Table 2.
Eye-movement features used for dyslexia detection and reading behavior analysis.
| Feature | Description | Studies reporting the feature (author-year) | Importance for dyslexia detection | Application in dyslexia research |
|---|---|---|---|---|
| Fixation metrics | ||||
| Fixation duration | Time (ms) the eyes remain stationary on a word/region (e.g., mean fixation, first-fixation duration, single-fixation duration) | Gran Ekstrand 2021; Rossier-Bisaillon 2025 (adults); Rossier-Bisaillon 2026 (moving-window children); Scaltritti 2019; Holmqvist Olander 2017; Le 2023; Park 2024; Raatikainen 2021; Appadurai 2021; Benfatto 2016; Pereira 2024 (SFD); Shalileh 2023; Vaitheeshwari 2024; Rello 2015; Vajs 2022; Svaricek 2025; Virlet 2024 (FFD); Peters 2021 (19–22, 25, 27, 31–33, 41–45, 48–51, 54) |
Longer/variable fixations reflect higher decoding/lexical processing load and slower reading | Used as core discriminative feature (ML screening), typographic/condition effects, intervention responsiveness (pre-post changes) |
| Fixation count | Number of fixations (overall, per AOI/word, or frequency) | Scaltritti 2019; Holmqvist Olander 2017; Le 2023; Park 2024; Raatikainen 2021; Appadurai 2021; Pereira 2024 (FC); Shalileh 2023 (FC); Vaitheeshwari 2024; Rello 2015; Bhargavi 2020; Vajs 2022; Svaricek 2025 (baseline); Peters 2021 (20–22, 25, 27, 31–33, 44, 48–51, 54) |
Higher fixation counts typically indicate inefficient decoding, rereading, or unstable attention allocation | Used for group comparisons (DD vs. controls), predictive models, and task/format optimization |
| Gaze duration | Sum of consecutive first-pass fixations on a word (often close to first-pass reading time/dwell time) | Rossier-Bisaillon 2025 (adults); Holmqvist Olander 2017 (dwell time); Pereira 2024 (FPRT/TFT); Virlet 2024 (GD); Vaitheeshwari 2024; Rello 2015 (reading time/visits); Shalileh 2023 (word-level totals); Bhargavi 2020 (fixation gaze duration) (25, 31–33, 42, 45, 49, 54) | Captures sustained processing time on a word; sensitive to lexical difficulty and decoding deficits | Used to quantify slowed lexical access and to show intervention effects (GD ↓ after effective training) |
| Saccadic movements | ||||
| Saccade length/amplitude | Distance moved between fixations (letters/pixels/degrees); shorter forward saccades often mean inefficient reading | Gran Ekstrand 2021; Rossier-Bisaillon 2025 (adults); Rossier-Bisaillon 2026 (moving-window children); Scaltritti 2019; Park 2024; Raatikainen 2021; Appadurai 2021; Benfatto 2016; Shalileh 2023 (first saccade amplitude); Vaitheeshwari 2024; Virlet 2024 (SA) (19, 20, 22, 27, 31, 33, 41–45) |
Shorter saccades + unstable forward progression often accompany decoding difficulty | Used in ML models, perceptual-span experiments, and intervention evaluation (SA ↑ toward typical after effective therapy) |
| Regression rate | Proportion/count of backward eye movements or re-reading; sometimes operationalized via second-pass/“go-past” measures | Gran Ekstrand 2021; Rossier-Bisaillon 2025 (adults; regressive fixation probability); Park 2024 (regression ratio); Appadurai 2021 (regression summaries); Benfatto 2016 (regressive features); Pereira 2024 (SPRT/s-pass time); Shalileh 2023 (regression in/out, go-past/paths); İleri 2025 (EOG-derived regressions); Bhargavi 2020; Latifoglu 2021 (re-reading detection); Peters 2021 (regression count in RAN) (22, 25, 27, 31, 32, 42, 43, 46) |
Regressions index comprehension/decoding breakdown and rereading strategies; frequently reported in dyslexic reading behavior | Used as diagnostic input features, to characterize reading behavior, and as EOG-based movement targets in AI pipelines |
| Gaze pattern analysis | ||||
| Scan path/gaze trajectory | Spatial sequence of fixations/saccades; often visualized as scanpaths/heatmaps or summarized as scanpath length/complexity | Holmqvist Olander 2017 (scanpaths); Le 2023 (scanpath + heatmap); Appadurai 2021 (scanpath length + scanpath images); Bhargavi 2020 (scanpath metrics); Vaitheeshwari 2024 (trajectory maps/saliency) (25, 27, 33, 49, 50) |
Captures global reading strategy differences (looping, skipping, unstable line tracking) | Used for qualitative discrimination, feature engineering (scanpath length/density), and deep learning via trajectory/saliency images |
| Gaze transition metrics | AOI-to-AOI transitions (e.g., text↔picture switching) or transition matrices between regions/sentences | Holmqvist Olander 2017 (text↔picture transitions); Raatikainen 2021 (AOI transition matrices) (20, 49) | Detects attentional control/strategy (integration vs. distraction; search behavior) | Used to quantify multimodal integration (illustrations) and to build ML features from AOI transition matrices |
| Fixation fractal dimension | Complexity/irregularity of gaze trajectory (often computed from fixation traces) | Vajs 2022 (21) | Dyslexic reading tends to show higher spatial/temporal complexity and less efficient patterns | Used as an interpretable “gaze complexity” marker and ML input |
| Fixation intersection coefficient | Counts/normalizes self-intersections of fixation gaze lines (spatial looping complexity) | Vajs 2022 (intersection coefficient); Vajs 2023 (self-intersection events as triggers) (21, 28) |
Captures rereading-like loops and spatial instability | Used as engineered interpretable features; can be computed without full fixation parsing (raw x-y) |
| Vertical alteration score (VAS) | Index of vertical instability (y-axis direction changes) around trigger events during reading | Vajs 2023 (28) | Sensitive to line-tracking instability and dysfluent scanning | Used as a single interpretable real-time screening feature (works even at lower sampling rates) |
| Physiological eye signals | ||||
| Blink rate/blink events | Blink count or rate during reading (often derived from EOG or eye-event streams) | İleri 2025 (blink rate); Bhargavi 2020 (blink events in pipeline) (25, 46) |
Proxy for cognitive load/visual strain; sometimes differs in DD | Used as auxiliary indicator and to compare display/typeface effects; sometimes included in engineered feature sets |
| Eye-movement signal energy (EOG) | Signal-level energy/effort measure from EOG channels (e.g., integrated squared amplitude) | İleri 2025 (46) | Reflects intensity/effort of eye movement behavior during reading | Used to quantify cognitive load and as supportive biomarker alongside deep learning scalogram inputs |
AOI, area of interest; DD, developmental dyslexia; EOG, electrooculography; ET, eye tracking; FC, fixation count; FFD, first fixation duration; FPRT, first-pass reading time; GD, gaze duration; RAN, rapid automatized naming; SA, saccade amplitude; SFD, single fixation duration; SPRT, second-pass reading time (re-reading time); TFT, total fixation time; VAS, vertical alteration score [feature in Vajs et al. (28)].