Prediction Models of Blood Glucose Change During Aerobic Exercise Using Machine Learning Techniques

• Published in: Exercise science (Seoul, Korea) Vol. 32; no. 3; pp. 295 - 303
• Main Authors: Oyama, Okimitsu, Choi, Seonggyu, Oh, Changgeun, Kim, Eunchan, Park, Dong-Hyuk, Oh, Minsuk, Park, Dae-hyun, Seo, Hye-Kyoung, Han, jungsun, Jeon, Dongiae, Kim, Seong-Hyok, Jeon, Justin Y
• Format: Journal Article
• Language: English
• Published: 01.08.2023
• ISSN: 1226-1726; 2384-0544
• DOI: 10.15857/ksep.2023.00318

PURPOSE: This study aimed to explore the relationship between blood glucose level changes and body characteristics during exercise and to present six models for predicting changes in blood glucose levels during exercise.METHODS: 148 healthy men and women (age: 31.9±9.7 year, fasting blood glucose: 102.1±14.1 mg/dL, p=.032) participated in the study, and 30 of them participated in the study. Eight variables were selected to build two prediction models: 24-hour ingested carbohydrates, age, blood glucose, heart rate changes, sex, skeletal muscle mass, heart rate recovery after exercise, and resting heart rate. Logistic regression and random forest classifier models were used to predict the changes in blood glucose levels during exercise.RESULTS: A total of six models were created for all participants, male and female. Random forest classification (training set: AUC=0.837, Youden index=0.66; validation set: AUC=0.730, Youden index=0.53) and logistic regression classification models (training set: AUC=0.807, Youden index=0.55; validation set: AUC=0.735, Youden index=0.57) were built.CONCLUSION: The random forest model showed good performance in classifying internal data, whereas the logistic regression classification model demonstrated relatively good performance in classifying validation data.