Anticipation Before Action: EEG-Based Implicit Intent Detection for Adaptive Gaze Interaction in Mixed Reality

Paper Title

Anticipation Before Action: EEG-Based Implicit Intent Detection for Adaptive Gaze Interaction in Mixed Reality

Publication Info

• Topic area: EEG-based intent detection for improving gaze interaction in Mixed Reality (MR).
• Keywords: EEG, SPN, Mixed Reality, gaze interaction, intent detection, deep learning, adaptive interfaces, Midas Touch problem, anticipatory uncertainty, neural decoding.

Background and Problem

• Problem / challenge: Mixed Reality interfaces struggle to distinguish between visual attention and intentional interaction, leading to the Midas Touch problem where unintended commands are triggered by gaze.
• Significance: Resolving this issue is critical for improving usability, reliability, and user experience in MR systems, enabling natural and intuitive interaction.
• Motivation and related work: Prior solutions rely on explicit confirmation signals (e.g., gestures, speech) or multimodal approaches, which can be fatiguing or cognitively demanding. Brain-computer interfaces (BCIs) have shown promise in detecting intent through EEG signals, particularly the Stimulus-Preceding Negativity (SPN), but previous studies were limited to controlled environments and did not explore SPN modulation by intention and feedback in realistic MR tasks.

Solution

• Proposed approach: Use SPN as an implicit neural marker of anticipatory uncertainty and intent during gaze-based MR interactions, combined with deep learning models for intent classification.
• Novelty:
  1. Demonstration of SPN as a robust marker of anticipatory processing in ecologically valid MR tasks.
  2. Refinement of SPN interpretation as reflecting anticipatory uncertainty rather than motor preparation.
  3. Successful decoding of user intention (Select vs. Observe) using deep learning models with accuracies up to 97% in person-dependent setups.
  4. Implications for adaptive MR design, including dynamic dwell-time adjustment, uncertainty-contingent confirmation, and personalized interaction profiles.
• Procedure and key techniques:
  • Conducted a 2 × 2 factorial experiment manipulating Intention (Select vs. Observe) and Feedback (With vs. Without).
  • Recorded EEG and eye-tracking data from 28 participants across three MR scenarios (App Launcher, Document Editor, Video Player).
  • Analyzed SPN amplitude using ERP methods and linear mixed models.
  • Applied deep learning models (e.g., EEGInceptionERP, EEGResNet) for person-dependent and person-independent intent classification.
  • Used LIME for interpretability analysis of neural network decisions.

Results

• Concrete findings:
  • SPN amplitude was strongest in Observe–No Feedback conditions, indicating heightened anticipatory uncertainty.
  • Intention and feedback interacted significantly: feedback reduced SPN during observation but had minimal impact during selection.
  • Person-dependent classification achieved accuracies ranging from 75% to 97%, with EEGInceptionERP performing best (mean accuracy 78.4%).
  • Person-independent classification achieved lower accuracy (best model: Deep4Net, mean accuracy 69.2%).
• Advantage over baselines:
  • SPN-based approach avoids reliance on explicit confirmation signals, reducing cognitive and physical fatigue.
  • Deep learning models demonstrated robust classification performance, particularly in person-dependent setups.
• Experiments / evaluation:
  • EEG recorded from 64 electrodes, analyzed for SPN amplitude in the -750 to 0 ms window.
  • Tasks included realistic MR scenarios with gaze-based interaction.
  • Deep learning models evaluated for intent classification using both person-dependent and person-independent datasets.
• Limitations and future work:
  • Did not test responses to unexpected feedback or error-related potentials.
  • Ambiguity between anticipatory uncertainty and intent needs further exploration.
  • Real-time implementation and hybrid person-dependent/independent models require validation.
  • Future work should integrate SPN analysis with multimodal measures (e.g., pupil size, electrodermal activity) and extend to multi-step MR tasks.

Summary

This study demonstrates that SPN is a reliable marker of anticipatory uncertainty during gaze-based MR interactions, modulated by intention and feedback. It refines SPN interpretation as reflecting uncertainty rather than motor preparation and shows that user intention can be decoded from EEG signals using deep learning, achieving up to 97% accuracy in person-dependent setups. These findings have direct implications for adaptive MR interface design, enabling dynamic dwell-time adjustment, uncertainty-sensitive confirmation strategies, and personalized interaction profiles. Future work should focus on real-time implementation, multimodal integration, and hybrid classification approaches for broader applicability.