Detection of Physical Activity Using Machine Learning Methods Based on Continuous Blood Glucose Monitoring and Heart Rate Signals

Abstract

Non-coordinated physical activity may lead to hypoglycemia, which is a dangerous condition for diabetic people. Decision support systems related to type 1 diabetes mellitus (T1DM) still lack the capability of automated therapy modification by recognizing and categorizing the physical activity. Further, this desired adaptive therapy should be achieved without increasing the administrative load, which is already high for the diabetic community. These requirements can be satisfied by using artificial intelligence-based solutions, signals collected by wearable devices, and relying on the already available data sources, such as continuous glucose monitoring systems. In this work, we focus on the detection of physical activity by using a continuous glucose monitoring system and a wearable sensor providing the heart rate-the latter is accessible even in the cheapest wearables. Our results show that the detection of physical activity is possible based on these data sources, even if only low-complexity artificial intelligence models are deployed. In general, our models achieved approximately 90% accuracy in the detection of physical activity.

Keywords: diabetes mellitus; machine learning; physical activity detection.

Figure 1

The ways of data extraction from the Ohio T1DM dataset for a given patient. Black arrows indicate 24 h long blocks according to the time stamps, midnight to midnight. Legend: blue blocks—CGM data available, exercise reported, no HR data available; green blocks—CGM data available, exercise reported, HR data available; transparent orange area—self-reported exercise, no CGM data available; transparent red area—probably an exercise event happened, but it was not reported.

Figure 2

Abstraction of feature extraction during operation. The v, $vp$, $vpp$, and $ap$ features originate from the d-kind features; thus, these are not listed here. In the case of the $d{p}_i$, only the first four values are represented as a demonstration; however, all sampled values are considered from the window during operation.

Figure 3

ROC curve of tested ML models for Ohio T1DM dataset using glucose features only.

Figure 4

ROC curves of various ML models obtained on D1namo dataset using glucose features only.

Figure 5

ROC curves of models on Ohio T1DM dataset using both blood glucose and heart rate features.

Figure 6

ROC curves of models obtained on D1namo dataset with blood glucose and heart rate features.

Figure 7

ROC curve of models on Ohio T1DM is the training data; the D1namo is the test with glucose and heart rate feature.

Figure 8

AUC of models in all use-cases—test results.