Datasets, Machine Learning Recognition Performance, and Key Issues in Nursing Care Activity Recognition: A Systematic Review

Yasuyuki Yoshida, Hiroyasu Miwa, Satoshi Nakae, Satoki Ogiso, Ryosuke Ichikari, Takashi Okuma

Abstract: Countries experiencing rapid aging face severe caregiver shortages, necessitating effective workforce optimization. Nursing care activity recognition (NCAR) offers a promising solution by automatically detecting caregiving tasks over time and visualizing workflows. Based on this information, NCAR can facilitate improvements in caregiving processes. This review systematically assessed 32 studies following PRISMA guidelines, analyzing dataset characteristics and recognition performance of algorithms in NCAR to identify factors influencing recognition accuracy. The review revealed significant challenges, including a lack of publicly available datasets, class imbalance, and decreased recognition accuracy in real-world environments. Additionally, sensor configuration and annotation quality were found to significantly impact recognition performance. The limited availability of field data restricts model generalization, highlighting the need for novel algorithms and efficient data collection methods. Expanding datasets, designing appropriate classes, and improving algorithms are critical steps for advancing NCAR research and practical applications. The findings of this study demonstrate the potential of NCAR in optimizing caregiving workflows and mitigating workforce shortages.