Predicting High vs Low Mother-Baby Synchrony with GRU-Based Ensemble Models



The early stages of life are paramount for the baby’s brain and emotional development, and the quality of interaction between mother and baby - measured as a dyadic synchrony score, is critical in that period. This study proposes the first machine learning prediction modelling approach, based on Gated Recurrent Unit - GRU ensemble models, to automatically differentiate high from low dyadic synchrony between mother and baby, using a dataset of videos capturing this interaction. The GRU ensemble models which were post-processed by maximising the Youden statistic in a ROC analysis procedure, show a good prediction capability on test samples, including a mean AUC of 0.79, a mean accuracy of 0.72, a mean precision of 0.87, a mean sensitivity of 0.64, a mean f1 performance of 0.72, and a mean specificity of 0.83. In particular the latter performance represents an 83% detection rate of the mother-baby dyads with low synchrony, suggesting these models’ high capability for automatically flagging such cases that may be clinically relevant for further investigation and potential intervention. A Monte Carlo validation procedure was conducted to accurately estimate the above mean performance levels, and to assess the proposed models’ stability. The statistical significance of the prediction ability of the models was also evaluated, i.e. mean AUC > 0.5 (p-value < 9.82 × 10–19), and future research directions were discussed.

Keywords: Automating mother-baby synchrony detection · Gated Recurrent Units - GRU · Ensemble learning · ROC analysis · Monte Carlo validation