Is this neonate feeling pain? Leveraging clinical knowledge towards high-precision Large Language Model-based neonatal pain assessment.

backgroundNeonates in intensive care undergo an average of 13 painful procedures daily, with untreated pain linked to structural brain alterations and long-term cognitive and behavioral impairments. Current pain assessment relies on subjective evaluation scales that vary according to infant characteristics, procedure type, and evaluator background, highlighting the need for more objective assessment methods.

methodsWe leverage a Vision-Language Model (VLM) for neonatal Automatic Pain Assessment (APA) and implemented novel prompt categories based on two approaches: (1) encouraging the model to retrieve clinical knowledge from its pretraining, and (2) providing information about clinically relevant facial features.

resultsWhen leveraging latent clinical knowledge, the model achieved a balance of precision (82.3%) and recall (73.2%). When assessing clinically relevant facial features, it reached perfect precision (100%) with lower recall (40.1%).

conclusionThis first application of VLMs for neonatal APA demonstrates superior performance compared to previous deep learning approaches. The model effectively retrieves latent clinical knowledge and performs best when provided with clinical context. When instructed with specific criteria and facial features, it achieved high precision with significantly reduced withdrawal rates compared to baseline prompts, highlighting the feasibility of this novel approach as a real-world evaluation of state-of-the-art technology through all its steps of development. IMPACT: This study pioneers the application of Vision-Language Models (VLMs) for Automatic Pain Assessment in neonates, offering a novel alternative to traditional deep learning approaches. We demonstrate that carefully designed prompts can leverage a model's latent clinical knowledge or guide it to assess specific facial features, without requiring fine-tuning. Our approach achieves perfect precision (100%) when assessing clinically relevant facial features, surpassing previous deep learning methods. This work opens new research directions for neonatal pain assessment using instruction-based AI systems that can incorporate clinical expertise through natural language prompts.