Nursing care activity recognition is vital for optimizing workflows and reducing documentation burden in elderly care. The SONAR dataset, recently released as a benchmark for nursing activities and recorded with wearable inertial sensors in a clinical environment, poses unique challenges due to noise, class imbalance, and fine-grained activity definitions. This study presents a systematic evaluation of classical ensemble learning methods on SONAR under deployment-oriented constraints for resource-limited microcontrollers. We assess the impact of sensor placement and model configuration on recognition performance. Among the evaluated algorithms, XGBoost consistently outperformed Extra Trees, LightGBM, and Random Forest, achieving 64.01 % accuracy and 61.21 % weighted F1. These findings highlight both the potential and the remaining challenges in developing reliable and resource-efficient nursing activity recognition systems.