Transfer learning for non-invasive glucose prediction under albumin interference in NIR spectroscopy

This study proposes a transfer learning framework for non-invasive glucose prediction using diffuse-reflectance near-infrared (NIR) spectroscopy, along with an in vitro phantom model that incorporates a pump-driven circulation system. Lipofundin and black ink were used to simulate blood-like scattering and absorption, respectively, to emulate realistic tissue conditions, while albumin was introduced as a representative spectral interferent. To investigate the model's adaptability under interference, a one-dimensional convolutional neural network (1D-CNN) pretraining strategy was evaluated with three datasets: solely non-interferent samples, incorporating a limited number of interferent samples, and combining both interferent and non-interferent samples. As a result, a model throughout pretraining and fine-tuning with data from combining both interferent and non-interferent samples yielded the best performance with an R² of 0.9115, RMSE of 10.5252 mg/dL, and MARD of 5.4679 %, respectively, highlighting its superior robustness and generalization ability in the presence of spectral interference. This approach provides a potential foundation for future applications involving real human data, particularly in scenarios where spectral variability may arise due to medication used in diabetic patients.