RaX-Crash: A Resource Efficient and Explainable Small Model Pipeline with an Application to City Scale Injury Severity Prediction

TL;DR

<3-5 sentence high-level summary> RaX-Crash presents a resource-efficient pipeline for city-scale injury severity prediction by unifying crash, person, and vehicle data into a compact feature schema and evaluating compact tree ensembles against small language models. A hybrid approach combines XGBoost/RF predictions with SHAP explanations and SLM-generated narratives to improve communication without sacrificing scalability. Results show tree ensembles outperform SLMs as predictors, with focal loss and class weighting improving fatal recall and SHAP highlighting human, timing, and location factors. The work demonstrates the viability of integrating interpretable tabular models with narrative explanations for urban safety analytics and points to future work on data quality, transferability, and deployment in policymaker-facing tools.

Abstract

New York City reports over one hundred thousand motor vehicle collisions each year, creating substantial injury and public health burden. We present RaX-Crash, a resource efficient and explainable small model pipeline for structured injury severity prediction on the official NYC Motor Vehicle Collisions dataset. RaX-Crash integrates three linked tables with tens of millions of records, builds a unified feature schema in partitioned storage, and trains compact tree based ensembles (Random Forest and XGBoost) on engineered tabular features, which are compared against locally deployed small language models (SLMs) prompted with textual summaries. On a temporally held out test set, XGBoost and Random Forest achieve accuracies of 0.7828 and 0.7794, clearly outperforming SLMs (0.594 and 0.496); class imbalance analysis shows that simple class weighting improves fatal recall with modest accuracy trade offs, and SHAP attribution highlights human vulnerability factors, timing, and location as dominant drivers of predicted severity. Overall, RaX-Crash indicates that interpretable small model ensembles remain strong baselines for city scale injury analytics, while hybrid pipelines that pair tabular predictors with SLM generated narratives improve communication without sacrificing scalability.

Figures (5)

• Figure 1: Overview of the NYC collision dataset used in RaX-Crash. Left: scale of crash, person, and vehicle tables (10M--22M records). Right: highly imbalanced severity distribution with fatal outcomes under 1%.
• Figure 2: Correlation matrix across top numerical features. Diagonal dominance and relatively weak cross feature correlations support the use of tree based models over heavily regularized linear baselines.
• Figure 3: RaX-Crash pipeline from raw NYC tables to hybrid small model prediction and SLM based narrative explanation.
• Figure 4: Effect of different imbalance methods on fatal recall. Focal loss provides the largest gain with stable accuracy.
• Figure 5: SHAP summary plot for the XGBoost model. Each point represents one event; colors encode feature value (blue = low, red = high), and horizontal position shows the impact on the predicted injury severity.