Transformer-Based Multi-Modal Dream Interpretation for Early Detection of PTSD

Mental health conditions like PTSD and depression often manifest in subconscious patterns, yet AI-based diagnostics rarely explore dream narratives. In this work, we propose a novel deep learning framework for analyzing free-text dream reports using transformer-based embeddings and emotion-aware classification. Leveraging the DreamBank dataset, we extract contextual embeddings using DistilBERT, perform unsupervised clustering to detect latent dream themes, and fine-tune a binary classifier to detect PTSD-like content. Our model achieves 85% classification accuracy, and cluster validity metrics such as a 0.22 silhouette score suggest meaningful separability in dream types. This study demonstrates the potential of dream analysis as a non-invasive AI tool for early mental health screening.