Transfer learning and data augmentation for glucose concentration prediction from colorimetric biosensor images

Abstract

A deep learning algorithm is introduced to accurately predict glucose concentrations using colorimetric paper sensor (CPS) images. We used an image dataset from CPS treated with five different glucose concentrations as input for deep learning models. Transfer learning was performed by modifying four established deep learning models—ResNet50, ResNet101, GoogLeNet, and VGG-19—to predict glucose concentrations. Additionally, we attempted to alleviate the challenge of requiring the large amount of training data by applying data augmentation techniques. Prediction performance was evaluated using coefficients of determination (R²), root mean squared error (RMSE), and relative-RMSE (rRMSE). GoogLeNet showed the highest coefficient of determination (R² = 0.994) and significantly lower prediction errors across all concentration levels compared with a traditional machine learning approach (two-sample t-test, p < 0.001). When data augmentation was performed using 20% of the entire training dataset, the mean prediction error was comparable to that of the original entire training dataset. We presented a novel approach for glucose concentration prediction using deep learning techniques based on transfer learning and data augmentation with image data. Our method uses images from CPS as input and eliminates the need for separate feature extraction, simplifying the prediction process.

Graphical Abstract