Design and experiment verification of a novel analysis framework for recognition of driver injury patterns: From a multi-class classification perspective

Detecting driver injury patterns is a typical classification problem. Crash data sets are highly skewed where fatalities and severe injuries are often less represented compared to other events. The severity prediction performance of the existing models is poor due to the highly imbalanced samples of different severity levels within a given dataset. This paper proposes a machine learning based analysis framework from a multi-class classification perspective for accurate recognition of the driver injury patterns. The proposed framework includes preprocessing, classification, evaluation and application of a given dataset. This framework is verified based on the three years single-vehicle ROR (run-off-road) crash records collected in Washington State from 2011 to 2013. At first, thirteen most important safety-related variables are recognized through random forests. Then, the four driver's injury severity levels viz., fatal/serious injury, evident injury, possible injury, and no injury are predicted by integrating the decomposed binary neural network models to achieve better performance. Finally, a sensitivity analysis is carried out to interpret variables’ impacts on the decomposed injury severity levels. The study shows that lack of restraint, female drivers, truck usage, driver impairment, driver distraction, vehicle overturn (rollover), dawn/dusk, and overtaking are the leading factors contributing to the driver fatalities or severe injuries in a single-vehicle ROR crash. Most of the findings are consistent with the previous studies. The experimental results validate the effectiveness of the proposed framework which can be further applied for pattern recognition in traffic safety research.