Development of an Interpretable AI-Assisted Framework for Contextual Interpretation of Continuous Glucose Monitoring Patterns
Abstract
Purpose
Continuous glucose monitoring (CGM) provides high-resolution time-series data reflecting real-world glucose dynamics. However, most existing CGM analytics emphasize absolute glucose values, numerical thresholds, or food-centric interpretations, often overlooking how glucose patterns evolve in response to non-nutritional physiological and behavioral factors such as sleep quality, breathing patterns, physical activity, circadian rhythm, and emotional stress. This work presents the conceptual development of an interpretable AI-assisted framework that translates clinically intuitive glucose pattern interpretation into a structured computational system.
Methods
Rather than attributing glucose variability solely to food intake, the framework focuses on temporal glucose dynamics, including rapid dips, rebound spikes, sustained plateaus, gradual trends, and oscillatory fluctuations, as meaningful indicators of underlying physiologic state. The system operates on post-processed CGM data exported in CSV format and analyzes glucose patterns in relation to contextual parameters such as time of day, sleep and rest periods, activity intervals, and breathing-related indicators. A rule-guided decision layer integrates glucose dynamics with contextual information to generate simple, human-readable, non-pharmacological behavioral feedback. An anonymized CGM dataset from a healthy adult following a nutrient-dense diet including adequate protein, fat, and complex carbohydrates is used solely as a worked example to demonstrate system logic, data flow, and interpretability. The framework does not perform diagnosis, prediction, or automated intervention, and its algorithmic structure remains unchanged across users.
Results
This work suggests that glucose stability may be better characterized by smooth temporal variation rather than isolated maximum or minimum values.
Conclusion
By emphasizing pattern dynamics over numerical ranges, the proposed framework provides a foundation for explainable AI systems that support early recognition of physiologic instability and promote preventive behavioral awareness using wearable sensor data.