Machine Learning-Based Classification and Dynamic Analysis of Tear Film Lipid Layer Using Smartphone-Based...

Area of Science:

• Ophthalmology
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• The tear film lipid layer (TFLL) plays a crucial role in maintaining ocular surface health.
• Current methods for TFLL assessment are often confined to clinical settings, limiting dynamic monitoring.
• Smartphone-based interferometry offers a potential avenue for accessible, self-acquired ocular imaging.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for classifying TFLL patterns using smartphone-based interferometer (SBI) images.
• To assess the model's capability for dynamic TFLL monitoring outside traditional clinical environments.
• To evaluate the model's accuracy in classifying TFLL patterns and estimating TFLL area.

Main Methods:

• A participant acquired tear film lipid layer (TFLL) videos using a smartphone-based interferometer (SBI).
• Images were processed, and 89,033 patches were annotated into four classes for ML model training.
• The ML model utilized Lab color and gray level co-occurrence matrix texture features, with cross-validation schemes employed.

Main Results:

• The ML model achieved an accuracy of 0.853 and a macro-F1 score of 0.755 in the all-days fold cross-validation.
• TFLL area estimates from the ML model showed strong correlation with manual measurements (r=0.969, P<0.001).
• The model successfully captured consistent postblink TFLL expansion dynamics.

Conclusions:

• The developed ML model demonstrates good accuracy in classifying TFLL patterns from self-acquired SBI images.
• The model can effectively replicate typical TFLL spreading dynamics postblink.
• This technology holds potential for dynamic monitoring in personalized eye care and home-based dry eye management.