Recognition of Endotracheal Suctioning Activities: A Feature Extraction and Ensemble Learning Approach Based on Pose Estimation Data

Endotracheal suctioning is a critical yet intricate procedure in healthcare, often associated with potential complications due to its invasive nature. The escalating demand for certified professionals proficient in this technique, particularly in the context of home healthcare, underscores the necessity for comprehensive research. However, there exists a dearth of studies investigating automated recognition and analysis of nursing activities during endotracheal suctioning. This study utilizes skeleton pose data to establish correlations between nursing activities and pose dynamics, utilizing a dataset comprising 26 nurses and nurse students. Feature extraction encompassed a diverse array derived from pose characteristics, movement patterns, and activity sequences. For classification purposes, three ensemble learning pipelines were proposed, incorporating both original features and synthetic features generated by the Synthetic Data Vault model. The optimal pipeline, a voting classification approach, achieved noteworthy performance metrics, attaining 89.9% and 90.4% for overall weighted F1 score and accuracy, respectively. Furthermore, ANOVA f-test results revealed the significance of statistical features such as relative position and movement in activity discernment, with sequential features significantly augmenting model efficacy. Notably, disparities in activity execution were observed between experienced nurses and students, particularly evident in high-skill invasive maneuvers during endotracheal suctioning. These findings offer valuable insights for nurse training applications and highlight the potential of automated analysis in healthcare settings.