Fast Recursive Data-driven Multi-resolution Feature Extraction For Physiological Signal Classification

This study presents a new approach to feature extraction for real-time classification of physiological signals. By using multiple resolutions for the analysis of the signal, the stability of large analysis windows is combined with the capability of small windows to reflect quick changes. A large number of generic features is extracted from each signal for each resolution. These are calculated recursively for each sample which makes them very efficient in terms of computation time; a version with low memory requirements is also provided. A labelled dataset is utilised to convert the generic features into taskand signal-specific features by means of a data-driven transform. The performance of the approach is evaluated on a database containing different stress levels collected in a simulated driving context. A recognition date of 89.8 % is achieved for online, user-independent classification of stress.