Development of risk prediction model for road traffic injury in drunk-drivers during festivals in Thailand

Abstract

Thailand is a middle-income country where the road traffic injury crisis has been one of the most serious public health concerns. Currently, machine learning (ML) algorithms are widely used for public health predictive analytics. Therefore, we developed the Multi- layer perceptron (MLP) classifier from the road traffic accident driver data in Thailand that aim to classify a high-risk driver who had severe injuries from road traffic accidents. However, the imbalanced data was a typical problem in public health data and also caused an "accuracy paradox" that the model intended to predict a majority class. Accurately detecting minority class was important especially in the public health data because it was associated with high impact events and serious adverse outcomes. Since the imbalanced data is unavoidable according to the nature of public health data. The rebalanced strategies or other data approaches were applied to encounter this problem. Subsequently, the over-sampling techniques were significantly improved discrimination performances of models comparing with under-sampling or without rebalancing approach. Therefore, this model can be applied as the personalized risk classification of driving under influence of alcohol and more precisely than using alcohol legal limit.