Efficient Pain Recognition via Deep Learning: A Comparative Analysis of Lightweight and Deep CNN Models

Abstract

Traditional methods for pain assessment often suffer from limitations such as subjectivity, observer variability, and challenges in evaluating non-verbal patients. These shortcomings have driven the development of automated pain recognition systems using deep learning techniques. Convolutional Neural Networks (CNNs) have shown strong potential by addressing issues like manual feature extraction and limited dataset availability. This study presents a comprehensive comparative analysis of seven prominent CNN architecturesResNet34, ResNet50, ResNet101, VGG19, DenseNet, EfficientNet (ENet), and SqueezeNet for automated classification of pain intensity from facial expressions. Each model is evaluated based on architectural design, parameter count, and effectiveness in extracting discriminative features from labeled facial pain datasets. Custom learning functions are integrated to enhance performance, and models are assessed using standard metrics including accuracy, F1-score, precision, and memory efficiency. Among all models, ResNet101 achieved the highest classification accuracy of 97.95%, demonstrating its superior ability to capture deep hierarchical features relevant to pain intensity. The findings of this study contribute to the advancement of intelligent, non-invasive pain recognition systems aimed at enhancing clinical decision-making and patient care, particularly for individuals unable to communicate pain effectively.