Pain Intensity Recognition from Facial Expression Using Deep Learning

Abstract

Pain weaves its way into daily life, turning ordinary tasks into challenges and testing our strength in unexpected ways. Pain can be observed in a human’s face and can be understood as pain intensity from patients’ verbal acquisition. However, for non-verbal patients—such as those in the ICU, individuals with mental challenges, or even AI—detecting and interpreting pain remains a complex challenge. To extend this, many researchers have done remarkable research and are still trying to find a solution with acceptable accuracy. This research paper presents a hybrid parallel model to detect the intensity of pain from facial expressions. Following the Prkachin and Solomon Pain Intensity (PSPI) metric, we considered 16 pain levels, which were divided into four subranges. We call these sub ranges as ” No Pain”, ”Mild Pain”, ”Moderate Pain”, and ”Severe Pain”. Our parallel feature fusion model consists of a fully connected network with inputs from two deep CNN models, one being VGG19, and the other model can be ResNet50, DenseNet121, or InceptionV3. Thus, we have 3 parallel feature fusion models (PFFM), respectively, PFFM-1, PFFM-2 and PFFM-3. Besides, we trained and evaluated our models using the McMaster Shoulder Pain dataset, where PFFM-2 emerged as the top performer, achieving an 82.54% accuracy in assessing pain intensity from facial expressions. By outperforming existing pain detection systems, this breakthrough bridges the gap between human perception and AI, enabling more precise and reliable pain interpretation.