Urinary tract infection (UTI) is among the most common and burdensome bacterial infections globally, affecting hundreds of millions annually and causing significant morbidity, mortality, and substantial healthcare costs (Foxman, 2010; Zi et al., 2024). Currently, microbial identification and antimicrobial susceptibility testing based on urine culture are widely regarded as the diagnostic gold standard. However, this method has a critical weakness—a prolonged turnaround time, typically 24–48 hours (Goebel et al., 2021; Chang et al., 2025). This diagnostic delay severely hinders the timely initiation of targeted antimicrobial therapy, compelling clinicians to rely heavily on empirical, broad-spectrum antibiotics in the absence of etiological evidence. This practice significantly contributes to the escalating global crisis of antimicrobial resistance, creating a vicious cycle that urgently needs to be broken (Dunne et al., 2022; Wagenlehner et al., 2022; Onyango et al., 2024).

To bridge this diagnostic gap, point-of-care biomarkers such as urine leukocyte esterase, nitrite, and microscopic white blood cell (WBC) and bacterial count (BACT) are widely used for rapid screening. However, numerous studies indicate that the diagnostic performance of these indicators is suboptimal, with sensitivity and specificity varying considerably across different patient populations, often leading to high rates of false negatives and false positives, resulting in misdiagnosis or missed diagnoses (Gu et al., 2020; Advani et al., 2024; Kristensen et al., 2025). Consequently, developing novel UTI diagnostic tools that are rapid (within hours), accurate, and capable of providing preliminary pathogen-level information has become a shared focus and urgent need in clinical microbiology, laboratory medicine, and infectious diseases (Bermudez et al., 2025).

In recent years, volatile organic compounds (VOCs) profiles released by microorganisms during growth and metabolism have emerged as a highly promising “metabolic fingerprint,” showing great potential for the rapid diagnosis and differential diagnosis of infectious diseases (Sethi et al., 2013; Kunze-Szikszay et al., 2021). Different pathogens possess unique metabolic networks and enzyme systems; when utilizing host nutrients for proliferation, they release specific types and concentrations of VOCs, forming a distinct “odor profile” that reflects the pathogen species, abundance, and host-pathogen interactions (Bos et al., 2013; Belizario et al., 2020; Ratiu et al., 2020). This diagnostic strategy based on volatile metabolites opens new avenues for non-invasive, rapid etiological identification. Among various VOCs analysis techniques, gas chromatography-ion mobility spectrometry (GC-IMS) is particularly suitable for high-throughput VOCs fingerprinting of clinical samples due to its high sensitivity, rapid analysis, and relative operational simplicity. Its successful application in emerging infectious diseases, respiratory infections, and sepsis underscores its significant potential in medical diagnostics (Drees et al., 2019; Nazareth et al., 2022; Zhao et al., 2025).

Although pioneering studies have preliminarily confirmed the potential of urine VOCs to distinguish UTI patients from healthy controls, the current research paradigm in this field faces significant limitations. First, most studies investigate VOCs profiles in isolation, failing to integrate them deeply with routine clinical indicators, thus preventing an accurate assessment of their incremental diagnostic value within existing workflows. Second, the powerful potential of VOCs for pathogen typing (e.g., discriminating Gram stain status and even specific species) remains underexplored, leaving the rich etiological information they contain largely undecoded. Finally, from a data analysis perspective, there is a widespread failure to fully leverage advanced machine learning algorithms to extract stable diagnostic features from high-dimensional, complex VOCs data, limiting model generalizability and clinical applicability (Guernion et al., 2001; Dospinescu et al., 2020).

Therefore, several key scientific questions remain to be answered through well-designed prospective studies: First, are there stable and reproducible differential patterns in urine VOCs profiles between patients classified as infected or non-infected based on the reference standard of urine culture? Second, can integrating high-dimensional VOCs information with routine clinical parameters construct a UTI diagnostic model with performance surpassing existing standards? Finally, can VOCs-based metabolic features reliably differentiate the Gram stain characteristics and specific types of causative pathogens, thereby providing clinicians with proactive, precise therapeutic guidance before traditional susceptibility results are available? This study aims to systematically address these questions through a prospective diagnostic accuracy study, ultimately promoting a shift towards faster, more precise, and more sustainable UTI diagnosis and management.

The clinical management of UTIs faces the dual challenge of diagnostic delays and the escalating threat of antimicrobial resistance. Although the traditional urine culture remains the gold standard, its protracted turnaround time compels empiric antibiotic use, exacerbating global antibiotic misuse and resistance selection pressure (Wang and LaSala, 2021; Zhu et al., 2021). Our study prospectively integrated GC-IMS technology with advanced machine learning to systematically evaluate the diagnostic and typing value of urine VOCs in addressing this challenge. Our findings consistently demonstrate that the urine VOCs landscape, shaped by pathogen-host interactions, serves as a stable diagnostic resource. The primary novelty of this study lies in establishing an integrated, translatable diagnostic framework that extends beyond the identification of individual biomarkers. This framework is built upon three key elements: First, through prospective multimodal fusion, we demonstrated the significant incremental value of adding VOCs profiles to routine clinical parameters, achieving a combined model with high diagnostic performance (test AUC = 0.914). Second, we applied machine learning to develop a directly applicable predictive model, moving beyond associative analysis. Third, by constructing a biochemical pathway network, we linked VOCs signatures to core microbial metabolism, providing a mechanism-informed interpretation. Collectively, this framework represents a step toward a more precise approach to anti-infective therapy.

This study successfully identified a characteristic VOCs profile significantly enriched in the infected group, with acetic acid, benzaldehyde, and propanoic acid being particularly prominent. The appearance of these molecules is not random but is underpinned by robust microbial metabolic logic, elevating VOCs analysis beyond the limitation of traditional biomarkers that merely suggest “inflammation is present” to enabling the non-invasive monitoring of “which pathogen is metabolically active and how.” The strong independent predictive value of acetic acid (multivariate OR = 3.27) aligns closely with Clark’s classical description of mixed-acid fermentation in E. coli (Clark, 1989). As the predominant UTI pathogens, E. coli and other Enterobacteriaceae convert pyruvate-derived acetyl-CoA to acetate via the phosphate acetyltransferase (PTA) and acetate kinase (ACK) pathways in the microaerobic environment of the urinary tract, concurrently generating ATP. Thus, markedly elevated urinary acetate levels can be interpreted as direct ‘chemical evidence’ of vigorous fermentative metabolism by pathogens in the urinary environment. The enrichment of benzaldehyde likely stems from pathogen metabolism of aromatic amino acids. The classic work by TABAK et al (Tabak et al., 1964), using targeted screening and cultivation of phenol-adapted bacteria, systematically evaluated their degradation capacity for aromatic compounds and demonstrated their immediate and efficient oxidation of benzaldehyde. This provides in vivo support for such metabolic capabilities in the clinical setting. The enrichment of propanoic acid also has a solid foundation in microbial metabolism. Research indicates that the human colonic microbiota and various pathogenic bacteria can produce propanoic acid through multiple pathways, including the succinate and acrylate pathways (Louis and Flint, 2017). Notably, acetone and 2-methyl-1-propanol exhibited protective effects (OR < 1), suggesting their potential role as ‘protective’ metabolic markers. This finding can be interpreted from two angles: First, a decrease in these metabolites might signal suppressed metabolic activity of certain commensal microbes in the host gut that confer colonization resistance—these symbionts help maintain gut homeostasis through their metabolic products, and their dysfunction may indirectly increase UTI risk. Second, they might directly participate in host immune or metabolic regulation, and their decreased levels could reflect a dysregulation of host-protective metabolic pathways during infection. Although the concept of ‘protective metabolic markers’ is nascent in the UTI field, its core premise—that a healthy metabolic state of the host and its microbiota can produce molecules that resist pathogen colonization—is supported by broader research on microbial ecology and host interactions (Lawley and Walker, 2013; Zechner and Kienesberger, 2024). Our findings contribute novel evidence for this framework within the specific context of urinary tract infections.

A central contribution of this study lies in revealing the multiplicative effect of multimodal data fusion. While the clinical-only (AUC 0.831) and VOC-only (AUC 0.850) models each have their strengths, they possess inherent limitations: the former is susceptible to non-infectious inflammation, limiting specificity, while the latter can be influenced by ‘background noise’ from factors like diet and host basal metabolism (Capuano et al., 2025). When combined, the model leverages powerful and complementary information dimensions—clinical indicators sketch the macroscopic outline of the host’s systemic and local inflammatory response, while VOCs depict the microscopic metabolic fingerprint of pathogen community activity. This deep integration allows non-linear algorithms like Random Forest to capture more complex and disease-specific patterns, ultimately achieving a significant leap in diagnostic efficacy. This approach aligns with the emerging trend in complex infection diagnostics (e.g., sepsis, pneumonia) advocating for multimodal data integration (Frondelius et al., 2024; Shao et al., 2024). Decision curve analysis further affirmed, from a clinical utility perspective, that the combined model provides a positive net clinical benefit across a wide threshold probability range (0.1–0.8), solidifying its potential for integration into clinical decision support systems to guide more precise initial antibiotic selection.

Beyond establishing diagnostic value, our study explored the application of VOCs for the deeper clinical need of etiological differentiation. PCA and t-SNE visualizations indicated a partial separation between Gram-negative and Gram-positive bacterial infections in the VOCs feature space (Figure 5), although with considerable overlap. This finding is consistent with previous research. Tait et al. (2014) explicitly noted that VOCs profiles from Gram-positive bacteria (Staphylococcus aureus) differed fundamentally in diversity and composition from those of Gram-negative bacteria (E. coli, K. pneumoniae), with the former being more limited and lacking specific long-chain alcohols produced by the latter. Similarly, Frerichs et al. (2023) confirmed that fecal VOCs profiles had different predictive power and characteristic metabolites for Gram-negative versus Gram-positive late-onset sepsis in preterm infants. Although the current performance of the VOC-based Gram stain classifier (AUC 0.800) is only moderate and is by no means a substitute for standard susceptibility testing, it provides compelling proof-of-concept for a revolutionary ‘rapid etiological triage’. During the 24–48 hour window awaiting culture results, a rapid VOCs report available within hours suggesting “probable Gram-negative rod infection” could immediately guide clinicians toward targeted, narrow-spectrum antibiotics, significantly optimizing initial empiric therapy—a critical step toward precision anti-infective treatment.

To move beyond a mere listing of discrete molecules, we integrated the identified significant VOCs into a coherent, biologically meaningful ‘metabolic story’ by constructing a biochemical pathway network. These molecules were significantly enriched in a few core pathways: short-chain fatty acid metabolism, aromatic amino acid metabolism, and ketone/alcohol metabolism. This networked, pathway-oriented analysis has dual significance: First, it substantially enhances the biological interpretability of our results, directly linking observed VOCs changes to known microbial biochemical pathways, thereby elevating our findings from purely ‘data-driven’ to possessing a ‘mechanism-informed’ dimension (Sevin et al., 2015). Second, it suggests that future diagnostic models could potentially move beyond reliance on the abundance of single, potentially volatile molecules to focus on the more robust assessment of metabolic pathway functionality, which may significantly improve model stability and generalizability across individual variations (e.g., diet, medication, comorbidities) and complex clinical settings (Matthews et al., 2016).

We openly acknowledge the limitations of our study. First, the diagnostic performance of our model was benchmarked against conventional urine culture, the widely accepted clinical gold standard. It is important to note that this reference standard itself is not infallible. Factors such as prior antibiotic use, the presence of fastidious microorganisms, or asymptomatic bacteriuria could lead to misclassification of some patients. For instance, true infections with culture-negative results would have been assigned to the control group, an unavoidable bias that may lead to an underestimation of the model’s true discriminatory power. Second, as a single-center investigation, its participant population and pathogen spectrum may reflect regional characteristics. The generalizability of our conclusions requires validation in multi-center, large-scale, prospective external cohorts encompassing broader geographical regions, diverse populations, and healthcare settings—a necessary step for translational diagnostic research (Yang et al., 2022). Third, while GC-IMS offers significant advantages in sensitivity and speed, its ability to precisely resolve isomeric compounds remains challenging, and standardized protocols are currently lacking. Future complementary analysis using gas chromatography-mass spectrometry (GC-MS) would provide higher specificity for the absolute quantification and precise identification of key biomarkers, facilitating the development of diagnostic models transferable across platforms (Capitain and Weller, 2021; Epping and Koch, 2023). For instance, the thermal desorption GC-MS approach used by Ahmed et al. (2023) offers a rigorous methodological reference for future precise identification and clinical translation of our biomarkers. Fourth, some laboratory parameters had missing data (PCT: 40.3%, CRP: 40.3%), which reflects real-world clinical practice where these tests are selectively ordered based on disease severity. However, our primary analyses focused on VOC biomarkers that had complete data for all participants. Fifth, our model did not account for all potential confounding sources of urinary VOCs, such as host metabolic states (e.g., diabetic ketosis), diet, or medications. Although we recorded and adjusted for diabetes status statistically, the potential influence of these factors may affect specificity. Sixth, the study did not clinically stratify patients into upper versus lower urinary tract infections. Differences in pathogen burden and host response could influence VOCs profiles. Seventh, while we focused on microbial origins, the identified VOCs may also originate from or be modulated by host metabolism. Disentangling this contribution remains a challenge. Eighth, our combined model incorporated established clinical markers (e.g., nitrite) alongside novel VOCs. While this could theoretically introduce incorporation bias, the significant performance gain with VOCs confirms their substantial incremental value.

Finally, although we successfully constructed high-accuracy VOC-based models for UTI diagnosis and pathogen typing, our attempt to explore their potential for discriminating more complex phenotypes—such as extended-spectrum β-lactamase (ESBL)-producing isolates—yielded suboptimal results (AUC < 0.6). This outcome partially aligns with the findings of Smart et al (Dospinescu et al., 2020), who noted that distinguishing resistance requires higher-resolution metabolic profiles and that resistance-related metabolic alterations might be masked by the inherent VOC differences between species. Future studies will require larger cohorts focused on specific pathogens (e.g., E. coli only) and potentially the integration of more sensitive targeted metabolomics or multi-omics data to successfully capture the volatile ‘fingerprint’ specific to antimicrobial resistance (Lammers et al., 2022).

In summary, through a methodologically rigorous prospective cohort study with appropriate non-parametric statistical analyses, we have established urine VOCs analysis via GC-IMS as a robust, non-invasive, rapid, and information-rich diagnostic tool for UTIs. The developed and validated clinical-VOCs combined model not only demonstrates excellent discriminatory capability but, more importantly, the pathogen-specific information and metabolic pathway context embedded within the VOCs signatures open new avenues for understanding infection mechanisms and advancing towards rapid pathogen differentiation. With further technological standardization and large-scale validation, diagnostic strategies incorporating VOCs analysis hold the potential to reshape the clinical pathway for UTI, ushering antimicrobial stewardship into a more precise, timely, and sustainable era.