Age estimation for disorder characterization from pediatric polysomnograms

Abstract

Estimating biological age via deep neural networks (DNNs) processing polysomnograms (PSGs) showed promising results in adults. While age estimation itself has limited clinical relevance, the residua between estimated biological and chronological age may serve as a proxy for a variety of health conditions. However, polysomnographic studies on infants and children with respect to age prediction are scarce. To address this gap, we studied a data set of $2097$ pediatric PSGs ($n=1971$; $43.9\text{\%}$ females; $0-18$ years) focusing on three disorders related to sleep dysfunctions. A DNN for age prediction was trained and five minutes of a PSG serving as input proved sufficient for age estimation, yielding a mean absolute error of $1.816$ years. Ablation experiments showed that the DNN’s decision-making was mainly based on brain signals. Moreover, we found systematic links between age estimation residua and two disorder clusters, namely cerebrovascular diseases and cerebral palsy. The distributions of residua differed significantly (two-sided t-test, $p<0.05$) between case and control group and the relative risk of being diagnosed was greater than 1 under the risk factor of having an absolute residuum larger than 1.8 years. For hyperkinetic disorders including attention deficit hyperactivity disorder (ADHD), such links could not be identified. Our analysis shows that systematic patterns in pediatric PSGs can be deciphered by DNNs and could provide new ways to profile disorder-specific sleep.