Background:AI-based triage models for emergency department (ED) overcrowding must ensure predictive performance, clinical safety, and interpretability. Prior studies suffer from target leakage by incorporating Emergency Severity Index (ESI) scores, or prioritize accuracy while failing to detect high-risk patients.Objective:To validate the clinical safety of an interpretable deep learning architecture (DA-V2) for ESI-free emergency department triage using outcome-based risk labels.Methods:Using MIMIC-IV-ED (N=407,735 adult visits), ESI scores were excluded from inputs and labels. Risk labels were defined by outcomes: High Risk (expired, transferred, or 30-day mortality), Medium Risk (admitted, survived >30 days), and Low Risk (discharged). DA-V2 integrates Squeeze-and-Excitation Feature Attention, Pre-LN Multi-Head Self-Attention, and Gated Residual Fusion, trained with focal loss and mixup augmentation in a five-seed ensemble. Seven baselines were evaluated under balanced and default class weighting, plus an MLP with DA-V2's identical pipeline as a training strategy control. Attention faithfulness was assessed via Spearman correlation with SHAP values.Results:DA-V2 achieved AUROC 0.7949 (95% CI: 0.7908-0.7991), significantly outperforming all balanced tree-based baselines (p<0.001). An MLP trained with the identical pipeline (MLP+ Enhanced) achieved statistically indistinguishable AUROC (0.7937, p=0.928) and equivalent High Risk (HR) Recall (74.2% vs. 74.1%), establishing focal loss, mixup, and seed ensembling as a generalizable training framework; DA-V2's unique advantage lies in verified interpretability. Default models-despite AUROC 0.796-0.800-detected fewer than 1% of high-risk patients. The dual verification protocol confirmed attention-SHAP alignment (Spearman rho=0.708, p < 10^-12). Ablation showed SE Feature Attention removal reduced HR Recall by 6.6 percentage points.Conclusions:DA-V2 demonstrates that clinically safe, interpretable AI for emergency triage is achievable without ESI scores. The training strategy drives clinical sensitivity, while DA-V2's attention mechanism uniquely enables verified interpretability, advancing deep learning as a trustworthy clinical decision support system.