ISSN :2582-9793
  • Subscribe

    • Epileptic Seizure Prediction on CHB-MIT EEG Using Soft Fusion Post-Processing of Top-K Predictions with CNN Architectures

    Original Research (Published On: 14-Nov-2025 )
    DOI : https://doi.org/10.54364/AAIML.2025.54254

    Asmaa Mohammad S Balamash, Ghadah Aldabbagh and Samar Alkhuraiji

    Adv. Artif. Intell. Mach. Learn., 5 (4):4575-4593

    1. Asmaa Mohammad S Balamash: King Abdulaziz University

    2. Ghadah Aldabbagh: Faculty of Computing and Information TechnologyKing Abdulaziz UniversityJeddah, Saudi Arabia

    3. Samar Alkhuraiji: Faculty of Computing and Information TechnologyKing Abdulaziz UniversityJeddah, Saudi Arabia

    Download PDF Here

    DOI: 10.54364/AAIML.2025.54254

    Article History: Received on: 04-Aug-25, Accepted on: 07-Nov-25, Published on: 14-Nov-25

    Corresponding Author: Asmaa Mohammad S Balamash

    Email: abalamash0003@stu.kau.edu.sa

    Citation: Asmaa Mohammad Balamas, et al. Epileptic Seizure Prediction on CHB-MIT EEG Using Soft Fusion Post-Processing of Top-K Predictions with CNN Architectures. Advances in Artificial Intelligence and Machine Learning. 2025;5(4):254. https://dx.doi.org/10.54364/AAIML.2025.54254


    Abstract

        

    Epileptic seizure prediction requires not only accurate detection but also reliable

    alarm generation to be clinically meaningful. We propose a Soft Fusion method

    based on Top-K averaging of preictal probabilities as a post-processing step to stabi-

    lize model outputs. Using the CHB-MIT scalp EEG dataset, we evaluate three back-

    bones—a baseline CNN, CNN with Convolutional Block Attention Module (CBAM),

    and CNN with Long Short-Term Memory (LSTM)—with STFT spectrogram inputs

    under a 5 min Seizure Prediction Horizon (SPH) and 30 min Seizure Occurrence Pe-

    riod (SOP). Across models, Soft Fusion improved sensitivity (macro-average up to

    92.8%) while reducing false prediction rates to about 0.8 events per hour (≈20/day),

    consistently outperforming the conventional k-of-n (Hard Fusion) rule. Although

    still above ideal clinical tolerance levels, these results highlight that post-processing

    can be as critical as architectural design in seizure prediction. Limitations include

    retrospective single-dataset evaluation and per-patient thresholding. Nonetheless,

    Soft Fusion represents a promising step toward more reliable and clinically usable

    seizure prediction systems.

    Statistics

       Article View: 1226
       PDF Downloaded: 15