Glucose data interpretation in pediatric diabetes using an artificial intelligence approach

Semi-automatic solutions to monitor and treat diabetes have been recently developed, including insulin pumps and continuous glucose monitoring devices. Integrating computational techniques with electrical, communication, and information systems offers significant opportunities. However, no decision support systems are capable of adequately managing and analyzing the data provided by these devices. As a result, the high specificity and complexity of the information generated cannot be effectively utilized in everyday clinical practice. Therefore, this paper proposes an artificial-intelligent-based approach to identify distinct patterns within the glucose readings of pediatric diabetic patients. The objectives are twofold: first, to cluster the data employing a dimensionality reduction technique based on autoencoders, and second, to classify the data using the labels derived from the clustering phase to profile the glycemic trends. Furthermore, the blind evaluation conducted by medical professionals on the clustering results has offered crucial clinical validation to the work carried out. The results highlight the effectiveness and reliability of the proposed approach, achieving a classification performance with accuracy values up to 98%. The data reduction step was fundamental to speed up the subsequent processes while improving the metrics. The medical evaluation allowed us to improve the work by finding a correspondence between experimental results and clinical value.