تشخيص و هشدار زودهنگام هيپوگليسمي شبانه در ديابت نوع 1 با استفاده از معماري LSTM-Transformer

Type 1 Diabetes Mellitus (T1DM) requires continuous glucose management an‎d intensive insulin therapy, yet nocturnal hypoglycemia remains a serious clinical problem because glucose can decline during sleep without the patient recognizing the early warning signs. Conventional Continuous Glucose Monitoring (CGM) systems mainly provide reactive threshold-based alarms, which may leave limited time for preventive intervention. This limitation, together with the nonlinear an‎d highly dynamic nature of glucose regulation, creates a need for intelligent prediction methods that can use information beyond glucose measurements alone. The present thesis therefore investigates the development of a multimodal deep learning framework for the early prediction of nocturnal hypoglycemia in individuals with T1DM. The proposed approach integrates CGM measurements with complementary physiological information obtained from wearable sensors. The methodology uses the D1NAMO dataset as the primary data source an‎d the OhioT1DM dataset for additional validation. The data are subjected to quality assessment, missing-value treatment, temporal synchronization, nocturnal extraction, normalization, an‎d fixed-length sliding-window construction. A two-hour observation window is used to represent the recent physiological state, while future-oriented labeling identifies whether hypoglycemia is expected within the prediction horizon. The predictive architecture combines a Bidirectional Long Short-Term Memory (BiLSTM) component with a Transformer-based representation module an‎d temporal attention. The BiLSTM is used to capture sequential dependencies in multimodal physiological signals, whereas the Transformer models broader temporal relationships through self-attention. The learned representations are subsequently integrated before the final classification layer estimates the probability of an impending nocturnal hypoglycemic event. The experimental results demonstrate that the proposed multimodal framework provides strong predictive performance on the independent test dataset. The final model achieved 97.53% accuracy, 79.01% recall (sensitivity), 98.38% specificity, a ROC-AUC of 0.984, an‎d a PR-AUC of 0.854. The confusion matrix showed that 207 of 262 hypoglycemic windows were correctly identified, while 55 were missed; among 5,683 non-hypoglycemic windows, 92 were incorrectly classified as positive. The results indicate that the model can distinguish future hypoglycemic an‎d non-hypoglycemic conditions with high overall discriminative capability while maintaining a relatively low false-alarm rate. Overall, the findings support the value of combining glucose information with complementary physiological signals an‎d of using recurrent an‎d attention-based architectures together for nocturnal hypoglycemia prediction. The proposed framework provides a methodological basis for future development of personalized, real-time, an‎d clinically deployable intelligent warning systems for safer diabetes management during sleep.

جعفر قيصري

امير اخوان بي تقصير , احسان روحاني