Pain assessment in infants via facial expressions analysis and deep learning

This research presents a deep learning-based approach for infant pain assessment and it was conducted in LINKS Foundation in collaboration with the Neonatal Unit of AO Ordine Mauriziano Hospital and the Pediatric Emergency Department of Regina Margherita Hospital, all based in Turin. Since young children are unable to communicate verbally the experience of pain, accurate pain assessment using validated tools is crucial to determine the most effective pain management strategies. Traditional pain assessment relies on the use of pain scales that consider behavioural, physiological, and contextual indicators, but none of these scales has yet emerged as the gold standard. The use of these tools is also limited by the lack of objectivity and repeatability of the assessment, which depends on the experience of healthcare professionals and the simultaneous monitoring of numerous parameters. For these reasons, automated pain assessment systems are highly desirable in clinical practice for more efficient recognition and management of pain. The aim of this thesis is to propose an objective and contactless computer vision approach for infant pain evaluation based on facial expression analysis. Video data related to two distinct infant populations are used: newborns undergoing heel stick procedures for blood sampling, and children aged 3-36 months admitted to the Pediatric Emergency Department with acute pain. Videos are accompanied by the corresponding pain scores assigned by healthcare professionals using traditional pain scales. Two separate datasets are created by extracting frames from the video recordings and labelling each frame based on the assigned score. Each dataset is used to train a specific convolutional neural network for binary pain classification. For both infant populations, the trained models achieve high accuracy in classifying images into pain or nonpain categories. Moreover, the application of a visual explanation technique shows that the networks' decisions are based on the facial regions most closely associated to the experience of pain. In conclusion, our promising findings pave the way for the development of an automated system that integrates, standardises, and improves human pain evaluation.