PREDICTING ANEMIA USING NON-INVASIVE MACHINE LEARNING TECHNIQUES

Abstract

Anemia represents a significant public health problem, especially in young children, where timely identification is essential to prevent serious developmental and health complications. Conventional diagnostic approaches are based on invasive blood tests, which can cause distress in children and are frequently unavailable in low-resource environments. This research investigates a non-invasive method of detecting anemia through the application of deep learning to images of the conjunctiva, palm, and fingernails, collected from children under five years of age in Ghana. Each modality exhibits distinct strengths and limitations: conjunctival images offer robust predictive features but carry an infection risk, fingernail images are small and challenging to analyze in young children, and palm images, while easy to capture, provide inferior contrast. To address these challenges, a multimodal fusion model is proposed that uses CNN-based feature extraction, attention mechanisms, and weighted late fusion via XGBoost. Explainability techniques such as SHAP and Grad-CAM are incorporated to enhance transparency and interpretability. To improve generalization to real-world data, a semi-supervised learning approach is introduced, in which confident pseudolabels from external datasets are merged with the labeled training data to train new models on the combined dataset. Experimental results demonstrate that the fusion approach significantly outperforms standalone models, achieving 94.22% accuracy and 0.9918 AUC. On a manually collected real-world test set of 30 images, the semi-supervised model achieved 83.3% accuracy and 81.5% F1-score, outperforming the baseline model and improving sensitivity to anemic cases. This study presents a scalable, explainable, and field-ready solution for early anemia screening, suitable for mobile and cost-effective health diagnostics in resource-limited settings.