Estimation of Hematocrit Volume Using Blood Glucose Concentration through Extreme Gradient Boosting Regressor Machine Learning Model.

Abstract

Lifestyle diseases such as cardiovascular disorders, diabetes, etc. affect the physiological metabolism and become chronic upon negligence. Diabetes is one of the key factors that is interlinked with a plethora of diseases. Health management can be achieved through balanced diet, physical exercise, and periodic examination of blood glucose level and hematocrit volume. Our study developed a model to estimate the hematocrit volume (red blood cells) from the correlation of the glucose concentration obtained from a glucometer by employing machine learning techniques. This Article explores the prediction of hematocrit volume in whole blood by applying various machine learning (ML) models such as linear regression (LR), support vector regressor (SVR), decision tree (DT), random forest regressor (RFR), artificial neural network (ANN), and extreme gradient boosting regressor model (XGBoost). We used amperometric signals generated from an electrochemical glucose sensor or glucose strip, which produces current values on glucose concentration. We estimated the hematocrit volume via processing of the amperometric signals to enhance diagnostic capabilities with the least error in the field of biomedical signal processing. The ML models were trained on the data set comprising 80% training set and 20% testing set in the Python programming language. The models were evaluated based on the metrics such as R-squared (R²), mean squared error (MSE), and root mean squared error (RMSE) values, and their reliability was assessed through the three validation mechanisms, namely, the relative error, K-fold cross-validation, and analysis of confidence interval. We observed that the XGBoost regression results were comparatively better than the LR and ANN results as corroborated through reliability analysis. It was concluded that XGBoost demonstrated 15% relative error between actual and predicted data and 68% accuracy with 6% standard deviation in the prediction obtained via a 5-fold cross-validation technique. The XGBoost model demonstrates comparatively better performance in terms of flexibility in tuning and interpretability options, which make it suitable for the regression task in the predictive biomedical analytics.