Comparison of machine learning algorithms for blood glucose prediction on AIDA simulator

Abstract

Predicting blood glucose (BG) values has an increasing interest along with the recent progress in processing capacity of computers and spreading of mobile devices. Inspired from existing research studies, this study aims to use a BG simulator program, AIDA, to generate BG values and make predictions. Thus, comparing results to existing studies has directed this objective to provide an in-silico testing. Other points in using a simulator instead of real patient data is that it is easy to collect data, and it disregards external factors like pregnancy or stress. For estimates with prediction horizons (PH) with 15,30 and 60 minutes, support vector regression (SVR), decision tree regression, Gaussian process regression, k-NN regression, random forest regression and for neural networks: recurrent neural network (RNN) with long short-term memory (LSTM) unit and neuro-fuzzy network and feed-forward neural network (FFNN)have been employed. Among multiple algorithms neuro-fuzzy network (ANFIS) has the best results with RMSE values of 1.19 mg/dl, 2.53mg/dl and 5.81mg/dl for 15,30 and 60 minutes prediction horizons (PH). The audience for this paper is the research community who works on BG prediction and looking for ways to design a model for an algorithm for their selected set of inputs. This study presents a guide to selecting an algorithm and build a model for in silico simulation. This research can be extended to real world data or converted into a tool to create benchmark tests for models with given features and hyperparameters.|Keywords : Diabetes Mellitus, Machine Learning, AIDA, Blood Glucose Prediction.