Similarity and variability of the detected TAA instances. We applied the k-means clustering on the normalized features of the detected TAA instances.

We applied the k-means clustering on the normalized features of the detected TAA instances. (A) Optimal numbers of clusters, assessed by the second derivative of sum of squared deviations (SSD, lower is better), silhouette score (higher is better), and Calinski-Harabasz score (higher is better). Taking the three criteria into account, we identify five clusters as optimal. (B) The clusters can be roughly described as: TAA instances with large duration (cluster 2), instances with low variance explained by first two PCA components, either with small slope (cluster 1) or large slope (cluster 4), and instances with high variance explained, either with small slope and low R² (cluster 3) or varying slope and high R² (cluster 5). The difference between the latter two clusters might not be meaningful, as the coefficient of determination R² does not convey useful information when the slope is small. (C) Detected TAA instances in individual subjects. The numbers refer to the S1 Table, each circle represents a detected TAA instance with coloring corresponding to the clusters in panel B, and each row represents one seizure. None of the clusters is specific to a single subject, and no subject is thus clear outlier from the rest of the data set.

(PDF)