Evaluation of Optimally Tuned K-Nearest Neighbors for 30-Minute Blood Glucose Prediction in Type 1 Diabetes Using OhioT1DM Dataset

Abstract

Diabetes is a long-term chronic medical condition with the potential to evolve into a global healthcare crisis, glycemic control is fundamental for the effective management of diabetes and the prevention of its associated complications. Forecasting future blood glucose levels (BGLs) for diabetic patients can help them avoid serious health problems. This study investigates the application of the KNN regression algorithm to predict future (BGLs), utilizing historical blood glucose measurements from twelve patients (six patients from the Ohio dataset version 2018 and six patients from the Ohio dataset version 2020) as the only input feature. Our proposed approach employed a methodology that utilized historical measurements to train predictive models. Specifically, we leveraged the following historical data points - (BGLs) at 4-hours, 8-hours, 12-hours, 16-hours, 20-hours, and 24-hours intervals - as input features to predict (BGLs) 30 minutes into the future. This study explores the impact of varying parameters of the KNN algorithm, such as the K value= [2,3,5,7,11], weights= ['uniform', 'distance'] and distance metric= ['euclidean', 'manhattan', 'minkowski'], on the performance of the model. Furthermore, we compared the obtained results of the KNN algorithm with other machine learning methods, including linear regression, Random Forests, Support Vector Machines, CatBoostRegressor, LightGBM, XGBoost, artificial neural networks and previous studies. Among these, KNN yielded the best results with optimal hyperparameters (k=2, Weights='distance', Metric='manhattan') in the tow version of datasets OhioT1DM V2018 and OhioT1DM V2020. The OhioT1DM V2018 dataset yielded optimal performance with an RMSE of 5.09 ± 0.91 mg/dl using a 24-hour window size, and an MAE of 2.42 ± 0.34 mg/dl with a 12-hour window size. For the OhioT1DM V2020 dataset, the best results were an RMSE of 5.56 ± 1.14 mg/dl with a 12-hour window size, and an MAE of 2.47 ± 0.34 mg/dl achieved using an 8-hour window size. This research confirms that KNN algorithm with optimal hyperparameters (k=2, Weights='distance', Metric='manhattan') can effectively predict blood glucose events, which will help prevent and reduce the occurrence of serious complications such as hypoglycemia and hyperglycemia.