Transformer models pre-trained on self-supervised tasks and fine-tuned on downstream objectives have achieved remarkable results across a variety of domains. However, fine-tuning these models for clinical predictions from longitudinal medical data, such as electronic health records (EHR), remains challenging due to limited labeled data and the complex, event-driven nature of medical sequences. While self-attention mechanisms are powerful for capturing relationships within sequences, they may underperform when modeling subtle dependencies between sparse clinical events under limited supervision. We introduce a simple yet effective fine-tuning technique, Adaptive Noise-Augmented Attention (ANAA), which injects adaptive noise directly into the self-attention weights and applies a 2D Gaussian kernel to smooth the resulting attention maps. This mechanism broadens the attention distribution across tokens while refining it to emphasize more informative events. Unlike prior approaches that require expensive modifications to the architecture and pre-training phase, ANAA operates entirely during fine-tuning. Empirical results across multiple clinical prediction tasks demonstrate consistent performance improvements. Furthermore, we analyze how ANAA shapes the learned attention behavior, offering interpretable insights into the model's handling of temporal dependencies in EHR data.
Amirahmadi et al. (Wed,) studied this question.