Transfer contrastive learning for Raman spectra data of urine: Detection of glucose, protein, and prediction of kidney disorders

Abstract

Raman spectroscopy, a non-invasive analytical technique, reveals significant potential in clinical diagnosis of kidney disorders by detecting key biomolecules in urine samples, especially glucose and protein. Although machine learning models have been widely applied for efficiently analyzing Raman spectral data, the high-dimensionality, imbalance and sample-scarcity of Raman spectral data still pose challenges to the models in achieving accurate detection. To address these challenges, we propose a novel deep learning model, TCRaman, which integrates transfer learning and contrastive learning for urine detection using Raman spectral data. As contrastive learning is capable of representation learning on imbalanced data, TCRaman first utilizes a pretrained contrastive learning model on a large labeled Raman spectral dataset of bacteria, to enhance the model’s capability to learn meaningful low-dimensional representations from high-dimensional Raman spectral data. Then, the pretrained model is finetuned on clinical urine Raman spectral data. This transfer learning framework is a foundation model that can break through the limitation of sample-scarcity on different downstream tasks. The experiments demonstrate the superiority of TCRaman compared with current state-of-the-art models. The results show that TCRaman achieves 91% accuracy on the detection of both glucose and protein, and 95% accuracy on the prediction of kidney disorders, highlighting the effectiveness of our proposed method in detecting urine Raman spectra. The proposed TCRaman method provides a promising way for accurate, rapid, and cost-effective detection for spectral data of biochemical samples.