Table 4.
Machine learning models used for dyslexia detection.
| Study | Input data/device | Reading or eye-movement task | Features used | Model type | Validation method (as reported by study) | Reported performance |
|---|---|---|---|---|---|---|
| Raatikainen et al. (20) (Finland) | Eye movement data (EyeLink 1,000, 1,000 Hz) | Internet search “question page” reading (10 trials) | Fixation duration/count; saccade amplitude/duration; AOI transition matrices; histogram features; RF-selected feature subset | RF feature selection → SVM (RBF) (class-weighted) | Repeated stratified 5-fold cross-validation with grid search for hyperparameter tuning | Best hybrid: Accuracy 89.7% ± 4.0, Recall 84.8% ± 14.0 |
| Appadurai and Bhargavi (27) (India) | Gaze points from Ober-2 dataset | Passage reading + comprehension questions | Fixation/saccade features via I-DA & I-VA; PCA + RFE-CV; also, fixation/scanpath images for CNN | SVM variants; Hybrid kernel SVM-PSO; RF/AdaBoost/LGBM/XGBoost; CNN | Repeated random train/test splits (80/20) combined with 10-fold cross-validation | Best reported: Hybrid SVM-PSO ~ 96%, XGBoost ~95% (CNN ~ 87–88%) |
| Nilsson Benfatto et al. (19) (Sweden) | Ober-2 eye tracker (100 Hz) | Silent short passage reading + comprehension | 168 low-level fixation/saccade parameters (progressive/regressive; horiz/vert; version/vergence) | Linear SVM + SVM-RFE feature selection | 10-fold stratified CV, repeated 100×; feature selection inside folds | Accuracy 95.6% ± 4.5%, Sens 95.5%, Spec 95.7% |
| Nerušil et al. (26) (Slovakia) | Horizontal gaze x-coordinate (avg L/R) from Ober-2 dataset | Same reading text task (Benfatto dataset) | Holistic: trimmed reading segment; time-signal + interpolation; DFT magnitude spectrum | CNN2/CNN3/CNN4 (2–4 conv layers) | Repeated random subsampling with multiple evaluation runs (100 iterations; train/validation split) | Best: Accuracy 96.6% ± 2.9; TPR 97.8% ± 2.1; TNR 95.4% ± 4.1 |
| Pereira et al. (32) (Portugal) | Eye tracking + cognitive scores (SMI iView X HI-SPEED, 1250 Hz) | Silent reading with word frequency/length manipulation | FC, SFD, FPRT, SPRT, TFT + cognitive tests; final predictors include Backwards Digit Span, Vocabulary, Coding, FC on long low-freq words | Multinomial logistic regression (SPSS) | No explicit validation strategy reported (overall classification accuracy provided) | 81.4% correctly classified |
| Shalileh et al. (31) (Russia) | EyeLink fixation report + word-level IA metrics + demographics (1,000 Hz) | Silent reading of 30 Russian sentences (+ subset comprehension) | Fixation duration/X/Y; IA metrics (FFD, TRT, FC, skip, regressions, first saccade amp, etc.) + demographics | Multiple models tested; best typically MLP (also CNN/RF/GB/SVM/LR etc.) | Bayesian hyperparameter tuning with stratified cross-validation; final performance reported using stratified 10-fold CV | Best (fix+demo): F1 = 0.912 ± 0.002; ROC-AUC = 0.983 ± 0.000; Best (IA + demo): F1 = 0.934 ± 0.005; ROC-AUC = 0.986 ± 0.001 |
| Vaitheeshwari et al. (33) (Taiwan) | VR eye tracking (HTC Vive Pro + Tobii VR4, 120 Hz) + text semantics + gaze maps | VR multi-page reading + comprehension | Fixation duration/count, saccade amplitude, gaze/trajectory metrics; saliency/trajectory maps; BERT text embeddings | SVM/RF/KNN/DT/NB/XGBoost/DNN + BERT + CNN + voting fusion | Validation approach not clearly specified; performance reported on pilot dataset | Fusion/voting model reported ~98% accuracy (pilot) |
| Rello and Ballesteros (54) (USA/Spain) | Tobii 1750 ET features + age | 12 short texts × 12 fonts + comprehension control | Reading time (sum of visits), mean fixation duration, # fixations (+ age) | Polynomial SVM (LIBSVM) | 10-fold CV keeping each subject’s readings in same fold | Accuracy 80.18% (subject-wise cross-validation applied) |
| İleri et al. (46) (Türkiye) | EOG (BIOPAC MP-36, 100 Hz) → CWT scalograms | Read-aloud Turkish texts varied by typeface/font size | Scalogram images (DL input); also reports blink rate/regressions/energy as indicators | Proposed DyslexiaNet CNN vs. AlexNet/ResNet50/MobileNetV2 | 5-fold CV (K = 5) | DyslexiaNet: ~99.968% accuracy (horizontal channel); vertical channel ~73.7% (performance should be interpreted with caution given dataset size and modality differences) |
| Bhargavi and Jothi Prabha (25) (India) | Ober-2 dataset + external Pupil Labs validation | Reading task (inherited dataset) | Fixation/saccade/blink event features; best subset includes avg. # fixations, avg. fixation gaze duration, avg. saccade duration, total # saccades | Hybrid kernel SVM-PSO + comparisons (SVM/RF/LR/KNN) | 10-fold cross-validation with additional train/test split (80/20); external validation described | Hybrid SVM-PSO 95.6% accuracy; external validation ~96.6% |
| Latifoglu et al. (47) (Türkiye) | EOG (BIOPAC MP36, 100 Hz) → STFT spectrograms | Reading 5-line text; targets rereading vs. line-skipping events | Event segments converted to STFT spectrogram images | 2D-CNN classifier | Validation strategy not clearly specified; performance evaluated on constructed dataset | Accuracy 99%, Sensitivity 100%, Specificity 98.18%, F-score 98.95% (event-type classification; not dyslexia classification) |
| Vajs et al. (21) (Serbia) | SMI RED-m ET (120 Hz) | Silent reading of 13 texts under 13 color configs | Conventional + proposed features (active reading time, fixation intersection coefficient/variability, fixation fractal dimension, saccade variability) | Logistic Regression, SVM, KNN, RF | Leave-one-subject-out cross-validation (30 folds) | Best: ACC = 0.94 (LR) using proposed features |
| Vajs et al. (28) (Serbia) | Raw x-y gaze coords (SMI RED-m 60 Hz; also 30 Hz) | Same colored-text reading paradigm | Single-feature models: SI-event feature, VAS-event feature, baseline reading time | Logistic Regression, SVM, KNN, RF | Leave-one-subject-out cross-validation with inner 5-fold cross-validation for hyperparameter tuning | Best: ACC 88.9% (VAS@60 Hz); best @30 Hz: ACC 87.8% |
| Svaricek et al. (51) (Czech Republic) | Fixation events (SMI RED 250, 250 Hz) → Fix-images | 3 reading tasks (aloud 2 min each) | Fix-images built from fixation x/y + dispersion x/y + duration; baseline conventional metrics also tested | ResNet18 (fine-tuned) + majority-vote ensemble across tasks | Stratified 5-fold CV, repeated; cross-dataset test also reported | Ensemble Accuracy 86.65%; cross-dataset test 86.11% |
Acc, accuracy; AUC, area under the receiver operating characteristic curve; CNN, convolutional neural network; CV, cross-validation; DD, developmental dyslexia; DFT, discrete Fourier transform; DL, deep learning; EOG, electrooculography; F1, F1-score; FC, fixation count; FFD, first fixation duration; FPRT, first-pass reading time; GD, gaze duration; KNN, k-nearest neighbors; LR, logistic regression; LSTM, long short-term memory (if reported); MLP, multilayer perceptron; NB, naïve Bayes; PSO, particle swarm optimization; RBF, radial basis function; RF, random forest; RFE, recursive feature elimination; ROC, receiver operating characteristic; SA, saccade amplitude; Sens, sensitivity; Spec, specificity; SFD, single fixation duration; SPRT, second-pass reading time; SVM, support vector machine; TFT, total fixation time; TNR, true negative rate (specificity); TPR, true positive rate (sensitivity); VAS, vertical alteration score.