Reliable and efficient methods to identify Parkinson’s disease (PD) from EEG signals are highly valuable for studying early disease markers and supporting timely intervention research. In this work, we propose a lightweight deep learning framework for PD detection using electroencephalogram (EEG) signals. The framework employs a novel spatial–frequency multiscale feature fusion strategy to characterize both local neural activity and global inter-channel connectivity. To achieve high accuracy with low computational cost, we design a hybrid network combining the DCFConv lightweight convolution module and the OSGAttn orthogonal global attention module. Experiments on the public Iowa PD-EEG dataset show that our method achieves 99.81% accuracy, 100% precision, and 99.62% sensitivity, while maintaining only 0.16 M parameters and 44.08 M FLOPs. The results demonstrate that the proposed approach provides an effective, efficient, and reliable solution for PD EEG signal classification.