A machine learning for transcutaneous bilirubinometer

Abstract

This research aimed to develop a reliable, non-invasive, and cost-effective prototype of a bilirubinometer for measuring bilirubin concentration. By using the principle of measuring light reflection spectrum from bilirubin sample as the input to a Machine Learning (ML) engine, the bilirubin concentration can be accurately predicted. The proposed bilirubinometer prototype consists of four main parts: light source, fiber probe, light spectrometer, and an Artificial Neural Network (ANN) that are specially engineered for the bilirubinometer construction. The light spectrometer and light source are controlled by a microcontroller platform called Raspberry Pi. In the data collection phase, the bilirubinometer prototype was used to measure the light reflection from the filter paper samples that were dripped in bilirubin solution at concentration 0, 10, 20, 30, 40, and 50 mg/dl. The spectrum of the reflected light was used for training the neural network in order to determine the bilirubin concentration. The multilayer error back-propagation learning method was chosen, and an additional experiment was conducted for neural network optimization. The neural network optimization varied the number of layers and number of neurons in each layer. From the results, the best neural network model has an average error value that predicted deviation from the actual concentration approximately 1.031 mg/dL or around 4.6% and the standard deviation of error value is 1.026.