This paper proposes MCLPD, a novel semi-supervised learning framework based on electroencephalography (EEG) for the early diagnosis of Parkinson's disease (PD). MCLPD integrates multi-view contrastive learning pretraining and lightweight supervised fine-tuning to improve robust and generalizable PD detection performance despite dataset differences. The pretraining stage performs self-supervised learning using an unlabeled dataset (UNM) and generates contrastive pairs through dual augmentation in the time and frequency domains, naturally integrating time-frequency information. The fine-tuning stage performs supervised learning using only a small amount of labeled data from the other two datasets (UI and UC). Experimental results show that MCLPD achieves F1 scores of 0.91 on the UI dataset and 0.81 on the UC dataset using only 1% of the labeled data, further improving performance to 0.97 and 0.87, respectively, using 5% of the labeled data. Compared to existing methods, MCLPD significantly improves cross-dataset generalization while reducing labeled data dependence.
