TABLE 4.
Machine learning applications used to predict clinical outcomes in epilepsy.
| References | Subjects | Imaging modality | Imaging features | Classifiers | Main outcomes |
| Paldino et al. (2014) | 32 MCD (P) | DTI | FA, MD | RF | Sens. = 1.00, Spec. = 0.954 to predict language impairment |
| Amarreh et al. (2014) | 20 EPI (P), 29 HC | DTI | FA, MD, AD, RD | SVM | ACC > 90% for EPI vs. HC, > 75% to discriminate remission |
| Memarian et al. (2015) | 20 TLE (10 R, 10 L) | sMRI | Morph (SBM) | SVM | ACC = 95% to predict seizure outcome after surgery, when combined with other clinical info/EEG |
| Munsell et al. (2015) | 70 TLE, 43 HC | DTI | Network measures | SVM | ACC = 0.80 for TLE vs. HC, 0.70 to predict seizure outcome after surgery |
| Bernhardt et al. (2015) | 141 TLE | T1 | SBM | k-means clustering (unsupervised) | Four data-driven distinct classes of TLE, associated with histopathology and seizure outcome |
| Gazit et al. (2016) | 76 EPI | Task-fMRI | BOLD | LR | 89% concordance with WADA to lateralize language hemisphere |
| Ibrahim et al. (2017) | 21 EPI (P) | rs-fMRI | FC | SVM | ACC = 0.86–0.88 to predict VNS response |
| He et al. (2017) | 56 TLE (30 R, 26 L), 28 HC | rs-fMRI | Network measures | SVM | ACC = 0.70 to predict seizure outcome after surgery |
| Paldino et al. (2017b) | 45 FE (P) | rs-fMRI | Network measures | RF | R = 0.95 to predict epilepsy duration |
| Paldino et al. (2017a) | 30 FE (P) | rs-fMRI | Network measures | RF | ML revealed no effect of motion parameters or general amnesia during the scan for IQ prediction |
| Zhang et al. (2018) | 24 FE | rs-fMRI | FC, Network measures | RF | 0.49 of fractional variation to predict IQ |
| Gleichgerrcht et al. (2018) | 50 TLE | DTI | Network measures | NN | PPV = 0.88, NPV = 0.79 to predict seizure outcome after surgery |
| Taylor et al. (2018) | 53 TLE (30 L, 23 R) | DTI | Network measures | SVM, E-net LR | ACC = 0.792 to retrospectively predict seizure outcome after surgery |
| He et al. (2018) | 50 TLE (25 R, 25 L), 30 HC | fMRI (task, rs) | Network measures | RF | ∼100% prediction for verbal fluency, improvement from traditional methods |
| Mithani et al. (2019) | 56 EPI (P) | DTI | FA | SVM | ACC = 0.83–0.89 to predict VNS responders, when combined with MEG |
| Yao et al. (2019) | 287 EPI | Routine MRI | Visual assessment | DT, RF, SVM, LR, XGBoost | AUC = 0.979 to predict AED responders, 0.918 for early responders when combined with clinical info/EEG. |
| Paldino et al. (2019) | 27 FE (P) | rs-fMRI | Network measures | RF | 0.34 fractional variation to predict IQ |
| Zhang et al. (2019) | 117 AVM | T2 | Location, Radiomics | LASSO | ACC around 0.800 to predict epilepsy occurrence |
| Li et al. (2019) | 10 TLE-HS (5 R, 5 L) | T1 | Morph (SBM, etc.) | RF, LR | Accurately predict the optimal trajectories for LITT |
| Ernst et al. (2019) | 46 GAD, 34 VGKC, 48 HC | T1 | Morph (GMV, etc.) | DT | Spec. = 0.87, Sens. = 0.80 for GAD vs. VGKC |
| La Rocca et al. (2020) | 53 TBI | T1 | Network measures | RF | AUC = 0.75 to predict seizure occurrence after TBI |
| Gleichgerrcht et al. (2020) | 168 TLE | T1, DTI | Network measures | NN | AUC = 0.88 to predict seizure outcome after surgery by BC |
| Höller et al. (2020) | 9 TLE, 19 MCI, 4 SCI, 18 HC | T1 | Morph (GMV, etc.) | SVM | Sens/Spec = 0.70–0.90 to predict cognitive decline over time, when combined with MRI, EEG, Neuropsychology. |
| Larivière et al. (2020) | 30 TLE-HS (R 19, L 11), 57 HC | T1, DTI, rs-fMRI | Connectivity distance | LR | ACC = 0.76 to predict seizure outcome after surgery |
| Mazrooyisebdani et al. (2020) | 27 TLE (R9, L18), 85 HC | rs-fMRI | FC | SVM, SVR | ACC = 0.81 for TLE vs. HC. R = 0.61–0.75 to predict neuropsychology |
| Akeret et al. (2020) | 923 brain tumors | T1, FLAIR | Anatomical features | DT, GLM, RF, GBM, NN, SVM, GAM | AUC = 0.79, ACC = 0.72 to predict seizure occurrence when combined with clinical info |
| Lee et al. (2020b) | 89 FE (P) | DTI | WM tract | CNN | ACC = 0.92 to predict functional deficit after surgery |
| Wang et al. (2020) | 205 LGG-related EPI | T2WI | Signal, shape, etc. | Novel radiomic nomogram | AUC = 0.863 to predict epilepsy type |
| Sinha et al. (2021) | 51 TLE, 29 HC | T1, DTI | Network measures | SVM | AUC = 0.84 to predict seizure outcome after surgery |
| Kini et al. (2021) | 89 TLE | FDG-PET | PET signal | RF | ACC = 0.71 to predict seizure outcome after surgery |
ACC, accuracy; AD, axial diffusivity; AUC, area under the ROC curve; AVM, arteriovenous malformation; BC, betweeness centrality; BOLD, blood-oxygen-level-dependent; CNN, convolutional neural network; DT, decision tree; DTI, diffusion tensor imaging; E-net LR,; EPI, epilepsy; FA, fractional anisotropy; FC, functional connectivity; FC, functional connectivity; FE, focal epilepsy; fMRI, functional MRI; GAD, anti-GAD-related encephalitis; GAM, generalized additive model; GBM, gradient boosting algorithm; GLM, generalized linear model; GMV, gray matter volumes; HC, healthy controls; L, left; LASSO, least absolute shrinkage and selection operator; LGG, low-grade gliomas; LITT, laser interstitial thermal therapy; LR, logistic regression; MCD, malformation of cortical development; MCI, mild cognitive impairment; MD, mean diffusivity; Morph, morphological features; NN, neural network; (P), pediatric cases; R, right; RD, radial diffusivity; RF, random forest; rs, resting-state; SBM, surface-based morphometry; SCI, subjective cognitive impairment; sMRI, structural MRI; SVM, support vector machine; SVR, support vector regression; TBI, traumatic brain injury; TLE, temporal lobe epilepsy; VGKC, anti-VGKC-related encephalitis.