Performance of artificial intelligence models in estimating blood glucose level among diabetic patients using non-invasive wearable device data

Abstract

Introduction

Diabetes Mellitus (DM) is characterized by impaired ability to metabolize glucose for use in cells for energy, resulting in high blood sugar (hyperglycemia). DM impacted 463 million individuals worldwide in 2019, with over four million fatalities documented. Blood glucose levels (BGL) are usually measured, as standard protocols, through invasive procedures. Recently, Artificial Intelligence (AI) based techniques have demonstrated the potential to estimate BGL using data collected by non-invasive Wearable Devices (WDs), thereby, facilitating monitoring and management of diabetics. One of the key aspects of WDs with machine learning (ML) algorithms is to find specific data signatures, called Digital biomarkers, that can be used in classification or gaging the extent of the underlying condition. The use of such biomarkers to monitor glycemic events represents a major shift in technology for self-monitoring and developing digital biomarkers using non-invasive WDs. To do this, it is necessary to investigate the correlations between characteristics acquired from non-invasive WDs and indicators of glycemic health; furthermore, much work is needed to validate accuracy.

Research Design & Methods

The study aimed to investigate performance of AI models in estimating BGL among diabetic patients using non-invasive wearable devices data An open-source dataset was used which provided BGL readings, diabetic status (Diabetic or non-diabetic), heart rate, Blood oxygen level (SPO2), Diastolic Blood pressure, Systolic Blood Pressure, Body temperature, Sweating, and Shivering for 13 participants by age group taken from WDs. Our experimental design included Data Collection, Feature Engineering, ML model selection/development, and reporting evaluation of metrics.

Results

We were able to estimate with high accuracy (RMSE range: 0.099 to 0.197) the relationship between glycemic metrics and features that can be derived from non-invasive WDs when utilizing AI models.

Conclusion

We provide further evidence of the feasibility of using commercially available WDs for the purpose of BGL estimation amongst diabetics.