A Comparison of Feature Selection and Forecasting Machine Learning Algorithms for Predicting Glycaemia in Type 1 Diabetes Mellitus

Type 1 diabetes mellitus (DM1) is a metabolic disease derived from falls in pancreatic insulin production resulting in chronic hyperglycemia. DM1 subjects usually have to undertake a number of assessments of blood glucose levels every day, employing capillary glucometers for the monitoring of blood glucose dynamics. In recent years, advances in technology have allowed for the creation of revolutionary biosensors and continuous glucose monitoring (CGM) techniques. This has enabled the monitoring of a subject’s blood glucose level in real time. On the other hand, few attempts have been made to apply machine learning techniques to predicting glycaemia levels, but dealing with a database containing such a high level of variables is problematic. In this sense, to the best of the authors’ knowledge, the issues of proper feature selection (FS)—the stage before applying predictive algorithms—have not been subject to in-depth discussion and comparison in past research when it comes to forecasting glycaemia. Therefore, in order to assess how a proper FS stage could improve the accuracy of the glycaemia forecasted, this work has developed six FS techniques alongside four predictive algorithms, applying them to a full dataset of biomedical features related to glycaemia. These were harvested through a wide-ranging passive monitoring process involving 25 patients with DM1 in practical real-life scenarios. From the obtained results, we affirm that Random Forest (RF) as both predictive algorithm and FS strategy offers the best average performance (Root Median Square Error, RMSE = 18.54 mg/dL) throughout the 12 considered predictive horizons (up to 60 min in steps of 5 min), showing Support Vector Machines (SVM) to have the best accuracy as a forecasting algorithm when considering, in turn, the average of the six FS techniques applied (RMSE = 20.58 mg/dL).