Orchestrating Attention: Bringing Harmony to the 'Chaos' of Neurodivergent Learning States

Adaptive learning systems optimize content delivery based on performance metrics but ignore the dynamic attention fluctuations that characterize neurodivergent learners. We present AttentionGuard, a framework that detects engagement-attention states from privacy-preserving behavioral signals and adapts interface elements accordingly. Our approach models four attention states derived from ADHD phenomenology and implements five novel UI adaptation patterns including bi-directional scaffolding that responds to both understimulation and overstimulation. We validate our detection model on the OULAD dataset, achieving 87.3% classification accuracy, and demonstrate correlation with clinical ADHD profiles through cross-validation on the HYPERAKTIV dataset. A Wizard-of-Oz study with 11 adults showing ADHD characteristics found significantly reduced cognitive load in the adaptive condition (NASA-TLX: 47.2 vs 62.8, Cohen’s d=1.21, p=0.008) and improved comprehension (78.4% vs 61.2%, p=0.009). Concordance analysis showed 84% agreement between wizard decisions and automated classifier predictions, supporting deployment feasibility. The system is presented as an interactive demo where observers can inspect detected attention states, observe real-time UI adaptations, and compare automated decisions with human-in-the-loop overrides. We contribute empirically validated UI patterns for attention-adaptive interfaces and evidence that behavioral attention detection can meaningfully support neurodivergent learning experiences.

💡 Research Summary

The paper introduces AttentionGuard, an adaptive user‑interface framework designed to support neurodivergent learners, especially adults with ADHD, by detecting real‑time attention states from privacy‑preserving behavioral signals and dynamically adjusting UI elements. The authors argue that current adaptive learning systems treat attention as a static or binary variable, ignoring the rapid fluctuations characteristic of ADHD (hyperfocus, drifting, understimulation, fatigue). To fill this gap they pose three research questions: (1) Can attention states be reliably detected from non‑intrusive interaction data and linked to clinical ADHD profiles? (2) Do attention‑adaptive UI patterns reduce cognitive load compared to static designs? (3) How do users perceive these real‑time adaptations?

Detection model – The system collects click rhythm, scroll velocity and reversals, mouse‑movement entropy, answer latency, revision frequency, tab visibility, focus/blur events, and back‑track frequency. Signals are aggregated over 30‑second sliding windows and compared to a personalized baseline established during a five‑minute calibration phase. A Random Forest classifier with balanced class weights predicts one of four states derived from ADHD phenomenology: Focused, Drifting, Hyperfocused, and Fatigued. Using the Open University Learning Analytics Dataset (OULAD, >32 k students, >10 M interactions) the model achieves 87.3 % overall accuracy, macro‑F1 = 0.84, AUC = 0.91, with per‑class F1 scores ranging from 0.78 to 0.89. Feature importance highlights click‑rate deviation, session‑duration ratio, and back‑track frequency. Cross‑dataset validation on the clinical HYPERAKTIV dataset (51 diagnosed ADHD, 52 controls) shows that the same behavioral features predict ADHD diagnosis with AUC = 0.81 and correlate with ASRS self‑reports (r = 0.47, p < .001), supporting construct validity.

Adaptive UI patterns – Based on the detected state, five UI adaptation patterns are applied:

1. Attention‑Responsive Chunking – dynamically changes content granularity (micro‑chunks during drifting, standard paragraphs while focused, extended sections during hyperfocus, review‑mode during fatigue).
2. State‑Aware Verification Timing – schedules comprehension checks according to cognitive availability (immediate during drifting, lightweight during focus, deferred to natural breakpoints during hyperfocus, no new checks during fatigue).
3. Bi‑Directional Scaffolding – reduces visual complexity and adds whitespace when overstimulated, injects novelty, curiosity hooks, and gamified elements when understimulated.
4. RSD‑Safe Feedback Rendering – mitigates rejection‑sensitive dysphoria by using neutral colors, constructive framing, distance‑based progress bars, and variable‑ratio micro‑rewards.
5. Temporal Landmark Navigation – replaces countdown timers with spatial journey metaphors, providing landmarks that make the passage of time tangible for users with impaired time perception.

All adaptations are visible, reversible, and can be disabled by the user, ensuring interface‑level transparency and preserving agency. An observer mode displays the inferred attention state, contributing signals, and the rationale for each adaptation, enabling real‑time auditing.

Evaluation – Two studies were conducted. Study 1 validates the detection model on OULAD and HYPERAKTIV, confirming high accuracy and clinical relevance. Study 2 is a Wizard‑of‑Oz pilot with 11 adult participants who self‑report ADHD or score ≥4 on the ASRS. Each participant completes two counterbalanced 25‑minute sessions: an adaptive condition (wizard triggers the five patterns) and a baseline static condition (fixed chunking and uniform verification). Cognitive load is measured with NASA‑TLX, comprehension with multiple‑choice questions, and task completion recorded. Results show a significant reduction in cognitive load (NASA‑TLX 47.2 vs 62.8, p = .008, Cohen’s d = 1.21) and higher comprehension (78.4 % vs 61.2 %, p = .009, d = 1.18). All participants finish the adaptive sessions, whereas two abandon the baseline. Concordance analysis reveals 84 % exact match (κ = 0.71) between wizard decisions and the automated classifier, indicating that the model’s predictions are reliable enough for deployment.

Contributions and limitations – The paper contributes (1) a validated, privacy‑preserving attention‑state classifier, (2) five neuroscience‑grounded UI patterns tailored to ADHD attention dynamics, and (3) empirical evidence that such adaptations lower cognitive load and improve learning outcomes. It also foregrounds user agency by making AI‑driven adaptations transparent and contestable. Limitations include reliance on online learning interaction data (generalization to mobile or offline contexts remains untested), a small pilot sample, and the fact that behavioral detection is not a clinical diagnostic tool. Future work should involve larger, more diverse cohorts, long‑term field studies, and multimodal fusion with physiological signals (EEG, PPG) to further boost detection fidelity.

Overall, AttentionGuard demonstrates that real‑time, privacy‑preserving attention detection combined with thoughtfully designed UI adaptations can meaningfully support neurodivergent learners, offering a promising direction for inclusive educational technology.