We presented the results in figures (Fig. 8, Fig. 9, and Fig. 10), and for readers' convenience in accessing precise performance values, we have included them in Table 2.
In Fig. 8, we present a bar plot comparing the accuracy of our model in each channel. It is the same for specificity and sensitivity, shown in Fig. 9 and Fig. 10 respectively.
Analyzing the figures (Fig. 8, Fig. 9, and Fig. 10) and referring to Table 2, it becomes evident that our model consistently achieved accuracies ranging from 89–99.52% across all 23 channels. Thus, we can highlight the model's independence from specific channel selection, providing flexibility for engineers and users to employ any channel without constraints on electrode positioning
However, as marked in Table 2, seven channels stand out for their exceptionally high accuracy: F4-C4, C4-P4, FP2-F8, T8-P8, FZ-CZ, CZ-PZ, and T8-P8. We strongly recommend these channels for use with our model.
When comparing the bar plots for channel FT9-FT10 with other channels in Fig. 10, it becomes apparent that this channel lacks sufficient information for distinguishing non-ictal from ictal states. With a sensitivity of 100% but a specificity of 77.88%, it results in an overall accuracy of 89%. Consequently, we do not recommend using this channel with our proposed model.
Similarly, we can see from Fig. 9 that the channel T7-FT9 shows a slight decrease in its ability to detect the ictal state, with a sensitivity of 93.33% but a specificity of 100%, resulting in an accuracy of 96.65%. Thus, the best channels to consider are those that achieved the highest accuracy of 99.52, with 14 others boasting accuracy levels between 96% and 97%.
In Fig. 11, we present the best electrodes highlighted in green, and it's noteworthy that they are all situated on the right side of the brain or in central positions, including FCZ, PZ, and CZ. This observation provides valuable insights into the significance of the right side and middle regions in automated epileptic seizure detection.
Conversely, the T7-FT9 channel, which exhibited a slight decrease in sensitivity, is located on the left side. The channel with the lowest accuracy is the difference potential between electrodes on the two sides. Thus, we can suggest a potential advantage for electrodes on the right side and indicate that shorter distances between electrodes may lead to better performance.
In general, the results emphasize our model's capability to detect seizures using data from a single channel. Its channel independence is evident, as it consistently maintains a high level of accuracy, with the lowest being 89%, underscoring the potential for cost-effective seizure onset detection solutions in the future.
Table 2
The results of validating our classifier in each channel
Electrode Names | Accuracy | Specificity | Sensitivity |
FP1-F7 | 98.56% | 97.12% | 100.00% |
F7-T7 | 97.61% | 99.04% | 96.19% |
T7-P7 | 98.09% | 100.00% | 96.19% |
P7-O1 | 97.61% | 100.00% | 95.24% |
FP1-F3 | 98.09% | 96.15% | 100.00% |
F3-C3 | 98.56% | 100.00% | 97.14% |
C3-P3 | 98.09% | 100.00% | 96.19% |
P3-O1 | 98.09% | 100.00% | 96.19% |
FP2-F4 | 98.56% | 97.12% | 100.00% |
F4-C4 | 99.52% | 100.00% | 99.05% |
C4-P4 | 99.04% | 100.00% | 98.10% |
P4-O2 | 98.09% | 97.12% | 99.05% |
FP2-F8 | 99.04% | 98.08% | 100.00% |
F8-T8 | 98.09% | 98.08% | 98.10% |
T8-P8 | 99.52% | 100.00% | 99.05% |
P8-O2 | 98.09% | 97.12% | 99.05% |
FZ-CZ | 99.52% | 100.00% | 99.05% |
CZ-PZ | 99.52% | 100.00% | 99.05% |
P7-T7 | 98.09% | 100.00% | 96.19% |
T7-FT9 | 96.65% | 100.00% | 93.33% |
FT9-FT10 | 89.00% | 77.88% | 100.00% |
FT10-T8 | 98.56% | 100.00% | 97.14% |
T8-P8 | 99.52% | 100.00% | 99.05% |