Real-world Classification of Caregiver Sensitivity to Infant Distress

This study introduces the first real-world binary audio classifier for caregiver sensitivity to infant distress, a key input for infant attachment and socioemotional development. We employ a multiscale temporal framework that integrates features at three levels: frames, events, and states. We evaluate our approach on a novel dataset of 387 annotated infant distress episodes, and compare our multiscale approach against a conventional frame-level baseline and a state-of-the-art pretrained audio transformer model. Our proposed multiscale temporal framework achieves a macro F1 score of $\mathbf{0. 7 6 0}$, outperforming the best baseline by $8.7 \%$. To systematically assess the contribution of each temporal scale, we conduct an incremental temporal scale analysis, showing multiscale integration enhances performance. Our findings demonstrate that integrating multiscale information improves performance over conventional frame-only models. We make the model publicly available, providing a foundation for basic science research on socio-emotional development and interventions to identify and support at-risk dyads.