Low cost point of care estimation of Hemoglobin levels

Anemia is a public health problem that affects populations in low-income and middle-income countries. World Health Organization defines anemia status by hemoglobin (Hb) concentrations. Our aim is to develop a method to estimate Hb levels in low resource settings. It is proposed to take an image of a drop of blood on a filter strip under controlled conditions and then estimate the Hb level of the blood using an image processing algorithm. In the first part of the paper we discuss the protocol used to simulate the controlled conditions needed to capture the image of the blood drop. In the second part of the paper, blood drop of 33 individuals were collected with the prototype. A classification tree and correlation based approach to feature selection was used to classify 4 levels of Hb (Class I - Hb above 12 g/dl, Class II - Hb between 10 and 12, Class III - Hb between 8 and 10, Class IV - Hb below 8) in the data. The features selected (`Y' from XYZ color space, `a' from Lab color space and `S' from HSI color space) from classification tree were used to train an artificial neural network. The data was partitioned into 17 training samples and 16 testing samples. The confusion matrix obtained on the testing set is desirable, with overall accuracy of 82%, sensitivity of 83% and specificity of 82%. This result on a small dataset is encouraging and shows that color image analysis of blood can be used to estimate Hb with the image being captured is under standard conditions.