A Low-Cost Test for Anemia Using an Artificial Neural Network

Abstract

BackgroundAnemia during pregnancy can complicate maternal and neonatal health and even lead to fatal consequences if not diagnosed early on. Around 99% of women who face maternal mortality are from middle or low-income countries. Early screening of anemia could facilitate improved health outcomes in pregnant women. Point of care techniques are preferred due to their ability to provide results rapidly and because they can be used by personnel with minimal or no training. Such techniques are especially useful in resource-constrained settings like rural parts of developing countries. ObjectivesThe aim of the study was to develop a tool using an Artificial Neural Network (ANN) to estimate hemoglobin values using color information recorded from blood sample images. Our method utilizes inexpensive consumables and a simple image acquisition setup that can be assembled easily. MethodsThis study explores a neural network model to estimate the hemoglobin content in an individual’s blood sample. Blood samples were collected from 86 volunteers and the images of blood drops were obtained using an image acquisition setup designed by the team. The color intensity values calculated from the blood drop images were used as feature descriptors for the samples. The features obtained from our samples were consequently fed to the Artificial Neural Network. ResultsOur neural network that gives the best result has the architecture of 11 neurons in each of the 5 layers. The best model gave estimated hemoglobin levels by analyzing color of blood samples with an accuracy of ±1.8 g/dl Limits of agreement (LOA) and bias 0.03 g/dl (with mean error of 0.75 g/dl). The model was subsequently tested with a validation set prepared from an additional 65 samples. The estimated hemoglobin levels gave an accuracy of +2 g/dl to -1.9 g/dl Limits of agreement (LOA) and bias 0.06 g/dl (with mean error of 0.78 g/dl). ConclusionOptimization of sensitivity and specificity has been able to achieve the sensitivity and specificity values as 95.5% and 52% respectively. These results are at par with the contemporary measurement techniques indicating that our method can be used as a workable screening technique itself.