Abstract: The (p, 2n) reaction excitation function is the vital data in both nuclear engineering and fundamental nuclear research, directly affects the accuracy and efficiency of applications such as medical radioisotope production, reactor neutronics design, and nuclear waste transmutation. However, the measured data are scarce and scattered for some nuclei or energy regions, the discrepancy of evaluated data are relatively significant, and the development of evaluation method for the (p, 2n) reaction excitation function is essential. In general, the systematics provides a convenient and reliable framework for predicting (p, 2n) reaction excitation function. However, its predictive accuracy remains limited. The neural network approach is a powerful tool for data analysis and modeling, while it lacks the solid theoretical foundation and relatively limited generalization capability. Therefore, it is necessary to develop a data evaluation method that integrates systematics with neural networks. In this work, an evaluation approach combining systematic formulas with a neural network (NN+SYS) was developed for the evaluation of (p, 2n) reaction excitation function. The NN+SYS framework consists of two main components: a feedforward neural network and a backpropagation optimization scheme. The feedforward neural network was used to model and predict two adjustable parameters C₀ and C₃. The predicted C₀ and C₃ parameters from the feedforward neural network, along with other input variables were substituted into systematics formula to obtain the preliminary (p, 2n) reaction cross sections, then compared with experimental data to construct a loss function Li that reflects the discrepancy between the prediction and the experimental data. Using the backpropagation method, the neural network was optimized to adjust the predicted C₀ and C₃ parameters, thereby Li was reduced until it converged to a reasonable range. Finally, the optimized C₀ and C₃ parameters and the corresponding predicted excitation function for the (p, 2n) reactions were obtained. Compared with the predictions based on local systematics parameters, the NN+SYS method is applicable to a broader range of nuclides and achieves higher prediction accuracy. Compared with the results obtained using global systematics parameters, NN+SYS shows a significant improvement in prediction accuracy, with an average enhancement of approximately 42.8%. In comparison with the neural network approach, the NN+SYS predictions are more consistent with physical constraints and exhibit stronger extrapolation capability. Furthermore, relative to the TENDL-2023 evaluated nuclear data library, the NN+SYS predictions are in better agreement with the available experimental data. The evaluation method of (p, 2n) reaction excitation function developed by integrating systematics and neural network can provide technical and data support for the Chinese Evaluated Nuclear Data Library (CENDL).