Trustworthy Glucose Prediction Using Federated Learning: Privacy Preserving as a Service

Diabetes is one of the most difficult chronic diseases to manage, especially among working individuals and elderly people with limited mobility. Without adequate time to monitor the blood sugar levels, many individuals find it particularly difficult to keep tabs on their glucose. This paper proposes a solution to this issue leveraging Deep Learning (DL) methods, particularly Long Short-Term Memory (LSTM) networks, which are a type of Recurrent Neural Networks (RNN) ideally designed for time series forecasting. To automate the process of diabetes management, the LSTM model is designed to predict glucose levels by being trained on active data collected from the users in real time. For the model to learn without the sensitive data being exposed, data privacy and protection mechanisms are established using Federated Learning (FL). This method allows sensitive information to remain securely stored within devices while ensuring the advantages of joint model training and reduced vulnerability risks from storing shared data. The healthcare centers provide the combination of LSTM and FL and host it in a cloud environment as a service or as web services enabling patients to receive blood glucose level predictions in real time without visiting the clinic. The proposed system shows an accuracy of LSTM with FL which results to 95% and only LSTM as such receives 98% with loss being minutely less of 1% compared to each other. The total error reduced using RMSE and MAE are 0.17 and 0.02 for plain LSTM and 0.29 and 0.09 for LSTM with FL.