Enhancing nutritional status prediction through attention-based deep learning and explainable AI

Accurate and interpretable prediction of child malnutrition remains a critical challenge, as existing AI models often lack the transparency needed for clinical adoption. This study introduces a deep learning framework enhanced with Multi-Head Attention (MHA) for nutritional status prediction, offering a novel contribution through the first direct head-to-head comparison of CNN-MHA and LSTM-MHA to evaluate the effectiveness of spatial feature learning versus sequential dependency modeling in structured anthropometric tabular data. Our framework integrates advanced preprocessing techniques, feature selection, and Explainable AI (SHAP), enabling clinically aligned and transparent predictions. Experimental results on a 9605-sample dataset reveal that CNN-MHA achieves superior performance (99.08 % accuracy) compared to LSTM-MHA (98.91 %), confirming that spatial modeling is better suited for this dataset type. SHAP-based feature attribution further validates WHO-standard z-scores as the most influential predictors, enhancing model credibility for clinical application. Additionally, the study introduces an IoT-enabled anthropometric data acquisition system, enhancing real-time monitoring and scalability. This research represents a significant methodological advancement in nutritional status prediction, addressing key gaps in feature prioritization, accuracy, and interpretability. By bridging the gap between high-accuracy AI and clinical transparency, this study advances AI-driven nutritional monitoring and offers a scalable, explainable framework for public health interventions. Future research should explore multi-modal data integration to further enhance generalizability and real-world applicability.