Time-frequency analysis combined with recurrence quantification for classification of onset of dementia using data from the oddball BCI paradigm

Reliable classification of EEG data based on a low number of electrodes is of great practical importance. Brain reactions to sensory stimuli may serve as digital biomarkers for detecting and monitoring the progression of dementia. Traditional approaches to the analysis of event related potentials (ERP) based on amplitude and latency variations do not have sufficient sensitivity to provide reliable biomarkers. Searching for a better approach we have combined time-frequency (TF) spectral representation of the EEG signal with recurrence quantification analysis (RQA). The non-linear features derived from the recurrence plots constructed in this way help to discriminate subtle differences of EEG signals. Auditory and visual stimuli were used in a single-trial oddball-type BCI experiments. Simple EEG using 16 electrodes was sufficient to achieve state-of-art classification accuracy of auditory and visual stimuli, using linear SVM method with selected RQA features. The differences between the standard linear amplitude ERP analysis and TF-RQA approach is also shown in the histograms of the SVM linear projection and in the Uniform Manifold Approximation and Projection (UMAP) plots. The small scale of these experiments and lack of longitudinal observations does not allow justifying the hypothesis linking mild cognitive impairment (MCI) with auditory response degradation, but the TF-RQA should be a good tool in creation of such biomarkers.