TL;DR
This paper presents a novel, low-cost system for detecting eye fatigue using single-channel electrooculography (EOG) and machine learning. By extracting key EOG features and training a decision tree classifier, the system demonstrates an objective link between EOG signals and self-reported eye fatigue, offering a promising avenue for daily eye health monitoring.
What Problem Does This Paper Address?
In our increasingly screen-dependent world, eye fatigue is a pervasive issue, yet objective and effective detection methods are scarce. Current approaches often rely on subjective self-assessments, which can be unreliable, or complex multimodal systems that are inconvenient and costly. The authors aim to bridge this gap by developing an accessible and accurate system for eye fatigue detection using a single, non-invasive EOG channel.
Methods
The core of this research lies in a machine learning pipeline designed to detect eye fatigue from EOG signals.
- Signal Acquisition: A custom, single-channel EOG device was developed, featuring three electrodes (one ground, two for EOGL and EOGR) placed strategically around the eyes. The system utilizes an ADS1191 chip for high-precision sampling and a CC2640 microcontroller for data processing and Bluetooth communication.
- Signal Pre-processing: Raw EOG signals undergo filtering to remove industrial frequency interference and baseline drift. This includes a 50 Hz notch filter and a 0.2–30 Hz band-pass filter.
- Feature Extraction: After pre-processing, five key features highly correlated with eye fatigue are extracted from the EOG signals. These include:
- Slow and Fast Eye Movement Features: Derived from spectral integration of EOG signals within specific frequency ranges (0.2-0.6 Hz for slow, 1-2 Hz for fast movements), reflecting different states of drowsiness and wakefulness.
- Energy Peak Frequency Features: Identified using the Welch spectrum estimation method, capturing the dominant frequencies in the EOG signal.
- Eye Movement Peak Ratio Feature: Calculated as the ratio of right eye movements to total eye movements, indicating directional eye activity.
- Permutation Entropy: A measure of signal regularity, reflecting the complexity of EOG patterns in different states.
- Machine Learning Model: A Classification and Regression Tree (CART) classifier is employed. The CART algorithm, known for its efficiency in both classification and regression, uses the Gini coefficient for feature selection. The model is trained on a dataset where EOG features are paired with self-reported fatigue scores from an industry-standard questionnaire.
Datasets
The study utilized a dataset generated from experiments with 12 subjects.
- Data Collection: Subjects completed an industry-standard visual fatigue questionnaire (based on the China Electronic Video Industry Association standard) to provide subjective fatigue labels. Simultaneously, their EOG signals were recorded using the developed single-channel device for 3 minutes per session.
- Dataset Size: After processing (downsampling and sampling), the dataset comprised:
- Training Data: 12,090 relaxed samples and 9,750 fatigue samples.
- Test Data: 4,095 relaxed samples and 5,070 fatigue samples.
- Benchmark: The study uses an industry-standard visual fatigue scale as the ground truth for labeling, which is a notable aspect as objective benchmarks for eye fatigue are scarce.
Key Insights and Ideas
- EOG as a Biomarker for Eye Fatigue: The research strongly supports the idea that EOG signals, which directly reflect eye movements, are a reliable objective indicator of eye fatigue. This moves beyond subjective self-reports.
- Feature Engineering is Crucial: Not all EOG features are equally informative for fatigue detection. The paper highlights the importance of carefully selecting and extracting specific features (like slow/fast eye movements, energy peaks, and movement ratios) that have a strong correlation with fatigue.
- Simplicity and Accessibility: The development of a single-channel EOG system is a significant step towards making eye fatigue detection more accessible and affordable for everyday use, contrasting with more complex multimodal approaches.
- CART's Effectiveness: The CART classifier proved effective in distinguishing between relaxed and fatigued states, demonstrating its suitability for this type of binary classification task.
Key Results
The study achieved promising results in classifying eye fatigue states:
- Classification Accuracy: The system achieved an overall accuracy of 61.6% in classifying between relaxed and fatigued states.
- Fatigue Classification Performance: The system performed better in classifying the fatigue state (50.5% TPR) than the relax state (58.1% Precision), suggesting that detecting the presence of fatigue is more robust than confirming the absence of it with this model.
- Correlation Verification: The results further verify the correlation between EOG signals and eye fatigue, demonstrating the potential of this approach.
Limitations and Future Work
The authors acknowledge several limitations and propose future directions:
- Limited Sample Diversity: The current study's sample was limited in age and eye health conditions. Future work will involve a more diverse subject pool to improve the generalizability of the model.
- Accuracy Improvement: While promising, the current accuracy can be improved. The authors suggest increasing the number of test samples and exploring deep learning models for potentially higher accuracy, especially with a larger dataset.
- Relax State Classification: The system's performance in classifying the "relax" state needs improvement.
- Real-time Application: The paper lays the groundwork for daily eye fatigue detection, implying future work could focus on real-time implementation and integration into user-friendly applications.
Summary updated