Medical University (GMUCR2025-02009).
Declaration of competing interest
The authors declare that they have no competing interests.
Acknowledgments
Not applicable.
Footnotes
Full list of author information is available at the end of the article
Supplementary data to this article can be found online at https://doi.org/10.1016/j.waojou.2025.101074.
Contributor Information
Zhangkai J. Cheng, Email: jasontable@gmail.com.
Baoqing Sun, Email: sunbaoqing@vip.163.com.
Appendix A. Supplementary data
The following is the Supplementary data to this article.
References
- 1.Brusselle G.G., Koppelman G.H. Biologic therapies for severe asthma. N Engl J Med. 2022;386:157–171. doi: 10.1056/NEJMra2032506. [DOI] [PubMed] [Google Scholar]
- 2.Global burden of 369 diseases and injuries in 204 countries and territories, 1990-2019: a systematic analysis for the global burden of disease study 2019. Lancet. 2020;396:1204–1222. doi: 10.1016/s0140-6736(20)30925-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Bourdin A., Brusselle G., Couillard S., et al. Phenotyping of severe asthma in the era of broad-acting anti-asthma biologics. J Allergy Clin Immunol Pract. 2024 Apr;12(4):809–823. doi: 10.1016/j.jaip.2024.01.023. [DOI] [PubMed] [Google Scholar]
- 4.Melén E, Zar HJ, Siroux V, et al. Asthma inception: epidemiologic risk factors and natural history across the life course. Am J Respir Crit Care Med. 2024 Sep 15;210(6):737–754. doi: 10.1164/rccm.202312-2249SO. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Althoff M.D., Ghincea A., Wood L.G., Holguin F., Sharma S. Asthma and three colinear comorbidities: obesity, OSA, and GERD. J Allergy Clin Immunol Pract. 2021;9:3877–3884. doi: 10.1016/j.jaip.2021.09.003. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Ye G., Baldwin D.S., Hou R. Anxiety in asthma: a systematic review and meta-analysis. Psychol Med. 2021;51:11–20. doi: 10.1017/s0033291720005097. [DOI] [PubMed] [Google Scholar]
- 7.Gans M.D., Gavrilova T. Understanding the immunology of asthma: pathophysiology, biomarkers, and treatments for asthma endotypes. Paediatr Respir Rev. 2020;36:118–127. doi: 10.1016/j.prrv.2019.08.002. [DOI] [PubMed] [Google Scholar]
- 8.Tillie-Leblond I, Montani D, Crestani B, et al. Relation between inflammation and symptoms in asthma. Allergy. 2009 Mar;64(3):354–367. doi: 10.1111/j.1398-9995.2009.01971.x. [DOI] [PubMed] [Google Scholar]
- 9.Ke J., Qiu F., Fan W., Wei S. Associations of complete blood cell count-derived inflammatory biomarkers with asthma and mortality in adults: a population-based study. Front Immunol. 2023;14 doi: 10.3389/fimmu.2023.1205687. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Singh S., Kumar P., Goyal J.P., Singh K. Relationship of peripheral blood counts with asthma control. Indian J Pediatr. 2023;90:555–559. doi: 10.1007/s12098-022-04262-5. [DOI] [PubMed] [Google Scholar]
- 11.di Palmo E, Cantarelli E, Catelli A, et al. The predictive role of biomarkers and genetics in childhood asthma exacerbations. Int J Mol Sci. 2021 Apr 28;22(9):4651. doi: 10.3390/ijms22094651. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Sterk P.J., Sinha A. Emerging complexity in the biomarkers of exacerbation-prone asthma. Am J Respir Crit Care Med. 2020;202:915–917. doi: 10.1164/rccm.202005-2004ED. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Meghji J, Mortimer K, Agusti A, et al. Improving lung health in low-income and middle-income countries: from challenges to solutions. Lancet. 2021 Mar 6;397(10277):928–940. doi: 10.1016/S0140-6736(21)00458-X. [DOI] [PubMed] [Google Scholar]
- 14.Wang Z, Li Y, Gao Y, et al. Global regional, and national burden of asthma and its attributable risk factors from 1990 to 2019: a systematic analysis for the Global Burden of Disease Study 2019. Respir Res. 2023 Jun 23;24(1):169. doi: 10.1186/s12931-023-02475-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Dabbs W., Bradley M.H., Chamberlin S.M. vol 109. American family physician; 2024. pp. 43–50. (Acute Asthma Exacerbations: Management Strategies). [PubMed] [Google Scholar]
- 16.Zhao M, Zhai H, Li H, et al. Age-standardized incidence, prevalence, and mortality rates of autoimmune diseases in adolescents and young adults (15–39 years): an analysis based on the Global Burden of Disease Study 2021.BMC Public. Health. 2024 Jul;24(1):1800. doi: 10.1186/s12889-024-19290-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Wang J, Zhang X, Zhang L, et al. Age-related clinical characteristics, inflammatory features, phenotypes, and treatment response in asthma. J Allergy Clin Immunol Pract. 2023 Jan;11(1):210–219.e3. doi: 10.1016/j.jaip.2022.09.029. [DOI] [PubMed] [Google Scholar]
- 18.García-Marcos L, Chiang CY, Asher MI, et al. Asthma management and control in children, adolescents, and adults in 25 countries: a Global Asthma Network Phase I cross-sectional study. Lancet Glob Health. 2023 Feb;11(2):e218–e228. doi: 10.1016/S2214-109X(22)00506-X. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Li Y., Yang T., Jiang B. Neutrophil and neutrophil extracellular trap involvement in neutrophilic asthma: a review. Medicine. 2024;103 doi: 10.1097/md.0000000000039342. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Porsbjerg CM, Townend J, Bergeron C, et al. Association between pre-biologic T2-biomarker combinations and response to biologics in patients with severe asthma. Front Immunol. 2024 Apr 19;15:1361891. doi: 10.3389/fimmu.2024.1361891. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Viinanen A, Aakko J, Lassenius MI, et al. Type 2 low biomarker stability and exacerbations in severe uncontrolled asthma. Biomolecules. 2023 Jul 13;13(7):1118. doi: 10.3390/biom13071118. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Zeng X, Qing J, Li CM, et al. Blood transcriptomic signature in type-2 biomarker-low severe asthma and asthma control. J Allergy Clin Immunol. 2023 Oct;152(4):876–886. doi: 10.1016/j.jaci.2023.05.023. [DOI] [PubMed] [Google Scholar]
- 23.Bleecker ER, Panettieri RA, Jr., Lugogo NL, et al. Dupilumab efficacy in patients with type 2 asthma with and without elevated blood neutrophils. J Immunol Res. 2023 Oct 19;2023:9943584. doi: 10.1155/2023/9943584. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Peters MC, Mauger D, Ross KR, et al. Evidence for exacerbation-prone asthma and predictive biomarkers of exacerbation frequency. Am J Respir Crit Care Med. 2020 Oct 1;202(7):973–982. doi: 10.1164/rccm.201909-1813OC. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Staggers K.A., Minard C., Byers M., Helmer D.A., Wu T.D. Metabolic dysfunction, triglyceride-glucose index, and risk of severe asthma exacerbation. J Allergy Clin Immunol. 2023;11:3700–3705.e3702. doi: 10.1016/j.jaip.2023.09.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Park J.W. Asthma phenotype with metabolic dysfunction. Yonsei Med J. 2022;63:1–7. doi: 10.3349/ymj.2022.63.1.1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Kuruvilla M.E., Lee F.E., Lee G.B. Understanding asthma phenotypes, endotypes, and mechanisms of disease. Clin Rev Allergy Immunol. 2019;56:219–233. doi: 10.1007/s12016-018-8712-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Guida G, Bagnasco D, Carriero V, et al. Critical evaluation of asthma biomarkers in clinical practice. Front Med (Lausanne) 2022 Oct 10;9:969243. doi: 10.3389/fmed.2022.969243. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Jackson P., Muyanja S.Z., Siddharthan T. Health equity and respiratory diseases in Low- and middle-income countries. Clin Chest Med. 2023;44:623–634. doi: 10.1016/j.ccm.2023.03.015. [DOI] [PubMed] [Google Scholar]
- 30.Levy ML, Bacharier LB, Bateman E, et al. Key recommendations for primary care from the 2022 Global Initiative for Asthma (GINA) update. NPJ Prim Care Respir Med. 2023;33(1):7. doi: 10.1038/s41533-023-00330-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Barne M. Gaps in asthma diagnosis and treatment in low- and middle-income countries. Frontiers in allergy. 2023;4 doi: 10.3389/falgy.2023.1240259. [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.
Supplementary Materials
Data Availability Statement
The datasets generated and analyzed during the current study are available from the corresponding author on reasonable request.