Disease risk prediction enables the screening of vulnerable populations and early preventive interventions to reduce disease incidence and mortality. With the rapid development of machine learning technologies, disease risk prediction based on machine learning has been widely used. However, machine learning is highly dependent on high-quality labeling information, and the label noise in medical data will bring severe challenges to the construction of high-performance disease risk prediction algorithms. In order to solve this problem, a noise robustness learning method based on a deep neural network and dynamic truncation loss function is proposed for disease risk prediction. The dynamic truncation loss function is introduced in this method, which combines the implicit weighting characteristics of the traditional cross entropy function and the label noise robustness of the mean square error loss function. By constructing a training loss lower bound and introducing a dynamic sample weighting mechanism to reduce the gradient of suspicious samples, the weight of possible noisy samples in the training process is limited, and the robustness of the model is further enhanced. By taking the stroke screening dataset as an example, the experimental results show that the proposed algorithm can achieve excellent prediction performance under each ratio of label noises, reduce the negative impact of label noises in disease risk prediction, and realize robust learning of data with label noises.