Data analysis of electronic nose technology in lung cancer: generating prediction models by means of Aethena. (1st June 2017)

• Record Type: Journal Article
• Authors: Kort, Sharina; Brusse-Keizer, Marjolein; Gerritsen, Jan-Willem; van der Palen, Job
• Is Part Of: Journal of breath research. Volume 11:Number 2(2017:Jun.)
• Publication Date: 2017-06-01
• Subjects: lung cancer -- electronic nose -- exhaled breath -- aeonose -- prediction models -- data analysis; Volatile organic compounds -- Analysis -- Periodicals; Clinical chemistry -- Periodicals; Bad breath -- Periodicals; Bad breath -- Treatment -- Periodicals; Bad breath -- Diagnosis -- Periodicals; 616.0756
• DOI: 10.1088/1752-7163/aa6b08
• Languages: English
• ISSNs: 1752-7155

Abstract

Abstract: Introduction . Only 15% of lung cancer cases present with potentially curable disease. Therefore, there is much interest in a fast, non-invasive tool to detect lung cancer earlier. Exhaled breath analysis using electronic nose technology measures volatile organic compounds (VOCs) in exhaled breath that are associated with lung cancer. Methods . The diagnostic accuracy of the Aeonose™ is currently being studied in a multi-centre, prospective study in 210 subjects suspected for lung cancer, where approximately half will have a confirmed diagnosis and the other half will have a rejected diagnosis of lung cancer. We will also include 100–150 healthy control subjects. The eNose Company (provider of the Aeonose™) uses a software program, called Aethena, comprising pre-processing, data compression and neural networks to handle big data analyses. Each individual exhaled breath measurement comprises a data matrix with thousands of conductivity values. This is followed by data compression using a Tucker3-like algorithm, resulting in a vector. Subsequently, model selection takes place after entering vectors with different presets in an artificial neural network to train and evaluate the results. Next, a 'judge model' is formed, which is a combination of models for optimizing performance. Finally, two types of cross-validation, being 'leave-10%-out' cross-validation and 'bagging', are used when recalculating the judge models. These judge models are subsequently used to classify new, blind measurements. Discussion. Data analysis in eNose technology is principally based on generating prediction models that need to be validated internally and externally for eventual use in clinical practice. This paper describes the analysis of big data, captured by eNose technology in lung cancer. This is done by means of generating prediction models with Aethena, a data analysis program specifically developed for analysing VOC data.