From class imbalance to structural rarity: a scoping review of rare-event and anomaly-aware learning for crash severity prediction
Year: 2026
Type: article
Abstract: Severe and fatal road traffic crashes represent a small yet disproportionately high-impact share of crash events, posing persistent challenges for predictive modeling due to extreme class imbalance and structural sparsity. Most existing crash severity models treat severe outcomes primarily as minority labels within imbalanced datasets, often overlooking the structural distinctiveness of rare high-impact crashes. This study presents a scoping review and conceptual synthesis of peer-reviewed studies on crash severity modelling, rare-event learning, class-imbalance mitigation, and anomaly-aware transportation safety approaches. The review searched Scopus and Web of Science for journal articles published between January 2010 and February 2026. After screening 270 records, 74 studies were included and analyzed. The review critically evaluates how the existing methods handle rare-event conditions, minority class metrics, and structural model capacity. Findings indicate that while imbalance mitigation techniques improve minority recall, they do not fully capture the structural rarity of severe crashes. Anomaly-aware and hybrid extreme value frameworks provide a complementary perspective by identifying high-risk configurations embedded within complex spatiotemporal systems, although only three of the 74 studies applied anomaly detection to a severity-related outcome. Accordingly, this review proposes an anomaly-guided rare-event crash severity modeling framework integrating spatiotemporal architectures with structural anomaly detection and rare-event-aware optimization intended to improve predictive reliability and policy relevance, offered as a roadmap for empirical testing.
Institution: King Fahd University of Petroleum and Minerals
Language: English
Topic: Traffic and Road Safety
Subfield: Safety, Risk, Reliability and Quality
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,400
OpenAlex ID: W7228486772
