A comparative evaluation of neural network classifiers for stress level analysis of automotive drivers using physiological signals

Automotive driving under unacceptable levels of accumulated stress deteriorates their vehicle control and risk-assessment capabilities often inviting road accidents. Design of a safety-critical wearable driver assist system for continuous stress level monitoring requires development of an intelligent algorithm capable of recognizing the drivers' affective state and cumulatively account for increasing stress level. Task induced modifications in rhythms of physiological signals acquired during a real-time driving are clinically proven hallmarks for quantitative analysis of stress and mental fatigue. The present work proposes a neural network driven based solution to learning driving-induced stress patterns and correlating it with statistical, structural and time-frequency changes observed in the recorded biosignals. Physiological signals like Galvanic Skin Response (GSR) and Photoplethysmography (PPG) were selected for the present work. A comprehensive performance analysis on the selected neural network configurations (both Feed forward and Recurrent) concluded that Layer Recurrent Neural Networks are most optimal for stress level detection. This evaluation achieved an average precision of 89.23%, sensitivity of 88.83% and specificity of 94.92% when tested over 19 automotive drivers. The biofeedback inferred about the driver's ongoing physiological state using this neural network based inference engine would provide crucial information to on-board safety embedded systems to activate accordingly. It is envisaged that such a driver-centric safety system will help save precious lives by way of providing fast and credible real-time alerts to drivers and their coupled cars.