Adaptive density peaks clustering: Towards exploratory EEG analysis

摘要

Finding appropriate cluster centers and determining the scope of influence explicitly associated with each center is at the very core of a successful clustering process, which has long been particularly difficult and important when handling bio-signals such as electroencephalography (EEG). Considering exploratory EEG analysis as a typical case, this study forms an adaptive density peaks clustering (ADPC) solution to the open problem based on the Density Peaks Clustering (DPC) algorithm. First, in order to optimize the cutoff distance (key parameter previously set manually) to adapt to various clustering tasks, an optimization function was constructed with the target dataset’s uncertainty that can be solved by the extended Pattern Search Algorithm (PSA). Second, ADPC automatically constructs a set of cluster centers by jointly ranking the local density and relative distance, and then fine-tuning the set by balancing the intra-set independence and the tendency as a center against extra-set competitors from the perspective of each candidate. An exploratory EEG analysis framework was then fostered by centering on ADPC. Benchmarks on public datasets show the superiority of ADPC over its mainstream counterparts in terms of effectiveness and adaptability. The case study on epileptic EEG indicates that (1) the framework achieves averages on Precision, Recall, and F1-score of 100%, 92.46%, and 95.92%, respectively, in seizure detection involving no a priori information, and (2) the key observations revealed through clustering match the accepted conclusions well. Overall, ADPC enables automated clustering, which is well adaptive to exploratory EEG analysis.