連続血糖測定および妊娠糖尿病リスク予測のためのハイブリッド深層学習フレームワーク

Gestational diabetes mellitus (GDM) affects almost 10%-12% of pregnancies worldwide, threatening maternal and fetal life. Continuous glucose monitoring (CGM) forms the backbone of managing GDM, and the current methodologies largely disregard physiological and behavioral factors, thereby greatly reducing accuracy and clinical interpretability.

A hybrid deep learning framework was developed by fusing CGM with multi-sensing modality data, including heart rate, activity levels, sleep patterns, and dietary intake. For data preprocessing, Kalman filtering was applied for temporal alignment, adaptive normalization provided outlier handling and imputation, while the CNN-BiLSTM backbone with attention was harnessed for feature extraction. A Multi-Task Attention Fusion Network (MTAFN) was used to predict glucose values and classify GDM risk simultaneously, while SHAP and dynamic smoothing contributed to interpretability sets.

The framework was validated on an extended OhioT1DM dataset with adaptations for pregnancy. It reached a glucose prediction RMSE of 9.8 mg/dL and a GDM risk classification accuracy of 93%. Compared to competitive approaches, the present solution attained a 25% better accuracy on interpretability and an improvement in sensitivity and specificity of about 4-6% across various physiological conditions.

The use of multi-sensing data increased prediction robustness by capturing complex physiological dependencies. The SHAP-based interpretability justified the predictions through a physiological lens. With an attention mechanism for feature weighting, it was possible to identify crucial variables like meal intake and nighttime variability in the workflow sets.

The hybrid framework proposed here is reliable for clinically interpretable continuous glucose monitoring and GDM risk predictions. Its application with high reliability can lead to integrating it within clinical protocols for real-time maternal care sets.