Interpretable Multiwavelength SERS Fingerprints of Human Urine for Ischemic Stroke Diagnosis

Abstract

Surface-enhanced Raman spectroscopy (SERS) can capture single-molecule-level component information from complex biological samples by providing their fingerprint profiles. However, increasing complexity and subtle variations in biological media can diminish the discrimination accuracy of traditional SERS excited by a single laser wavelength. Here, we demonstrate a multiwavelength SERS strategy for urine detection, aiming to achieve accurate diagnosis of ischemic stroke (IS). This strategy can acquire more comprehensive and unique chemical information on complex samples by capturing SERS fingerprints under multiple excitation wavelengths and vertically stacking them. Then, a convolutional neural network (CNN) classifier specifically designed for spectral data achieved an accuracy rate of 85.0% and an area under the curve (AUC) of 92.8% in recognizing IS. Furthermore, interpretation of neural net responses in the trained CNN model using a full-gradient algorithm highlights Raman spectral ranges that are most important to the diagnosis of IS. By correlating the important feature ranges selected by machine learning (ML) with the feature ranges of known biomolecules (such as lysine, arginine, glutamic acid, and hypoxanthine), we verified that the ML model effectively identified the Raman features of IS-related molecules and used a weighted combination of these features for the diagnosis of IS. Meanwhile, based on the multiwavelength SERS spectra stacking strategy, more effective information was extracted for the diagnosis of IS.