Blood glucose prediction using wearable sensors and dietary logs

Abstract

Diabetes is a significant and growing health issue that currently affects over half a billion people worldwide. One way to mitigate this problem is to focus on reducing the risk of prediabetic patients with lifestyle interventions, which can be enhanced by tracking the person with a continuous glucose monitor (CGM). Unfortunately, these devices are still expensive, so this thesis aims to develop a supervised machine learning model for predicting continuous blood glucose values using data from wearable devices, biological data, and dietary logs.

This study reviews the relevant literature, outlines the data collection methods, and applies two machine learning techniques: Random Forest Regressor and Long Short-Term Memory (LSTM) network. The Random Forest Regressor outperformed the LSTM network, achieving an average root mean square error (RMSE) of 0.81 mmol/L and a mean absolute percentage error (MAPE) of 11.1\%. These findings show that blood glucose predictions can be achieved using non-invasive methods, providing useful insights into how food and physical activities affect glucose levels. However, the accuracy of the prediction values is not yet good enough to be a replacement for an actual CGM.

These results highlight the potential for cost-effective and accessible tools to teach about the effects of nutrition and exercise, which could potentially reduce the incidence of diabetes by helping individuals make informed lifestyle choices. Future research should focus on refining these models with larger and more diverse datasets, expanding its feature with additional variables such as heart rate, stress levels, and sleep patterns, in order to improve the accuracy and reliability of predictions. Ultimately, this work contributes to the broader goal of utilizing technology to enhance preventive healthcare and improve the quality of life for individuals at risk of developing diabetes.