Rapid and accurate detection of ST-elevation myocardial infarction (STEMI) in prehospital settings is essential for timely triage and reperfusion therapy. We developed an end-to-end Multi-Scale Hybrid Transformer (MS-Hybrid Transformer) to classify study-defined STEMI from 10-second 12-lead ECGs and evaluated its generalizability on an independent prehospital dataset. Development used hospital-acquired ECGs from Long Beach Medical Center (259 STEMI, 571 No-STEMI) with ECG-specific randomized data augmentation and patient-independent 10-fold cross-validation. Models from the 10 folds were ensembled and externally evaluated on prehospital ECGs from Carolinas HealthCare System (147 STEMI, 341 No-STEMI). The ensemble achieved a sensitivity of 0.755, specificity of 0.833, precision of 0.661, F1-score of 0.705, accuracy of 0.809, AUROC of 0.853, and AUPRC of 0.732. Gradient-weighted Class Activation Mapping (Grad-CAM) provided temporally localized explanations of model predictions, with beat-averaged visualizations highlighting clinically relevant ST-T waveform regions across multiple leads. These findings demonstrate good discrimination on an independent prehospital cohort and support accurate and interpretable STEMI detection using the MS-Hybrid Transformer.