Abstract:
The goal of this project is to evaluate various machine learning and deep learning models on patients from the D1NAMO dataset to predict blood glucose level based on heart rate values and additional features extracted from the smart watch. These features will then be processed with feature extraction techniques to identify the features that are most strongly correlated with the glucose values.
The suggested system will be patient-dependent and operate in the following manner: each patient will need to get the two devices which are the CGM and the smart watch, while the CGM device measures glucose levels over the course of two weeks, the smartwatch simultaneously measures heart-rate measurements, and both devices will be calibrated together simultaneously.
After The data is done being extracted from the two devices mentioned, the data will be taken to be processed, they will be processed by removing any noise that might interfere with the ECG signal, also due to the time measuring differences between the two devices, moving average techniques needed to be applied on the ECG readings in order to be compatible with the glucose’s time readings. This technique is applied because the CGM device measures the glucose values every 5 minutes while the smartwatch measures heart rate every millisecond.
The Processed and calibrated data will now be used as a training dataset for the different models that will be used throughout this project. It was found that among the different machine learning models that were used, the extra trees regressor model scored the least MSE (mean squared error) and among the deep learning, neural network models, the ANN (artificial neural network) model scored the least MSE.