Reference method for the development of domain action recognition classifiers: The case of medical consultations

Abstract

Advances in human action recognition and interaction recognition enable the reliable execution of action classification tasks through machine learning algorithms. However, no systematic approach for developing such classifiers exists and since actions vary between domains, appropriate and usable datasets are uncommon. In this paper, we propose a reference method that assists non-experts in building classifiers for domain action recognition. To demonstrate feasibility, we instantiate it in a case study in the medical domain that concerns the recognition of basic actions of general practitioners. The developed classifier is effective, as it shows a prediction accuracy of 75.6% for the medical action classification task and of more than 90% for three related classification tasks. The study shows that the method can be applied to a specific activity context and that the resulting classifier has an acceptable prediction accuracy. In the future, fine-tuning of the method parameters will endorse the applicability to other domains.