Objective In order to address the challenge of rapid diagnosis in pulmonary diseases, this paper proposes a cross-modal fusion method based on the Cross-modal Transformer (CMT) that integrates respiratory sounds (RS) and electrical impedance tomography (EIT), with the aim of improving the accuracy and robustness of multi-classification tasks. RS encodes the acoustic characteristics of the airways via the Meyrieh spectrogram, whilst EIT depicts regional lung ventilation distribution through spatio-temporal image sequences. The two modalities are naturally complementary in terms of functional and spatial information, providing a physiological basis for multimodal fusion diagnosis.Methods A dual-branch feature extraction framework was constructed, employing Convolutional Neural Networks (CNNs) to extract local features from RS time-frequency spectra and EIT ventilation images, whilst utilising Bidirectional Long Short-Term Memory Networks (BiLSTMs) to model the temporal dependencies across modalities. The development of a transformer-based cross-modal attention fusion module represents a significant advancement in the field. The module utilises a multi-head self-attention mechanism and gated convolutional units to achieve deep semantic alignment and complementary information fusion between RS acoustic features and EIT spatial ventilation features. The model employs an end-to-end joint optimisation strategy, with a cross-entropy loss function supervising the overall training process, and utilises a time-synchronisation mechanism to ensure strict alignment of the dual-modal inputs within the respiratory cycle. The proposed CMT method was subjected to systematic experimental validation on two datasets: The BRACETS dataset (three-class classification, 795 samples) and the CleftPalate dataset (two-class classification, 549 samples) are the focus of this study. Quantitative evaluation was performed using accuracy, balanced accuracy (BAcc) and macroF1 score.Results On the BRACETS dataset, the CMT method achieved an accuracy of 87.21%, a balanced accuracy (BAcc) of 88.24%, and a macroF1 score of 87.40%. In comparison to the optimal baseline method, DCNN, the macroF1 score exhibited an enhancement of 8.73 percentage points, thereby substantiating a substantial performance superiority. The findings of the ablation experiments suggest that the three core modules – CNN, BiLSTM and Transformer – all contribute to performance enhancements. Upon the removal of these modules, the MacroF1 score decreased by 1.84%, 1.13% and 2.20%, respectively. Among these, the Transformer cross-modal fusion module had the most significant impact, validating its crucial role in the interaction of heterogeneous modal information. Hyperparameter sensitivity analysis indicates that the optimal parameter configuration is a sequence length of T=128 and a feature dimension of d=128, achieving a good balance between classification performance and computational efficiency. Sequences that are insufficiently extensive or dimensions that are unduly limited impede the capacity to adequately represent temporal dynamics, whilst excessively protracted sequences may engender superfluous information. On the CleftPalate dataset, the CMT method achieved an accuracy of 96.58%, a BAcc of 96.60%, and a MacroF1 of 96.56%, representing a further improvement of 2.51 percentage points compared to the DCNN. This result validates the model's generalisation capability across different data distributions and task scenarios. The fusion representation learned by CMT has been shown to exhibit tighter intra-class cohesion and clearer inter-class separation boundaries, as evidenced by feature visualisation and case studies (Smith et al., 2022). The system has the capacity to adaptively aggregate effective features when the reliability of multimodal information is uneven, and can maintain correct classification even when ambiguity exists in a single modality.Conclusion The proposed method effectively achieves deep fusion of heterogeneous modal information from RS and EIT, fully exploiting the physiological complementary relationship between respiratory airflow acoustic features and regional lung ventilation distribution. The model displays excellent classification performance and stability across various data distributions and task scenarios, thus providing a novel technical pathway for the non-invasive intelligent diagnosis of pulmonary diseases.