Akademik Veri Yönetim Sistemi
Karadeniz Fen Bilimleri Dergisi, cilt.15, sa.1, ss.152-170, 2025 (Hakemli Dergi)
Motor imagery (MI)
classification using EEG signals has gained popularity, playing an
essential role in developing technologies such as brain-computer
interfaces (BCIs). This paper proposes novel approaches using the
Stockwell transform (S-transform) to encode signals into images in
time-frequency space and classify them by feeding them to pre-trained
Inception-ResNet-V2, AlexNet, and SqueezeNet CNNs. High
subject-to-subject and session-to-session signal variability hinder the
recognition of MI tasks. Most literature has studied within-subject
performance. This study conducted experiments using a
leave-one-subject-out cross-validation strategy, investigated
inter-subject variation's effect and contributed by evaluating the
model's performance and generalization ability. At the same time,
different sessions and the presence or absence of feedback were
assessed, and the results were analyzed. The results are encouraging,
considering the difficulty of classifying MI and inter-subject
differences. For a cue-based paradigm and non-feedback signals, the
results are between 62.1-80.8%; for signals with smiley feedback, the
results are between 57.1-96.3%; and for signals with and without
feedback are between 56.8-91.4%. These findings highlight the potential
of combining the S-transform with CNNs, offering valuable insights into
inter-subject variability in EEG-based BCI applications.