Table 2 . Classification algorithms, validation methods, and performance metrics of machine learning models for MCI detection using wearable EEG

From: Wearable EEG devices in the detection of mild cognitive impairment: a systematic review

Author (Year)

Classification algorithms

Validation method

Accuracy

Performance metrics

Performance ranking

Rutkowski et al. (2023a)33

LR, LDA, SVM, RF, DFNN

LOOCV

0.81–0.95

AUC: 0.82–0.94; F1: 0.85–0.95; Recall: 0.85–0.95

Reminiscent interior oddball task > Emotion Assessment Evaluation Task > Emotion Assessment Learning Task

Boudaya et al. (2024)34

SVM (Gaussian kernel), GB, RF, KNN, DT

10-fold CV

0.75–0.91

NR

KNN(k=5) > KNN(k=3) > KNN(k=1) > SVM > RF > GB > DT

Wu et al. (2023)35

DT, RF, SVM, XGBoost

LOOCV

0.74–0.83

F1: 0.74–0.81; Precision: 0.72–0.81; Recall: 0.74–0.80

SVM > XGBoost > RF > DT

Li et al. (2023)36

Soft Voting, GBDT, RF, NB, DT, k-NN

10-fold CV

0.68–0.82

F1: 0.68–0.83; Precision: 0.69–0.84; Recall: 0.68–0.82

Soft Voting > GBDT > RF > NB > DT > kNN

Chai et al. (2023)37

SVM, RF, XGBoost

10-fold CV

0.80–0.86

F1: 0.79–0.86; Kappa: 0.60–0.72

RBF SVM > XGBoost > RF

Perez-Valero et al. (2022)30

SVM, LR

LOSO CV

0.94

F1: 0.93–0.96; Precision: 0.98; Recall: 0.88

SVM > LR

Lee et al. (2022a)38

SVM

LPO-CV

0.56–0.74

Sensitivity: 0.50–0.70; Specificity: 0.62–0.87

Imec (8ch) > EPOC X (14ch) > DSI-7 (7ch) > Insight (5ch) > Focusband (2ch)

Lee et al. (2022b)39

RF

LOOCV

0.46–0.66

F1: 0.42–0.64; Precision: 0.44–0.65; Recall: 0.40–0.64; MCC: −0.09-0.31

Resting-state EEG + Selective attention EEG > Resting-state EEG + Working memory EEG > Resting-state EEG + Depth perception EEG > Resting-state EEG

Boudaya et al. (2022)40

LR, SVM (RBF kernel), KNN, DT, RF

10-fold CV

0.75–0.87

NR

RF > LR > SVM > KNN > DT

Jiang et al. (2019)32

LR

Randomized CV

0.62

Sensitivity: 0.44; Specificity: 0.76; AUC: 0.67

-

Jin-Young et al. (2022)41

LR

ROC Curve Analysis

0.67–0.68a

AUC: 0.67–0.68

Model with high beta power > Model with gamma power

Chen et al. (2023)42

SVM, KNN, DT

LOOCV

0.61–0.75

Sensitivity: 0.63–0.75; Specificity: 0.63–0.78

Single channel: ERP (latency) > ERP (all) > ERP (voltage)

Multiple channels: ERP > PSD > PSDE

Rutkowski et al. (2021a)43

LR, LDA, linearSVM, rbfSVM, polySVM, sigmoidSVM, RF, DFNN

10-fold CV

0.54–0.92b

NR

RFC > FNN > rbfSVM > linearSVM > LDA > shrinkageLDA > polySVM > LR > sigmoidSVM

Rutkowski et al. (2022)44

LR, LDA, linearSVM, rbfSVM, polySVM, sigmoidSVM, RFC, FNN

10-fold CV

0.45–0.85a

Best AUC median results (approaching 0.9)

RFC > rbfSVM > FNN > shrinkageLDA > polySVM > LDA > linearSVM > LR > sigmoidSVM

Rutkowski et al. (2023b)45

LR, LDA, linearSVM, RFC, DFNN

LOOSCV

0.64–0.95b

NR

linearSVM > DFNN > LDA > LR > RFC

Rutkowski et al. (2021b)46

LR, LDA, rbfSVM, polySVM, sigmoidSVM, RFC, FNN

10-fold CV

0.51–0.88b

NR

FNN > RFC > rbfSVM > polySVM > LDA > LR > sigmoidSVM

Meghdadi et al. (2021)51

LDA

LOOCV

0.60–0.80a

NR

Trained (No CV) > LOOCV

Segaert et al. (2022)47

ROC

NR

0.87–0.90a

NR

Single-word retrieval signature > Binding signature

Xue et al. (2024)48

DT, RF, SVM (RBF kernel), XGboost

10-fold CV

0.72–0.83b

NR

Cognitive Task EEG (Beta band) > Resting-state EEG (Theta band)

Jiang et al. (2025)49

LR

ROC

0.68–0.83a

Sensitivity: 0.794–0.962 Specificity: 0.376–0.602

OPL8 > OPL7 > PPL7 > OCL4 > PLL4 > PRL3

Hata et al. (2025)50

Customized Transformer (Deep Learning)

10-fold Cross-Validation & Holdout Set

0.79–0.86c

AUC:0.81–0.86

Sensitivity: 0.75–0.80 Specificity: 0.83–0.92

-

  1. AUC Area Under the Receiver Operating Characteristic Curve, CV Cross-Validation, DFNN Deep Feedforward Neural Network, DT Decision Tree, EEG Electroencephalography, ERP Event-Related Potential, F1 F1 Score, FNN Feedforward Neural Network, GB Gradient Boosting, GBDT Gradient Boosting Decision Tree, KNN/k-NN K-Nearest Neighbors, LDA Linear Discriminant Analysis, LOOCV Leave-One-Out Cross-Validation, LOOSCV Leave-One-Out Subject Cross-Validation, LOSO CV Leave-One-Subject-Out Cross-Validation, LPO-CV Leave-Pair-Out Cross-Validation, LR Logistic Regression, MCC Matthews Correlation Coefficient, MCI Mild Cognitive Impairment, NB Naive Bayes, NR Not Reported, Precision Precision (Positive Predictive Value), PSD Power Spectral Density, PSDE Power Spectral Density Entropy, Recall Recall (Sensitivity, True Positive Rate), RF Random Forest, RFC Random Forest Classifier, ROC Receiver Operating Characteristic, Sensitivity Sensitivity (Recall, True Positive Rate), Specificity Specificity (True Negative Rate), SVM Support Vector Machine (kernels mentioned: Gaussian, RBF linear, polynomial, sigmoid), XGBoost Extreme Gradient Boosting.
  2. aDenotes the Area Under the Receiver Operating Characteristic Curve (AUC) is reported in the “Accuracy” column.
  3. bIndicates values were approximated based on graphical data presented in the source publication.
  4. cDenotes Balanced Accuracy (bACC) is reported in the “Accuracy” column.
Back to article page