Poor treatment adherence remains a persistent healthcare challenge, costing hundreds of billions annually and compromising patient outcomes. Traditional health-economics models assume latent behavioral preferences — feedback learning, risk aversion, and loss aversion—that are rarely observed directly at the neural level. We propose a framework integrating Vision Transformer (ViT) attention patterns from electroencephalography-based event-related potential (EEGERP) images into structural health-economics models of compliance. Using three open-access OpenNeuro datasets spanning risk/ambiguity, reward-punishment learning, and delay feedback learning, ViT models classify trial-level economic choices and estimate subject-level preference parameters (risk aversion ρ, loss aversion λ, feedback learning rate). Attention maps consistently identify medial prefrontal, anterior cingulate, and parietal regions associated with high loss aversion — neural systems linked to adherence behavior. ViT achieves cross-dataset ROC-AUC up to 0.88 for risk/ambiguity and 0.85 for reward sensitivity, outperforming EEG-feature SVMs, 1D-CNNs, and ResNet baselines. In structural discrete choice models, ViT-derived neural states significantly predict adherence-proxy decisions (p < 0.001), explaining an additional 12% of variance beyond task attributes. Interpretable neural biomarkers from ERP images enable improved personalized adherence prediction and intervention design. Potentially deployable in 10-minute EEG sessions with 14 ms inference, this framework supports a scalable precision-medicine pipeline for addressing costly non-adherence.