Predicting Child-Bicyclist Crash Severity Under Severe Class Imbalance: A Benchmark of XGBoost and Tabular Deep Learning

Objectives Child-bicyclist crash-severity data are strongly imbalanced, making ordinary accuracy potentially misleading. This study compared three tabular deep-learning models with machine learning and statistical reference models under a train–validation–test design. Methods The analysis used 2394 Texas child-bicyclist crashes from 2017 to 2022 classified as killed or incapacitating injury (KA), possible or non-incapacitating injury (BC), or other/no injury (O). The original records were partitioned before feature selection, preprocessing, and imbalance treatment into training ( n = 1454), validation ( n = 475), and untouched test ( n = 465) sets. Random Forest and XGBoost rankings were calculated using the training partition only, yielding a common 10-predictor input. Validation performance was used to select class weighting for XGBoost and training-only Synthetic Minority Over-sampling Technique (SMOTEN) for the three deep-learning models. The final primary-model estimates were averaged across five seeds. Results Adaptive Relation-Modeling (ARM)-Net had the highest primary-model accuracy (0.662 ± 0.020) and macro-F1 (0.335 ± 0.011); XGBoost had the highest primary-model balanced accuracy (0.340 ± 0.008); and FT-Transformer had the highest primary-model macro-ROC-AUC (0.540 ± 0.006). However, the dummy-majority accuracy was 0.787 because BC dominated the test set and multinomial logistic regression exceeded all four primary models in balanced accuracy (0.378), macro ROC−AUC (0.561), and macro average precision (0.361). Minority-class discrimination was weak, bootstrap intervals overlapped substantially, and SHapley Additive exPlanations (SHAP) rankings showed limited stability relative to cross-seed permutation importance. Conclusions Class-balanced performance was limited across all models. The evaluated models should be treated as transparent benchmarks rather than operational prediction tools; larger, externally validated datasets with richer exposure and behavioral information are needed. From an AI-for-transportation-safety perspective, the study contributes a leakage-controlled benchmarking protocol and an explanation-stability check, showing that advanced tabular deep learning models should be judged against simple baselines and class-balanced metrics before their outputs are used to support safety decisions.