Engineers from Centro Universitário FEI and pediatricians from Unifesp (Federal University of São Paulo) developed in partnership an artificial intelligence tool capable of identifying the level of pain in newborns admitted to ICUs (Intensive Care Units). The technology uses multimodal language and vision models (vision-language models), which integrate images and texts to interpret babies’ facial expressions with more precision and less subjectivity.
“As pain is a subjective phenomenon and the baby cannot yet communicate verbally, it essentially depends on the observation of third parties. In neonatal ICUs, we use pain scales, but they are very subjective. Interpretations can vary depending on the emotional state of the observer, as a doctor, a nurse or a more distressed mother may have different perceptions. In this context, the artificial intelligence tool can help reduce this subjectivity and support”, says Ruth Guinsburg, professor of neonatal pediatrics at the Unifesp and general coordinator of the Neonatal ICU at Hospital São Paulo.
The research, funded by Fapesp, was published in the journal Pediatric Research and demonstrated that the artificial intelligence system excels in identifying states of pain and comfort. Furthermore, the model does not need to be trained separately for each task, which expands its clinical applicability.
“Until recently, classic machine learning models were used that required a huge and specific database for each task, in addition to the need for complex pre-processing of images. With the arrival of multimodal language models, such as ChatGPT and Gemini, for example, it became possible to use pre-trained models on a huge amount of internet data to solve specific medical tasks more quickly”, explains Carlos Eduardo Thomaz, professor at FEI.
According to Guinsburg, a baby admitted to one can be subjected to up to 13 painful procedures per day, such as punctures, catheter insertion, surgeries and intubations. “These interventions are vital, but they cause pain. Therefore, it is essential to balance clinical need and suffering, as poorly managed pain can leave lasting consequences”, he highlights.
She says that until the 1990s it was believed that newborns did not feel pain because they were neurologically immature. “Today the exact opposite is known: because they are neurologically immature, they are even more vulnerable to the adverse effects of painful stimuli,” he says.
Therefore, researchers believe that the AI tool can be an ally to transform subjective signals into objective parameters, functioning as a “balancer” in clinical assessment.
The expectation is that, in the future, the tool will be able to issue alerts in real time, acting as a pain monitor alongside cardiac and respiratory devices. And it could also support safer painkiller prescriptions.
“In the developing brain, both untreated pain and excess medication can be neurotoxic. The challenge is to hit the target: treat when there is pain and stop it when it stops”, highlights Guinsburg.
For engineer Lucas Pereira Carlini, a member of the team, the impact of AI goes beyond technical performance. “We are always looking for more precision, but it is important to remember: what does each percentage point of accuracy represent for a baby?”, he concludes.
