TY - GEN
T1 - LMU-EEG
T2 - 21st IEEE Biomedical Circuits and Systems, BioCAS 2025
AU - Baharlouei, Danial
AU - Ahrar, Alireza
AU - Assaad, Maher
AU - Azghadi, Mostafa Rahimi
AU - Amirsoleimani, Amirali
N1 - Publisher Copyright:
© 2025 IEEE.
PY - 2025
Y1 - 2025
N2 - This paper presents an energy-efficient seizure detection framework leveraging Legendre Memory Unit (LMU)based models, designed to address the challenges of long-range temporal modeling in EEG time-series data. LMUs offer a hardware-friendly alternative to conventional recurrent architectures by employing a fixed, linear time-invariant (LTI) memory system based on orthogonal Legendre polynomials, enabling precise temporal representation with minimal computational overhead. Our pipeline integrates signal preprocessing, frequency domain feature extraction, Random Forest-based EEG channel selection, class balancing via SMOTE, and LMU-based classification. Evaluated on the CHB-MIT Scalp EEG dataset, the proposed system achieves high sensitivity (average 97.26%) and low false detection rates (average 0.0329) across multiple subjects. These results demonstrate that LMUs not only outperform traditional models like LSTMs in detection accuracy and temporal capacity, but also maintain advantages critical for real-time, lowpower neurotechnologies.
AB - This paper presents an energy-efficient seizure detection framework leveraging Legendre Memory Unit (LMU)based models, designed to address the challenges of long-range temporal modeling in EEG time-series data. LMUs offer a hardware-friendly alternative to conventional recurrent architectures by employing a fixed, linear time-invariant (LTI) memory system based on orthogonal Legendre polynomials, enabling precise temporal representation with minimal computational overhead. Our pipeline integrates signal preprocessing, frequency domain feature extraction, Random Forest-based EEG channel selection, class balancing via SMOTE, and LMU-based classification. Evaluated on the CHB-MIT Scalp EEG dataset, the proposed system achieves high sensitivity (average 97.26%) and low false detection rates (average 0.0329) across multiple subjects. These results demonstrate that LMUs not only outperform traditional models like LSTMs in detection accuracy and temporal capacity, but also maintain advantages critical for real-time, lowpower neurotechnologies.
KW - LSTM
KW - Legendre Memory Unit
KW - Linear Time-Invariant
KW - Random Forest
KW - SMOTE
UR - https://www.scopus.com/pages/publications/105033220363
U2 - 10.1109/BioCAS67066.2025.00062
DO - 10.1109/BioCAS67066.2025.00062
M3 - Conference contribution
AN - SCOPUS:105033220363
T3 - Proceedings - 21st IEEE Biomedical Circuits and Systems, BioCAS 2025
SP - 249
EP - 253
BT - Proceedings - 21st IEEE Biomedical Circuits and Systems, BioCAS 2025
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 16 October 2025 through 18 October 2025
ER -