Complementary Breath Profiling Using eNose and Gas Chromatography-Mass Spectrometry for Noninvasive Triage of Indeterminate Pulmonary Nodules
Waqar Ahmed Afridi et al.
ABSTRACT
Breath analysis may improve the noninvasive triage of CT-detected pulmonary nodules. We prospectively evaluated two complementary breath platforms, an electronic nose (eNose) and gas chromatography-mass spectrometry (GC-MS), in patients with newly detected nodules. The eNose cohort included 89 participants (55 malignant and 34 benign), and the GC-MS cohort included 75 participants (52 malignant and 23 benign). For each platform, three logistic regression models were assessed: clinical-only, biomarker-only, and combined clinical + biomarker. For eNose, apparent discrimination improved from the clinical-only model (AUC 0.67; misclassification rate [MCR] 40.4%) to the sensor-only model (AUC 0.86; MCR 22.5%), with further improvement in the combined clinical + sensor model (AUC 0.88; MCR 22.5%). Internal validation showed attenuation, with the combined eNose model achieving a leave-one-out cross-validation (LOOCV) AUC of 0.67. For GC-MS, LASSO logistic regression showed strong apparent discrimination for the VOC-only (AUC 0.97; MCR 4%), although internal validation indicated attenuation of performance (bootstrap-corrected AUC 0.64; LOOCV AUC 0.65). The combined clinical + VOC model remained comparably high (AUC 0.96; MCR 8%), outperforming the clinical-only benchmark (AUC 0.73; MCR 32%). These findings support complementary, noninvasive breath-based approaches for pulmonary nodule triage and warrant larger multicenter validation studies.
摘要
呼气分析可能改善CT检出肺结节的无创分诊。我们前瞻性评估了两种互补的呼气平台——电子鼻(eNose)和气相色谱-质谱联用(GC-MS)——在新检出结节患者中的应用。eNose队列包括89名参与者(55名恶性,34名良性),GC-MS队列包括75名参与者(52名恶性,23名良性)。对每个平台评估了三种逻辑回归模型:仅临床、仅生物标志物以及临床+生物标志物联合。对于eNose,表观区分度从仅临床模型(AUC 0.67;误分类率[MCR] 40.4%)提高到仅传感器模型(AUC 0.86;MCR 22.5%),并在临床+传感器联合模型中进一步提高(AUC 0.88;MCR 22.5%)。内部验证显示性能衰减,联合eNose模型的留一交叉验证(LOOCV)AUC为0.67。对于GC-MS,LASSO逻辑回归显示仅VOC模型具有强表观区分度(AUC 0.97;MCR 4%),但内部验证表明性能衰减(bootstrap校正AUC 0.64;LOOCV AUC 0.65)。临床+VOC联合模型仍保持相当高的水平(AUC 0.96;MCR 8%),优于仅临床基准模型(AUC 0.73;MCR 32%)。这些发现支持将互补的无创呼气方法用于肺结节分诊,并有必要开展更大规模的多中心验证研究。
2 Results结果
2.1 Breath eNose呼气电子鼻
A total of 98 participants with CT-detected pulmonary nodules (LN01-LN98) were enrolled. After the quality-control filters and data-availability criteria were applied, 89 participants were included in the eNose analyses, and 75 were included in the GC-MS analyses
共纳入98名CT检出肺结节的参与者(LN01-LN98)。应用质量控制过滤和数据可用性标准后,89名参与者纳入eNose分析,75名纳入GC-MS分析。
5.4 ENose Analysis电子鼻分析
5.4.1 Device Setup and Acquisition设备设置与采集
The Cyranose 320 was used to analyze exhaled breath samples. This portable device contains 32 carbon black polymer composite chemiresistors that adsorb exhaled VOCs to varying degrees, producing a characteristic differential electrical pattern. Each participant provided a single full-breath exhalation that was processed twice; only the second pass was retained for analysis, with the first used to condition the sensors (swelling/relaxation) and stabilize responses across the array. Device operation followed the Amsterdam protocol with local adaptations [50].
使用Cyranose 320分析呼出气样本。该便携式设备包含32个炭黑聚合物复合化学电阻器,它们以不同程度吸附呼出气VOCs,产生特征性的差分电学模式。每位参与者提供一次完整呼气,该呼气被处理两次;仅保留第二遍用于分析,第一遍用于调节传感器(溶胀/松弛)并稳定阵列响应。设备操作遵循Amsterdam协议并进行了本地化调整[50]。
5.4.2 Pre-Processing and Feature Engineering预处理与特征提取
The raw sensor outputs were exported to JMP Pro, log-transformed, and then within-subject mean-centered to reduce inter-individual baseline offsets before modeling.
原始传感器输出导出至JMP Pro,进行对数转换,然后在建模前进行受试者内均值中心化,以减少个体间基线偏移。
5.4.3 Model Families and Tuning模型类型与调参
Three logistic regression models were constructed: (i) clinical-only, (ii) eNose sensor-only, and (iii) combined clinical + eNose sensor. The clinical covariates included age, sex, smoking history, and ECOG status. The Brock score and nodule diameter were omitted because they were missing in historical controls. Sensor selection utilized adaptive LASSO with parameter tuning via LOOCV. Model discrimination was evaluated using ROC curves and corresponding AUCs, with 95% CIs, MCRs, and AUC comparisons among the three models. We used adaptive LASSO logistic regression to reduce overfitting and obtain an interpretable subset of predictors.
构建了三种逻辑回归模型:(i) 仅临床;(ii) 仅eNose传感器;(iii) 临床+eNose传感器联合。临床协变量包括年龄、性别、吸烟史和ECOG状态。Brock评分和结节直径被省略,因为其在历史对照中缺失。传感器选择使用自适应LASSO,并通过LOOCV进行参数调优。使用ROC曲线及相应AUC评估模型区分度,并计算95%CI、MCR以及三个模型间的AUC比较。我们使用自适应LASSO逻辑回归以减少过拟合,并获得可解释的预测变量子集。

