ROC analysis of a fatigue classifier for vehicular drivers

Few systems have been developed for the detection of fatigue / stress level of a vehicular driver in order to monitor and control the alertness level for preventing road accidents. Physiological parameters of the body vary with respect to every minute variation in mental and physical states and this work utilizes select such parameters. The present work is a part of the BITS Life-Guard research project that aims to design a simple wearable computing system using noninvasive type of physiological parameter-based sensors for the detection of fatigue / stress level of a driver and alerts the driver in time so as to avoid any accident. The skin conductance and the oximetry pulse signals of the vehicular drivers have been recorded for various states. The features extracted were subsequently used to design multilayer perceptron neural network (MLP NN) to fetch an optimal set of performance measures. The analysis of one hidden layer and two hidden layer MLP NN using Receiver Operating Characteristics (ROC) analysis has been carried out. A two-state classifier has been designed using MLP NN and the classifier performance has been analyzed using the ROC method and the independent validation method. The ROC parameters used for analysis include `Area Under ROC Curve´ (AROC), `Area Under Convex Hull of ROC Curve´ (AHROC), Sensitivity(S), Specificity(R), Percentage Classification Accuracy (PCLA), Mean Square Error (MSE) etc. The work establishes the correlation of fatigue with Skin Conductance (SC) and Oximetry Pulse (PO) and consequently presents a design for MLP NNs for the detection of fatigue level of a vehicular driver.