TABLE 1.
The methods for classifying epilepsy using machine learning in MEG data.
| References | Moda-lity | Problem that was solved in that study | Database | Data acquisition | Source locali-zation | Fea-tures | Classifi-cation | Performance metrics | |||||||||
| Sample size | Age range |
Sex (M:F) |
Source | Total dura-tion | Cha-nnels |
Seg-ment/ epoch length |
Fre-quency samp-ling | Pre-process-ing |
Sensiti-vity (%) |
Specifi-city (%) |
Accu-racy (%) |
||||||
| Khalid et al., 2015 | MEG | EP vs HC | 18 EPs, 15 HCs | / | / | KFMC | 15 min | more than 300 | / | 1,000 Hz | / | / | Standard deviation | LDA | 94.4 | 100 | 95.7 |
| Khalid et al., 2016a | MEG | EP vs HC | 35Eps, 35HCs | / | / | KFMC | 15 min | 306 | / | 1,000 Hz | / | / | Energy of the Delta and Theta component | Threshold method | 96.66 | 98.66 | / |
| Matsubara et al., 2018 | MEG | mTLE vs HCs | 25 left mTLE, 14 right mTLE, 32 HCs | 20–68 | females | Kyushu University | at least 120 evoked responses were counted | 306 | / | 1,000 Hz | band-pass filter: 0.1–330 Hz | minimum norm estimate (MNE) software | phase-locking factor (PLF) and phase-locking value (PLV) | LDA | 82.1 | 81.3 | 81.7 |
| left TLE vs right TLE | 92 | 92.9 | 92.3 | ||||||||||||||
| Wu et al., 2018 | MEG | left / right TLE vs HCs | 15 left/15 right TLEs, 15 HCs | 15–62 | / | Nanjing Brain Hospital, Nanjing Medical University | 30 min | 275 | 120 s | 1,200 Hz | band-pass filter: 1–4 Hz | standardized low resolution brain electro-magnetic tomography (sLORETA) was based on minimum-norm estimation (MNE) | nodal degree, betweenness centrality, and nodal efficiency | RBF-SVM | / | / | 77.38 |
| left TLE vs right TLE | / | / | 88.1 | ||||||||||||||
| Alotaiby et al., 2019 | MEG | EPs vs HCs | 32 EPs, 32 HCs | / | / | KFMC | ≈ 19 min | 306 | 1 min | 1,000 Hz | band-pass filter: 0.03–330 Hz | / | 8 statistical features | RBF-SVM | 99.35 | 95.47 | / |
| Gautham et al., 2022 | MEG | left TLE vs HCs | 54 TLE, 21 HCs | 15–37 | 36:18 | the MEG Research Centre at NIMHANS, Bangalore, India | 5 min | 306 | / | 2,000 Hz | down-sampled to 500 Hz | beamformer | phase amplitude coupling (PAC) | SVM | / | / | 92.92 |
| right TLE vs HCs | / | / | 93.54 | ||||||||||||||
| left TLE vs right TLE | / | / | 92.04 | ||||||||||||||
| Wang et al., 2022 | MEG | CPS vs SPS | 16 SPS, 16 CPS | 17–38 | 13:19 | Nanjing Brain Hospital, Nanjing Medical University | 40 min | 275 | 120 s | 1,200 Hz | low-pass filter:70 Hz, high-pass filter: 1,000 Hz, notch-filter: 50 Hz, down-sampled to 100 Hz | partial canonical correlation/ coherence (PCC), Fieldtrip |
resting state functional connectivity features | SVM | 81.1 | 81.54 | 81.37 |
| Soriano et al., 2017 | MEG | EPs vs HCs | 14 frontal focal EPs, 14 idiopathic generalized EPs, 14 HCs |
16–52 | 1:1 | the University General Hospital of Ciudad Real | 10 min | 306 | 5 s | 1,000 Hz | band-pass filter: 0.1–330 Hz | / | total and relative power | ELM | 93 | 86 | 90 |
| generalized vs focal epilepsy | spectral densities (PSD), the phase-locking value (PLV) and the phase-lag index (PLI) | 1 | 86 | 93 | |||||||||||||
| Bhanot et al., 2022 | MEG | localize the brain region from where the seizure originated | 15 EPs | 14–34 | 9:6 | the MEG Research Centre at NIMHANS, Bangalore, India | 12 min | 306 | / | 2,000 Hz or 50,000 Hz | band-pass filter: 0.1–100 Hz, down-sampled to 250 Hz | / | short-time permutation entropy (STPE), gradient of STPE (GSTPE), short-time energy (STE), and short-time mean (STM) | RUSBoost | 93 | 93 | 93.4 |
EPs, Epileptic patients; HCs, Healthy controls; TLE, temporal lobe epilepsy; SPS, simple partial seizure; CPS, complex partial seizure; SCI, spinal cord injury patients; NNI, National Neuroscience Institute; KFMC, King Fahad Medical City, Riyadh, Saudi Arabia.