Automated human activity analysis is essential for intelligent healthcare systems. However, existing visual methods still struggle to accurately analyze diverse and complex activities. Moreover, few studies specifically address older adults, who are particularly vulnerable to emergencies. This paper introduces an Attention-based Encoder-Decoder Convolutional Long-Short Term Memory (ConvLSTM) model for activity analysis and elderly emergency detection using skeleton frames. Experiments were conducted on a challenging dataset of various activities associated with geriatric emergencies. Results demonstrate that integrating the attention mechanism enhances the ConvLSTM ability to capture motion dynamics by focusing on the most informative skeleton features. In addition, a comparative evaluation with state-of-the-art methods confirms the effectiveness of the proposed method in detecting elderly emergencies.