Table 6.
Age- and sex-segregated metrics for the unsupervised IC classification algorithm, the LS-SVM approach by Hunyadi et al., and the CNN deep learning approach.
| Metric | Algorithm | 0 < Age ≤ 5 (N = 20) | 5 < Age ≤13 (N = 18) | 13 < Age ≤18 (N = 14) | P-value fixed effects on age | Men (N = 23) | Women (N = 29) | P-value fixed effects on sex | Key observations |
|---|---|---|---|---|---|---|---|---|---|
| Noise vs. network/SOZ performance metrics | |||||||||
| Accuracy | EPIK | 69.4% (±9%) | 74% (±8.2%) | 71.7% (±5.8%) | 0.32 | 73.8% (±5.3%) | 70% (±9 . 4 % ) | 0.04 † | CNN gives the best RSN identification accuracy for all age categories. followed closely by EPIK |
| LS-SVM | 55.8% (±11.5%) | 63.7% (±7.7%) | 65.3% (±8.4%) | 0.004† | 63.6% (±9.5%) | 59.1% (±10.6%) | 0.06 | LS-SVM is poorest in identifying RSN since it only considers SOZ markers in ICs. | |
| CNN | 73.2% (±4.5%) | 76.1% (±0.6%) | 80.2% (±5.8%) | ~0 | 72.8% (±8.2%) | 77.4% (±4.7%) | 0.09 | Success of CNN can be attributed to availability of a significant number of normal RSN ICs (n = 2,427) | |
| Precision | EPIK | 74.9% (±16.2%) | 73.6% (±13.7%) | 66.5% (±10%) | 0.048 | 73.5% (±11.5%) | 71.1% (±16%) | 0.27 | |
| LS-SVM | 55.6% (±32.4%) | 52.8% (±15.9%) | 46.5% (±18%) | 0.3 | 54.8% (±22.2%) | 50.1% (±25.4%) | 0.24 | ||
| CNN | 68.2% (±11.7%) | 75.2% (±1.5%) | 75.4% (±7.3%) | ~0 | 69.2% (±13%) | 75.91% (±15.1%) | 0.3 | ||
| Sensitivity | EPIK | 63% (±18%) | 76.6% (±9.3%) | 76.8% (±9.7%) | 0.001 † | 75.2% (±10.7%) | 68.43% (±16.9%) | 0.047 | |
| LS-SVM | 27.5% (±25.9%) | 50.8% (±26.9%) | 55.1% (±22%) | 0.001† | 52.6% (±28.9%) | 35.4% (±24.9%) | 0.012† | ||
| CNN | 86.09% | 81.5% | 85.96% | 78% | 79.5% | ||||
| Specificity | EPIK | 79% (±13.8%) | 72.7% (±17.1%) | 68.2% (±8.4%) | 0.01 † | 73.4% (±13.5%) | 74.2% (±15.2%) | 0.41 | |
| LS-SVM | 80.5% (±17.4%) | 68.8% (±23%) | 70% (±10%) | 0.035† | 71.3% (±17%) | 75.5% (±19.9%) | 0.2 | ||
| CNN | 60.5% | 70.8% | 75% | 67.9% | 75.31% | ||||
| SOZ identification metrics | |||||||||
| Accuracy | EPIK | 87.5% (±7.6%) | 83.5% (±9.6%) | 82.2% (±6.1%) | 0.025 † | 84.6% (±6.7%) | 84.7% (±9.3%) | 0.48 | EPIK has the best performance for SOZ localizing IC identification |
| LS-SVM | 85.3% (±6.6%) | 77.2% (±9.4%) | 78.6% (5.7%) | 0.008† | 79.5% (±8.8%) | 81.6% (±7.7%) | 0.17 | EPIK has consistent performance across age. | |
| CNN | 75.5% (±27.7%) | 75.3% (±26.6%) | 76.5% (±21%) | 0.8 | 71% (±28.2%) | 73% (±30.2%) | 0.44 | EPIC has the best performance for children of age <5 years. This is a key benefit because it is known that earlier surgery for epilepsy yields better surgical and developmental outcomes. | |
| Precision | EPIK | 76.7% (±16.3%) | 75.2% (±14.4%) | 69.2% (±9.9%) | 0.07 | 76.3% (±10.7%) | 72.5% (±16.5%) | 0.17 | |
| LS-SVM | 62.5% (±17.2%) | 56.9% (±15%) | 51.4% (±15%) | 0.06 | 54.7% (±15.7%) | 55% (±16.3%) | 0.2 | ||
| CNN | 53.8% (±50.2%) | 50% (±51.4%) | 45% (±50%) | 0.035 | 45.9% (±49.9%) | 54.4% (±50%) | 0.54 | ||
| Sensitivity | EPIK | 86.8% (±8 . 8%) | 89.4% (±6.8%) | 90.4% (±7.6%) | 0.085 † | 88.1% (±7.1%) | 89.1% (±8.5%) | 0.34 | |
| LS-SVM | 58.8% (±33.9%) | 74.6% (±19.5%) | 87.8% (±23%) | 0.001† | 78.83% (±25.6%) | 66.7% (±30.4%) | 0.065 | ||
| CNN | 11.1% (±3%) | 0 (±0) | 0 (±0) | 0.001 | 20% (±5%) | 3.44% (±3%) | 0.002 | ||
| Specificity | EPIK | 86.6% (±12.8%) | 79.9 % (±15.6%) | 77.8% (±8%) | 0.02 † | 81.7% (±12.1%) | 82% (±14.2%) | 0.47 | |
| LS-SVM | 86.4% (±14.1%) | 73.2% (±21.8%) | 74.9% (±9.3%) | 0.015† | 75.3% (±17%) | 81.5% (±17%) | 0.09 | ||
| CNN | 74.9% (±27.6%) | 75.3% (±26.6%) | 76.1% (±22%) | 0.4 | 75.37% (±28.2%) | 76.59% (±30%) | 0.8 | ||
†
Indicates that the result has a p-value of < 0.05 and is statistically significant. Bold value refers to our technique EPIK's results.