Deep Learning Approach to Eye Blink Detection: Toward Enhanced Communication in Critical Care Settings

This technical feasibility study presents a CNN-based approach for automated eye blink detection as a foundation for future ICU patient communication systems, establishing computational viability toward enhanced communication in critical care settings. We developed and validated a CNN model using a dataset of 5,200 grayscale eye region images with balanced open and closed eye states, systematically divided into training (70%), validation (15%), and testing (15%) sets. The model incorporates three convolutional blocks with progressive filter sizes of 32, 64, and 128 filters respectively. Through threshold optimization (θ=0.6), the model achieved 99.7% accuracy, 100.0% precision, 99.3% recall, and 99.7% F1-score on the validation dataset. Robustness evaluation under simulated lighting variations demonstrated performance stability with accuracy ranging from 99.0% to 100.0% across different brightness conditions. Real-time implementation demonstrates 2.3ms model inference time with overall system performance of 10 FPS. This controlled laboratory validation establishes technical feasibility without clinical validation in actual ICU environments. Comprehensive comparative analysis with alternative AI architectures and extensive clinical validation studies are required before practical deployment in critical care settings.