A trimodal CNN fusing spatial, temporal, and spectral EEG features achieved automated seizure detection with 0.834 accuracy, 0.875 ROC-AUC, 0.751 sensitivity, and 0.839 specificity.
Does a trimodal CNN fusion model improve automated seizure detection in clinical EEG recordings?
Fusing spatial, temporal, and spectral EEG representations in a single CNN provides robust automated seizure detection on heterogeneous clinical recordings.
Absolute Event Rate: 0% vs 0%
Introduction: Automated seizure detection is useful in critical care environments without continuous EEG review but is challenged by non-stationarity, class imbalance, and noise. We sought to enable robust, patient-independent seizure recognition via late fusion of three feature representations, raw waveforms, frequency spectra, and spatial channel mosaics. Methods: We used the Temple University Hospital (TUH) EEG Seizure Corpus v2.0.3 (1,643 sessions: 579 patients for training, 53 for development (dev), and 43 for evaluation (eval)). We notch-filtered at 60 Hz, band-pass filtered 1.5-35 Hz, converted to 16-channel bipolar montages, resampled at 200 Hz, segmented 12-s windows (271,751 train, 38,384 dev, 130,259 eval), and labelled by maximal temporal overlap with expert annotations. Our CombinedNet model used 3 parallel branches: 1) an EfficientNet-B5 operating on 2×8-channel mosaic “images”; 2) a 6-layer temporal convolutional neural network (CNN) on raw windows; and 3) a 4-layer 1D CNN on global spectral FFT magnitudes. Feature vectors were concatenated and passed to a 2-layer classifier. Each window was perturbed on-the-fly with 1 or more stochastic transforms: 1) random circular time-shifts up to ±10% of window length; 2) additive Gaussian noise (σ=0.002×signal range); 3) amplitude scaling drawn from U(0.9, 1.1); (4) full time-reversal; and (5) a smooth temporal mask zeroing a 50-sample band using a double-sigmoid taper to preserve contextual cues at the edges. To combat class imbalance, seizure windows were always eligible, while only 30% of background windows were augmented. Training used class-balanced sampling, AdamW, cosine learning rate annealing, and early stopping based on dev macro-F1. Results: In the dev set, CombinedNet yielded macro-F1=0.655, Se=0.651, and Sp=0.872. In the blinded eval set, the frozen model yielded accuracy=0.834, macro-F1=0.615, balanced accuracy=0.795, ROC-AUC=0.875, Se=0.751, and Sp=0.839. Conclusions: Fusing spatial, temporal, and spectral EEG representations in a single network substantially improved seizure detection on heterogeneous clinical recordings without handcrafted heuristics. The automated pipeline generalized across unseen patients, offering a building block for real-time seizure-monitoring systems.
Zabihi et al. (Sun,) reported a other. A trimodal CNN fusing spatial, temporal, and spectral EEG features achieved automated seizure detection with 0.834 accuracy, 0.875 ROC-AUC, 0.751 sensitivity, and 0.839 specificity.