Fluctuations in blood glucose concentration directly influence the body's internal milieu, resulting in altered bioimpedance characteristics. Recognizing the imperative need of continuous blood glucose monitoring for optimized diabetes care, this article explores a novel, noninvasive method leveraging array bioimpedance and graph neural networks. Concretely, we first extract graph-structured data from bioimpedance measurements using the four-electrode acquisition technology and an array electrode. Then, we propose a differential principal neighborhood aggregation (PNA) graph neural network, which integrates differential computation, positional normalization, and PNA, to process the graph-structured data and solve the problem of blood glucose classification. Finally, we evaluate our system with in vitro agar simulation experiments, with the goal of accurately identifying glucose concentrations from 0 to 10 g/l. Our model achieved 95.32% accuracy, 95.30% precision, and 95.18% recall through five-fold cross validation, which outperforms current graph neural network algorithms, and shows promising potential for practical applications.