.png){kind=icon}
.png){kind=icon}
Return
Development and validation of a machine learning-based diagnostic prediction model for adult eosinophilic asthma
.png){kind=icon}
OA
AI
M
Z
H
T
Z
Y
Y
M
Y
Y
N
Y
DOI：10.1186/s12931-026-03852-7.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Adult eosinophilic asthma (EA) is a predominant phenotype of severe asthma. Induced sputum cytology, the diagnostic gold standard, is limited by complex procedures and poor feasibility. Non-invasive type 2 markers perform unsatisfactorily, and existing machine learning models lack interpretability and clinical accessibility. This study aimed to construct a machine learning (ML)-based prediction model for EA to enable accurate phenotyping and personalized precision treatment of asthma. A single-center retrospective study was performed at a tertiary hospital in China. A total of 734 asthmatic patients who underwent induced sputum cytology were screened, and 503 were ultimately included. Data preprocessing, missing value imputation, and feature normalization were conducted within the training set to avoid data leakage. Ten ML algorithms were developed with hyperparameter tuning and internal cross-validation for model optimization. Model performance was assessed using AUC, calibration curves, DCA, accuracy, sensitivity, specificity, precision, F1-score, and confusion matrix. The SHapley Additive exPlanations (SHAP) method was used to improve model interpretability. Our results showed that the K-Nearest Neighbors (KNN) model outperformed the other 9 ML algorithms. After feature selection, an interpretable final KNN model incorporating 6 features was developed, achieving the following performance metrics: AUC = 0.814, accuracy = 0.780, sensitivity = 0.845, specificity = 0.697, precision = 0.780, and F1-score = 0.811. Fractional exhaled nitric oxide (FeNO50) was the most important predictor, followed by blood eosinophil count (bEOS). A web-based prediction tool with a probability threshold of 0.391 was developed for clinical use. The developed ML model holds promise as a non-invasive alternative to induced sputum cytology for EA diagnosis. The corresponding web-based prediction tool is accessible at https://mm-zhou-2026.shinyapps.io/adult-eosinophilic-asthma-prediction/ and may help clinicians achieve rapid phenotyping and personalized precision treatment for adult asthma.
Keywords:
Eosinophilic asthma
Induced sputum cytology
Blood eosinophil count
FeNO50
Machine learning
Prediction model
Journal
IF:
5
Papers:
803
Citations:
1.5W
Organization
Cited Papers