Understanding How Low Vision People Read Using Eye Tracking

The paper investigates whether and how commercial eye trackers can be used with low vision users to collect high-quality gaze data and to understand their detailed gaze behaviors during reading. Low vision encompasses heterogeneous conditions (e.g., central/peripheral vision loss, blurry vision) that lead to reading difficulties such as reduced visual span, line-switching challenges, and slower speed. While screen magnifiers and accessibility features are commonly used, they introduce new challenges (e.g., reduced field of view, higher cognitive load). The advance and ubiquity of eye tracking may enable fine-grained detection of gaze behaviors (e.g., line deviations, return sweeps) and support targeted, gaze-based assistive technologies. However, standard eye-tracker calibration and data collection procedures often fail for low vision due to visibility, eye appearance, and usage differences. The study aims to close this gap by designing an accessible calibration interface and a dominant-eye-based data collection strategy, then using them to characterize low vision gaze during reading and assess impacts of visual conditions and magnification modes.

The related work covers three areas: (1) Reading with low vision: Low vision users have varied reading performance depending on condition and severity. Screen magnifiers and enlarged text can help but reduce usable field of view, slow reading, and increase cognitive load, particularly for line switching and context maintenance. (2) Eye tracking and gaze: Modern eye trackers estimate gaze via pupil and corneal reflections and require user-dependent calibration (kappa angle). In reading, fixations and saccades encode perceptual and cognitive processing; return sweeps mediate line transitions; regressive saccades aid comprehension; microsaccades can correct long saccades. Standard calibration paradigms may be inaccessible to low vision users. (3) Eye tracking for low vision: Prior optometry/vision science studies often use high-end or custom trackers to examine PRL and reading behavior in macular degeneration, linking fixation stability and saccade patterns to performance. HCI efforts are limited; earlier work explored larger calibration targets and gaze-guided magnification, with challenges such as data loss and nystagmus. With improved consumer trackers, it is timely to revisit low vision feasibility and design gaze-based assistive technologies.

Study goal: Evaluate feasibility of collecting reliable gaze from low vision users with a commercial eye tracker using an accessible calibration and dominant-eye data collection; compare gaze patterns with sighted controls; analyze effects of visual acuity/field; evaluate magnification modes. Participants: 40 total (20 low vision: 14F/6M, age 19–86, mean 58.3±22.1; 7 legally blind; varied conditions; one wore glasses; recruited via clinic and university postings. 20 sighted: 7F/13M, age 21–51, mean 31.1±9.5; VA ≥ 20/40; some used glasses/contacts). Apparatus: Tobii Pro Fusion (120 Hz) eye tracker on a 24-inch display (1920×1200). Web interface (React) for reading tasks and adjustable calibration; Tobii Pro SDK (Python); Flask server for data communication. Accessible calibration: Modified Tobii Pro Lab 9-point calibration and 4-point validation with adjustable high-contrast target sizes (36–256 px). For low vision, targets were enlarged until visible without squinting. Repeated calibration if validation error >1° or poor compliance observed. Dominant-eye-based data collection: Used monocular (dominant eye) gaze when participants indicated dominant eye and showed asymmetric behaviors or one eye was untrackable; otherwise binocular (as for sighted). Reading tasks: Six passages (sixth-grade difficulty; mean 185±8 words) from CLEAR corpus. Reading aloud to align gaze and speech. Sighted and low vision read in regular mode; low vision could also read with lens magnifier and full-screen magnifier (Windows-like behavior). Low vision could adjust font size (up to 16×), weight, and color. Four low vision participants did not need magnification and used regular mode only. Order of modes counterbalanced via Latin Square. Practice given before data collection. Visual function tests: Visual acuity measured with ETDRS logMAR charts at 5 ft for each eye, recording lowest visible line. On-screen visual field screening (simplified Octopus): central fixation; random stimuli (white circle, 0.1 s, 1.7°, grid spacing 5°, 9×5 positions spanning ~40°×20° at 65 cm); SPACE press upon detection. Procedure: Consent; demographic and low-vision interview; acuity testing; accessible calibration and validation; reading tasks with eye tracking and audio recording; exit interview; visual field screening. Analysis: Validated calibration accuracy via 4-point validation (mean angular error). Data loss computed as % invalid gaze points. Aligned gaze and audio timestamps (VOSK) to confirm return sweeps vs spoken line endings. Eye-movement events classified with REMoDNaV. Measures: fixation count and mean duration; saccade count and normalized length (in letter-width units), regressive saccades, revisitation rate (regressive saccade following a forward saccade + fixation), line switching (mean number of searched lines), smooth pursuits (count and mean duration; for moving text/magnifier). Statistical tests: Shapiro-Wilk for normality; ANOVA or aligned rank transform (ART) ANOVA for nonparametric factorials; Tukey/ART contrasts for post-hoc; effect sizes via partial eta². Factors included Vision (Sighted vs LowVision), VisualAcuity (Low vs High; threshold 20/100 better eye), VisualField (Limited vs Intact), MagnificationMode (Regular, Lens, FullScreen), and Order (to validate counterbalancing).

H1: Eye tracker feasibility and data quality - Four-point validation accuracy: No significant difference between sighted and low vision (ART F(1,38)=0.61, p=0.44, ηp²=0.02), indicating accessible calibration is effective. - Data loss: No significant difference (ART F(1,38)=2.61, p=0.11, ηp²=0.06); means: LowVision 4.62% (SD 6.89) vs Sighted 1.35% (SD 1.22). Causes of loss included obscured pupils, moving closer than optimal working distance, and head tilt for functional fields. - Gaze–reading alignment: Return sweep times highly correlated with spoken line switches for both groups (r≈1.00, p<0.001), validating gaze data. H2: Low vision vs sighted in regular mode - Reading time: Low vision significantly slower than sighted (ART, large effect; exact F omitted in text segment but reported significant). - Fixations: More fixations (ART F(1,78)=8.30, p=0.005, ηp²=0.10) and shorter mean fixation duration (ART F(1,78)=57.8, p<0.001, ηp²=0.43) for low vision. - Saccades: More forward saccades (ART F(1,78)=4.54, p=0.04, ηp²=0.05) with shorter normalized forward saccade length (ANOVA, significant; F partially truncated but reported as significant). More regressive saccades (ART F(1,78)=5.50, p=0.02, ηp²=0.07). Revisitation rate: no significant difference (ART F(1,78)=1.48, p=0.23). - Line switching: Low vision searched more lines (ART F(1,76)=11.64, p=0.001, ηp²=0.13). They spent more fixation time in first 10% of each line: 14.87%±4.30% vs sighted 11.27%±2.34% (ART F(1,76)=26.93, p<0.001, ηp²=0.26). H3: Effects of visual abilities (regular mode) - Fixations: Low acuity → more fixations (ART F(1,36)=5.18, p=0.03, ηp²=0.13) and shorter duration (ART F(1,36)=6.01, p=0.02, ηp²=0.14). Limited visual field → more fixations (ART F(1,36)=4.69, p=0.04, ηp²=0.12) and shorter duration (ART F(1,36)=7.88, p=0.01, ηp²=0.18). Interaction: combined low acuity + field loss had shortest fixation durations (post-hoc contrasts significant vs other groups). - Saccades: Forward saccade length shorter with low acuity (ANOVA F(1,36)=12.00, p=0.001, ηp²=0.25) and with limited field (ANOVA F(1,36)=5.70, p=0.01, ηp²=0.30). Number of forward/regressive saccades and revisitation rate: no significant effects of acuity/field. - Line switching: No significant effect of field; trend for acuity (ART F(1,34)=3.62, p=0.07, ηp²=0.10). H4: Magnification mode effects (n=14 used all modes) - Reading time: Magnifier modes slower than regular increased-font (ART F(2,26)=16.45, p<0.001; Lens vs Regular t(26)=-5.31, p<0.001; FullScreen vs Regular t(26)=-4.53, p<0.001); no difference Lens vs FullScreen (p=0.71). - Fixations: Mean fixation duration shorter with Lens and FullScreen than Regular (ART F(2,55)=38.60, p<0.001; Lens>shorter than FullScreen, t(55)=-2.84, p=0.02). Fixation count: no significant effect. - Saccades: More regressive saccades with Lens and FullScreen than Regular (ART F(2,26)=6.19, p=0.01). Higher revisitation rate with Lens and FullScreen than Regular (ART F(2,26)=6.62, p=0.005). No difference between Lens and FullScreen for these two. - Window size effects (Lens): Normalized window width positively correlated with forward saccade length (r=0.72, p<0.001) and negatively with reading time (r=-0.67, p<0.001). Line switching: more searched lines with Lens than Regular (ART, post-hoc t(26)=-3.35, p=0.01). Normalized window height negatively correlated with searched lines (r=-0.60, p<0.001). - Smooth pursuit: Mode effect on count (ART F(2,55)=88.08, p<0.001, ηp²=0.76) and duration (ANOVA F(2,26)=24.38, p<0.001, ηp²=0.65). Lens and FullScreen > Regular in count and duration; Lens > FullScreen in count; similar duration between magnifiers.

Findings show that commercial eye trackers can collect reliable gaze data from low vision users when calibration and data collection are accessibly designed (adjustable target size, dominant-eye selection), enabling fine-grained analysis of reading behaviors. Low vision readers exhibit more but shorter fixations, more and shorter forward saccades, more regressions, and greater difficulty at line starts, consistent with reduced perceptual span and increased effort in line switching. Visual acuity and field loss independently and jointly reduce information per fixation and forward saccade length. Magnifier-based reading introduces additional control burdens leading to shorter fixations, more regressions/revisitations, increased smooth pursuits, and slower reading; wider/taller magnifier windows can mitigate some costs by improving context and perceptual span. Implications for eye-tracking technology: calibration should adapt to visual abilities (e.g., adjustable number/size/duration/color of targets), and systems should consider dominant-eye data when inter-ocular asymmetries exist. Eye trackers should tolerate low-vision reading postures (closer viewing, head tilt), and provide real-time feedback on data capture status and quality. Assistive design implications: provide real-time gaze-aware support for line following/switching (e.g., highlight current/next line, adaptive line spacing, reminder of key words at line transitions), detect prolonged fixations/revisitations to trigger local text-to-speech for hard words, and develop gaze-controlled, context-aware magnifiers that automatically adjust window size based on gaze state (e.g., enlargement near line switches, reduction during steady reading).

A study with 20 low vision participants and 20 sighted controls demonstrated that a commercial eye tracker, paired with an adjustable calibration interface and dominant-eye-based data collection, can capture gaze from low vision users with quality comparable to sighted controls. Analyses uncovered distinctive gaze patterns in low vision reading and showed how visual abilities and magnification modes influence fixations, saccades, line switching, and smooth pursuits. These insights motivate accessible calibration practices, improvements in eye-tracker usability for low vision, and gaze-based assistive features for reading. Future work will target specific low vision subgroups, age-matched controls, and silent-reading scenarios to extend generalizability.

- Heterogeneous low vision cohort with varied diagnoses, acuity, and field loss may reduce statistical power; some subgroups (e.g., nystagmus) remain challenging for tracking. - Age differences between low vision and sighted groups could confound comparisons; future work should use age-matched controls. - Reading aloud was used for alignment; findings may not fully generalize to silent reading, which typically has faster rates and different fixation dynamics. - Screen-based setup imposes posture and distance constraints that may conflict with natural low-vision reading habits.