Time Series Based Short Term T1DM Prediction of Librepro CGM Sensor Data: A Novel Ensemble Method
Rekha Phadke1, Varsha Prasad2, H C Nagaraj3

1Rekha Phadke*, ECE Department, Nitte Meenakshi Institute of Technology, Bangalore, India.
2H C Nagaraj, ECE Department, Nitte Meenakshi Institute of Technology, Bangalore, India.
3Varsha Prasad, ECE Department, Nitte Meenakshi Institute of Technology, Bangalore, India.
Manuscript received on July 20, 2019. | Revised Manuscript received on August 10, 2019. | Manuscript published on August 30, 2019. | PP: 1695-1704 | Volume-8 Issue-6, August 2019. | Retrieval Number: F8420088619/2019©BEIESP | DOI: 10.35940/ijeat.F8420.088619
Open Access | Ethics and Policies | Cite | Mendeley
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

Abstract: As per statistics over 30 million in India have been diagnosed with diabetes. There is an enormous need and development to be made to recognize the possible fluctuation of blood glucose before hand with minimal errors and thereby enabling proactive decision making.. The present work details out the algorithms used for glucose prediction and makes a relative assessment of glucose prediction of Librepro Continuous Glucose Monitoring (CGM) sensor data of Type 1 Diabetes Mellitus (T1DM) subjects. For the development and evaluation of the model, 10 days observation data of 10 different subjects with T1DM recorded at every 15 minutes time interval is considered. The model’s predictive performance is evaluated for one step ahead (15 minutes prediction horizon), two step ahead (30 minutes prediction horizon) and three step ahead (45 minutes prediction horizon) under univariate glucose prediction model. A novel hybrid data driven model which combines both linear regression and auto regression method is designed and developed for glucose prediction. This novel data driven model gave satisfactory performance metrics of MAPE value of 3.22 and RMSE of 7.38 mg/dl over the complex ARIMA model which requires proper selection of parameters to be chosen beforehand. In this paper an attempt has been made by the author to propose an ensemble method towards data driven model for glucose prediction under time series forecasting.
Keywords: ARIMA, CGM Sensor, Linear Regression, Overlapping forward window rolling technique, Time Series Forecasting.