Feature-enhanced residual signals with parallel hybrid network and novel loss for noninvasive blood glucose detection

Abstract

Diabetes remains a major global health challenge. According to the International Diabetes Federation (IDF), approximately 10.5% of the world’s population is affected, with nearly half unaware of their condition. Therefore, accurate and accessible blood glucose level (BGL) monitoring remains critical. This paper presents a noninvasive BGL estimation approach using in vivo residual infrared (IR) signals collected from patients’ fingertips, which is inherently challenging due to weak glucose-related signatures and interference from other physiological factors. To address this, we propose a tailored feature engineering strategy to uncover hidden relationships between residual signals and BGL. The proposed method improves BGL range discrimination, as indicated by increased Spearman correlations and consistent gains in Kendall and Pearson correlations, demonstrating its effectiveness. Furthermore, we propose a task-specific hybrid neural network architecture, consisting of parallel 1D convolutional and dense blocks to capture both local spectral–temporal patterns and global nonlinear relationships. A distribution-aware loss function is also introduced to reduce prediction bias toward the dataset mean, improving sensitivity across the full BGL range. For validation, the proposed model is compared with random forest (RF) and boosting-based methods, showing superior performance with a 14.11% mean absolute percentage error (MAPE). Clinical reliability is evaluated using Clarke error grid (CEG) analysis, where 84.7% of predictions fall within the accurate zone and 15.3% within the clinically acceptable zone. These results demonstrate the potential of the proposed approach for practical, low-cost, and portable noninvasive BGL monitoring.