Consensus machine learning for interpretable road-user fatality risk analysis: Evidence from the Canadian National Collision Database

Objectives Fatal road traffic collisions account for a small proportion of reported crashes but impose a disproportionate societal and public health burden. This study evaluates interpretable machine-learning models and proposes a consensus machine learning (CML) framework to identify robust predictors associated with fatal collision risk using Canadian collision data. Methods Person-level records from the Canadian National Collision Database (NCDB) covering 2014-2021 were used to classify fatal and nonfatal col…

By the Truvace deskOctober 19, 20262 min read1 sourcePeer-reviewedTRV-2026-1943· cite & verify

Consensus machine learning for interpretable road-user fatality risk analysis: Evidence from the Canadian National Collision Database — "A pony grazing at Monks Farm oblivious to the motorway traffic behind - geograph.org.uk - 2940138" by Ian Greig is licensed under CC BY-SA 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by-sa/2.0/.

In brief

Researchers used person-level records from the Canadian National Collision Database for 2014-2021 to train balanced random forest, XGBoost and neural network models that separate fatal from nonfatal collision outcomes. They then combined permutation-based feature-importance rankings from the three models into a consensus framework.

The result matters because fatal crashes are a small share of reported collisions but carry a large public health burden, and a stable ranking of risk factors can inform where engineering, enforcement and education effort goes. The authors present the approach as a complementary decision-support tool, not a replacement for existing safety planning.

Main points

XGBoost performed best with ROC-AUC 0.932 and PR-AUC 0.714, ahead of balanced random forest (0.898 and 0.585) and the neural network (0.886 and 0.587).
Collision configuration, traffic control, collision hour and number of vehicles involved ranked consistently among the most influential predictors across all three models.
The consensus framework named collision configuration, road-surface conditions, time of collision, number of vehicles, roadway configuration and weather as leading predictors.
Crash characteristics and traffic and environmental conditions were the two key feature categories for predicting fatalities.

The gain

Aggregating feature rankings across three models gave a consistent, interpretable list of fatal-collision predictors that could support road-safety planning.

The rundown

The data came from person-level NCDB records covering 2014-2021, and all models used a common preprocessing framework. Performance was assessed with ROC-AUC and PR-AUC, the latter being informative when fatal outcomes are rare.

Roadway alignment, vehicle characteristics and safety-device use also contributed to prediction, but their relative importance varied between models. The consensus method was designed to highlight predictors that stayed important across algorithms and across reporting years.

The authors conclude that the findings support engineering, policy, enforcement and education initiatives aimed at reducing fatal collisions. This is a predictive and associational analysis of historical records, not a test of any intervention.