In addressing the challenges of real-time monitoring and dynamic risk assessment in elderly chronic disease management, this study proposes an innovative prediction framework leveraging multi-source sensor data. First, a multi-dimensional data collection system integrating wearable devices and environmental sensors is established to capture physiological parameters, behavioral patterns, and environmental factors. Next, a spatiotemporal feature engineering approach combines sliding window segmentation and LightGBM-based feature selection to extract clinically meaningful dynamic indicators. A hybrid prediction model is then developed, integrating LSTM for time-series dependency learning and an attention mechanism to prioritize critical health events. Finally, a Shapley value-based interpretability module generates personalized intervention strategies. Experimental results demonstrate the model’s superiority, achieving 88.4% accuracy and 0.92 AUC in predicting hypertensive crises, outperforming CNN-based methods. Environmental correlation analysis reveals temperature and humidity significantly influence blood pressure (r = 0.63) and glucose (r = -0.58). This framework provides a robust, interpretable solution for proactive chronic disease management in aging populations.