Traditional survival analysis methods often fail to capture the temporal complexity inherent in longitudinal treatment data, resulting in limitations in modeling time-dependent outcomes. Although recent deep learning approaches have improved predictive performance, most still represent patient histories as static feature vectors, overlooking the sequential nature of care. This study proposes ViT-Survive, a novel application of the Vision Transformer (ViT) for survival analysis, aiming to predict the time to treatment dropout in HIV care. Each patient’s treatment history is reformulated as a sequence of temporal patterns, where each visit serves as a standalone unit comprising clinical and behavioral indicators. Through self-attention across visits, the model learns dynamic temporal patterns that reflect health trajectories and behavioral engagement. A classification token (CLS token) is trained within a Cox proportional hazards framework to summarize the treatment sequence and predict dropout timing. The model was evaluated on 4,153 anonymized longitudinal HIV treatment records collected in Ho Chi Minh City from 2022 to 2024. Experimental results demonstrate superior performance over existing statistical and deep learning models, achieving a C-index of 0.726. This approach opens a new direction for individualized dropout prediction using time-to-event healthcare data, contributing to proactive and personalized HIV care.