Machine Learning-Based Non-Invasive Diagnosis of Anemia in Children Using Palm Image Analysis

Anemia, particularly iron-deficiency anemia, is a critical global health concern, with a high prevalence among children under six years of age. Early and non-invasive detection can significantly improve health outcomes. This study proposes a computer vision and machine learning framework for anemia screening and hemoglobin (Hb) level prediction using palmar images from pediatric subjects. The region of interest (palm) was segmented using a U-Net model, achieving a Dice coefficient of 0.96. Images were processed across RGB, CIELab, and HSV color spaces to extract key color features, including red fraction, erythema index, and normalized a-component. For anemia classification, multiple machine learning models were evaluated, with Logistic Regression, Gradient Boosting, and a custom Convolutional Neural Network (CNN) achieving the highest test accuracies of approximately 94.5% and 95.53%, respectively. For hemoglobin regression, a Random Forest model in the CIELab color space achieved a coefficient of determination (R2) of 0.95. The Pearson correlation coefficient in the Lab color space was 0.98 for the Random Forest algorithm and 0.94 for the Linear Regression algorithm. The analysis, supported by SHAP values, identified red-related color features as the most significant predictors. The model demonstrated robust performance across different skin tones, with particularly high accuracy (R2 = 0.9926) in darker-skinned individuals, who constituted the majority of the studied Iranian population. The results confirm that pallor analysis of palmar images using artificial intelligence techniques offers a reliable, non-invasive, and effective tool for pediatric anemia screening and hemoglobin estimation, with strong potential for point-of-care applications.