Implementation of a Real-Time EDA Processing Tool to Assess Cognitive Workload

Real-time monitoring of cognitive load is a key topic in the prevention of workplace injuries and in the assessment of operator well-being in high-intensity contexts. The final objective of this study is to develop a hardware and software tool capable of acquiring, processing and analyzing the electrodermal activity (EDA) signal, to detect different states of cognitive workload (CWL). The hardware system consisted of safety helmet where some extra instrumentation have been installed: a Raspberry Pi 5, to which a camera and an accelerometer were connected, and a Shimmer GSR 3+ sensor, which was used for the detection of the EDA signal. Specifically, a set of Python scripts was developed to manage communication via serial port with the sensor and to perform real-time processing. The signal was segmented into 20-second windows, with 50% overlap, in order to detect rapid changes in the signal. Then, it was filtered using three different techniques. First a low-pass filter (5Hz cutoff) was applied to remove high-frequency noise, followed by downsampling at 16 Hz, and finally a moving average filter with 1s window was applied to smooth the signal. Ultimately, the cvxeda algorithm was used to decompose the signal into its phasic and tonic components through convex optimizazion. The whole system was tested in an experimentation undertaken in a logistics laboratory (ResLog) at the Politecnico di Torino. The laboratory simulates an automatic logistics warehouse. Specifically, the experimental protocol consisted of five phases, each lasting 10 minutes and organized in order of increasing difficulty. The experiment involved 32 participants, and for each window, four features were computed and normalized to the baseline value. Three of these features are related to the phasic component of the signal, (average value, maximum peak amplitude, average peak amplitude), while the fourth is calculated as the average value of the tonic component of the signal. The data collected online were then used offline for the training of various classifiers (Random Forest, KNN, SVM, XGBoost). The findings indicate that the so-processed EDA information has the ability to differentiate between different levels of arousal and cognitive load. The Random Forest and XGBoost classifiers demonstrated the best performance, particularly in the context of binary classifications. Conversely, their performance decreased in multiclass classifications.