使用粪便挥发性有机物通过可解释电子鼻进行牛酮症的无创产前预测

Highlights
•Non-invasive prenatal ketosis prediction in cows using fecal VOCs and e-nose
•95.9% accuracy achieved three weeks before calving via an interpretable KNN model
•Sensor S8 identified as dominant predictor using LIME and permutation analysis
•“Microbiota → VOCs → sensor” pathway quantified by two-stage PLS (R²=0.483)
•E-nose signals driven by microbial VOCs like pentanal and pyridine derivatives

亮点
• 利用粪便挥发性有机物和电子鼻技术，实现奶牛产前无创酮症预测
• 通过可解释的KNN模型，在产前三周达到95.9%的准确率
• 结合LIME和置换重要性分析，确认传感器S8为主要的预测因子
• 通过两阶段偏最小二乘法量化“微生物群→挥发性有机物→传感器”通路（R²=0.483）
• 电子鼻信号由戊醛和吡啶衍生物等微生物挥发性有机物驱动

Abstract
Bovine ketosis is a prevalent periparturient metabolic disorder severely impairing dairy cow health and farm profitability. This study developed an interpretable electronic nose (e-nose) system for noninvasive prenatal ketosis prediction in Holstein cows using fecal volatile organic compounds (VOCs). Samples collected at 3, 2, and 1 weeks prepartum were postnatally classified as healthy or ketotic based on blood β-hydroxybutyrate ≥ 1.2 mmol/L. Multivariate analysis revealed significant VOC profile separation as early as 3 weeks prepartum. Least Absolute Shrinkage and Selection Operator (LASSO) identified sensors S6, S8, and S14 as key discriminators, and the K-nearest neighbors (KNN) model achieved 95.9% accuracy and area under the curve (AUC) of 0.98 through nested 10-fold cross-validation, with interpretability analysis confirming S8 as dominant. To uncover the biological basis of e-nose signals, we integrated headspace solid-phase microextraction-gas chromatography mass spectrometry and metagenomics, identifying 21 differential VOCs and distinct gut microbiota. A two-stage partial least squares (PLS) model quantified the gut microbiota → VOCs → sensor pathway (R² = 0.483, p = 0.02), validated by stability selection. Semi-quantitative correlation analysis revealed key cascade links, for example: CAG 791 sp. 002358875 → pentanal → S8 (|ρ| = 0.870, p = 0.000237), supporting the PLS-derived mechanistic pathway. This work presents a novel, traceable, and mechanism-interpretable e-nose framework for predicting bovine ketosis risk three weeks prepartum. This approach not only expands the application of commercial sensors in livestock health monitoring but also provides a generalizable paradigm for the intelligent, biologically grounded sensing of complex odors in precision livestock farming.

摘要
奶牛酮症是一种常见的围产期代谢性疾病，严重损害奶牛健康及牧场经济效益。本研究利用荷斯坦奶牛的粪便挥发性有机物，开发了一种可解释的电子鼻系统，用于产前无创预测酮症。分别于产前3周、2周和1周采集样本，然后根据产后血β-羟丁酸≥1.2 mmol/L的标准，将奶牛分为健康组和酮症组。多变量分析显示，最早在产前3周，挥发性有机物谱即出现显著分离。通过最小绝对收缩和选择算子识别出传感器S6、S8和S14为关键区分因子，采用嵌套10折交叉验证的K近邻模型实现了95.9%的准确率和0.98的曲线下面积，可解释性分析确认S8为主要预测因子。为了揭示电子鼻信号的生物学基础，我们整合了顶空固相微萃取-气相色谱质谱联用技术和宏基因组学，鉴定出21种差异挥发性有机物和独特的肠道微生物群。一个两阶段偏最小二乘模型量化了“肠道微生物群→挥发性有机物→传感器”通路（R²=0.483，p=0.02），并通过稳定性选择进行验证。半定量相关分析揭示了关键的级联关联，例如：CAG 791 sp. 002358875 → 戊醛 → S8（|ρ| = 0.870，p = 0.000237），支持了偏最小二乘法推导的机制通路。本研究提出了一种新颖、可追溯且机制可解释的电子鼻框架，用于在产前三周预测奶牛酮症风险。该框架不仅扩展了商用传感器在畜牧健康监测中的应用，也为精准畜牧业中复杂气味的智能化、基于生物学的感知提供了一个可推广的范式。

Introduction
Ketosis is a common metabolic disease in dairy cows during the periparturient period and is primarily driven by a negative energy balance after parturition [1]. It leads to excessive fat mobilization, hepatic β oxidation, and ketone overproduction [2], which collectively reduce milk yield, impair reproduction, and increase the risk of secondary disorders such as mastitis and displaced abomasum [3], [4]. As a major global disease burden, ketosis is estimated to cause annual economic losses of 18.2 billion dollars [5], which severely affects progress toward sustainable dairy farming. Importantly, ketosis-related metabolic disruption begins approximately 3 weeks before calving [1], yet current detection tools fail to capture this early window. Postpartum diagnostic approaches, such as blood β-hydroxybutyrate (BHB) testing and ketone strips, identify only established disease [6], [7], and prenatal nonsensor methods are often limited by environmental interference or high operational cost [8], [9]. Therefore, a scalable and noninvasive early warning strategy for the detection of pre-symptomatic risk remains a critical need.

引言
酮症是奶牛围产期常见的代谢性疾病，主要由分娩后的能量负平衡驱动[1]。它导致过度的脂肪动员、肝脏β氧化和酮体过量产生[2]，从而降低产奶量、损害繁殖性能，并增加患乳房炎和真胃变位等继发性疾病的风险[3,4]。作为一种主要的全球性疾病负担，酮症每年造成的经济损失估计达182亿美元[5]，严重影响可持续奶业的发展进程。重要的是，与酮症相关的代谢紊乱大约在产前三周就开始出现[1]，而目前的检测手段无法捕捉这一早期窗口。产后诊断方法，如血β-羟丁酸检测和酮试纸，只能在疾病已发生时识别[6,7]；产前的非传感器方法通常受到环境干扰或高操作成本的限制[8,9]。因此，一种可扩展、无创的预警策略，用于检测症状前风险，仍然是迫切需要的。

Electronic nose (e-nose) technology that utilizes volatile organic compounds (VOCs) provides a promising alternative because it is noninvasive, rapid, and relatively low cost [10], [11]. In recent years, e-nose platforms have been increasingly adopted for food quality control [12], environmental air monitoring [13], and clinical disease screening [14]. The use of chemical sensor arrays for noninvasive detection through analysis of complex biological volatilomes is also expanding, as shown by recent studies and reviews [15], [16]. In veterinary medicine, e-nose applications are still relatively limited but have been explored in several animal-health-related contexts, including respiratory disease diagnosis, mastitis detection, large-animal physiological monitoring, and livestock-related gas sensing [17], [18], [19], [20].

利用挥发性有机物的电子鼻技术提供了一种有前景的替代方案，因为它无创、快速且成本相对较低[10,11]。近年来，电子鼻平台越来越多地被用于食品质量控制[12]、环境空气监测[13]和临床疾病筛查[14]。通过分析复杂生物挥发性代谢组，使用化学传感器阵列进行无创检测的应用也在不断扩大，如近期研究和综述所示[15,16]。在兽医学领域，电子鼻的应用仍然相对有限，但已在多个动物健康相关背景下进行了探索，包括呼吸系统疾病诊断、乳房炎检测、大型动物生理监测以及畜牧相关的气体传感[17,18,19,20]。

The Cyranose® 320 e-nose applied in this work contains an array of 32 chemoresistive sensors composed of conductive carbon black particles dispersed within hydrophobic polymer matrices. When exposed to VOCs, these polymers selectively swell through van der Waals interactions and differential partitioning, which increases interparticle spacing and generates a measurable resistance shift, reflecting a transduction process grounded in polymer gas thermodynamics [21], [22]. The diverse polymer chemistries, including polystyrene and polyethylene oxide, provide complementary selectivity toward aldehydes, ketones, esters, and other chemical classes [22], making the system well-suited for characterizing complex substrates such as feces. This study, therefore, aims to develop an e-nose-based prenatal prediction strategy for ketosis in dairy cows.

本研究所用的Cyranose® 320电子鼻包含一个由32个化学电阻传感器组成的阵列，这些传感器由分散在疏水聚合物基质中的导电碳黑颗粒构成。当暴露于挥发性有机物时，这些聚合物通过范德华相互作用和差异分配发生选择性溶胀，从而增加颗粒间距并产生可测量的电阻变化，反映了基于聚合物气体热力学的转导过程[21,22]。不同的聚合物化学性质（包括聚苯乙烯和聚环氧乙烷）对醛、酮、酯及其他化学类别提供了互补的选择性[22]，使得该系统非常适合表征粪便等复杂基质。因此，本研究旨在开发一种基于电子鼻的奶牛酮症产前预测策略。

Ketosis induces systemic metabolic reprogramming and gut microbial disturbances that alter volatile metabolite production, which yields detectable changes in fecal odor profiles before parturition [9]. As endpoints of host microbiota co-metabolism, VOCs released from feces, breath, or milk provide informative markers of metabolic status. Feces offer particular advantages because they contain high concentrations of VOCs, exhibit greater stability, and allow stress-free sampling that is suitable for application at the farm scale. Our prior research employed the Cyranose® 320 e-nose for auxiliary postpartum ketosis diagnosis and achieved 88.8% diagnostic accuracy [23]. However, that work addressed postpartum detection, whereas the present study focuses on prenatal prediction, which is a more challenging yet clinically valuable scenario. To the best of our knowledge, no existing study has investigated whether fecal VOCs can enable prenatal prediction of ketosis.

酮症会诱导全身代谢重编程和肠道微生物紊乱，从而改变挥发性代谢物的产生，使得产前粪便气味特征发生可检测的变化[9]。作为宿主-微生物群共代谢的终产物，从粪便、呼吸或牛奶中释放的挥发性有机物提供了代谢状态的信息标志物。粪便具有特别的优势，因为它含有高浓度的挥发性有机物，表现出更大的稳定性，并且允许无应激采样，适合在牧场规模上应用。我们先前的研究使用Cyranose® 320电子鼻进行产后酮症辅助诊断，达到了88.8%的诊断准确率[23]。然而，该工作侧重于产后检测，而本研究则聚焦于产前预测，这是一个更具挑战性但也更具临床价值的场景。据我们所知，目前尚无研究探讨粪便挥发性有机物是否能够实现酮症的产前预测。

Headspace solid phase microextraction gas chromatography mass spectrometry (HS-SPME- GC/MS) can detect differential volatile metabolites in feces with high analytical sensitivity, and metagenomic sequencing can reveal the composition and metabolic functions of microbiota that generate these metabolites. When combined, these approaches provide traceable molecular and microbial evidence that links biological processes to sensor responses [24]. In recent years, explainable artificial intelligence methods such as Local Interpretable Model-Agnostic Explanations (LIME) and Shapley Additive Explanations (SHAP) have demonstrated value in human disease diagnostics using exhaled breath [25]. Their application in livestock health monitoring, however, remains limited. In addition, reliance on interpretability algorithms alone is insufficient to establish a causal sequence. A complete mechanistic chain can be clarified only when metabolomics identifies differential VOCs and metagenomics reveals the microbial pathways that drive these alterations, thereby outlining the sequence of gut microbiota disturbance, metabolic pathway shift, VOC profile change, and sensor response.

顶空固相微萃取-气相色谱质谱联用技术可以高分析灵敏度检测粪便中的差异挥发性代谢物，而宏基因组分测序能够揭示产生这些代谢物的微生物群的组成和代谢功能。这两种方法结合使用时，可以提供可追溯的分子和微生物证据，将生物过程与传感器响应联系起来[24]。近年来，局部可解释模型无关解释和沙普利可加性解释等可解释人工智能方法已在利用呼出气进行人类疾病诊断中展现出价值[25]。然而，它们在畜牧健康监测中的应用仍然有限。此外，仅依赖可解释性算法不足以建立因果序列。只有当代谢组学鉴定出差异挥发性有机物，宏基因组学揭示驱动这些变化的微生物通路，从而勾勒出肠道微生物群紊乱→代谢通路转变→挥发性有机物谱变化→传感器响应的完整序列时，才能阐明完整的机制链。

Distinct from studies focusing on sensor material synthesis or hardware optimization, we integrated a commercial e-nose with HS-SPME-GC/MS metabolomics and metagenomic sequencing to establish an interpretable, verifiable, and traceable multimodal framework for prenatal ketosis prediction in dairy cows. This integrated approach supports noninvasive early warning based on fecal VOCs, identifies key microbiota metabolite interactions that underlie VOC alterations, and, through the combined use of LIME, permutation importance, and two-stage partial least squares (PLS) regression, quantifies the microbiota to VOCs to sensor pathway that links biological mechanisms to e-nose signals. Rather than developing new sensor hardware, this work establishes a novel integrative framework that transforms commercial e-nose data into mechanistically grounded diagnostics. By tracing signals from gut microbiota to specific VOCs and sensor responses, we enable reliable ketosis prediction three weeks prepartum. This study redefines the value of off-the-shelf e-noses in precision livestock farming, shifting the focus from mere pattern recognition to biologically interpretable early warning systems for complex metabolic disorders.

不同于那些专注于传感器材料合成或硬件优化的研究，我们整合了商用电子鼻与顶空固相微萃取-气相色谱质谱联用代谢组学和宏基因组分测序，建立了一个可解释、可验证、可追溯的多模式框架，用于奶牛产前酮症预测。这种综合方法支持基于粪便挥发性有机物的无创早期预警，识别了构成挥发性有机物变化基础的关键微生物-代谢物相互作用，并通过结合使用局部可解释模型无关解释、置换重要性和两阶段偏最小二乘回归，量化了将生物学机制与电子鼻信号联系起来的“微生物群→挥发性有机物→传感器”通路。本研究并非开发新的传感器硬件，而是建立了一种新颖的整合框架，将商用电子鼻数据转化为基于机制的诊断。通过追踪从肠道微生物群到特定挥发性有机物再到传感器响应的信号，我们实现了产前三周的可靠酮症预测。本研究重新定义了现成电子鼻在精准畜牧业中的价值，将重点从单纯的模式识别转向用于复杂代谢紊乱的、具有生物学可解释性的早期预警系统。