<p>The current smart elderly care system is difficult to fully mine complex nonlinear data relationships, ignores user experience and long-term adaptability, and makes it difficult for the system to meet diverse health needs. This article combines improved Bayesian optimization with deep belief networks (DBNs) to design a smart elderly care health monitoring system. In the data acquisition stage, the STM32 minimum system is used with a variety of sensors to monitor the body temperature, blood oxygen, blood pressure, heart rate, and exercise status of the elderly in real-time. For the noise and outliers in the original data, low-pass filters and Kalman filters are used to reduce noise, and statistical methods are used to remove abnormal data to ensure data accuracy. To enhance the feature extraction efficiency, the improved Bayesian optimization method is used, combined with dynamic weight update and high-dimensional search strategies, to efficiently explore DBN model hyperparameters and optimize model configuration. In the inference stage, the optimized DBN is used for deep feature recognition and classification prediction to ensure a rapid response to health changes. Meanwhile, a user-friendly interface is designed to intuitively display health information and provide personalized health advice and risk warnings. The study findings demonstrate that the average response and feature extraction time of the system under the improved Bayesian optimization and DBN method in this article are 4.07 and 3.00 milliseconds, respectively, and the average accuracy and recall are 95.99% and 92.59%, respectively. This shows that combining improved Bayesian optimization and DBN can greatly enhance the smart elderly care monitoring system performance, provide precise and reliable health management services for the elderly, and meet the needs of complex data processing.</p>