Virtual Readability Lab (VRL)

• Virtual Readability Lab is a VR system that provides standardized, automated diagnostic tasks mimicking traditional neuropsychological reading assessments.
• It integrates consumer-grade HMDs, custom scene design, and ML pipelines to capture detailed metrics like reading speed, accuracy, and psychometric responses.
• The platform employs rigorous statistical analyses and adaptable cross-linguistic protocols, establishing it as a benchmark for immersive language and cognitive research.

A Virtual Readability Lab (VRL) is an integrated virtual reality (VR) system for precise measurement and analysis of reading performance, cognitive traits, and accessibility in immersive environments. Utilizing consumer head-mounted displays (HMDs), custom scene design, and ML pipelines, the VRL provides standardized, automated frameworks for deploying validated text-based diagnostic tasks (e.g., dyslexia assessment) and collecting rich interaction data including timing, accuracy, and psychometric responses. The VRL formalizes the stimulus presentation, gaze/navigation affordances, metric extraction, and statistical/ML validation methodologies required for research-grade investigations into language processing, cognitive workload, and accessibility in VR contexts (Materazzini et al., 2 Sep 2025).

1. VR Environment Architecture and Hardware

VRL implementations employ commodity HMDs with 6-degree-of-freedom (DOF) head tracking. For example, Materazzini et al. (2025) utilized Meta Quest 2 hardware with “Out of the Box” software, providing all interaction via head orientation (“gaze”) and a single controller button—free locomotion and hand-gesture selection are disabled. The VR environment closely models physical neuropsychological reading tests (specifically the BDA 16–30 adult dyslexia battery), presenting a “reading wall” with a customizable text panel (white background, black text, ~1.5 m viewing distance), and a row of three virtual buttons below for task responses. Ambient, uniform scene lighting is utilized with the field of view matching the HMD’s native (Quest 2, ~100° horizontal) optics. Text size and font are live-adjustable to accommodate accessibility requirements, and a contextual “guide” character delivers instructions. Environmental noise is recorded as a Boolean flag (Materazzini et al., 2 Sep 2025).

Component	Implementation Details	System Example
HMD	6-DOF, head-tracked, wrist controller (1 button)	Meta Quest 2
Software / Engine	Unity 3D, validated test stimulus	“Out of the Box”
Text-Panel	Adjustable font/size, 1.5m distance, high contrast	BDA 16–30 passages

2. Experimental Tasks, Navigation, and Data Capture

The VRL exposes subjects to Silent Reading (SR) tasks comprised of nine standardized text passages from the BDA 16–30 battery (translated/localized as required). Stimulus layout employs multi-line sans-serif text (Arial, adjustable) on a white panel. The gaze-based pointer—a central white dot—acts as the selection cursor; target alignment is confirmed by a single controller button. For oral comprehension sub-items, a speech-to-text module prompts repetition, triggering audio capture upon successful detection. Translational user movement is disabled; all scene navigation is via head rotation.

Reading performance is timestamped for each SR sub-task ($t_{SR,i}$), with global SR duration $T_{SR} = \sum_{i=1}^9 t_{SR,i}$. Accuracy is binary per item ($a_{SR,i} \in \{0,1\}$), and overall error count $E_{SR}=9-\sum_{i=1}^9 a_{SR,i}$. The VRL concurrently administers a self-esteem instrument (Rosenberg RSES, 10 items, Likert 1–4; global score $S_{RSES} = \sum_{j=1}^{10} r_j$). Self-esteem is categorized (high: $30 \leq S_{RSES} \leq 40$, medium: $26 \leq S_{RSES} \leq 29$, low: $S_{RSES} < 25$), and global RSES response time $T_{RSES}$ is recorded.

3. Statistical and Machine Learning Analysis

Statistical analysis in VRL-based protocols adheres to an a priori significance threshold $\alpha=0.05$. For interval metrics ($T_{SR}, T_{RSES}$), an independent-samples t-test is executed:

$$t = \frac{\bar x_1 - \bar x_2}{\sqrt{s_1^2/n_1 + s_2^2/n_2}}$$

with Welch's degrees of freedom. For non-parametric/binary measures ($E_{SR}, S_{RSES}$), the Mann–Whitney U test is applied:

$U = n_1 n_2 + \frac{n_1(n_1 + 1)}{2} - R_1$

Detection of statistically significant differences in reading speed between dyslexic and non-dyslexic subjects ($p < 0.001$ in Italian, pooled samples) is observed, while accuracy ($E_{SR}$) and self-esteem ($S_{RSES}$) differences are not significant ($p > 0.05$) (Materazzini et al., 2 Sep 2025).

The ML pipeline employs 41 raw features, ultimately reduced via correlation analysis to four predictors: $E_{SR}, T_{SR}, T_{RSES}, S_{RSES}$. Preprocessing includes removal of missing values and default scaling where required (e.g., SVM “scale” gamma). The classifiers include Logistic Regression (L1/L2, liblinear or L-BFGS solvers), SVM (linear, degree-2 polynomial, RBF kernels; gamma=“scale”/“auto”), k-NN ($k = 3, 5, 7$), Decision Trees (“gini”/“entropy” criteria, max_depth = 10, 15), and Random Forest (n_estimators = 10–40, same criteria and depth as DT). Train/test split is stratified 80/20. Grid-search and cross-validation (implicit 5-fold) determine optimal hyperparameters.

Performance is reported via standard ACC and $F_1$:

$$\mathrm{ACC} = \frac{TP+TN}{TP+TN+FP+FN}, \quad F_1 = 2\frac{\mathrm{precision}\times \mathrm{recall}}{\mathrm{precision}+\mathrm{recall}}$$

The Italian VRL sample achieves SVM-RBF ACC = 87.5%, $F_1 = 85.7\%$; Spanish RF-gini yields ACC = 66.6%, $F_1 = 66.6\%$; pooled RF-entropy reaches ACC = 75.0%, $F_1 = 71.4\%$ (Materazzini et al., 2 Sep 2025).

4. Cross-Linguistic and Task Adaptation

The VRL paradigm enforces identical interaction paradigms, feature specifications, and data collection schemas across languages by translating the original clinical passages (BDA 16–30) to the target language (here, Italian and Spanish). This enables systematic study of orthographic transparency, text complexity, and linguistic factors influencing reading metrics and dyslexia classifier performance. Notably, VRL results in high discriminatory accuracy for Italian (transparent orthography), with degradation for Spanish (less transparent, lower statistical power for RSES time).

Recommendations for adapting VRL include increasing cohort size (for Spanish), developing parallel passages for repeated measures, integrating language/orthographic features directly into the ML covariate set, and expanding VR-derived feature sets to include low-level behaviors (head-movement variance, gaze-dwell time, button-hold) and potential addition of HMD eye-tracking for fine-grained oculomotor metrics (Materazzini et al., 2 Sep 2025).

5. Comparison with Other VR Reading Frameworks

VRL is distinct from generic VR reading interfaces as its primary goal is the clinical-grade, quantitative evaluation of reading and psycholinguistic performance. In contrast, systems such as VRDoc focus on optimizing reading experience for knowledge workers via gaze-driven UI primitives—Gaze Select-and-Snap for document targeting and positioning (cone angle $\varphi_{\text{thresh}}=2^\circ$, dwell $t_d=0.5$ s), Gaze MagGlass for dynamic text magnification (radius $R=3^\circ$, $M_{max}=1.5\times$, parabolic profile), and Gaze Scroll for navigation with exponential velocity scaling (Lee et al., 2022). While VRDoc demonstrates significant reductions in workload (NASA-TLX, SUS improvements, p<0.01) and increased reading efficiency (~11% total completion time, p=0.015), it does not implement clinical or diagnostic metric capture, nor does it utilize machine-learning-based classification or standardized psychometric assessment as in the VRL.

System	Goal	Key Features	ML/Clinical Metrics
VRL	Diagnosis/Study	Standardized tasks, analytics	Full ML/statistics
VRDoc	Usability	Gaze control, magnification, nav	None

6. Implications and Refinement Strategies

Empirical application of the VRL demonstrates that VR-derived reading speed ($T_{SR}$) is a robust indicator for dyslexia in transparent orthographies (Italian) but is less sensitive in lower-powered or more opaque samples (Spanish, $p=0.063$ for speed). No VR-derived differences are observed in error counts ($E_{SR}$) or self-esteem ($S_{RSES}$). For future extensions, the VRL blueprint recommends increasing participant numbers for underpowered subgroups, developing parallel text sets for test–retest reliability, and explicitly modeling orthographic complexity in ML. The integration of additional VR metrics (e.g., head or gaze dynamics) and physiologic monitoring (e.g., eye-tracking) promises to enhance discriminatory power and enable finer-grained analysis of reading behaviors and cognitive effort (Materazzini et al., 2 Sep 2025).

A plausible implication is that VRL can serve as a reference implementation for other sensory or cognitive testing batteries in VR, provided each test’s structure, environmental controls, and scoring systems are equivalently mapped to the virtual domain. Expanding on the VRL framework may also facilitate research on accessibility, usability, and the neurocognitive correlates of reading in immersive settings.