Objective: To develop an Adaptive Modality Gating Multi-Head Cross-Attention Fusion Model (AM-GMCA-Swin-S) based on Swin-S Transformer and explore its application value in non-invasive prediction of MGMT methylation status. Materials and Methods: A retrospective cohort of 328 adult diffuse glioma patients admitted to Deyang People’s Hospital from January 2011 to January 2025 was collected, including 230 cases with MGMT methylation positive and 98 cases with negative status, for model training and internal validation. A total of 82 cases selected from the publicly available multi-center ReMIND dataset were used as an independent external test set (59 methylation positive, 23 negative). After standardized preprocessing of T2WI, DWI, and 3D-T1CE images, 27-channel 2.5D image matrices covering axial, coronal, and sagittal planes were constructed. Clinical features, intratumoral and peritumoral radiomic features were extracted simultaneously. Four comparative models were established: semantic feature baseline model (LR), pure imaging modal baseline model (Native Swin-S), fixed-weight cross-attention fusion model (FW-CA-Swin-S), and AM-GMCA-Swin-S model. Multiple indicators, including PR-AUC, ROC-AUC, sensitivity, specificity, and F1-score were adopted to evaluate model discrimination efficacy. DeLong test was used to compare inter-group differences in AUC values, and continuous net reclassification improvement (continuous NRI) as well as integrated discrimination improvement (IDI) were calculated to quantify reclassification benefits. Results: Stratified 5-fold internal cross-validation demonstrated that the LR baseline model achieved the optimal overall discriminative performance, while AM-GMCA-Swin-S obtained the highest specificity. In independent external validation, AM-GMCA-Swin-S exhibited the best comprehensive performance, with PR-AUC of 0.856, specificity of 0.696, sensitivity of 0.678, and F1-score of 0.755. Its sample reclassification and risk stratification capacity were significantly superior to FW-CA-Swin-S (P<0.05). Compared to the LR baseline model (PR-AUC = 0.830), although there was no statistically significant difference in overall discriminative performance (P>0.05), AM-GMCA-Swin-S showed better individualized risk stratification. The Native Swin-S model had the worst discriminative performance and exhibited severe prediction bias. Conclusions: AM-GMCA-Swin-S can dynamically adjust the weights of different modality features es based on the tumor heterogeneity of individual samples, achieving adaptive fusion of multi-modal information and significantly improving the model’s comprehensive predictive performance. It remarkably enhances sample reclassification and individualized risk stratification capabilities, offering a new technical reference for the non-invasive assessment of MGMT methylation status and individualized diagnosis and treatment in adult diffuse gliomas.