Detection of functional state after alcohol consumption by classification and machine learning technics

Machine learning (ML) technics have been recently used to detect emotion and predict crash severity. This research work aims at assessing different classifications and machine learning technics in predicting the alcohol consumption and associated functional states. 28 young drivers were tested for a 45 min drive with a blood alcohol concentration (BAC) of 0.0, 0.2 and 0.5g/m. Subjective functional states were analysed using Thayer’s scale and NASA-TLX. The physiological parameters (electroencephalogram, electrodermal and cardiac activity) and driver simulators parameters (speed, lateral positioning and wheel steering) were acquired during the three alcohol sessions. Data were analysed on 10s temporal windows without superposition nor gap. Two analyses using classification and ML technics were used: to determine both capacity of the algorithms to detect alcohol consumption (BAC level) and functional states (effort, performance and alertness) from NASA and Thayer’s scales. Different algorithms were trained using 10 folds cross validation technics using Weka (University of Waikato, NZ). Using both vehicle and physiological data was beneficial for BAC prediction and ROC area of the top three algorithms were found between 0.62 and 0.72 with higher results for Random Forest (RF) algorithms. In functional states prediction, results were similar for all effort, performance and alertness predictions with ROC area reaching 0.75 for RF. Once algorithm setting optimized, performances for BAC prediction reached 0.73 while were, lower than for functional states prediction with ROC area of 0.91 when pairing data. Such results could help in the strategy for detecting alcohol consumption in drivers.