Selection of consistent breath biomarkers of abnormal liver function using feature selection: a pilot study

Abstract

PurposeBreath profiling has gained importance in recent years as it is a non-invasive technique to identify biomarkers for various diseases. Breath profiling of abnormal liver function in individuals for identifying potential biomarkers in exhaled breath could be a useful diagnostic tool. The objective of this study was to identify potential biomarkers in exhaled breath that remain stable and consistent during different physiological states, including rest and brief workouts, intending to develop a non-invasive diagnostic tool for detecting abnormal liver function.MethodOur study employed a gas chromatography and mass-spectrometer quantified dataset for analysis. Machine learning techniques, including feature selection and model training, were used to rank and evaluate potential biomarkers' contributions to the model's performance. Statistical methods were applied to filter significant and consistent biomarkers. The final selected biomarkers were iterated for all possible combinations using machine learning algorithms to determine their accuracy range. Furthermore, classification models were used to evaluate the performance metrics of the biomarkers and compare models.ResultThe final selected biomarkers, including 2-Myristynoyl Pantetheine, Pterin-6 Carboxylic Acid, Methyl Mercaptan, N-Acetyl Cysteine, and Butyric Acid, exhibited stable levels in exhaled breath during different physiological states. They showed high accuracy and precision in detecting abnormal liver function. Our machine learning models achieved an accuracy rate ranging from 0.7 to 0.95 in all conditions, with precision, recall, prediction probability, and a 95% confidence interval ranging from 0.84 to 0.94, using various combinations of these biomarkers.ConclusionOur statistical and machine learning analysis identified significant and potential biomarkers that contribute to the detection of abnormal liver function. These biomarkers were consistent across different physiological states of the body in both patient and healthy groups. The use of breath samples and feature selection machine learning methods proved to be an accurate and reliable approach for identifying these biomarkers. Our findings provide valuable insights for future research in this field and can inform the development of non-invasive and cost-effective diagnostic tests for liver disease.