Explainable electroencephalography-based attention-deficit/hyperactivity disorder detection model with a combination of ternary pattern and twin wavelet transform

Figure 1 schematic of the recommended model. EEG: Electroencephalography; TWT: Twin wavelet transform; kNN: K-nearest neighbors; NCA: Neighborhood component analysis; IMV: Iterative majority voting; CTP: Combination ternary pattern; ADHD: Attention-deficit/hyperactivity disorder.

Figure 2 Schema of twin wavelet transformation. H: High-pass filter band; L: Low-pass filter band.

Figure 3  Sample signal and its wavelet bands.

Figure 4  Frequency band analysis of wavelet decomposition showing the first six levels.

Figure 5 Graphical overview of the developed combination ternary pattern. EEG: Electroencephalography.

Figure 6  Confusion matrices of channel 8 (left) and the sixth iteration of majority voting (right).

Figure 7 Receiver operating characteristic curve. ROC: Receiver operating characteristic; AUC: Area under the curve.

Figure 8 Channel 8. A: Distribution of features on the selected feature vector generated from channel 8, stratified by signal/wavelet bands and textural/statistical features; B: Classification performance obtained for various ablation models using channel 8 electroencephalography input. The channel-wise classification was performed using k-nearest neighbors and a bagged tree with 10-fold cross-validation. kNN: K-nearest neighbors; BT: Bagged tree.

Figure 9 Comparison with deep learning. 1D-CNN: One-dimensional convolutional neural network; BilSTM: Bidirectional long short-term memory network.