Dual-mode Identification of Ischemic Stroke based on Urine SERS Spectra and Carotid B-Ultrasound

Achieving noninvasive high-frequency monitoring of ischemic stroke (IS) remains a major clinical challenge for timely intervention and precise secondary prevention. Establishing precise correlations between patients' systemic microscopic molecular fingerprints and localized macroscopic organ pathological events is essential to overcome the limitations of single- modal detection and enhance the efficacy of clinical risk assessment. However, due to the complexity of heterogeneous data, effectively integrating the cross-dimensional “molecular-imaging” data remains a critical bottleneck in achieving this goal. Here, we present a method for accurately identification of IS that utilized machine learning (ML) based methods to surface-enhanced Raman spectroscopy (SERS) of urine (one-dimensional) and carotid artery B-ultrasound images (CBI) (two-dimensional). Through a simple ML workflow, we analyzed 10,100 SERS spectra and 481 CBI images from 101 participants. This technology achieved an outstanding identification accuracy of 92% and an AUC value of 0.95, significantly outperforming the evaluation results of SERS spectra alone (AUC 0.92) or CBI alone (AUC 0.88). In addition, SERS spectra combined with liquid chromatography-mass spectrometry (LC-MS) technology identified significant intergroup differences in the levels of arginine, lysine, and aspartic acid between the HC group and the IS group. The multi-dimensional data fusion strategy proposed in this study effectively bridges the information gap between traditional molecular detection and clinical phenotypes by systematically correlating micro-fluidic biomarkers with macro-organ imaging features. This approach provides a novel, non-invasive, and highly accurate tool for risk stratification and clinical decision-making in IS.

Keywords: ischemic stroke, urine, surface-enhanced Raman spectroscopy (SERS), carotid artery B-ultrasound image (CBI), machine learning,

缺血性脑卒中（ischemic stroke, IS）的无创高频监测仍是实现及时干预与精准二级预防的重大临床挑战。构建患者全身微观分子指纹与局部宏观器官病理事件间的精准关联，是突破单模态检测局限、提升临床风险评估效能的核心前提。然而，受限于异质性数据的复杂性，有效整合跨维度“分子-影像”数据仍是达成该目标的关键瓶颈。

本研究提出一种可精准识别缺血性脑卒中的方法：基于机器学习（machine learning, ML）技术，对尿液来源的一维表面增强拉曼光谱（surface-enhanced Raman spectroscopy, SERS）与颈动脉B超图像（carotid artery B-ultrasound image, CBI，二维）分别进行建模分析。通过简洁的机器学习工作流程，本研究对101名受试者的10100条SERS光谱与481幅CBI图像开展了分析。该技术取得了92%的优异识别准确率与0.95的曲线下面积（Area Under Curve, AUC）值，显著优于仅采用单一SERS光谱（AUC=0.92）或单一CBI（AUC=0.88）的评估结果。

此外，结合液相色谱-质谱联用（liquid chromatography-mass spectrometry, LC-MS）技术的SERS光谱分析，还发现健康对照组（Healthy Control, HC）与缺血性脑卒中组间的精氨酸、赖氨酸与天冬氨酸水平存在显著组间差异。本研究提出的多维度数据融合策略，通过将微流体生物标志物与宏观器官影像特征进行系统性关联，有效填补了传统分子检测与临床表型间的信息鸿沟。该方法为缺血性脑卒中的风险分层与临床决策提供了一种全新、无创且高精度的辅助工具。

关键词：缺血性脑卒中（ischemic stroke, IS）、尿液、表面增强拉曼光谱（surface-enhanced Raman spectroscopy, SERS）、颈动脉B超图像（carotid artery B-ultrasound image, CBI）、机器学习（machine learning, ML）