Clustering of Infant Poses as a Novel Metric of Data Diversity for Automated General Movement Assessments Applications.

Monitoring General Movements (GMs) in infants between 6-20 weeks post-term age serves as a reliable prognostic tool for assessing neurodevelopmental quality and determining the risk of neurological impairments in early infancy. Recent studies have demonstrated the power of deep-learning algorithms for markerless pose-estimation and tracking of infants' anatomical landmarks in clinically recorded videos from handheld devices. However, the performance of these algorithms on 'out-of-domain' unseen recordings is largely reliant on using training sets containing a diverse range of poses. This research demonstrates the use of an established clustering algorithm, k-means with silhouette analysis to determine the optimal number of clusters (k), on infant pose data. This method promises to be a convenient, objective metric to assess pose diversity within a set of video frames. We show that the results of this process agree well with another method of measuring pose similarity, the cosine similarity index.Clinical relevance- Findings hold promise for the identification and selection of a set of diverse poses for inclusion in the training set of machine-learning algorithms that are crucial for designing generalized automated General Movement (GM) tracking machine-learning classifiers to tackle challenges associated with 'out-of-domain' unseen poses.