Diabetes is a common chronic condition occurs due to imbalance of blood glucose levels in the body.In this paper, a forecasting model based on deep learning algorithm is proposed for accurate blood glucose level prediction in both short (15 minute) and long (30 minute) forecasting scenarios. The proposed model is based on a Multiheaded Convolution Neural Network (MHCNN) and multiple convolutional layers for extracting useful features providing meaningful information for pattern formation. The model is tested and trained on 30 subjects that includes 10 subjects of three different categories namely adult, adolescent and child on UVA/Padova dataset. The proposed model is tested against three types of Long Short-Term Memory (LSTM) networks namely Vanilla LSTM, Layered LSTM and Bidirectional LSTM. In addition, the performance of MHCNN is compared with CNN exhibiting the role of multiple heads in architecture.The MHCNN outperformed existing LSTM models and CNN in terms of accuracy and time of execution, according to preliminary experimental results. MHCNN is much faster than other models.

Index Terms-Diabetes, Convolution Neural Network, Multiheaded, LSTM, Deep Learning, Forecasting