Objective This study aimed to develop and validate a comprehensive machine learning framework combining radiomic features, deep features extracted from CT images, and clinical variables to predict and stratify asthma severity into three clinically defined categories. Materials and methods CT imaging and clinical data from 1365 patients across three medical centers were retrospectively collected. Lung segmentation was performed manually by two expert radiologists using 3D Slicer, with reproducibility assessed via the Dice Similarity Coefficient. A total of 215 radiomic features were extracted following IBSI standards, and 512 deep features were obtained through a custom attention-based autoencoder. All features were filtered using ICC (>0.75), correlation analysis (r < 0.9), and LASSO regression. Five machine learning models-XGBoost, LightGBM, Autoencoder, GAT, and a Hybrid Deep Attention-LSTM-were trained using stratified 70/10/20 train-validation-test splits. SHAP analysis, nomogram construction, and decision curve analysis (DCA) were applied for model interpretability and clinical relevance assessment. Results The Hybrid Deep Attention-LSTM model achieved the highest classification performance, with 95.7 % accuracy and 98.0 % AUC on the test set. All models using combined features outperformed radiomic-only or deep-only models. SHAP analysis identified FEV1, BMI, shortness of breath, and nighttime symptoms as key clinical predictors. The nomogram provided interpretable risk visualization, while DCA confirmed the hybrid model's superior net benefit across threshold probabilities compared to standard machine learning models. Conclusions The study presents a reproducible, interpretable, and high-performing model for asthma severity classification using integrated radiomic, deep, and clinical features, with potential for real-world clinical deployment in personalized asthma care.