Implementation of mathematical and machine learning models in handheld glucometer to predict whole blood viscosity

Abstract

Blood viscosity depends mainly on hematocrit and shear rate, and its low-cost and real-time estimation in handheld devices can enable rapid cardiovascular diagnostics and might revolutionize the translational research in healthcare. We developed a synergistic shear-dependent blood viscosity model by integrating and parametrically optimizing the Krieger–Dougherty (KD) and Carreau–Yasuda (CY) formulations to accurately match experimental observations. Furthermore, we calculated relative blood viscosity η(rel) for the amperometric dataset by incorporating the KD–CY model. The amperometric dataset contains peak current (Ip), corresponding time to the peak current (Tp), glucose concentration (Gc), and hematocrit volume (Hv). Besides, we developed a predictive model by utilizing several machine learning models to predict blood viscosity. The models were evaluated through the analysis of statistical parameters, namely, R-squared (R2) score, mean squared error (MSE), and root-mean squared error (RMSE). Moreover, the reliability of the models was tested via relative error (RE) and K-fold cross-validation techniques. It was concluded that the XGBoost and ANN outperformed other models because they have very relative errors (nearly 1%–3%). Our finding was consolidated by the 5-fold cross-validation technique, wherein the R-squared (R2) scores for XGBoost and ANN were observed to be 0.998 ± 0.000 and 0.999 ± 0.000, respectively.