brunch

You can make anything
by writing

C.S.Lewis

by KAIST ICLAB Sep 25. 2024

CGM-Glucose Monitoring Project

Summer '24 Internship Experience Report

Hello, my name is Gyeongju Lee. I’d like to share my internship experience at KAIST IC Lab. Overall, my internship at ICLab was a truly meaningful experience. I worked on glucose monitoring research which focused on using CGM (Continuous Glucose Monitoring) sensors and wearable sensor data to help prevent or manage diabetes. 


Diabetes is a chronic condition that significantly affects individuals' quality of life, often leading to long-term health complications. Despite the availability of various tools and technologies to monitor glucose levels, many patients struggle to manage their condition effectively due to a lack of personalized insights and proactive guidance. Current solutions primarily focus on reactive care rather than preventing spikes in glucose levels or guiding behavior changes before complications arise.


My research topic addresses this gap by focusing on predicting future glucose trends using CGM sensor data and wearable devices. By providing users with personalized, data-driven insights, we can help them make informed lifestyle choices to better manage their health (see related work by Pai et al., 2024 [3] and Barth et al., 2024 [4]).

The 8-week project was divided into two parts: the first four weeks and the last four weeks. 

 

※ CGM (Continuous Glucose Monitoring): A sensor that continuously tracks blood glucose levels in real-time, helping users monitor and manage their glucose fluctuations.

Fig1. CGM (Continuous Glucose Monitoring) Sensor


Research

During the first half of my internship, I tested collecting real-time blood glucose data by attaching a CGM (Continuous Glucose Monitoring) sensor. The CGM sensor continuously measures blood glucose levels when attached to the skin. To collect and store this data, I used an open-source mobile app called 'xDrip' (see [5]) along with a database called 'InfluxDB'. This helped me understand how the CGM sensor works and how to collect and analyze the data effectively.


In addition, as part of my basic research, I explored various books and papers to learn about lifestyle habits that help control blood glucose and prevent diabetes. I read books like "The Diabetes Code" and "당뇨병의 정석", which gave me valuable insights into scientific evidence and healthy habits for managing blood sugar. I also looked into how previous studies used different kinds of information to help prevent diabetes. This foundational research helped me understand the broader context of my work and provided important guidance for the data collection and analysis process.


Fig2. CGM Data Collection and Storage Process


In the second half of the internship, I developed an analysis pipeline to predict glucose levels based on data from CGM sensors, wristband wearable devices, and meal records (see the original data published via Bent et al., 2021 [1] and Cho et al., 2023 [2]). Wristband and smartwatch-type wearable devices provide various sensor data, including accelerometry, electrodermal activity (EDA), temperature, and photoplethysmography (PPG). This diverse set of sensor data allows for a more comprehensive analysis of factors influencing glucose levels, such as physical activity, stress, and physiological responses.


First, I enhanced the quality of the dataset through data preprocessing, addressing missing values to ensure data integrity. Following preprocessing, the feature extraction process involved referencing previous studies to identify key factors such as diet, stress, exercise, circadian rhythm, and behavioral habits. The data was then normalized to ensure uniform behavior of all variables within the model.


Afterward, I built predictive models for glucose levels using both a population-based approach with leave-one-subject-out cross-validation and a personalized approach with partial personalized cross-validation. These splitting methods were applied to construct an XGBoost regression model that predicts future glucose levels based on the extracted features.


This internship allowed me to deepen my understanding of data analysis and gain practical experience in applying machine learning techniques to real-world datasets.



Fig3. Wearable Sensor Data and Meal Log-based Glucose Prediction



Lessons Learned 

During these eight weeks, I learned a lot from my mentor, Professor Uichin Lee, who guided me throughout the project. I learned the importance of building solid arguments based on previous studies and the need to have a clear rationale when setting research directions. I also realized the importance of organizing research findings in a well-structured document. Professor Lee showed great passion and joy in research, and I was inspired to follow that example.


The advice I received from lab members also opened up new perspectives for me. There were boarding and mid-term presentations. During the presentations, lab members asked me valuable questions and gave suggestions on what I could study further and how I could expand my research. This feedback helped me understand my strengths and weaknesses in both my research and presentations.


Towards the end of the internship, I had a chance to talk with Professor Lee. He emphasized that persistence and continuous passion are key to successful research. He also explained that mentoring is not just about the mentor giving instructions but about both the mentor and the mentee learning from each other and growing together.


This internship was a valuable experience because I could work on research that aligned with my interests in HCI (Human-Computer Interaction), machine learning, and sensor data analysis. I truly enjoyed every moment of the eight weeks, immersed in topics that fascinate me.




Bio of the Author


Gyeongju Lee is currently an undergraduate student at the School of AI Convergence, Soongsil University. He is focusing on HCI, AI, machine learning, and data science. His research interests include topics such as HCI, wearable technology, and sensor data.


References

[1] Bent, B., Cho, P.J., Henriquez, M. et al. Engineering digital biomarkers of interstitial glucose from noninvasive smartwatches. npj Digit. Med. 4, 89 (2021). https://doi.org/10.1038/s41746-021-00465-w

[2] Cho, P., Kim, J., Bent, B., & Dunn, J. (2023). BIG IDEAs Lab Glycemic Variability and Wearable Device Data (version 1.1.2). PhysioNet. https://doi.org/10.13026/zthx-5212

[3] Pai, A., Santiago, R., Glantz, N. et al. Multimodal digital phenotyping of diet, physical activity, and glycemia in Hispanic/Latino adults with or at risk of type 2 diabetes. npj Digit. Med. 7, 7 (2024). https://doi.org/10.1038/s41746-023-00985-7

[4] Clara-Maria Barth, Jürgen Bernard, and Elaine M. Huang. 2024. "It's like a glimpse into the future": Exploring the Role of Blood Glucose Prediction Technologies for Type 1 Diabetes Self-Management. In Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI '24). Association for Computing Machinery, New York, NY, USA, Article 135, 1–21. https://doi.org/10.1145/3613904.3642234

[5] xDrip+ Documentation. https://xdrip.readthedocs.io/en/latest/

Image source: Connected in Motion Blog, https://www.connectedinmotion.ca/blog/tech-update/dexcom-g7-appears-on-health-canadas-list-of-approved-devices/




작가의 이전글 Reproducible Affect Prediction
브런치는 최신 브라우저에 최적화 되어있습니다. IE chrome safari