In recent years, breath analysis has been used as a tool for lung cancer detection and many gas sensors were developed for this purpose. Although they are fabricated with advanced materials, for now, gas sensors are still limited in their medical application due to their unfavorable performance. Here, we hypothesized that a combination of diverse types of sensors could aid in improving the detection performance. We fabricated an e-nose based on 10 gas sensors of 4 types and directly tested it using samples from 153 healthy participants and 115 lung cancer patients, without gas pre-concentration. Additionally, we studied and compared five feature extraction algorithms. The extracted features were then used in 2 optimized clustering algorithms and 3 supervised classification strategies, and their performance was investigated. As a result, "breath-prints" for all subjects were successfully obtained. The combined features extracted by LDA and Fast ICA formed the best feature space. Within this feature space, both clustering algorithms grouped all "breath-prints" into exactly 2 clusters with an Adjusted Rand Index greater than 0.95. Among the 3 supervised classification strategies, random forest with 3-fold cross validation showed the best performance with 86.42% of mean classification accuracy and 0.87 of AUC, which was somewhat better than many recently reported sensor arrays. It can be concluded that, the diversity of sensors may play a role in improving the performance of the e-nose though to what extent still requires evaluation.

中文翻译：

---

使用自制电子鼻识别肺癌：一项全面研究。

近年来，呼气分析已被用作检测肺癌的工具，并且为此目的开发了许多气体传感器。尽管它们是用先进的材料制成的，但由于其性能不佳，目前气体传感器在医学上仍然受到限制。在这里，我们假设多种传感器的组合可以帮助提高检测性能。我们基于10种4种类型的气体传感器制造了一个电子鼻，并使用来自153名健康参与者和115名肺癌患者的样本直接对其进行了测试，而没有进行气体预浓缩。此外，我们研究并比较了五种特征提取算法。然后将提取的特征用于2种优化的聚类算法和3种监督分类策略，并对它们的性能进行了研究。结果是， ” LDA和Fast ICA提取的组合特征形成了最佳特征空间。在此功能空间内，两种聚类算法均将所有“呼吸印迹”分组为调整后的兰德指数大于0.95的正好两个聚类。在3种监督分类策略中，具有3倍交叉验证的随机森林表现出最佳性能，平均分类准确度为86.42％，AUC为0.87，这比许多最近报道的传感器阵列要好一些。可以得出结论，尽管仍需要评估到何种程度，传感器的多样性可能会在改善电子鼻的性能方面发挥作用。LDA和Fast ICA提取的组合特征形成了最佳特征空间。在此功能空间内，两种聚类算法均将所有“呼吸印迹”分组为调整后的兰德指数大于0.95的正好两个聚类。在3种监督分类策略中，具有3倍交叉验证的随机森林表现出最佳性能，平均分类准确度为86.42％，AUC为0.87，这比许多最近报道的传感器阵列要好一些。可以得出结论，尽管仍需要评估到何种程度，传感器的多样性可能会在改善电子鼻的性能方面发挥作用。在3种监督分类策略中，具有3倍交叉验证的随机森林表现出最佳性能，平均分类准确度为86.42％，AUC为0.87，这比许多最近报道的传感器阵列要好一些。可以得出结论，尽管仍需要评估到何种程度，传感器的多样性可能会在改善电子鼻的性能方面发挥作用。在3种监督分类策略中，经过3倍交叉验证的随机森林表现出最佳性能，平均分类准确度为86.42％，AUC为0.87，这比许多最近报道的传感器阵列要好一些。可以得出结论，尽管仍需要评估到何种程度，传感器的多样性可能会在改善电子鼻的性能方面发挥作用。