120 Blood Glucose Prediction Using Near Infrared Spectroscopy and Machine Learning Mahmud Iwan Solihin 1 , Fahri Heltha 2 , Muzaiyanah Hidayab 3 1,3 Fac. of Engineering, UCSI University Malaysia Cheras, Kuala Lumpur, 56000 2 Fakultas Teknik, Universitas Syiah Kuala Darussalam, Banda Aceh, 23333 e-mail: fheltha@unsyiah.ac.id I. INTRODUCTION Over 425 million people are currently living with Diabetes. Diabetes is a disease in which the body lose its ability to maintain blood glucose (BG) at healthy levels. Hence, diabetic patients must control their BG levels by external means/device. Most of the current methods that are used to measure BG today are invasive, which rely on extracting blood by piercing the body with needle and causes pain to the patient [1]. The application of spectroscopy for BG levels prediction as in other applications is basically based on fundamental concept of Lamberts beer law. Absorption spectra of chemical species (atoms, molecules, or ions) are generated when a beam of electromagnetic energy (i.e. light) is passed through a sample, and the chemical species absorbs a portion of the photons of electromagnetic energy passing through the sample. Lamberts beer law states that that the absorptive capacity of a dissolved substance is directly proportional to its concentration in a solution. The relationship can be expressed as shown in Eq. (1) [2].            (1) Where:  = absorbance  = the molar extinction  = length of the path light must travel in the solution in centimeters  = concentration of a given solution Spectroscopy has been widely applied nowadays into many fields as non-invasive/non-destructive quality assessment of substance [3][4][5][6]. Particularly, Near Infra-Red (NIR) spectroscopy can be used non-invasive quantitative measurement of the chemical properties of a material using this property stated in lamberts beer law. Several successful researches have been carried out to measure BG non-invasively using NIR spectroscopy. Chemometric methods such as Partial Least Squares (PLS) regression is usually used in studies to find a linear relationship between blood glucose levels and NIR spectra [7]. However, almost all these studies used relatively large and seemingly un-practical spectrometers. This project aimed at developing a predictive model which can estimate BG levels by using a practical handheld micro NIR spectroscopic device, i.e. micro NIR spectrometer. The prediction of BG levels from obtained NIR spectra data is then performed using machine learning (ML) approach. Different methods of ML used in this study to see the effectiveness of the BG prediction. The predictive regression results are then evaluated based on the coefficient of the determination (   ) and RMSE (Root Mean Squared Error) value. II. DATA COLLECTION AND METHODOLOGY Samples of NIR spectra and blood glucose levels were collected from 45 participants for this project. NIR spectra data that were collected using micro NIR spectroscopic instrument with wavelength range from 900-1700nm. Some participants have taken the data for more than once or twice at different time so that there are overall 90 data Abstract— Blood glucose management would ease diabetic patients and may cut the cost involved in their treatments. This project aims at developing a non-invasive blood glucose prediction model using machine learning based on NIR (near infrared) spectroscopy (spectra) data. NIR spectra data and blood glucose levels were collected from 45 participants, resulting 90 samples (75 samples for training and 15 samples for testing) in this project. These samples were then used to develop a predictive model using Support Vector Machine (SVM), Artificial Neural Networks (ANN) and AdaBoost ensemble algorithm. The results obtained from this project indicate that the handheld micro NIR has potential use for rapid non-invasive blood glucose monitoring. The coefficient of determination (R 2 ) obtained for training and testing dataset are 0.814, 0.832 and 0.828 for SVM, ANN and AdaBoost respectively. Keywords: blood glucose, near infrared spectroscopy, machine learning, regression