Interpretable machine learning predicts Victorian zonal crash counts

An interpretable machine-learning framework that predicts five-year crash counts across 3,111 analytical zones in Victoria, Australia, could help engineers prioritize locations for further safety investigation. The study used Extreme Gradient Boosting, or XGBoost, to combine 138 candidate predictors describing population, mobility, transit infrastructure, land use and road networks. Under nested cross-validation, the tuned model produced a pooled Poisson deviance of 1.460, corresponding to deviance skill of 0.374 relative to a training-mean predictor. It also achieved a mean absolute error of 0.902 crashes, a root mean squared error of 2.079 crashes and a Pearson correlation of 0.670 with observed counts. These results show useful internal predictive performance for retrospective screening, but they do not establish future crash reduction, causal effects or the safety benefits of a particular infrastructure intervention.

The dataset covered total crashes recorded in each analytical zone from January 2014 through December 2018. The outcome was highly uneven: 1,649 zones, or 53.0%, recorded no crashes, the median was zero, the 75th percentile was one and the maximum was 83. This sparse, right-skewed distribution motivated a count-Poisson objective and evaluation with mean Poisson deviance, alongside mean absolute error, root mean squared error, prediction bias, correlation and coefficient of determination. Five outer validation folds and three inner folds separated hyperparameter selection from held-out performance assessment, while all preprocessing was confined to each development set. This design reduced the risk of using an observation to tune a model that later evaluated it, although it was still row-wise internal validation rather than temporal forecasting or geographically independent testing.

Tuning compared grid search, random search and Bayesian optimization with a Tree-structured Parzen Estimator, with each method evaluating 24 configurations in every outer-development set under the same inner-fold criterion. Bayesian optimization produced the lowest inner Poisson deviance in four development sets, while random search was marginally better in the fifth. The resulting fold-specific models used learning rates from 0.0569 to 0.0829, between 164 and 453 trees and maximum depths from two to five. On the outer validation data, tuning lowered pooled Poisson deviance from 1.744 for untuned XGBoost to 1.460 and reduced root mean squared error from 2.300 to 2.079 crashes. Mean absolute error, however, increased slightly from 0.888 to 0.902, showing that optimizing a count-based loss did not improve every error measure. Equal numbers of configurations were tested, but recorded durations were not directly comparable because the searches differed in parallel execution and in which hyperparameters they varied.

The model’s overall performance concealed substantial errors in the upper tail. Across all zones, the tuned predictions summed to 2,766.5 crashes, compared with 3,036 observed, and the mean prediction bias was negative 0.087 crashes. The 85 zones with at least five crashes were especially difficult: they contained 853 observed crashes but received predicted totals of only 322.4. Their mean absolute error was 6.736 crashes and their mean bias was negative 6.242 crashes. In the highest predicted-count decile, the mean prediction was 2.709 crashes and the observed mean was 3.344. Conversely, the model overpredicted zero-crash zones by an average of 0.671 crashes and underpredicted positive-count zones by 0.942. These opposing tendencies partly masked the high-count weakness in the aggregate bias, underscoring why ranking accuracy and correlation should be considered alongside calibration.

Two high-dimensional count-regression benchmarks behaved poorly in one outer fold. Ridge-regularized Poisson and negative binomial models generated predictions of approximately 14,885 and 11,416 crashes for a held-out zone that had 70 crashes. The observation lay outside the training range for three retained predictors, including an extreme standardized value for the number of signalized intersections, and the models extrapolated sharply through their log links without convergence warnings. The authors described this as severe extrapolation in the fitted benchmark specifications, not evidence that count-regression models generally fail. In the primary analysis, including a road-length descriptor dated 2019 changed deviance only from 1.460 to 1.456, but the variable was excluded from the main model because it postdated the crash period. Median rather than mean imputation, retention of duplicate columns and replacement of the count objective with squared-error loss also produced limited changes under fixed configurations; the squared-error variant generated 50 negative predictions and therefore had no defined Poisson deviance.

Feature attribution methods identified road and transit inventory as the strongest source of predictive information. In pooled SHAP analysis, the RoadNetwork group had the largest mean absolute grouped contribution, 0.3480 on the log expected-count scale, and ranked first in every outer fold. Transit stops ranked second, with a value of 0.1325. Joint permutation of a feature group also raised held-out Poisson deviance, by 0.741 for RoadNetwork and 0.093 for NumTransitStops; the Spearman correlation between the SHAP and permutation rankings across 14 groups was 0.846. Among individual predictors, speed entropy, metro bus stops and the length of 60 km/h roads had the largest pooled mean absolute SHAP values, at 0.1296, 0.1008 and 0.0960. Uns signalized intersections and total road length followed. Speed entropy here describes the balance of road length among posted-speed categories, not variation in actual vehicle speeds, while transit-stop counts are infrastructure inventories rather than measures of pedestrian or passenger exposure. Correlated and mathematically derived descriptors also mean that importance cannot be assigned uniquely to one variable.

Retrospective ranking demonstrated how the predictions might concentrate attention, although it did not test prospective field performance. The 156 highest-ranked zones, representing 5% of the total, contained 793 observed crashes, or 26.1% of all crashes, and included 43 of the 85 zones with at least five crashes. Expanding the priority set to 10% captured 1,042 crashes, or 34.3% of the total, and 52 of the high-count zones, or 61.2%. At 20% and 30% of zones, the observed crash shares rose to 48.8% and 60.7%, while the corresponding shares of high-count zones reached 76.5% and 91.8%. At the 10% threshold, untuned XGBoost captured 32.2% of crashes, an exploratory model based only on area, population and total road length captured 27.3%, and a direct road-length ranking captured 24.9%. These comparisons used the same 2014–2018 outcomes used for predictive evaluation, and the thresholds were not optimized against them, so they demonstrate retrospective concentration rather than successful future hotspot detection.

The results position the model as a screening step before detailed engineering assessment, not as a decision system for construction projects. In pre-construction planning, elevated predicted counts and feature attributions can help teams decide which zones require scrutiny of road hierarchy, intersection configuration, speed-zone composition, transit interfaces and pedestrian facilities. Strategic planners could combine the output with crash histories, traffic exposure, equity objectives, feasibility and budgets when ordering safety studies. A separate construction-phase application would require data on lane closures, taper geometry, work-zone status, operating speeds, traffic demand, worker and pedestrian exposure and compliance behavior, none of which are present in the analytical dataset. Quantifying the return on a safety investment would also require treatment costs and credible estimates of intervention effectiveness rather than predictions from this cross-sectional model.

Several limitations constrain how far the findings can be generalized. The study used one regional dataset, included no direct measure such as annual average daily traffic or vehicle-kilometres traveled, and lacked a verified row-to-zone key, coordinates or polygons. Spatially separated validation was therefore impossible, so nearby or otherwise dependent zones could appear in different folds. The 95% intervals estimated from 2,000 paired bootstrap resamples were conditional on the fitted fold models and treated rows as exchangeable; they did not include model-refitting uncertainty, dependence from overlapping training sets or unknown spatial correlation, and they were not prediction intervals for new zones or future counts. Observation periods for many predictors were incompletely documented, although the explicitly postdated 2019 descriptor was removed. Calibration was weaker for high-count zones, predictive attribution was constrained by correlated variables, and the observational design could not identify causal effects. External and temporal validation, authentic geographic linkage, direct exposure measures, work-zone-specific data and treatment-outcome studies are required before the framework can support broader deployment or investment decisions.

Subject of Research: Interpretable machine-learning prediction of five-year crash counts in 3,111 Victorian zones

Article Title: Integrating safety analytics into construction management: An interpretable machine learning approach for zonal crash-count screening

Article References: Nguyen, T.-H. T., Nguyen, T.-A., Nguyen, L. T., Nguyen, T.-D., & Ly, H.-B. (2026). Integrating safety analytics into construction management: An interpretable machine learning approach for zonal crash-count screening. Results in Engineering, 32, Article 113628. https://doi.org/10.1016/j.rineng.2026.113628

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113628

Keywords: road safety, XGBoost, crash counts, machine learning, nested validation, model interpretability, zonal screening, infrastructure planning, Victoria, count data, Integrating, safety