Boosting pathology detection in infants by deep transfer learning from adult speech

Anonymous

Abstract: Can knowledge extracted from adult speech help improve the performance models developed for infant cry? This work investigates this question in the context of pathology detection in newborns. The analysis of infant crying patterns to detect pathology is of interest as it opens the possibility of more accessible diagnostic tools in resource-constrained settings. Classical machine learning approaches leveraging features extracted as Mel frequency cepstral coefficients, have supported the viability of the infant cry as a diagnostic input, but performance is not yet at a level of clinical utility. The application of deep learning models has been limited due to the unavailability of large infant cry databases which are costly to acquire. This work argues that the transfer of useful knowledge from adult speech is possible because it is driven by the same underlying physiologic process as that of infants. Our experiments demonstrate that on the task of predicting perinatal asphyxia from infant cry, such transfer learning provides an overall improvement of 13.5\% in F1 score over a model trained from random initialization.

Keywords: transfer learning, infant cry, keyword spotting, pathology detection, MFCC