Prediction of Elderly Behavior Patterns Based on Big Data in Smart Elderly Care

This paper proposes a hybrid model that integrates spatiotemporal graph convolutional network (STGCN) and federated learning (FL) to solve the core problem of elderly behavior pattern prediction in smart elderly care. By constructing a multimodal dataset with a time span of 3 years (including sensor data, health indicators and environmental parameters of 1024 elderly people), the experimental results show that the STGCN-FL model achieves an accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 2. 3 \%}$</tex> in the behavior pattern prediction task, which is 18.6 % higher than the traditional LSTM model; the F1 score in abnormal behavior detection reaches 0.941, which is 12.7 % higher than the Transformer model. In particular, the federated learning framework reduces the risk of privacy leakage when sharing data across institutions by 89.4 %, while keeping the model performance loss less than 2.1 %. By introducing a dynamic weight allocation mechanism, the model can effectively capture the spatiotemporal dependency of behavior patterns, and the MAE is only 0.183 in long-term prediction (72 hours), which is significantly better than other comparison algorithms. The research results provide theoretical support and technical solutions for building a distributed smart elderly care system.