Hematocrit estimation from compact single hidden layer feedforward neural networks trained by evolutionary algorithm

Hematocrit is expressed as the percentage of red blood cells in the whole blood; it is the most highly influencing factor for measuring glucose in the whole blood by handheld devices. This paper presents hematocrit estimation from transduced current curves by using single hidden layer feedforward neural networks (SLFNs). These transduced current curves are produced by glucose-oxidase reaction in electrochemical biosensors which is used in glucose measurements. Points of the current curve sampled at frequency of 1OHz are used as the input features for the networks. Applying neural networks to hematocrit estimation has also proposed in our previous works. However, in this paper, the SLFN is trained by evolutionary least-squares extreme learning machine (ELS-ELM) algorithm in which the input weights and hidden layer biases are determined based on the differential evolution (DE). Experimental results show that the accuracy of hematocrit estimation on ELS-ELM can be improved, from which it can be used to reduce the dependency of hematocrit in measurement of glucose values by handheld devices.