Supervised learning techniques for stress detection in car drivers

In this paper we propose the application of supervised learning techniques to recognize stress situations in drivers by analyzing their Skin Potential Response (SPR) and the Electrocardiogram (ECG). A sensing device is used to acquire the SPR from both hands of the drivers, and the ECG from their chest. We also consider a motion artifact removal algorithm that allows the generation of a single cleaned SPR signal, starting from the two SPR signals, which could be characterized by artifacts due to vibrations or movements of the hands on the wheel. From both the cleaned SPR and the ECG signals we compute some statistical features that are used as input to six Machine Learning Algorithms for stress or non-stress episodes classification. The SPR and ECG signals are also used as input to Deep Learning Algorithms, thus allowing us to compare the performance of the different classifiers. The experiments have been carried out in a firm specialized in developing professional car driving simulators. In particular, a dynamic driving simulator has been used, with subjects driving along a straight road affected by some unanticipated stress-evoking events, located at different positions. We obtain an accuracy of 88.13% in stress recognition using a Long Short-Term Memory (LSTM) network.