A new framework uses autoencoder with contrastive learning to analyze segment-level Mandarin speech for cognitive impairment detection. It achieves stable, competitive performance across four datasets, with significant improvements in three-class classification, especially under limited labeled data conditions.