Machine learning-enhanced behavioural approach to detecting reactions to sound in infants and toddlers: proof-of-concept study.

Every child tested had essentially normal hearing, apart from one mild conductive loss, so the method has not been tried on the children it would be meant for: those with hearing loss or developmental conditions that make behavioural testing difficult. The algorithm made no left/right decision for most responses, so the majority of trials produced no result to judge. Only one video frame, 2 seconds after each sound, was analysed, so responses that started earlier, ran later or were simply slow would be missed. The sample is small: 58 children, one hospital, a single session each, and the toddler test set produced only about four to six decisions per measure, which the authors say is too few for meaningful statistical tests. Many head turns and gazes went in the opposite direction to the sound, and the assistant who kept re-centring the child may have interrupted some responses. The authors describe this as a proof of concept and say larger datasets and more complex methods are needed before any clinical use.

The authors reported no potential conflict of interest. The article does not state who funded the work.

Do not read this as a working test of a young child's hearing. The algorithm stayed silent in most trials, the infant head-turn results were not significantly better than guessing, and the toddler results rest on a handful of decisions. Every child tested had normal hearing except one with a mild conductive loss, so nothing here shows how the method behaves in the children who are hardest to test.