Human stress has become a serious problem affecting people’s life. The current state of sensor technology allows developing systems measuring physical symptoms reflecting people’s stress level. However, the detection of stressful events using the skin conductance (SC) and finger temperature (FT) data is a challenging task due to varieties of patterns in the data. Also each person handles stress in a different way.
A big part of the thesis work is to compare the proximity of the results from FT sensor and SC sensor data using two artificial intelligence techniques (AI): neural networks (NN) and case base reasoning (CBR). More precise the features extracted from the data are: FT slope and SC slope that detects the angle of change of the signal and skin conductance response (SCR) which is the change in SC within a short period of time. These feature data are stored in a database and matched against old cases in the CBR case library.
In the NN case, weights are automatically been assigned to each feature to produce a result. These three features are applied separately to the CBR and to the NN. The results from the combined experiments are expected to derive to the most efficient “set” (sensor/features/A.I. technique) that automatically produce an answer closest to that of a clinician’s expert diagnosis.
The current thesis experimental workflow proved that using both CBR and NN with SC and FT slope features can give up to 83% successful results. Unfortunately the SCR features didn’t give the same results and it is suggested as a future work to complete the algorithm of feature extraction. A big contribution of this thesis work is that the FT & SC slope algorithm shows a promising performance on both CBR and NN. Another contribution is on the SCR feature extraction where the biggest part of the work has been done using Matlab code.
Source: Mälardalen University
Author: Belogiannis, Theodoros