Predicting Endotracheal Suctioning Activities Based On Skeletal Coordinates: A Novel Approach Combining Handcrafted Feature Engineering and Ensemble Learning

Endotracheal Suctioning (ETS) is a critical procedure typically carried out to patients with artificial airways in intensive care units. It is an extremely delicate procedure requiring adequate expertise to minimize the risk of complication and ensure patient safety. Although numerous studies have applied machine learning techniques for the purpose of nurse-care activity recognition, little works focus specifically on endotracheal suctioning activity recognition. Utilizing a model proficient in categorizing endotracheal suctioning activities can serve as a valuable tool for evaluating nursing proficiency. In this paper we use skeletal coordinate data obtained from the 6th ABC Challenge (Activity Recognition of Nurse Training Activity) and apply machine learning techniques to predict activities associated with ETS. Following some data pre-processing steps including data smoothing, we extracted joint angles, velocity and several time-domain features and performed data augmentation. We split the data into training and test set and finally, we employed an ensemble model for classifying the ETS activities. The ensemble model yielded superior results than conventional machine learning approaches with an accuracy of 68% and a macro average f1-score of 71 %.