Crash Severity Prediction for Road Safety Policy: Random Forest Analysis in Bangladesh

Abstract

Road traffic accidents kill over 12,000 Bangladeshis annually, costing 2% of gross domestic product. This study develops a machine learning classifier for accident severity prediction using 750 crashes from Khulna, Bangladesh. A Random Forest algorithm trained on driver demographics, environmental conditions, and crash dynamics achieved 82.1% accuracy and 75% fatal-outcome recall on held-out test data, substantially exceeding Logistic Regression (71.3%) and Support Vector Machine (78.0%). Feature importance analysis revealed speed (16.2%) and major contributing cause (14.4%) as dominant predictors, substantially outranking licensing status (3.9%) and gender (0.9%), reflecting developing-country accident characteristics diverging from developed-country models. Stratified 80/20 train-test splitting with SMOTE-based class imbalance handling addressed the overrepresentation of non-fatal outcomes (78.7%) relative to fatal cases (21.3%). Cross-validation (5-fold stratified) confirmed generalization (79.8% ± 2.3% accuracy). The model's 81.2% fatal prediction precision enables evidence-based emergency resource deployment. Key finding: speed enforcement represents the dominant modifiable risk factor in Bangladeshi traffic contexts, supporting government policy prioritization. The study provides empirical validation for South Asian-specific transportation safety interventions and establishes a methodology for developing-country accident severity prediction adapted to local traffic environments, vehicle type distributions, and causation patterns distinct from developed economies.