Figure 2

Representation of the number of clustering solutions obtained for each seizure and feature group. Four clustering methods were applied to UMAP-reduced three-dimensional datasets (KM: K-means clustering for \(k=2, 3, 4\), AH: agglomerative hierarchical clustering for \(k=2, 3, 4\), GMM: Gaussian mixture models for \(k=2, 3, 4\) and HDBSCAN). Parameter tuning was performed for UMAP (ten values of nearest neighbours and nine values of minimum distance). The final reduced data obtained before each seizure’s onset and for each feature reduction method corresponds to the maximum DI obtained among the computed clustering solutions.