Reliable network screening requires estimates that distinguish persistent crash burden from random variation and support decisions under limited investigation capacity. This study evaluates whether flexible machine-learning prior means improve Empirical Bayes screening beyond a conventional negative-binomial benchmark. Publicly accessible Montana crash records for 2019–2023 were linked to a fixed 2019 traffic-segment inventory. The resulting network contained 6,169 segments and 78,560 matched crash records. Model development used 2019–2020 counts, while 2021 counts supplied an independent site-history update. Entire corridors were separated into development, calibration and test groups. Negative-binomial regression, Poisson-criterion random forests and Poisson-loss gradient boosting supplied prior means. A Poisson–gamma model, with dispersion estimated on separate calibration corridors, defined the EB update. The untouched evaluation comprised 1138 test segments and 3,888 crashes during 2022–2023. Boosting reduced standalone root mean squared error from 4.521 to 4.419 crashes per segment over two years. Conventional EB nevertheless achieved lower mean absolute error than boosting-assisted EB, 1.875 versus 1.887. Selecting the highest-ranked 10% of test sites captured 50.95% and 50.75% of future crashes, respectively. The paired corridor-bootstrap difference for boosting-assisted EB was −0.21 percentage points, with a 95% interval from −0.68 to 0.85. These findings support EB updating but provide no evidence that the tested machine-learning integration improves conventional EB screening on this network. Interpretation is restricted to reported crash frequency under fixed baseline exposure. Severity information, contemporary exposure and external-network validation remain necessary for broader deployment.