Design of AI-Powered Pain Monitoring System for Inpatient Care

Deep learning image classification has received high interest in the biomedical and psychology literature for the identification of pain. This work introduces a comparative study of two CNN architectures to classify facial expressions of pain based on grayscale images. The initial method uses an ordinary CNN architecture with three convolutional layers, max pooling, and dropout to extract features and classify. The second method employs a ResNet-like CNN with residual links to facilitate gradient flow and enhance feature learning. Both models were trained and tested on a database of four pain states: Algometer, Laser, Neutral and Posed Pain. Experimental results indicate that the ResNet-based model has better classification performance than the common CNN, which proves its capability in identifying subtle differences in pain expressions. The paper emphasizes the necessity of sophisticated deep learning methods for computerized pain assessment and presents guidelines for fine-tuning CNN structures for bio-medical image processing.