Audio Feature Analysis for Acoustic Pain Detection in Term Newborns

INTRODUCTION: Crying newborns signal a need or discomfort as part of the innate communication system. Exposure to pain is related to infants' unfavorable neurodevelopmental outcomes. There is a tremendous need for more objective methods to assess neonatal pain. An audio analysis of acoustic utterances could provide specific information on the patient's pain level.

METHODS: We analyzed 67 videos of 33 term-born newborns recorded during a planned capillary blood sample, including the stimuli, non-noxious thermal stimulus, short noxious stimulus, and prolonged unpleasant stimulus, between December 2020 and March 2021. Two expert raters evaluated the infants' pain responses using the Neonatal Facial Coding System (NFCS). The mean values of 123 timbre features of the recorded audio data were analyzed by using specific toolboxes and libraries from the following programming environments: MIRtoolbox (MATLAB), MiningSuite (MATLAB), Essentia (Python), AudioCommons timbral models (Python), and Librosa (Python).

RESULTS: The NFCS values were significantly higher during the short noxious stimulus (p < 0.001) and prolonged unpleasant stimulus (p < 0.001) than during the non-noxious thermal stimulus, whereas NFCS values during the short noxious stimulus and prolonged unpleasant stimulus were similar (p = 0.79). Brightness, roughness, percussive energy, and attack times were identified as the features having the highest impact on the NFCS.

CONCLUSION: This hypothesis-generating study identified several salient acoustic features highly associated with pain responses in term newborns. Our analysis is an encouraging starting point for the targeted analysis of pain-specific acoustic features of neonatal cries and vocalizations from the perspective of real-time acoustic processing.