Mutual Information based Feature Selection for Nurse Care Activity Recognition

Human activity recognition is a challenging task as performing the activities varies from person to person. For the last few years, many complex methods have been proposed to identify human activities from sensor readings. To date, several studies have been conducted successfully to identify simple activities and many commercial applications have also been developed. However, reliable recognition of complex activities is still an active research area. The nurse care activity dataset can be treated as a complex activity recognition dataset. Researchers have proposed many solutions to identify nurse care activity by extracting numerous handcrafted features or using the Spatio-temporal graph convolution method. However, some of these features may be noisy, redundant, or even distract the classifier performance. In this paper, we propose a feature selection strategy to select important features from the handcrafted features. We claim that the simple classifier can provide satisfactory performance once the important features are selected and the noisy ones are eliminated. Experiments demonstrate that our proposed approach achieves 87.93% accuracy and 87.97% f1-score in test data. This is a significant improvement over state-of-the-art approaches on this benchmark dataset and thereby establishing our claim.