Improving Glucose Prediction : Using Wearable Device Data in Machine Learning Systems

Sammanfattning

Managing blood glucose levels is a significant daily challenge for individuals with Type 1 Diabetes (T1D). Predicting blood glucose fluctuations using continuous glucose monitoring (CGM) data is essential for reducing complications. The data from these CGM devices are becoming more easily available, creating better opportunities to train machine learning models to predict glucose fluctuations and help patients manage their diabetes. This thesis investigates the potential for fully autonomous blood glucose prediction models without manual inputs. By analyzing how current machine learning models could benefit from incorporating additional contextual data, such as heart rate (HR) and time of day, collected from wearable devices, the study aims to give better insights into what can be done to improve performance in everyday scenarios. Three machine learning models were evaluated: a baseline LSTM model using CGM data, a similarly structured LSTM model that used heart rate in addition to CGM data, and a new CNN-LSTM using heart rate. The models were tested using data collected during everyday activities and their performance was measured using Root Mean Squared Error (RMSE) and Clarke Error Grid (CEG) analysis to assess clinical relevance. The results indicated that the new CNN-LSTM model achieved higher prediction accuracy than both the baseline and the LSTM with added HR. In entire everyday scenarios, the new model showed consistently lower RMSE values across all subjects without significantly increasing size, thus remaining small enough for use on mobile devices. Some results were inconclusive when testing the models on data collected from specific activities recorded months later, indicating the need for further research on how well the models generalize over extended periods and whether another training approach could help maintain accuracy. The study also shows that data availability plays a significant role in model performance, far outweighing the benefits of contextual data and model choices.