EEG records brain activity and contains two kinds of features: traits (relatively stable within a recording session, e.g., subject identity, age, sex) and states (instantaneous signatures of external stimuli such as visual, auditory, or motor information). This work compares signal‑processing features and foundation‑model embeddings in how they capture these aspects. Using a hierarchy metric and k‑nearest‑neighbor purity on a BCI dataset, we find that most embeddings form a macro‑trait hierarchy: trait‑level clusters contain micro‑state sub‑clusters. Linear‑probe results show that this hierarchy reduces state separability and harms BCI performance. To address this, we apply normalization techniques that suppress trait information in the embeddings, yielding roughly a $10\%$ gain in balanced accuracy for state classification. Further fine‑tuning experiments reveal that downstream state‑decoding tasks reshape the embedding space from a macro‑trait to a macro‑state hierarchy, improving state discrimination.
Review: While a macro‑trait hierarchy aids subject identification, it obscures subtle stimulus‑driven variations. Trait‑reduction via normalization or task‑specific fine‑tuning can markedly boost state decoding performance in BCI applications.