DeepTriage-CN: integrating clinical text with vital signs for emergency department admission prediction in an aging population.

Emergency department (ED) triage of older adults is challenging because standard early warning scores are often insensitive to atypical presentations. We developed DeepTriage-CN, a late-fusion framework combining frozen BERT-Chinese embeddings of nurse-recorded chief complaints with structured triage vital signs via an XGBoost classifier. The model was trained on 8000 adult ED visits and validated temporally on 2000 independent encounters. DeepTriage-CN achieved an AUROC of 0.865, statistically indistinguishable from the tabular deep-learning comparator TabNet (0.867; DeLong p = 0.42). The multimodal model significantly outperformed traditional clinical scores (NEWS2, 0.772; ESI, 0.760) and vitals-only baselines (XGBoost, 0.858; all DeLong p < 0.001). In older patients (age ≥ 65 years), the model maintained an AUROC of 0.852 versus 0.710 for ESI and 0.741 for NEWS2, though performance remained comparable to TabNet (0.835; p = 0.21). Under simulated 30% informative missingness with Gaussian noise, DeepTriage-CN retained 95.4% of its baseline AUROC compared with 83.3% for TabNet and 82.8% for vitals-only XGBoost. However, at the Youden-optimal classification threshold (0.28), positive predictive value was 0.50, meaning one in every two generated alerts is a false positive. These findings identify a specific operational envelope in which the multimodal approach confers situational advantage-primarily when structured physiological data are degraded-while confirming that text embeddings provide no statistically significant incremental discriminative gain over optimized tabular deep-learning architectures in the general population. Prospective multicenter validation with clinically actionable endpoints is required before this framework can inform ED triage practice.