RESEARCH · RESEARCH · #834
Auditable, calibrated selective-deferral framework for dementia-related crash severity prediction (arXiv:2609.22694v1)
This arXiv preprint presents a reproducible, leakage-controlled evaluation and a selective-deferral framework for dementia-related crash severity prediction using 4,781 Texas crash records. The authors compare structured, narrative, fusion, calibrated fusion, BERT-family, and local LLM baselines; report Gemma (leakage-controlled) as highest observed macro-F1 = 0.545 (95% CI [0.507, 0.583]), the best calibrated fusion model with macro-F1 = 0.522 and ECE = 0.033, and show that selective deferral (70% coverage) raises macro-F1 to 0.573 and lowers severity cost to 0.577 while deferring cases for proposed human review.
KEY POINTS
- This arXiv preprint presents a reproducible, leakage-controlled evaluation and a selective-deferral framework for dementia-related crash severity prediction using 4,781 Texas crash records.
- The authors compare structured, narrative, fusion, calibrated fusion, BERT-family, and local LLM baselines; report Gemma (leakage-controlled) as highest observed macro-F1 = 0.545 (95% CI [0.507, 0.583]), the best calibrated fusion model with macro-F1 = 0.522 and ECE = 0.033, and show that selective deferral (70% coverage) raises macro-F1 to 0.573 and lowers severity cost to 0.577 while deferring cases for proposed human review.
- Reframes crash-severity modeling as a decision-aware, auditable triage problem and shows selective deferral plus calibration can improve reliability and enable human review for safety-critical public-sector workflows.
WHY IT MATTERS
Reframes crash-severity modeling as a decision-aware, auditable triage problem and shows selective deferral plus calibration can improve reliability and enable human review for safety-critical public-sector workflows.