MIRERC 030/2025 Adaptive Concept Drift Detection Technique for Glucose Monitoring in Diabetic Patients.

Abstract

Effective diabetes management depends on accurate glucose monitoring. However, this process is complicated by the dynamic nature of glucose levels, which fluctuate due to factors such as diet, medication, and individual metabolic responses. These fluctuations contribute to concept drift, a phenomenon in which the relationship between input variables—such as glucose levels, dietary habits, and medication—and outcomes like blood sugar control and complications evolve. As a patient’s condition, lifestyle, or treatment response changes, historical data becomes less reliable for predicting future trends. Consequently, failing to account for these shifts can lead to inaccurate insulin dosage recommendations or misclassification of risks, underscoring the need for continuous monitoring and adaptive predictive models. Several existing techniques are employed for glucose
monitoring and drift detection, yet each has notable limitations. Threshold- based methods rely on fixed glucose level thresholds to detect abnormalities. While simple and widely used, these methods struggle to adapt to gradual changes and individual patient variations. Statistical Process Control (SPC) techniques, such as Shewhart charts and CUSUM (Cumulative Sum), excel at detecting abrupt fluctuations but are less effective in identifying slow, evolving trends. Machine learning models, increasingly used for glucose prediction, often assume that past patterns remain valid, making them susceptible to accuracy degradation as concept drift occurs. Similarly,moving average techniques help smooth glucose variations but fail to actively detect shifts in the underlying data distribution, leading to delayed responses to critical changes. To address these challenges, this research introduces a novel adaptive concept drift detection technique tailored for glucose monitoring in diabetic patients. By building upon advanced principles from existing methods such as ADWIN (Adaptive Windowing) and the Drift Detection Method (DDM), this study aims to develop an innovative approach specifically designed to detect subtle shifts in glucose patterns. This study employs a quasi-experimental design using existing CGM datasets to develop an adaptive concept drift detection technique. Purposive sampling selects datasets with drift patterns, and machine learning algorithms are applied for data preprocessing, noise reduction, and drift detection. Data analysis involves statistical and performance evaluation metrics, including accuracy, precision, recall, and F1-score,  to compare the developed technique with existing methods. The findings aim to enhance glucose monitoring by
improving drift detection and adaptation accuracy.