Deep Learning-Based Eye Rubbing Detection Using Wrist-Based Wearable Devices to Enable Rigorous Study of Risk Factors...

Area of Science:

• Ophthalmology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Eye rubbing is a significant factor in the development and progression of corneal ectasias like keratoconus.
• Objective and reliable methods for quantifying eye rubbing are crucial for research and clinical practice.

Purpose of the Study:

• To develop and validate an artificial intelligence (AI)-enabled tool for detecting eye rubbing using sensor data from wrist-based wearable devices.
• To assess the performance of various deep learning models for accurate eye rubbing classification.

Main Methods:

• An automated system was designed incorporating sensor data acquisition, preprocessing, and a deep learning classification model.
• Six deep learning architectures, including 1D and 2D CNN-LSTM models and an ensemble, were evaluated.
• Two datasets (time and frequency domains) comprising 8,640 time-series recordings and 112,320 scalogram images from 20 subjects were utilized.

Main Results:

• The 1D CNN-LSTM model achieved high performance, with an F1-score of 95.27% and AUC of 98.26% during cross-validation.
• On the testing set, the system maintained strong performance with an F1-score of 92.54% and AUC of 96.70%.
• The model demonstrated real-time capabilities, requiring only 15.32 milliseconds for inference per segment.

Conclusions:

• The AI-powered system reliably detects eye rubbing behaviors.
• This tool has significant potential to assist ophthalmologists and researchers in studying the role of eye rubbing in keratoconus and other corneal diseases.