Attention Deserves Human Insight: An AHP-Enhanced Attention Mechanism for Interpretable and Expert-Guided AI in Skin-Related Neglected Tropical Disease Diagnosis

The development of artificial intelligence (AI) tools for diagnosing skin-related neglected tropical diseases (NTDs) is strongly limited by the scarcity of annotated clinical data. In such contexts, purely data-driven models often struggle to generalize and remain difficult to interpret, which limits their practical use in clinical settings. In this work, we explore whether integrating human expert knowledge can improve both the reliability and interpretability of attention-based models under data-constrained conditions. We propose an AHP-enhanced attention mechanism that incorporates feature importance weights elicited from experts using the Analytic Hierarchy Process (AHP). These weights are directly embedded into the attention computation, guiding the model toward clinically meaningful patterns. The approach is evaluated using a synthetic dataset designed to reflect the challenges of real-world data scarcity in skin NTD diagnosis. We compare our method with a standard data-driven attention model and a random-weighted variant. While classification accuracy remains comparable across models, the proposed approach shows clear advantages in terms of stability and training behavior. It achieves more consistent performance across cross-validation folds and exhibits smoother convergence. In addition, the learned attention patterns and latent representations are better aligned with expert-defined diagnostic priorities. Although the study relies on a controlled experimental setting, it provides a strong proof of concept. The results suggest that embedding structured expert knowledge into attention mechanisms is a promising direction for developing more trustworthy and interpretable AI systems in low-resource medical domains. • AHP-enhanced attention integrates expert knowledge into AI models • Improves stability and interpretability under extreme data scarcity • Ensures more consistent performance across validation settings • Aligns attention patterns with clinically relevant features • Demonstrates a robust framework for expert-guided medical AI