Wearable X AI To Predict Blood Glucose According To Your Daily Rhythm

3 main points
✔️ We have proposed the GluMarker, a comprehensive next-day blood glucose control prediction model that takes into account a wide range of factors, including dietary intake.
✔️ It achieves higher prediction accuracy than conventional methods and identifies important digital biomarkers.
✔️ From the identified biomarkers, we have identified lifestyle factors that affect glycemic control, such as corrected insulin administration and blood glucose status on the previous day.

GluMarker: A Novel Predictive Modeling of Glycemic Control Through Digital Biomarkers
written by Ziyi Zhou, Ming Cheng, Xingjian Diao, Yanjun Cui, Xiangling Li
(Submitted on 19 Apr 2024)
Comments: Published on arxiv.
Subjects: Artificial Intelligence (cs.AI)

code：  

The images used in this article are from the paper, the introductory slides, or were created based on them.

Introduction

Digital biomarkers are a new type of medical indicator derived from data collected from digital devices such as wearable devices and smartphones. Unlike pathology markers derived from traditional biological samples, digital biomarkers can capture behavioral patterns, physiological rhythms, and environmental factors in detail.

In recent years, digital biomarkers have attracted attention in a variety of fields, including chronic disease management and mental health monitoring. In the field of diabetes, in particular, digital biomarkers have been studied to help individuals manage their blood glucose levels by utilizing data obtained from continuous glucose monitoring (CGM) devices and insulin pumps.

However, most previous studies have focused on insulin dosage and individual blood glucose levels, and have yet to assess comprehensive glycemic control. Therefore, this study proposes a digital biomarker modeling method that can predict overall blood glucose status on the following day, taking into account lifestyle factors such as dietary intake.

Related Research

Traditional research has taken the following approaches to diabetes management using digital biomarkers

Bartolome et al. [6] propose a computing framework for identifying digital biomarkers from CGM and insulin pump data to manage glycemic control in diabetic patients.

Bent et al [16] developed an approach that allows noninvasive monitoring and prediction of personalized interstitial glucose (interstitial glucose) levels from smartwatch sensor data and dietary records. An important contribution is the individualized definition of hyperglycemia (PersHigh), hypoglycemia (PersLow), and normal (PersNorm) based on individual blood glucose variability.

However, these previous studies have tended to be limited to insulin administration and blood glucose data and have yet to assess comprehensive glycemic control. Some studies are limited to predicting short-term glycemic variability.

Thus, conventional research has had the challenge of failing to capture the wide range of factors involved in blood glucose control, and the application of digital biomarkers to comprehensive diabetes management has been called for.

Proposed Method (GluMarker)

In this paper, we propose a framework called GluMarker that takes into account a wide range of factors, including dietary intake, to predict the overall glycemic control status on the following day.

First, as a preprocessing step, the data are divided into multiple intervals based on the distribution of the input data (Figure 2). For example, dietary intake is divided into 0-120 units, 120-200 units, etc. Each interval is represented as a digital biomarker to discretize continuous value data.

The model structure of GluMarker (Figure 1) is described next. We have two branches with parallel input of continuous value data and discretized digital biomarker data. The continuous branch is trained with a CNN and the discrete branch is trained with a tightly coupled layer of feature representations.Then, an attention mechanism (cross-attention) is applied to the two feature representations to effectively integrate them. Finally, the integrated features are used to predict the state of blood glucose control (good/moderate/poor).

Thus, GluMarker is characterized by its ability to efficiently fuse input data of different natures, continuous and discrete, through the introduction of parallel branches and attention mechanisms. Higher prediction performance than conventional methods can be expected.Furthermore, by extracting important features from the model, daily lifestyle factors that affect blood glucose control on the following day are revealed.

Experiment

Figure 4 shows the ROC curves of the prediction performance of the four models (proposed method GluMarker, Linear SVC, Naive Bayes, and MLP) for glycemic control status.

GluMarker exhibited the highest AUC values (area under the curve) in all three categories: good, moderate, and poor. It performed particularly well in predicting poor control status, with an AUC=0.85. This is a practically important result given the reality of the difficulty of proper glycemic control by diabetic patients.

Figure 5 shows the extent to which each characteristic affects the three glycemic control categories. For example, we see that the previous day's corrected insulin dose contributes most to good control, while the previous day's hyperglycemia time (TAR) is a major factor in poor control.

In addition, Figure 6 visualizes the top 10 digital biomarkers that are important in predicting glycemic control. For good control, "no corrected insulin the day before" was most important. On the other hand, for poor controls, "TAR=0% the day before" and "10-20 units of corrected insulin on the day" were at the top of the list. The latter is a counter-intuitive finding, but it is thought to be due to the large number of patients with poor glycemic control in this data set.

Based on these results, GluMarker not only achieved higher prediction performance than conventional methods, but also was able to extract important factors affecting blood glucose control from daily lifestyle habits. In particular, the importance of clinically meaningful indicators such as the presence or absence of corrected insulin administration, food intake, and the previous day's blood glucose status is an interesting finding.These digital biomarkers can serve as reference indicators for physicians' daily glycemic management of their patients. Furthermore, they can be expected to help patients improve their own lifestyle through the provision of personalized feedback.

Conclusion

In this study, we proposed the GluMarker framework to accurately predict the overall glycemic control status on the following day, taking into account a wide range of factors such as the amount of food consumed. Along with outperforming conventional methods, we identified clinically useful digital biomarkers such as corrected insulin dosage and previous day's blood glucose status.

In the future, they plan to integrate data on stress, exercise, and behavioral habits to assess performance in various patient populations. Further combined with individualized blood glucose prediction models, it will provide more comprehensive personalized diabetes management.

• 
• 
• 

Categories related to this article

• Medical Care
• Machine Learning
• Deep Learning
• Medical
• Natural Language Processing
• Prediction
• Neural Network