Machine Learning Based Clinical Study of Target Breath Biomarkers Concentration Profile for Diagnosis of Liver Disease
Rakesh Kumar Patnaik 1, Yu-Chen Lin 1, Ashish Agarwal 1, Ming-Chih Ho 2and , J. Andrew Yeh 1
Presenting author: Rakesh Kumar Patnaik
1. INEMS, National Tsing Hua University, Taiwan
2. Department of Surgery, NTU Hospital, Taiwan
Diagnosis at an early stage may reduce the risk of disease progression and improves the mortality rate. Liver function impairment dysregulates the volatile organic compounds that are found in the exhaled breath. The combination of breath biomarkers and Machine Learning (ML) holds a lot of potential to improve the healthcare sector. Breath samples were collected following a unique breath collection protocol and three different biomarkers were quantified from athletes (control group) and liver patients from the NTU Hospital. Breath biomarkers have a significant difference in the concentration between the groups. The breath test data are used to distinguish the mild liver patients and healthy controls using isoprene and limonene as a dual biomarker combination with an ensemble boosting classification model. Different breath test ML models yield a classification accuracy of 60% to 87%. We have found the panel of target biomarkers strongly related to the liver disease scores with DMS added with said target biomarkers. In the regression model, the R-square value between the actual clinical score and predicted clinical score is found to be 0.78, 0.82, and 0.85 for Child-Pugh’s, APRI, and MELD scores respectively. The data imbalance problems can be successfully overcome by various sampling techniques is verified. This finding holds a promise of prediction to help diagnosis, routine measurement and mass-screening using limitless stock of exhaled breath. If this model is practiced and studied on a bigger dataset then it may unfold many possibilities that will help the health care sector in a whole new different way.