Identification of Epileptogenic and Non-epileptogenic High-Frequency Oscillations Using a Multi-Feature Convolutional Neural Network Model

Figure 1. Architecture of the convolutional neural network (CNN) model. The CNN model consists of input, convolutional, batch normalization, max pooling, fully connected, dropout, and softmax layers.

Figure 2. The smoothed pseudo Wigner–Ville distribution (SPWVD). (A) The original waveform of a 400 ms epoch. (B) Left: the first normalization of the SPWVD transformation (0–500 Hz). In this figure, all of the SPWVD transformation values are normalized by dividing all time points by the maximum SPWVD value. The normalized value (Ŵtf) is the SPWVD index. Middle: Re-normalized SPWVD. The Ŵtf in the left is re-normalized by dividing all time points in each HFO by the maximum SPWVD index in all frequencies. Right: Adjusted SPWVD (ASPWVD). The Ŵtf in the left were adjusted by dividing all HFOs time points by the median maximum SPWVD index. (C) Curve C for re-normalized SPWVD. This is curve C for this epoch from 80 to 500 Hz. The small rectangle in the figure is the enlargement of curve C for FR (200–500 Hz). (D) Curve C for the ASPWVD. This is curve C with the median maximum SPWVD index for all epochs from 80 to 500 Hz. The small rectangle in the figure is the enlargement of curve C for FR (200–500 Hz). (E) Ripple (80–200 Hz) SPWVD spectrum images. Left: the first normalization of the SPWVD spectrum image. Middle: Re-normalized SPWVD spectrum image. Right: the ASPWVD spectrum image. (F) FR…

Figure 3. Four ripple image transformations. Left: Transformations of a ripple. Right: Network inputs in the form of two-dimensional 96 × 96-pixel images that represent various transformations. (A) Transformations with fixed scales. From top to bottom are original waveforms, filtered (80–200 Hz) waveforms, wavelet spectrum images, and ASPWVD spectrum images. (B) Transformations with fitted scales. From top to bottom are original waveforms, filtered (80–200 Hz) waveforms, wavelet spectrum images, and re-normalized SPWVD images.

Figure 4. Four fast ripples (FRs) image transformations. Left: Transformations of a FR. Right: Network inputs in the form of two-dimensional pictures of 96 × 96 pixels that represent various transformations. (A) Transformations with fixed scales. From top to bottom are original waveforms, filtered (200–500 Hz) waveforms, wavelet spectrum images, and ASPWVD spectrum images. (B) Transformations with fitted scales. From top to bottom are original waveforms, filtered (200–500 Hz) waveforms, wavelet spectrum images, and re-normalized SPWVD spectrum images.

Figure 5. Network inputs in the form of two-dimensional pictures of ripples. (A) Two-dimensional pictures (96 × 96 × 1 pixel) of five ripples in different channels of EZs. Upper and lower four panels are fixed and fitted scales. The four panels are original waveforms, filtered (80–200 Hz) waveforms, wavelet spectrum images, and SPWVD spectrum images. (B) Two-dimensional pictures (96 × 96 × 1 pixel) of ripples in different channels of non-EZs. Upper and lower four panels are fixed and fitted scales. The four panels are original waveforms, filtered (80–200 Hz) waveforms, wavelet spectrum images, and SPWVD spectrum images.

Figure 6. Network inputs in the form of two-dimensional pictures of FRs. (A) Two-dimensional pictures (96 × 96 × 1 pixel) of five FRs in different channels of EZs. Upper and lower four panels are fixed and fitted scales. The four panels are original waveforms, filtered (200–500 Hz) waveforms, wavelet spectrum images, and SPWVD spectrum images. (B) Two-dimensional pictures (96 × 96 × 1 pixel) of ripples in different channels of non-EZs. Upper and lower four panels are fixed and fitted scales. The four panels are original waveforms, filtered (200–500 Hz) waveforms, wavelet spectrum images, and SPWVD spectrum images.

Figure 7. Convolutional neural networks (CNN) accuracies for ripples classification. (A) Accuracies of CNNs for ripples classification using various transformations. The CNN network using the fixed-scale ASPWVD images provides the highest accuracy (p < 0.05). Except for original ripple waveforms, the results from CNN models that are trained using fixed-scale images are significantly better than those from models trained using fitted-scale images (p < 0.05). (B) CNN training and validation accuracies and losses are displayed. Asterisk (*) means a statistically significant difference.

Figure 8. CNN accuracies for FRs classification. (A) Accuracies of CNNs for FRs classification using various transformations. The CNN using the fixed-scale ASPWVD images provides the highest accuracy (p < 0.05). The results from CNN models that are trained using fixed-scale images are significantly better than those from models trained using fitted-scale images (p < 0.05). (B) CNN training and validation accuracies and losses are displayed. Asterisk (*) means a statistically significant difference.