Abstract:  
Abstract: Current blood glucose detection methods are often invasive， causing inconvenience and potential risks. A non-invasive method using wearable devices to collect physiological data and user-input dietary information for real-time high blood glucose prediction is proposed. To improve prediction accuracy， a deep learning-based time series classification model， FA-CGRNet， is developed. Physiological data are preprocessed through denoising and resampling. Statistical features are extracted. A residual convolutional network is designed to extract and fuse features through convolution and residual connections. A feature enhancement module is utilized to calculate feature weights and perform feature selection. Finally， an LSTM model is employed to extract long-term dependency features from time series data. The model is tested on a public dataset from the BIG IDEAs Lab at Duke University. In the field of non-invasive blood glucose detection， the feature extraction method and network model presented in this study demonstrated superior performance compared to existing time series classification models. The model’s ability to distinguish between positive and negative cases is significantly enhanced. Notably， the weighted F1 score is improved by over 6.2%， while the AUC is increased by more than 2.5%. These results underscore the effectiveness of the proposed approach in advancing non-invasive blood glucose monitoring techniques.

Key words: Key words: non-invasive hyperglycemia prediction, residual convolutional, LSTM, feature enhancement module, neural network