Nurse Care Activity Recognition: A Cost-Sensitive Ensemble Approach to Handle Imbalanced Class Problem in the Wild

Abstract

Nurse care activity recognition is a recent but demanding study topic in human activity recognition (HAR) since it has high class imbalance and intra-class variability problem depending on both the patient and the receiver. Although traditional imbalance learning approaches are offered to address this issue, they have several limitations: 1) important information is lost while using undersampling approaches, and 2) oversampling methods shifts the class distribution, exposing the model to overfitting and over-optimism. To tackle this problem, we present a framework for a cost-sensitive hybrid optimum ensemble classifier. Our technique modified the class weights to impose higher misclassification costs on minority classes. Then we combined different weighted base classifier outputs with a stacked generalization method. This technique allowed us to leverage the skills of a varied collection of high-performing base learners while balancing each other's limitations and strengths. By utilizing our hybrid technique, we achieved 70.8% average cross-validation balanced accuracy in classifying 28 activities conducted by nurses in real-life settings. We developed this algorithmic pipeline for "The 3rd Nurse Care Activity Recognition Challenge" at ACM UbiComp 2021.

Cost-sensitive learning for imbalanced medical data: a review

“…A noteworthy observation in this study is that some papers combined two development tools, resulting in their double counting in both respective tool categories. Among them, three papers (1.7%) utilised both LibSVM and MATLAB (Liu et al 2018;Razzaghi et al 2015;Prashanth and Roy 2018), while one paper (0.6%) employed Python alongside MAT-LAB (Rahman et al 2021a). Additionally, one study (0.6%) leveraged the combined capabilities of Python and Weka (Wu et al 2020).…”

Trustworthy Long-Tailed Classification

“…In modern applications of long-tailed classification, the semantic importance of "tailed" data often implies more penalty in the circumstance of predicting tailed samples as head (Sengupta et al, 2016;Rahman et al, 2021;Yang et al, 2022). Besides, the lack of training samples in tailed classes has been empirically proved to be the bottleneck of classification performance (Li et al, 2022).…”

“…Conventional models trained on long-tailed data often report significant performance drops compared with the results obtained on balanced training data (Wang et al, 2022). Besides, for some real-world applications, the risk of classifying tailed samples as head (which is a common type of mistakes) is obviously more severe than that of classifying head samples as tail (which is less common) (Sengupta et al, 2016;Rahman et al, 2021;Yang et al, 2022). The significant performance drop and the "tail-sensitivity risk" limit the application of ML models to long-tailed classification.…”