Abstract:
Hypoglycemia, defined as a blood glucose level below 70 mg/dL, is a serious risk for individuals with Type 1 Diabetes Mellitus (T1DM), potentially causing severe outcomes such as seizures or unconsciousness if not addressed in time. Continuous Glucose Monitoring (CGM) systems, though less invasive than finger-prick testing, suffer from a physiological lag of 5–20 minutes between blood and interstitial glucose levels, limiting their ability to warn patients early.
This study proposes a predictive model to anticipate hypoglycemic events ahead of time, helping patients take preventive actions like carbohydrate intake. Using time-series CGM
data, the model explores both univariate (CGM-only) and multivariate (including insulin and carbohydrate intake) inputs. It also addresses the challenge of data imbalance, with
a focus on achieving high precision and recall to reduce false alarms.
Results show that univariate models perform comparably to multivariate ones, making them practical for real-world use. Regression-based models also generalize better across test conditions than classification models. The model’s clinical validity is supported by Clarke error grid analysis, where over 98% of predictions fall in safe zones (A and B). This approach supports safer, proactive diabetes management through timely, CGM-based hypoglycemia prediction.
