Feasibility of Electroencephalography-Based Detection of Single-Flash Microperimetry Stimuli: A Proof-of-Concept Study

This proof-of-concept study demonstrates the feasibility of using a BiLSTM deep learning model to detect single-flash microperimetry stimuli from occipital EEG signals in healthy participants, achieving up to 80% accuracy even without hardware-level synchronization.

The Gist

Imagine you are trying to listen to a friend whisper a secret in a very noisy, crowded room. Usually, to hear them clearly, you might ask them to shout the same secret over and over again so you can filter out the noise. This is how traditional brain scanning for vision works: it flashes a light many times and averages the brain's reaction to find a clear signal.

But what if your friend can only whisper the secret once, and they can't shout? And what if you don't have a perfect microphone to record the exact moment they spoke?

That is the challenge this study tackled. Here is a simple breakdown of what the researchers did, using everyday analogies.

The Problem: The "Subjective" Eye Test

Doctors use a tool called Microperimetry to check how well the back of your eye (the retina) sees things. It's like a game where a tiny light flashes on a screen, and you have to press a button when you see it.

• The Catch: This game relies entirely on you telling the doctor, "I saw it!"
• The Issue: If you are tired, distracted, elderly, a child, or have trouble communicating, you might miss the light or press the button by mistake. The doctor then doesn't know if you actually have a blind spot or if you just weren't paying attention.

The Solution: Listening to the Brain's "Whisper"

The researchers asked: Can we listen to the brain directly to know if the light was seen, without asking the patient to press a button?

They used an EEG headset (a cap with sensors on the head) to listen to the electrical "whispers" of the visual cortex (the part of the brain that sees).

The Hurdles: A Noisy Room and a Blurry Clock

The researchers faced two big problems:

1. The "One-Shot" Rule: Unlike standard tests that flash lights rapidly, this medical device flashes a light for a longer time (200 milliseconds) and only once per spot. It's like trying to hear a single clap in a thunderstorm. The signal is very weak.
2. The "Blurred Clock": The medical eye device and the brain scanner didn't have a perfect digital handshake to sync up. The researchers had to manually match the video of the light flashing with the brain data later. It was like trying to sync two watches by looking at them through a foggy window; they knew the times were roughly the same, but not exact.

The Magic Tool: The "Time-Traveling Detective" (BiLSTM)

To solve this, they didn't use standard math. They used a type of Artificial Intelligence (Deep Learning) called a BiLSTM.

Think of this AI as a Time-Traveling Detective:

• Standard AI looks at a moment in time and says, "Is there a signal right now?"
• The Time-Traveling Detective looks at the moment before, the moment during, and the moment after. It understands that a brain signal isn't just a spike; it's a story that unfolds over time.
• Because the brain's reaction to a single flash is messy and slow, this detective looks at the whole "story" of the 600 milliseconds surrounding the flash to decide: Did the brain react to the light, or was it just random noise?

The Results: A Promising First Step

They tested this on two healthy people with normal vision.

• The Bright Lights: When the flashes were bright (like a flashlight in a dark room), the AI could tell if the brain saw the light about 80% of the time.
• The Dim Lights: When the flashes were dim (like a candle in a windy room), it was much harder. The AI got confused more often, but it still found some signal.
• The Best Spot: They found that sensors placed right at the back of the head (the "Occipital" sensors) worked best, just like placing a microphone right next to the speaker.

Why This Matters

This study is a Proof-of-Concept. It's like building a prototype car that runs on a bumpy dirt road. It's not ready to drive on the highway yet, but it proves the engine works.

The Future Vision:
Imagine a future where an eye doctor can test a child with autism or an elderly patient with dementia. Instead of asking, "Did you see the light?", the doctor puts on a headset, and the computer says, "Yes, the brain registered the light at this spot."

• This removes the need for the patient to be perfect.
• It makes the test objective (based on facts, not feelings).
• It could help detect diseases earlier and more accurately.

The Bottom Line

The researchers showed that even with imperfect equipment and a "noisy" environment, AI can listen to the brain's reaction to a single flash of light. While it's not ready for the doctor's office tomorrow, it opens the door to a future where eye exams are fairer, more accurate, and don't rely on the patient's ability to press a button.

Technical Summary

1. Problem Statement

Microperimetry (MP) is a critical clinical tool for mapping retinal sensitivity, particularly for macular diseases like Age-Related Macular Degeneration (AMD) and Diabetic Macular Edema (DME). However, MP relies heavily on subjective patient responses (button presses), which introduces significant limitations:

• Reliability Issues: Test results are compromised by patient fatigue, inattention, cognitive decline, or lack of cooperation, especially in pediatric, elderly, or cognitively impaired populations.
• Lack of Objective Metrics: There is currently no robust, objective method to verify if a stimulus was actually perceived by the visual cortex during standard MP testing without relying on patient feedback.
• Technical Constraints: Existing objective perimetry methods (e.g., mfVEP, mfPOP) often require specialized hardware, proprietary stimulus designs, or extensive signal averaging, making them incompatible with standard clinical MP workflows.
• The Specific Challenge: Standard clinical MP devices (like the Nidek-MP1) use single-flash, long-duration stimuli (200 ms) rather than the rapid flickering or patterned stimuli used in traditional Visual Evoked Potential (VEP) testing. This makes detecting neural responses in single trials difficult due to low Signal-to-Noise Ratio (SNR) and the absence of hardware-level synchronization between the stimulus and EEG recording.

2. Methodology

This proof-of-concept study aimed to determine if deep learning could detect cortical responses to individual MP stimuli using standard EEG under non-ideal conditions.

• Participants & Data Acquisition:

  • Subjects: 2 healthy participants (12 trials total).
  • Stimuli: Single-flash stimuli presented via a Nidek-MP1 device at two intensity levels:
    • High Intensity: 0 dB (127 cd/m²) or 4 dB (50.6 cd/m²).
    • Low Intensity: 16 dB (3.2 cd/m²).
  • EEG Hardware: OpenBCI 8-channel ultracortex headset (sampling rate: 250 Hz). Electrodes placed over parietal (P3, P4), parieto-occipital (PO3, PO4, POz), and occipital (O1, O2, Oz) regions.
  • Synchronization: No hardware trigger was used. Synchronization was achieved offline by recording MP screen videos and manually aligning stimulus timestamps with EEG data using DataVyu software. This introduced an estimated temporal uncertainty of ~250 ms.

• Signal Processing:

  • Preprocessing: Common Median Reference (CMR) re-referencing to remove global noise.
  • Filtering: Band-pass Butterworth filter (4–49 Hz) to remove slow drifts and muscle/electrical artifacts.
  • Normalization: Z-score normalization applied to all channels.
  • Segmentation: Data was segmented into 600-ms windows.
    • Stimulus segments: 600 ms ending at the stimulus offset.
    • Non-stimulus segments: 600 ms windows from periods between stimuli.
    • Rationale: The 600 ms window accounts for the 200 ms stimulus duration plus potential neural latency and synchronization lag, avoiding post-perceptual motor processes.

• Deep Learning Model:

  • Architecture: A stacked Bidirectional Long Short-Term Memory (BiLSTM) network.
  • Configuration: Two sequential BiLSTM layers followed by dropout layers to prevent overfitting.
  • Objective: Binary classification of EEG segments as "Stimulus-Present" vs. "Stimulus-Absent."
  • Training Strategy: Data was split within each trial (70% training, 15% validation, 15% testing) using an ADAM optimizer.

3. Key Contributions

• Novel Paradigm: First demonstration of detecting single-flash, long-duration MP stimuli using EEG without hardware synchronization or trial averaging.
• Deep Learning Application: Application of BiLSTM models to decode non-repetitive, variable-intensity visual stimuli where traditional peak-detection (e.g., P100) fails.
• Hardware Feasibility: Validation that low-cost, portable 8-channel EEG systems can capture relevant visual cortex activity, challenging the notion that dense arrays are strictly necessary for this application.
• Objective Registration: Proposes a workflow for patient-independent stimulus registration, potentially allowing for "catch trials" or blind-spot testing without diverting patient attention.

4. Results

• Classification Performance:
  • High-Intensity Stimuli: Detection accuracy approached 80% (range: 70.97% – 80.00%). Specificity was consistently high (>80%), indicating the model reliably identified non-stimulus periods.
  • Low-Intensity Stimuli: Performance was more variable and lower (accuracy range: 58% – 79%), with significant inter-trial variability. Sensitivity dropped significantly in some trials (as low as 7%), highlighting the difficulty of detecting faint neural signals.
• Electrode Configuration:
  • Occipital Dominance: The O1 and O2 (lateral occipital) electrode pair consistently outperformed other combinations, aligning with the anatomy of the primary visual cortex.
  • Oz Limitations: Adding the central Oz electrode often degraded performance, likely due to susceptibility to eye/muscle artifacts.
  • Parietal Channels: In some outlier cases, parietal channels or broader montages performed better, suggesting potential involvement of attentional networks, but occipital pairs remained the most robust on average.
• Model Behavior: The model exhibited a trade-off between sensitivity and precision. High specificity was observed across trials, but sensitivity varied, likely due to class imbalance (more non-stimulus data) and low SNR.

5. Significance and Future Directions

• Clinical Implication: This study suggests that EEG could serve as an objective adjunct to subjective MP, particularly for populations unable to provide reliable responses (e.g., children, elderly, cognitively impaired). It could reduce fixation losses and improve test reliability by automating stimulus verification.
• Limitations:
  • Sample Size: Only two participants; results are not generalizable to the broader population.
  • Synchronization: The ~250 ms timing error from offline synchronization limits the precision of neural response mapping.
  • Generalization: The model was trained and tested within the same trial, potentially leading to optimistic performance estimates due to shared noise characteristics.
• Future Work:
  • Implementation of hardware-level synchronization (e.g., photodiodes) to reduce timing uncertainty to the millisecond range.
  • Expansion to larger cohorts and diverse patient populations (e.g., AMD, DME).
  • Exploration of hybrid architectures (CNN-LSTM) to better capture spatiotemporal features.
  • Development of real-time classification systems for clinical integration.

Conclusion: The study successfully demonstrates the feasibility of using deep learning to detect single-flash microperimetry stimuli via occipital EEG, even under constrained acquisition conditions. While not yet ready for clinical deployment, it establishes a foundational pathway for objective, patient-independent functional vision testing.