To address the challenges of vulnerability to gradient inversion attacks, high communication overhead, and performance degradation caused by data heterogeneity in federated learning-based smart healthcare systems, this paper proposes a personalized federated learning framework with selective aggregation based on hypernetworks, namely pFedHNSA (Personalized Federated HyperNetworks with Selective Aggregation) . The proposed method constructs an intelligent diagnosis system under the federated learning paradigm, where a hierarchical hypernetwork deployed on each hospital client dynamically generates personalized local model parameters via embedding vectors, thereby avoiding the direct transmission of complete model parameters or gradients and reducing privacy leakage risks from a structural perspective. Furthermore, we introduce a PCA-based embedding generation mechanism, which takes the trained local model parameters as input and performs dimensionality reduction to replace the traditional second-order backpropagation update, improving the stability and representational capability of embeddings. In addition, we design a selective aggregation strategy for the shared layers of the hypernetwork, where the system uploads only the public layers for global aggregation, significantly reducing communication overhead. Experiments conducted on CIFAR-10, Fashion-MNIST, and PathMNIST datasets under various non-IID settings demonstrate that the proposed method outperforms HyperFL, DPFedAvg, and CENTAUR in terms of model accuracy and communication efficiency while effectively defending against gradient inversion attacks, achieving a favorable trade-off among privacy preservation, communication cost, and model performance.