Comparative Analysis of Predictive Interstitial Glucose Level Classification Models.

Area of Science:

• Biomedical Engineering
• Data Science in Healthcare
• Diabetes Technology

Background:

• Continuous glucose monitoring (CGM) offers real-time glucose level alerts, crucial for managing diabetes during meals and activity.
• CGM systems present challenges including sensor lag and data interpretation, necessitating advanced predictive models.
• Predictive glucose classification models are vital for optimizing insulin dosing and daily diabetes management.

Purpose of the Study:

• To evaluate the efficacy of three predictive models: ARIMA, logistic regression, and LSTM for glucose level classification.
• To assess model performance in predicting hypoglycemia, euglycemia, and hyperglycemia at 15-minute and 1-hour horizons.

Main Methods:

• Compared autoregressive integrated moving average (ARIMA), logistic regression, and long short-term memory (LSTM) networks.
• Evaluated prediction accuracy for hypoglycemia (<70 mg/dL), euglycemia (70-180 mg/dL), and hyperglycemia (>180 mg/dL).
• Calculated precision, recall, and accuracy using confusion matrices for 15-minute and 1-hour prediction intervals.

Main Results:

• ARIMA models showed underperformance in predicting hyper- and hypoglycemia across both time horizons.
• Logistic regression demonstrated superior performance for 15-minute predictions, achieving high recall rates for all glucose classes.
• LSTM models outperformed logistic regression for 1-hour predictions, particularly for hyper- and hypoglycemia.

Conclusions:

• Model selection for glucose prediction depends on clinical application requirements and desired prediction horizon.
• Logistic regression is optimal for short-term (15-min) glucose level predictions, especially hypoglycemia.
• LSTM models are more effective for longer-term (1-hour) glucose level forecasting, suggesting potential for advanced diabetes management.