Objective Conventional pulmonary function testing remains the primary clinical approach for evaluating respiratory dysfunction, but its reliability is strongly dependent on subject cooperation and test execution. In addition, global spirometric indices provide limited information on regional ventilation heterogeneity, which is important for understanding the spatial characteristics of pulmonary functional abnormalities. To address these limitations, this study proposed a dual-branch fusion method based on electrical impedance tomography (EIT) for the identification of pulmonary function abnormalities and evaluated its feasibility in screening and auxiliary classification tasks.Methods A total of 507 valid EIT-pulmonary function examination samples were included, comprising 244 samples with normal pulmonary function and 263 samples with abnormal pulmonary function. During forced expiration, thoracic EIT boundary voltage signals were collected, and corresponding conductivity image sequences were reconstructed to represent both respiratory dynamic changes and intrapulmonary ventilation distribution. For each valid examination sample, 64 consecutive frames from the expiratory phase were selected as the model input. A dual-branch architecture was then constructed, in which one branch used boundary voltage signals to characterize temporal respiratory dynamics, while the other branch used reconstructed conductivity images to describe spatial ventilation patterns. Residual network (ResNet) models with different depths, including ResNet18, ResNet50, and ResNet101, were employed for feature extraction. The extracted voltage and image features were concatenated to generate a unified spatiotemporal representation. Several machine learning classifiers, including support vector machine, k-nearest neighbor, decision tree, linear discriminant analysis, and Stacking ensemble learning, were further applied for classification. Five-fold cross-validation was used to evaluate model performance in binary classification between normal and abnormal pulmonary function, three-class classification among normal, obstructive, and non-obstructive pulmonary function patterns, and four-class classification involving multiple ventilatory dysfunction types. Accuracy and macro-averaged F1-score were used as the main evaluation metrics. Feature extraction, feature standardization, classifier training, and model evaluation were independently completed within each fold to reduce the risk of information leakage.Results On average, the fused voltage-conductivity channel outperformed the single voltage and single conductivity-image channels. The average accuracy and F1-score of the fusion channel reached 0.835 ± 0.053 and 0.765 ± 0.073, respectively, representing improvements of 6.5% and 9.7% compared with the voltage-only channel. Among the different network depths, ResNet18 achieved the best overall performance, suggesting that a relatively shallow residual network may be more suitable for learning EIT-based spatiotemporal features, probably because the key pathological information in EIT is mainly reflected by low-frequency temporal changes and large-scale regional ventilation differences rather than fine image textures. Among the classifiers, the Stacking strategy achieved the highest comprehensive performance, with an accuracy of 0.830 8 ± 0.059 9 and an F1-score of 0.821 9 ± 0.060 7, indicating that ensemble learning can further improve the robustness of multi-class recognition. The overall advantage of the fusion channel also supports the complementary value of temporal boundary-voltage dynamics and spatial conductivity-distribution information in characterizing pulmonary function abnormalities.Conclusion These findings demonstrate that dual-branch fusion of EIT voltage signals and conductivity images can effectively characterize spatiotemporal features associated with abnormal pulmonary ventilation. By integrating temporal respiratory dynamics with regional ventilation distribution, the proposed framework reduces dependence on a single predefined EIT index and provides a more comprehensive representation of pulmonary function abnormalities. The proposed method offers a non-invasive and radiation-free approach for pulmonary function abnormality screening, auxiliary classification, and bedside dynamic assessment, and may support the future development of wearable or portable EIT-based respiratory monitoring systems.