Iron Deficiency Anemia Diagnosis for Young Children: Image-Driven Comparative Analysis of Recurrent and Projected Neural Networks

Iron deficiency anemia (IDA) is a major public health issue, especially affecting young children, and demands efficient, non-invasive diagnostic methods. IDA, marked by low red blood cell (RBC) counts due to insufficient iron, can cause fatigue and hinder cognitive development, particularly in resource-limited areas. Traditional diagnostic approaches, which involve invasive and costly blood tests, are often impractical for children due to discomfort, the need for skilled personnel, and variability in results. Machine learning offers a promising alternative by using non-invasive imaging and automated analysis to provide quick and accurate diagnoses. This study explores the application of machine learning for diagnosing IDA using palpable palm images collected from hospitals in sub-Saharan Africa. The dataset comprises images of children aged five and under, categorized into anemic and non-anemic groups based on corresponding hemoglobin (HGB) values provided by laboratory tests. Techniques including image preprocessing (e.g., resizing, gray-scaling, mean-variance normalization), segmentation, feature extraction (e.g., color, texture with GLCM, edge detection), and data augmentation were employed to enhance dataset quality and balance class sizes. Comparative analysis is conducted using a set of recurrent networks, namely, Gated Recurrent Units (GRU), Projected GRU (PGRU), Long Short-Term Memory (LSTM), and Projected LSTM (PLSTM) models to assess diagnostic accuracy. All algorithms' hyperparameters were tuned based on generalization errors as an objective function of a Bayesian optimization algorithm. Results demonstrate the efficacy of GRU in accurately classifying IDA, with a recall of 100<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0</sup>/0, offering a promising approach for diagnosis in pediatric populations.