Data Sheet 1_Urine volatile organic compounds profiling via GC-IMS combined with machine learning: a powerful diagnostic and pathogen differentiation tool for urinary tract infections.docx

BackgroundThe diagnostic delay associated with standard urine culture necessitates rapid, accurate alternatives for urinary tract infection (UTI) management. Volatile organic compounds (VOCs) emitted by microbes represent a promising source of metabolic biomarkers for infection diagnosis.
ObjectiveTo develop and validate a diagnostic model for UTI by integrating urine VOCs profiles obtained via gas chromatography-ion mobility spectrometry (GC-IMS) with clinical features using machine learning.
MethodsWe conducted a prospective cohort study of 258 adults with suspected UTI. Clean-catch midstream urine samples were collected for clinical urinalysis, culture (reference standard), and GC-IMS-based VOCs analysis. VOCs and clinical data were used to train and test machine learning models (Logistic Regression, Random Forest, Support Vector Machine). Model performance was assessed by area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and decision curve analysis.
ResultsAmong 258 enrolled patients, 152 (58.9%) were culture-positive. We identified 11 differentially expressed VOCs between infected and non-infected groups, with acetic acid, benzaldehyde, and furan being the most significant (Bonferroni-adjusted p < 0.05). A Random Forest model integrating both VOCs and clinical features demonstrated superior performance (AUC of 0.914, with an accuracy of 82.1% (95% CI: 71.8-89.8%), sensitivity of 87.0%, specificity of 75.0%, and an F1-score of 0.851) compared to models using clinical-only (AUC 0.831) or VOC-only (AUC 0.850). Multivariate analysis confirmed acetic acid (OR 3.27) and benzaldehyde (OR 4.95) as strong independent predictors of UTI. Furthermore, VOCs profiles allowed moderate discrimination between Gram-positive and Gram-negative bacterial infections (AUC 0.800) and exhibited pathogen-specific patterns.
ConclusionThe integration of urine VOCs profiles obtained by GC-IMS with routine clinical parameters using machine learning achieves high diagnostic accuracy for UTI and shows potential for rapid pathogen differentiation. This strategy could improve UTI diagnostics, enabling faster, more precise antibiotic therapy.

背景 标准尿液培养所带来的诊断延迟，使得尿路感染（urinary tract infection, UTI）诊疗亟需快速且精准的替代方案。微生物释放的挥发性有机化合物（volatile organic compounds, VOCs）是感染诊断中颇具前景的代谢生物标志物来源。
研究目标 本研究旨在通过机器学习方法，整合气相色谱-离子迁移谱（gas chromatography-ion mobility spectrometry, GC-IMS）获取的尿液VOCs特征与临床指标，构建并验证尿路感染的诊断模型。
研究方法 本研究针对258名疑似尿路感染的成年人开展了一项前瞻性队列研究。收集清洁中段尿样本，分别进行临床尿液分析、培养（金标准）以及基于GC-IMS的VOCs分析。利用VOCs数据与临床数据训练并测试机器学习模型，包括逻辑回归、随机森林与支持向量机。通过受试者工作特征曲线下面积（area under the receiver operating characteristic curve, AUC）、灵敏度、特异度以及决策曲线分析评估模型性能。
研究结果 入组的258名患者中，152名（58.9%）培养结果呈阳性。本研究在感染组与非感染组之间鉴定出11种差异表达的VOCs，其中乙酸、苯甲醛与呋喃的差异最为显著（邦费罗尼校正后P<0.05）。相较于仅使用临床指标（AUC 0.831）或仅使用VOCs特征（AUC 0.850）的模型，整合VOCs与临床特征的随机森林模型展现出更优的性能：AUC为0.914，准确率达82.1%（95%置信区间：71.8%~89.8%），灵敏度为87.0%，特异度为75.0%，F1分数为0.851。多变量分析证实，乙酸（比值比OR=3.27）与苯甲醛（OR=4.95）是尿路感染的强独立预测因子。此外，VOCs特征可对革兰氏阳性与革兰氏阴性细菌感染进行中等程度的区分（AUC 0.800），且呈现出病原体特异性的表达模式。
研究结论 利用机器学习方法整合GC-IMS获取的尿液VOCs特征与常规临床指标，可实现尿路感染的高精准诊断，且具备快速区分病原体的潜力。该策略有望优化尿路感染的诊断流程，助力实现更快速、更精准的抗生素治疗。