Deep representation learning identifies associations between physical activity and sleep patterns during pregnancy and prematurity

Preterm birth (PTB) is a leading cause of infant mortality. This study uses wearable device data (181,944 h from 1083 patients) and a novel deep learning time-series classification architecture ('series2signal') to model the progression of pregnancy using gestational age (GA). Novel interpretability algorithms integrate unsupervised clustering and feature attribution to interpret the model's relationship with sleep, activity, and clinical variables. The model outperforms other machine learning methods, showing that deviations from normal physical activity and sleep patterns during pregnancy are strongly associated with PTB. Model error is negatively correlated with interdaily stability, and higher importance is attributed to sleep in predicting higher-than-actual GA. The findings suggest that continuous monitoring and interventions targeting activity and sleep habits could mitigate PTB risk, particularly in low- and middle-income countries.

Neal G. Ravindra, Camilo Espinosa, Eloïse Berson, Thanaphong Phongpreecha, Peinan Zhao, Martin Becker, Alan L. Chang, Sayane Shome, Ivana Marić, Davide De Francesco, Samson Mataraso, Geetha Saarunya, Melan Thuraiappah, Lei Xue, Brice Gaudillière, Martin S. Angst, Gary M. Shaw, Erik D. Herzog, David K. Stevenson, Sarah K. England, Nima Aghaeepour