This study models the injury severity of riders in motorcycle crashes occurring on state and national highways of Tamil Nadu, India. Accurate prediction and identification of factors affecting injury severity of an accident help various stakeholders such as traffic engineers, law-making bodies, law enforcement agencies and emergency services to take appropriate measures. Both statistical and machine learning models have been used for modelling injury severity, and the choice between the two has been primarily dictated by the purpose of modelling, namely prediction or interpretation. When interpretations are the primary focus of the study, then statistical models were used, and machine learning models were used when predictive performance is crucial. This study bridges this dichotomy by using machine learning models to model the injury severity and subsequently uses interpretable machine learning techniques to derive interpretations from them. The interpretations derived from the study provides fresh insights into the effects of variables and interactions between them in determining the injury severity of a crash.


Moreover, crash datasets used for modelling injury severity are noisy in nature. This can be attributed to errors while manually entering data by the police and the inherent subjectivity in deciding whether an injury is minor or grievous. To handle this issue, the study proposes a new framework by which noisy data points can be incorporated into the modelling process without discarding them. A semi-supervised decision tree is built (part of the framework) on a dataset formed by combining the clean data points with noisy data points after deleting their labels. The results show that the semi-supervised decision tree applied as a part of the framework has a performance comparable to that of a supervised random forest and is easily interpretable.