In emergency response, dispatch speed and trauma-center activation depend on accurate severity classification. Current classifiers face two problems: extreme class imbalance and a semantic gap that leaves numerical models blind to textual severity cues. Resampling methods adjust class distributions but add no new information, while LLM-based hybrids exhibit feature dilution, where numerical priors override semantic reasoning. We propose SAFE (Semantic-Augmented Fusion Ensemble), a framework that routes features through parallel branches: XGBoost for numerical data and a Small Language Model for text. Structured records are enriched into narratives with severity-predictive keywords. The branches merge through class-adaptive probability fusion, governed by an analytically derived condition that preserves minority-class detections against majority-biased priors. On the US Accidents dataset and UK road accident records, Severe Recall rises from 30.7% (RF + SMOTE) to 91.2%, with overall accuracy reaching 83.3%; Serious Recall reaches 54.5% against 33.8% (XGBoost + SMOTE-ENN) on UK data. Keyword enrichment is essential: its removal collapses recall regardless of model size. SAFE enables severity-aware triage using only structured records that transportation agencies already collect. Deployment efficiency remains practical. SAFE achieves 188.4 ms mean per-sample latency at 5.3 samples/s on consumer hardware (Qwen3-4B INT8, 6.41 GB memory footprint), supporting operational batch classification of incident records.
Tariq Alsahfi (Sun,) studied this question.