作者
Abhishek Panthri,V. Vijayabaskar,N Subbarayudu,J Lenin,R Arthi,Akhil Gupta
摘要
An effective forecasting of blood glucose can facilitate intelligent management of diabetes due to its ability to track the temporal fluctuations caused by different factors including physiological, dietary, and behavioral. The task is to design an intelligent deep learning model that teaches complex temporal dependencies within the multimodal glucose-related data. The approach involves temporal feature extraction using Mamba, multimodal feature fusion, representation learning using attention mechanism, glucose prediction, and analysis of interpretability. The uniqueness of the approach consists in combining state space temporal modelling with glucose, insulin, carbohydrate, physical activity, physiological, sleep, and stress data. The combination of the full Mamba, fusion, and attention setup results in the MAE of 4.18, RMSE of 6.27, R2of 0.941, and inference speed of 18.6 ms. Robustness testing leads to the R2of 0.911 for 20% noise level, whereas CGM glucose has 0.872 sensitivity and 34.8% relative influence. The GlucoBench dataset includes adult patients from UK and USA, having measurements every 5 minutes, with such features as glucose, insulin bolus and basal dose, carbohydrates, meals, physical activity, heart rate, skin temperature, galvanic skin response, sleep, stress, medications, HbA1c, age, gender, weight, carb ratio