Lin Yu-Tzu, Lin Hsiu-Hsien, Chen Chih-Hao, Tseng Kun-Hao, Hsu Pang-Chien, Wu Ya-Lun, Chang Wei-Cheng, Liao Nai-Shun, Chou Yi-Fan, Hsu Chun-Yi, Liao Yu-Hui, Ho Mao-Wang, Chang Shih-Sheng, Hsueh Po-Ren, Cho Der-Yang
Department of Medical Laboratory Science and Biotechnology, China Medical University, Taichung, Taiwan.
Department of Laboratory Medicine, China Medical University Hospital, China Medical University, Taichung, Taiwan.
J Microbiol Immunol Infect. 2025 Feb;58(1):77-85. doi: 10.1016/j.jmii.2024.11.014. Epub 2024 Nov 29.
Rapid and accurate identification of bacteria is required in order to develop effective treatment strategies. Traditional culture-based methods are time-consuming, while MALDI-TOF MS is expensive. The Raman spectroscopy, due to its relatively cost-effectiveness, offers a promising alternative for bacterial identification. However, its clinical utility still requires further validation.
In this study, the artificial intelligent Raman detection and identification system (AIRDIS) was implemented to identify bacterial species, including Staphylococcus aureus (n = 1290), Enterococcus faecium (n = 1020), Klebsiella pneumoniae (n = 1366), Pseudomonas aeruginosa (n = 1067), and Acinetobacter baumannii (n = 811). Raman spectra were collected, preprocessed, and analyzed by machine learning (ML).
After training on 24,420 Raman spectra from 1221 isolates and testing on 4333 isolates, the AIRDIS demonstrated an area under the curve (AUC) of 0.99 for Gram classification, with accuracies of 97.64 % for Gram-positive bacteria and 98.86 % for Gram-negative bacteria. Spectral differences between Gram-positive and Gram-negative bacteria were linked to structural variations in their cell walls, such as peptidoglycan and lipopolysaccharides. At the species level, S. aureus, E. faecium, K. pneumoniae, P. aeruginosa, and A. baumannii were identified with high accuracy, ranging from 94.76 % to 96.88 %, with all species achieving an AUC of 0.99.
Validation with a large number of clinical isolates demonstrated Raman spectroscopy combined with ML excels in identification of five bacterial species associated with multidrug resistance. This finding confirms the clinical utility of the system while laying a solid foundation for the future development of antimicrobial resistance prediction models.
为制定有效的治疗策略，需要快速准确地鉴定细菌。传统的基于培养的方法耗时，而基质辅助激光解吸电离飞行时间质谱（MALDI-TOF MS）成本高昂。拉曼光谱因其相对的成本效益，为细菌鉴定提供了一种有前景的替代方法。然而，其临床实用性仍需进一步验证。
在本研究中，实施了人工智能拉曼检测与鉴定系统（AIRDIS）来鉴定包括金黄色葡萄球菌（n = 1290）、粪肠球菌（n = 1020）、肺炎克雷伯菌（n = 1366）、铜绿假单胞菌（n = 1067）和鲍曼不动杆菌（n = 811）在内的细菌种类。收集拉曼光谱，进行预处理，并通过机器学习（ML）进行分析。
在对来自1221株分离菌的24420条拉曼光谱进行训练并对4333株分离菌进行测试后，AIRDIS在革兰氏分类中的曲线下面积（AUC）为0.99，革兰氏阳性菌的准确率为97.64%，革兰氏阴性菌的准确率为98.86%。革兰氏阳性菌和革兰氏阴性菌之间的光谱差异与它们细胞壁的结构变化有关，如肽聚糖和脂多糖。在种水平上，金黄色葡萄球菌、粪肠球菌、肺炎克雷伯菌、铜绿假单胞菌和鲍曼不动杆菌的鉴定准确率很高，范围从94.76%到96.88%，所有菌种的AUC均为0.99。
大量临床分离菌的验证表明，拉曼光谱结合机器学习在鉴定与多重耐药相关的五种细菌方面表现出色。这一发现证实了该系统的临床实用性，同时为未来抗菌药物耐药性预测模型的发展奠定了坚实基础。
