"A Multimodal Framework For Anaemia Screening Using Images And Clinical Features: A Comprehensive Survey And Methodological Proposal"

Abstract

Anaemia is a major global health burden that urgently demands accessible and non-invasive screening solutions. Traditional single modality diagnostic methods often fail to capture subtle physiological cues, leading to reduced sensitivity and limiting their reliability in real world clinical settings. To overcome these limitations, we survey recent deep-learning approaches for non-invasive anaemia detection and propose a high efficiency multimodal framework t hat combines visual features extracted by EfficientNetV2 with structured clinical data such as age, hemoglobin level, and gender. The model addresses class imbalance using SMOTE and leverages a Multi Head Self Attention fusion layer to dynamically weight and integrate information across modalities, capturing complex interdependencies that simple early or late-fusion strategies often miss. Moreover, explainable AI tools, including Grad-CAM and SHAP,
are embedded within the pipeline to provide both visual and quantitative interpretability, enabling clinicians to understand
the basis of predictions and fostering trust in the system. This unified, attention d riven multimodal approach offers a scalable
and robust framework for non-invasive anaemia screening, with potential to enhance early detection and support clinical decision
making.